Professional Documents
Culture Documents
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
sureshsatapathy@ieee.org
Suresh Chandra Satapathy K. Srujan Raju
•
Editors
123
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Preface
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Team IC3T 2015
Chief Patrons
Advisory Committee
vii
sureshsatapathy@ieee.org
viii Team IC3T 2015
Program Chairs
Dr. A. Govardhan
Dr. K. Srujan Raju
Prof. G. Srikanth
Dr. Suresh Chandra Satapathy
Conveners
Organizing Committee
Program Committee
Ms. V. Swapna
Mrs. J. Srividya
Mr. Narasimha Rao
Mr. K. Murali
sureshsatapathy@ieee.org
Team IC3T 2015 ix
Finance Committee
Prof. G. Srikanth
Mr. D. Venkateshwarlu
Publicity Committee
Exhibition Committee
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
sureshsatapathy@ieee.org
x Team IC3T 2015
Sponsorship Committee
Mr. K. Bharath
Mr. E. Uma Shankar
Mr. P. Kranthi Rathan
Mr. Ch. Narendar
Registrations Committee
Mrs. P. Satyavathi
Mrs. K. Shrisha
Mrs. K. Mohana Lakshmi
Mrs. K. Jeji
Cultural Committee
sureshsatapathy@ieee.org
Team IC3T 2015 xi
sureshsatapathy@ieee.org
xii Team IC3T 2015
sureshsatapathy@ieee.org
Team IC3T 2015 xiii
sureshsatapathy@ieee.org
Contents
xv
sureshsatapathy@ieee.org
xvi Contents
sureshsatapathy@ieee.org
Contents xvii
sureshsatapathy@ieee.org
xviii Contents
sureshsatapathy@ieee.org
Contents xix
sureshsatapathy@ieee.org
xx Contents
sureshsatapathy@ieee.org
Contents xxi
sureshsatapathy@ieee.org
About the Editors
xxiii
sureshsatapathy@ieee.org
xxiv About the Editors
sureshsatapathy@ieee.org
Gesture Controlled Wireless Device
for Disabled
1 Introduction
sureshsatapathy@ieee.org
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.
Switches Microcontroller
Display (LCD)
Accelerometer
Zigbee Transmitter
sureshsatapathy@ieee.org
Gesture Controlled Wireless Device for Disabled 3
Zigbee M1
Microcontroller L293D
LCD 2 M2
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.
3.2 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
sureshsatapathy@ieee.org
4 Shantanu et al.
3.3 Receiver
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.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Implementation of Genetic Algorithm
for Optimization of Network Route
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.
1 Introduction
sureshsatapathy@ieee.org
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.
1.2 Optimization
sureshsatapathy@ieee.org
Implementation of Genetic Algorithm for Optimization … 9
sureshsatapathy@ieee.org
10 K.K. Sharma and I. Kaur
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).
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
sureshsatapathy@ieee.org
Implementation of Genetic Algorithm for Optimization … 11
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
sureshsatapathy@ieee.org
12 K.K. Sharma and I. Kaur
0.03
Green
0.02 Red
0.01
0
Objective Function = 1/TDF
sureshsatapathy@ieee.org
Implementation of Genetic Algorithm for Optimization … 13
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.
sureshsatapathy@ieee.org
14 K.K. Sharma and I. Kaur
References
sureshsatapathy@ieee.org
An Investigation of Gabor PCA
and Different Similarity Measure
Techniques for Image Classification
Keywords Image retrieval PCA MPCA KPCA GPCA G-vectors Score
features Similarity measures GRNN PNN
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
An Investigation of Gabor PCA and Different Similarity … 17
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
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):
sureshsatapathy@ieee.org
18 N. Hemavathi et al.
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.
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).
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.
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
sureshsatapathy@ieee.org
An Investigation of Gabor PCA and Different Similarity … 19
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.
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.
sureshsatapathy@ieee.org
20 N. Hemavathi et al.
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.
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
sureshsatapathy@ieee.org
An Investigation of Gabor PCA and Different Similarity … 21
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.
sureshsatapathy@ieee.org
22 N. Hemavathi et al.
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.
sureshsatapathy@ieee.org
An Investigation of Gabor PCA and Different Similarity … 23
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)
sureshsatapathy@ieee.org
24 N. Hemavathi et al.
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)
sureshsatapathy@ieee.org
A Low-Power High-Speed Double
Manchester Carry Chain with Carry-Skip
Using D3L
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.
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
A Low-Power High-Speed Double Manchester Carry Chain … 27
h0 ¼ g0 þ c1 : ð2Þ
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
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.
sureshsatapathy@ieee.org
28 J. Asha et al.
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
sureshsatapathy@ieee.org
A Low-Power High-Speed Double Manchester Carry Chain … 29
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.
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.
sureshsatapathy@ieee.org
30 J. Asha et al.
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
sureshsatapathy@ieee.org
A Low-Power High-Speed Double Manchester Carry Chain … 31
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.
sureshsatapathy@ieee.org
32 J. Asha et al.
2500
2000
1500
1000
500
0
8-bit 16-bit 32-bit 64-bit
No.of bits
sureshsatapathy@ieee.org
A Low-Power High-Speed Double Manchester Carry Chain … 33
8000
PDP v/s no. of bits
3000
2000
1000
0
8-bit 16-bit 32-bit 64-bit
No.of bits
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.
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Study and Analysis
of Electrocardiography Signals
for Computation of R Peak Value
for Sleep Apnea Patient
Keywords Sleep apnea Electrocardiography (ECG) Polysomnography (PSG)
R peak value
1 Introduction
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Study and Analysis of Electrocardiography Signals … 37
2 Sleep Apnea
3 Corpus
sureshsatapathy@ieee.org
38 M. Sahu et al.
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.
sureshsatapathy@ieee.org
Study and Analysis of Electrocardiography Signals … 39
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.
sureshsatapathy@ieee.org
40 M. Sahu et al.
sureshsatapathy@ieee.org
Study and Analysis of Electrocardiography Signals … 41
sureshsatapathy@ieee.org
42 M. Sahu et al.
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.
sureshsatapathy@ieee.org
Study and Analysis of Electrocardiography Signals … 43
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)
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Design and Implementation
of Two-Wheeled Self-Balancing Vehicle
Using Accelerometer and Fuzzy Logic
Keywords Fuzzy logic controller Self-balancing Fuzzy rule base accelerometer
Arduino UNO R3
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Design and Implementation of Two-Wheeled Self-Balancing … 47
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.
sureshsatapathy@ieee.org
48 S.S. Babu and A.S. Pillai
• 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_
sureshsatapathy@ieee.org
Design and Implementation of Two-Wheeled Self-Balancing … 49
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
":#
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.
sureshsatapathy@ieee.org
50 S.S. Babu and A.S. Pillai
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.
sureshsatapathy@ieee.org
Design and Implementation of Two-Wheeled Self-Balancing … 51
sureshsatapathy@ieee.org
52 S.S. Babu and A.S. Pillai
5 Result Analysis
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.
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
sureshsatapathy@ieee.org
Design and Implementation of Two-Wheeled Self-Balancing … 53
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.
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)
sureshsatapathy@ieee.org
A Time Efficient Secret Image Sharing
Scheme for Group Authentication System
Without Pixel Expansions
sureshsatapathy@ieee.org
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].
sureshsatapathy@ieee.org
A Time Efficient Secret Image Sharing Scheme … 57
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.
sureshsatapathy@ieee.org
58 S. Biswas et al.
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.
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
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
sureshsatapathy@ieee.org
A Time Efficient Secret Image Sharing Scheme … 59
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):
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:
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 division method (2) is now used to generate the last share K3:
K3 ¼ ð256 qÞ þ K ðK1 þ K2 Þ
sureshsatapathy@ieee.org
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
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
4 Experimental Results
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.
sureshsatapathy@ieee.org
A Time Efficient Secret Image Sharing Scheme … 61
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
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
A Time Efficient Secret Image Sharing Scheme … 63
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
sureshsatapathy@ieee.org
Imputation of Missing Gene Expressions
for DNA Microarray Using Particle
Swarm Optimization
Keywords Microarray
Swarm intelligence Particle swarm optimization
(PSO) Imputation Row averaging etc.
sureshsatapathy@ieee.org
66 C. Panse et al.
1 Introduction
sureshsatapathy@ieee.org
Imputation of Missing Gene Expressions for DNA Microarray … 67
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.
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]:
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.
sureshsatapathy@ieee.org
68 C. Panse et al.
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 Proposed Methodology
sureshsatapathy@ieee.org
Imputation of Missing Gene Expressions for DNA Microarray … 69
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.
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
;
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Imputation of Missing Gene Expressions for DNA Microarray … 71
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
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
sureshsatapathy@ieee.org
72 C. Panse et al.
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.
sureshsatapathy@ieee.org
Imputation of Missing Gene Expressions for DNA Microarray … 73
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
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.
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Deep Belief Network Based Part-of-Speech
Tagger for Telugu Language
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.
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Deep Belief Network Based Part-of-Speech Tagger … 77
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.
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.
sureshsatapathy@ieee.org
78 M. Jagadeesh et al.
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.
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.
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
sureshsatapathy@ieee.org
Deep Belief Network Based Part-of-Speech Tagger … 79
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.
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
sureshsatapathy@ieee.org
80 M. Jagadeesh et al.
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.
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
sureshsatapathy@ieee.org
Deep Belief Network Based Part-of-Speech Tagger … 81
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.
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.
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.
w2
w1 w3
sureshsatapathy@ieee.org
82 M. Jagadeesh et al.
65 Threshold
100
60
200
55
300
50
400
45
40
50 100 200 400 600 800 1000 1300 1500
Epochs
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.
sureshsatapathy@ieee.org
Deep Belief Network Based Part-of-Speech Tagger … 83
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)
sureshsatapathy@ieee.org
84 M. Jagadeesh et al.
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)
sureshsatapathy@ieee.org
On Context Awareness for Multisensor
Data Fusion in IoT
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.
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
sureshsatapathy@ieee.org
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.
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].
sureshsatapathy@ieee.org
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).
sureshsatapathy@ieee.org
88 S. Gite and H. Agrawal
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
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].
sureshsatapathy@ieee.org
On Context Awareness for Multisensor Data Fusion in IoT 89
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
sureshsatapathy@ieee.org
90 S. Gite and H. Agrawal
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
sureshsatapathy@ieee.org
On Context Awareness for Multisensor Data Fusion in IoT 91
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.
• 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.
sureshsatapathy@ieee.org
92 S. Gite and H. Agrawal
5 Conclusion
sureshsatapathy@ieee.org
On Context Awareness for Multisensor Data Fusion in IoT 93
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)
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
98 S. Ahuja et al.
5 Our Methodology
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:
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.
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
sureshsatapathy@ieee.org
CRiPT: Cryptography in Penetration Testing 99
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.
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.
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.
sureshsatapathy@ieee.org
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”);
}
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.
sureshsatapathy@ieee.org
CRiPT: Cryptography in Penetration Testing 101
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
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
sureshsatapathy@ieee.org
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
: 2 +1
:3
(continued)
sureshsatapathy@ieee.org
CRiPT: Cryptography in Penetration Testing 103
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
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.
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.
sureshsatapathy@ieee.org
104 S. Ahuja et al.
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
sureshsatapathy@ieee.org
CRiPT: Cryptography in Penetration Testing 105
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)
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Simultaneous Localization and Mapping
for Visually Impaired People for Outdoor
Environment
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.
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
110 A. Joshi et al.
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.
sureshsatapathy@ieee.org
Simultaneous Localization and Mapping … 111
4 Experimental Results
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
sureshsatapathy@ieee.org
112 A. Joshi et al.
Fig. 7 Obstacle 1
sureshsatapathy@ieee.org
Simultaneous Localization and Mapping … 113
Fig. 8 Obstacle 2
Fig. 9 Obstacle 3
sureshsatapathy@ieee.org
114 A. Joshi et al.
Fig. 10 Obstacle 4
6 Conclusion
sureshsatapathy@ieee.org
Simultaneous Localization and Mapping … 115
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
sureshsatapathy@ieee.org
Automatic ECG Image Classification
Using HOG and RPC Features
by Template Matching
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.
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
sureshsatapathy@ieee.org
118 V. Rathikarani et al.
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].
sureshsatapathy@ieee.org
Automatic ECG Image Classification … 119
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.
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).
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
sureshsatapathy@ieee.org
120 V. Rathikarani et al.
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
sureshsatapathy@ieee.org
Automatic ECG Image Classification … 121
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.
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
sureshsatapathy@ieee.org
122 V. Rathikarani et al.
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.
3 Experimental Results
The ECG image of normal and abnormal patients namely Arrhythmia, Myocardial
Infarction, and Conduction blocks are taken from the physiobank databases.
3.1 Database
sureshsatapathy@ieee.org
Automatic ECG Image Classification … 123
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
sureshsatapathy@ieee.org
124 V. Rathikarani et al.
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).
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).
sureshsatapathy@ieee.org
Automatic ECG Image Classification … 125
4 Conclusion
References
sureshsatapathy@ieee.org
Enhancement of Fuzzy Rank
Aggregation Technique
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Enhancement of Fuzzy Rank Aggregation Technique 129
2 Background
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 :
sureshsatapathy@ieee.org
130 M.Z. Ansari et al.
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.
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.
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.
sureshsatapathy@ieee.org
Enhancement of Fuzzy Rank Aggregation Technique 131
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.
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
sureshsatapathy@ieee.org
132 M.Z. Ansari et al.
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.
0.5 1
sureshsatapathy@ieee.org
Enhancement of Fuzzy Rank Aggregation Technique 133
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.
Propsed Technique
Modified Shimura
Shimura
MFO
Borda
sureshsatapathy@ieee.org
134 M.Z. Ansari et al.
6 Conclusion
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)
sureshsatapathy@ieee.org
Enhancement of Fuzzy Rank Aggregation Technique 135
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)
sureshsatapathy@ieee.org
Intelligent Telecommunication System
Using Semantic-Based Information
Retrieval
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Intelligent Telecommunication System … 139
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.
4 Proposed Work
sureshsatapathy@ieee.org
140 E. Ajith Jubilson et al.
Fig. 1 Voice-to-text
conversion using natural
language processing
sureshsatapathy@ieee.org
Intelligent Telecommunication System … 141
Fig. 3 Architecture diagram for semantic-based information retrieval through intelligent agents
sureshsatapathy@ieee.org
142 E. Ajith Jubilson et al.
5 Role of Agents
6 Conclusion
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)
sureshsatapathy@ieee.org
Intelligent Telecommunication System … 143
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)
sureshsatapathy@ieee.org
Minimizing Excessive Handover Using
Optimized Cuckoo Algorithm
in Heterogeneous Wireless Networks
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
sureshsatapathy@ieee.org
146 S.A. Kumar and K.E.S. Murthy
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
sureshsatapathy@ieee.org
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.
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.
sureshsatapathy@ieee.org
148 S.A. Kumar and K.E.S. Murthy
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.
sureshsatapathy@ieee.org
Minimizing Excessive Handover Using Optimized Cuckoo Algorithm … 149
Target
Network1
Home
Target Selected
Network
Network 2 Network
Target
Network n
(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.
sureshsatapathy@ieee.org
150 S.A. Kumar and K.E.S. Murthy
sureshsatapathy@ieee.org
Minimizing Excessive Handover Using Optimized Cuckoo Algorithm … 151
4 Simulation Results
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.
Stop
sureshsatapathy@ieee.org
152 S.A. Kumar and K.E.S. Murthy
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.
sureshsatapathy@ieee.org
Minimizing Excessive Handover Using Optimized Cuckoo Algorithm … 153
sureshsatapathy@ieee.org
154 S.A. Kumar and K.E.S. Murthy
sureshsatapathy@ieee.org
Minimizing Excessive Handover Using Optimized Cuckoo Algorithm … 155
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)
sureshsatapathy@ieee.org
Audio Songs Classification Based
on Music Patterns
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
sureshsatapathy@ieee.org
158 R. Sharma et al.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
160 R. Sharma et al.
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.
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.
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)
sureshsatapathy@ieee.org
Audio Songs Classification Based on Music Patterns 161
(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]:
1
http://www.raaga.com.
sureshsatapathy@ieee.org
162 R. Sharma et al.
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.
4 Result Analysis
sureshsatapathy@ieee.org
Audio Songs Classification Based on Music Patterns 163
Fig. 2 Artificial neural network model for the song classification system
sureshsatapathy@ieee.org
164 R. Sharma et al.
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.
sureshsatapathy@ieee.org
Audio Songs Classification Based on Music Patterns 165
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)
sureshsatapathy@ieee.org
166 R. Sharma et al.
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)
sureshsatapathy@ieee.org
Software Reliability Based on Software
Measures Applying Bayesian Technique
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.
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
sureshsatapathy@ieee.org
168 A. Senathi et al.
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].
sureshsatapathy@ieee.org
Software Reliability Based on Software Measures … 169
4 Proposed Methodology
The various steps for developing a model for software reliability prediction are as
follows:
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.
sureshsatapathy@ieee.org
170 A. Senathi et al.
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.
sureshsatapathy@ieee.org
Software Reliability Based on Software Measures … 171
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 %.
sureshsatapathy@ieee.org
172 A. Senathi et al.
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.
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
sureshsatapathy@ieee.org
Software Reliability Based on Software Measures … 173
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
sureshsatapathy@ieee.org
On Context Awareness and Analysis
of Various Classification Algorithms
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
On Context Awareness and Analysis … 177
2 Algorithms Used
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.
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.
sureshsatapathy@ieee.org
178 U. Nanda et al.
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
sureshsatapathy@ieee.org
On Context Awareness and Analysis … 179
1X n
1X n
MAE ¼ fi yi ¼ ei ;
n i¼1 n i¼1
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.
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
sureshsatapathy@ieee.org
180 U. Nanda et al.
Fig. 1 Comparison of algorithms with algorithm on x-axis and classified instances on y-axis
sureshsatapathy@ieee.org
On Context Awareness and Analysis … 181
4 Conclusion
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
sureshsatapathy@ieee.org
Neural Network-Based Automated
Priority Assigner
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
sureshsatapathy@ieee.org
184 H. Bhasin et al.
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
186 H. Bhasin et al.
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.
sureshsatapathy@ieee.org
Neural Network-Based Automated Priority Assigner 187
sureshsatapathy@ieee.org
188 H. Bhasin et al.
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
sureshsatapathy@ieee.org
Neural Network-Based Automated Priority Assigner 189
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
sureshsatapathy@ieee.org
190 H. Bhasin et al.
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)
sureshsatapathy@ieee.org
Design of IMC Controller for TITO
Process with Dynamic Closed-Loop
Time Constant
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.
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Design of IMC Controller for TITO Process … 193
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
where M1 ðsÞ, M2 ðsÞ, d12 ðsÞ and d21 ðsÞ are given by Eq. (5).
d21 G21
r2 + c2 + u2 +
C2 G22
- + +
y2
sureshsatapathy@ieee.org
194 P.K. Paul et al.
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
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
sureshsatapathy@ieee.org
Design of IMC Controller for TITO Process … 195
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
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
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
sureshsatapathy@ieee.org
196 P.K. Paul et al.
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
Wood and Berry model [13] represents a process for separating methanol from
water and is given by Eq. (20).
sureshsatapathy@ieee.org
Design of IMC Controller for TITO Process … 197
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)
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
Time (sec) Time (sec)
sureshsatapathy@ieee.org
198 P.K. Paul et al.
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].
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
(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
Time (sec) Time (sec)
sureshsatapathy@ieee.org
Design of IMC Controller for TITO Process … 199
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
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)
sureshsatapathy@ieee.org
200 P.K. Paul et al.
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)
sureshsatapathy@ieee.org
All Optical SOA-MZI-Based Encryption
Decryption System Using Co Propagating
Optical Pulses and CW Wave at 40 Gb/s
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.
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
All Optical SOA-MZI-Based Encryption Decryption System … 203
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.
sureshsatapathy@ieee.org
204 V. Agarwal and V. Chaurasia
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.
Decryption of encrypted data can be achieved with XOR logic operation of Ci and
Ki, i.e.,
Pi ¼ Ci XOR Ki ð1Þ
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
sureshsatapathy@ieee.org
All Optical SOA-MZI-Based Encryption Decryption System … 205
Fig. 2 Eye diagram of input signal before (above) and after encryption (below)
sureshsatapathy@ieee.org
206 V. Agarwal and V. Chaurasia
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).
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
sureshsatapathy@ieee.org
All Optical SOA-MZI-Based Encryption Decryption System … 207
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
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)
sureshsatapathy@ieee.org
208 V. Agarwal and V. Chaurasia
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)
sureshsatapathy@ieee.org
Exploiting Common Nodes in Overlapped
Clusters for Path Optimization in Wireless
Sensor Networks
Keywords Cluster overlap Common nodes Latency Path optimization WSN
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
212 D. Rao BV et al.
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.
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:
sureshsatapathy@ieee.org
Exploiting Common Nodes in Overlapped Clusters … 213
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
sureshsatapathy@ieee.org
214 D. Rao BV et al.
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
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.
sureshsatapathy@ieee.org
Exploiting Common Nodes in Overlapped Clusters … 215
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
sureshsatapathy@ieee.org
216 D. Rao BV et al.
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
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)
sureshsatapathy@ieee.org
Exploiting Common Nodes in Overlapped Clusters … 217
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)
sureshsatapathy@ieee.org
Maximizing Availability and Minimizing
Markesan for Task Scheduling in Grid
Computing Using NSGA 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
sureshsatapathy@ieee.org
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
3 Related Work
sureshsatapathy@ieee.org
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.
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
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
sureshsatapathy@ieee.org
222 D.P. Sahu et al.
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):
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.
sureshsatapathy@ieee.org
Maximizing Availability and Minimizing Markesan … 223
6 Conclusion
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)
sureshsatapathy@ieee.org
224 D.P. Sahu et al.
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)
sureshsatapathy@ieee.org
Moving Object Detection for Visual
Surveillance Using Quasi-euclidian
Distance
Keywords Gaussian mixture model Quasi-euclidian distance Object detection
Visual surveillance Morphological filter
1 Introduction
sureshsatapathy@ieee.org
226 D.K. Yadav and K. Singh
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.
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]
sureshsatapathy@ieee.org
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.
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:
sureshsatapathy@ieee.org
228 D.K. Yadav and K. Singh
lk ¼ ð1 aÞ lk þ a It ð2Þ
wk ¼ ð1 aÞ wk þ a ð4Þ
wk ¼ ð1 aÞ wk ð5Þ
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.
sureshsatapathy@ieee.org
Moving Object Detection for Visual Surveillance … 229
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).
sureshsatapathy@ieee.org
230 D.K. Yadav and K. Singh
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.
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:
sureshsatapathy@ieee.org
Moving Object Detection for Visual Surveillance … 231
sureshsatapathy@ieee.org
232 D.K. Yadav and K. Singh
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)
sureshsatapathy@ieee.org
Moving Object Detection for Visual Surveillance … 233
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)
sureshsatapathy@ieee.org
IoTA: Internet of Things Application
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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,
sureshsatapathy@ieee.org
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.
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).
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
sureshsatapathy@ieee.org
IoTA: Internet of Things Application 239
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).
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
sureshsatapathy@ieee.org
240 S. Ahuja et al.
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).
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
sureshsatapathy@ieee.org
IoTA: Internet of Things Application 241
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).
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:
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Table 1 Database snapshot
Snapshot description Snapshot
Disease table
Doctor table
IoTA: Internet of Things Application
Patient table
sureshsatapathy@ieee.org
HCS table
243
Table 2 Snapshots of form to receive input in healthcare system
244
(continued)
sureshsatapathy@ieee.org
S. Ahuja et al.
IoTA: Internet of Things Application 245
Table 2 (continued)
sureshsatapathy@ieee.org
246 S. Ahuja et al.
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
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.
sureshsatapathy@ieee.org
IoTA: Internet of Things Application 247
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)
sureshsatapathy@ieee.org
Improving Performance of Wireless
Mesh Networks Through User
Association Method
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
sureshsatapathy@ieee.org
250 G.V. Kumar and C.S. Bindu
sureshsatapathy@ieee.org
Improving Performance of Wireless Mesh Networks … 251
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
sureshsatapathy@ieee.org
252 G.V. Kumar and C.S. Bindu
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.
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
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.
sureshsatapathy@ieee.org
Improving Performance of Wireless Mesh Networks … 253
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
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 .
sureshsatapathy@ieee.org
254 G.V. Kumar and C.S. Bindu
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
sureshsatapathy@ieee.org
Improving Performance of Wireless Mesh Networks … 255
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.
sureshsatapathy@ieee.org
256 G.V. Kumar and C.S. Bindu
sureshsatapathy@ieee.org
Improving Performance of Wireless Mesh Networks … 257
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)
sureshsatapathy@ieee.org
258 G.V. Kumar and C.S. Bindu
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
sureshsatapathy@ieee.org
Characterization of Human Fingernails
Using Iterative Thresholding
Segmentation
Keywords Fingernail biometric Multispectral data Intensity transformation
Contrast enhancement Morphological segmentation Iterative thresholding Nail
region of interest
sureshsatapathy@ieee.org
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].
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.
sureshsatapathy@ieee.org
Characterization of Human Fingernails … 261
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.
sureshsatapathy@ieee.org
262 N.S. Kumuda et al.
2.1 Preprocessing
Fig. 2 a Gray of original color. b Gamma corrected R, G, B. c Edge detection for (b)
sureshsatapathy@ieee.org
Characterization of Human Fingernails … 263
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)
sureshsatapathy@ieee.org
264 N.S. Kumuda et al.
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
N.S. Kumuda et al.
Characterization of Human Fingernails … 267
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.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
14-1 Square Circular
17-1 Square Circular
N.S. Kumuda et al.
Characterization of Human Fingernails … 269
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)
sureshsatapathy@ieee.org
270 N.S. Kumuda et al.
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)
sureshsatapathy@ieee.org
Confidential Terms Detection
Using Language Modeling Technique
in Data Leakage Prevention
1 Introduction
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Confidential Terms Detection Using Language Modeling Technique … 273
sureshsatapathy@ieee.org
274 P. Subhashini and B.P. Rani
• 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
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.
sureshsatapathy@ieee.org
Confidential Terms Detection Using Language Modeling Technique … 275
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]:
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.
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.
sureshsatapathy@ieee.org
276 P. Subhashini and B.P. Rani
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
3 Evaluations
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
sureshsatapathy@ieee.org
Confidential Terms Detection Using Language Modeling Technique … 277
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.
sureshsatapathy@ieee.org
278 P. Subhashini and B.P. Rani
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)
sureshsatapathy@ieee.org
Confidential Terms Detection Using Language Modeling Technique … 279
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/
sureshsatapathy@ieee.org
ECG-Driven Extraction of Respiration
Rate Using Ensemble Empirical
Mode Decomposition and Canonical
Correlation Analysis
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
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.
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:
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 :
sureshsatapathy@ieee.org
284 V. Kumar and G. Singh
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
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
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.
sureshsatapathy@ieee.org
ECG-Driven Extraction of Respiration Rate … 285
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
sureshsatapathy@ieee.org
286 V. Kumar and G. Singh
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 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
sureshsatapathy@ieee.org
ECG-Driven Extraction of Respiration Rate … 287
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.
sureshsatapathy@ieee.org
288 V. Kumar and G. Singh
400
380
360
340
320
300
280
260
240
0 1000 2000 3000 4000 5000 6000
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)
sureshsatapathy@ieee.org
ECG-Driven Extraction of Respiration Rate … 289
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
sureshsatapathy@ieee.org
An Efficient On-Chip Implementation
of Reconfigurable Continuous Time Sigma
Delta ADC for Digital Beamforming
Applications
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).
sureshsatapathy@ieee.org
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
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].
sureshsatapathy@ieee.org
An Efficient On-Chip Implementation … 293
Analog
Input signal
ANALOG
ADC
Digital
Signal
Processing
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.
sureshsatapathy@ieee.org
294 A.H.V. Palagiri et al.
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
–
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
sureshsatapathy@ieee.org
An Efficient On-Chip Implementation … 295
Analog
Input signal Differential Pin
12 bits 12 bits
1 bit
LVPECL CFIR
COMP BUFFER CIC PFIR
R
C
LV Digital
PE Output
CL
sureshsatapathy@ieee.org
296 A.H.V. Palagiri et al.
sureshsatapathy@ieee.org
An Efficient On-Chip Implementation … 297
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.
sureshsatapathy@ieee.org
298 A.H.V. Palagiri et al.
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
sureshsatapathy@ieee.org
An Efficient On-Chip Implementation … 299
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
sureshsatapathy@ieee.org
Random Forest for the Real Forests
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.
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.
sureshsatapathy@ieee.org
302 S. Agrawal et al.
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
sureshsatapathy@ieee.org
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
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).
sureshsatapathy@ieee.org
304 S. Agrawal et al.
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.
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
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
sureshsatapathy@ieee.org
Random Forest for the Real Forests 305
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:
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.
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/.
sureshsatapathy@ieee.org
306 S. Agrawal et al.
sureshsatapathy@ieee.org
Random Forest for the Real Forests 307
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).
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
sureshsatapathy@ieee.org
308 S. Agrawal et al.
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.
sureshsatapathy@ieee.org
Random Forest for the Real Forests 309
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)
sureshsatapathy@ieee.org
Hybrid GA and PSO Approach
for Transmission Expansion Planning
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.
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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).
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.
This inequality constraint is required to check the maximum power flow for each
path with limit and is formulated as follows:
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.
There should be specified limit for generated power and is represented as follows:
sureshsatapathy@ieee.org
314 S. Sisodia et al.
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.
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.
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.
sureshsatapathy@ieee.org
Hybrid GA and PSO Approach for Transmission Expansion Planning 315
This section gives overview of GA, PSO, and HGAPSO algorithms and their
implementation to TEP problem.
sureshsatapathy@ieee.org
316 S. Sisodia et al.
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.
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:
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.
sureshsatapathy@ieee.org
Hybrid GA and PSO Approach for Transmission Expansion Planning 317
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.
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.
sureshsatapathy@ieee.org
318 S. Sisodia et al.
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
300
250
200
150
0 10 20 30 40 50 60 70
No. of iteration
2250
2200
2150
2100
2050
2000
1950
1900
0 10 20 30 40 50 60 70
No. of iteration
sureshsatapathy@ieee.org
Hybrid GA and PSO Approach for Transmission Expansion Planning 319
Objective function
1100
1000
900
800
700
600
500
0 10 20 30 40 50 60 70
No. of iteration
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
S. Sisodia et al.
Hybrid GA and PSO Approach for Transmission Expansion Planning 321
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
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)
sureshsatapathy@ieee.org
322 S. Sisodia et al.
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)
sureshsatapathy@ieee.org
Pattern Detection Framework
for MRI Images and Labeling
Volume of Interest (VoI)
1 Introduction
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
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
Pattern Detection Framework for MRI Images … 325
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
sureshsatapathy@ieee.org
326 R. Snehkunj et al.
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
sureshsatapathy@ieee.org
Pattern Detection Framework for MRI Images … 327
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.
sureshsatapathy@ieee.org
328 R. Snehkunj et al.
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
sureshsatapathy@ieee.org
Pattern Detection Framework for MRI Images … 329
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.
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).
sureshsatapathy@ieee.org
330 R. Snehkunj et al.
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.
sureshsatapathy@ieee.org
Pattern Detection Framework for MRI Images … 331
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)
sureshsatapathy@ieee.org
332 R. Snehkunj et al.
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)
sureshsatapathy@ieee.org
A Distributed Spanning Tree-Based
Dynamic Self-Organizational Framework
for Web Server
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
336 J. Amudhavel et al.
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.
3.1.1 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.
sureshsatapathy@ieee.org
A Distributed Spanning Tree-Based Dynamic Self-Organizational … 337
3.1.2 Joblist
The joblist consists of all the requests or jobs created by the client to the server.
3.1.3 Dispatcher
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.
sureshsatapathy@ieee.org
338 J. Amudhavel et al.
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)
sureshsatapathy@ieee.org
A Distributed Spanning Tree-Based Dynamic Self-Organizational … 339
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)
sureshsatapathy@ieee.org
Recursive Ant Colony Optimization
Routing in Wireless Mesh Network
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.
sureshsatapathy@ieee.org
342 J. Amudhavel et al.
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].
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
344 J. Amudhavel et al.
5 2 3 4
5 8
3 6 7 8 9
4
7
7 5
sureshsatapathy@ieee.org
Recursive Ant Colony Optimization Routing … 345
6
8
4 2 3
1 6
4
1 5
8
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
sureshsatapathy@ieee.org
346 J. Amudhavel et al.
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.
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.
sureshsatapathy@ieee.org
Recursive Ant Colony Optimization Routing … 347
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
sureshsatapathy@ieee.org
348 J. Amudhavel et al.
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.
2 Proposed Methodology
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
sureshsatapathy@ieee.org
Recursive Ant Colony Optimization Routing … 349
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
sureshsatapathy@ieee.org
350 J. Amudhavel et al.
• Step 4
In subproblem S2 recursive method is used to find the shortest path based on
low cost path.
• Step 5
In subproblem S3 recursive method is used to find the shortest path based on
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
sureshsatapathy@ieee.org
Recursive Ant Colony Optimization Routing … 351
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:
International Conference on Green Computing Communication and Electrical Engineering
(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:
International Conference on Green Computing Communication and Electrical Engineering
(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
sureshsatapathy@ieee.org
Image Processing Representation Using
Binary Image; Grayscale, Color Image,
and Histogram
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.
1 Introduction
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
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.
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.
sureshsatapathy@ieee.org
356 V.M. Mohan et al.
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.
sureshsatapathy@ieee.org
Image Processing Representation Using Binary Image … 357
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
sureshsatapathy@ieee.org
358 V.M. Mohan et al.
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.
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.
sureshsatapathy@ieee.org
Image Processing Representation Using Binary Image … 359
4.2 Grayscale
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.
sureshsatapathy@ieee.org
360 V.M. Mohan et al.
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.
sureshsatapathy@ieee.org
Image Processing Representation Using Binary Image … 361
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)
sureshsatapathy@ieee.org
A Vector Space Model Approach
for Web Attack Classification
Using Machine Learning Technique
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
366 B.V. Ram Naresh Yadav et al.
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.
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.
sureshsatapathy@ieee.org
A Vector Space Model Approach for Web Attack Classification … 367
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.
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
sureshsatapathy@ieee.org
368 B.V. Ram Naresh Yadav et al.
countðt; dÞ
tf ðt; dÞ ¼ P ð2Þ
v2d countðv; dÞ
jDj
idf ðtÞ ¼ log ð3Þ
jfd 2 D : t 2 dgj
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:
sureshsatapathy@ieee.org
A Vector Space Model Approach for Web Attack Classification … 369
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.
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.
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.
sureshsatapathy@ieee.org
370 B.V. Ram Naresh Yadav et al.
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.
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.
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.
sureshsatapathy@ieee.org
A Vector Space Model Approach for Web Attack Classification … 371
sureshsatapathy@ieee.org
372 B.V. Ram Naresh Yadav et al.
5 Conclusions
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
sureshsatapathy@ieee.org
A Vector Space Model Approach for Web Attack Classification … 373
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)
sureshsatapathy@ieee.org
Formal Verification of Secure
Authentication in Wireless Mesh Network
(SAWMN)
1 Introduction
sureshsatapathy@ieee.org
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,
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
378 N. Singh and H. Saini
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.
sureshsatapathy@ieee.org
Formal Verification of Secure Authentication … 379
[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
sureshsatapathy@ieee.org
380 N. Singh and H. Saini
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.
sureshsatapathy@ieee.org
Formal Verification of Secure Authentication … 381
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
sureshsatapathy@ieee.org
382 N. Singh and H. Saini
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.
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:
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:
CertBA ¼ tPht
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:
sureshsatapathy@ieee.org
Formal Verification of Secure Authentication … 383
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:
MR ¼ PRMG PhMR
CertMG
sureshsatapathy@ieee.org
384 N. Singh and H. Saini
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.'
sureshsatapathy@ieee.org
Formal Verification of Secure Authentication … 385
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).
sureshsatapathy@ieee.org
386 N. Singh and H. Saini
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.
sureshsatapathy@ieee.org
Formal Verification of Secure Authentication … 387
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)
sureshsatapathy@ieee.org
388 N. Singh and H. Saini
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)
sureshsatapathy@ieee.org
Visual K-Means Approach for Handling
Class Imbalance Learning
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Visual K-Means Approach for Handling Class Imbalance Learning 391
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.
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.
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.
sureshsatapathy@ieee.org
392 Ch.N. Santhosh Kumar et al.
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.
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.
sureshsatapathy@ieee.org
Visual K-Means Approach for Handling Class Imbalance Learning 393
5 Results
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
sureshsatapathy@ieee.org
394 Ch.N. Santhosh Kumar et al.
Table 2 Summary of tenfold cross-validation performance for precision on all the datasets
Datasets K-means Density FF EM Hire VKM
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
Datasets K-means Density FF EM Hire VKM
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
Datasets K-means Density FF EM Hire VKM
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
sureshsatapathy@ieee.org
Visual K-Means Approach for Handling Class Imbalance Learning 395
Table 5 Summary of tenfold cross-validation performance for FP Rate on all the datasets
Datasets K-means Density FF EM Hire VKM
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
Datasets K-means Density FF EM Hire VKM
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
6 Conclusion
sureshsatapathy@ieee.org
396 Ch.N. Santhosh Kumar et al.
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
sureshsatapathy@ieee.org
A Framework for Discovering Important
Patterns Through Parallel Mining
of Protein–Protein Interaction Network
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.
1 Introduction
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
sureshsatapathy@ieee.org
398 S. Dasgupta and B. Saha
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.
sureshsatapathy@ieee.org
A Framework for Discovering Important Patterns … 399
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.
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,
sureshsatapathy@ieee.org
400 S. Dasgupta and B. Saha
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.
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
sureshsatapathy@ieee.org
A Framework for Discovering Important Patterns … 401
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:
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.
sureshsatapathy@ieee.org
402 S. Dasgupta and B. Saha
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)
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.
}
sureshsatapathy@ieee.org
A Framework for Discovering Important Patterns … 403
4 Performance Analysis
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
sureshsatapathy@ieee.org
404 S. Dasgupta and B. Saha
Scaleup
1
0
1 2 3 4
Number of machines
15
1 core
SizeUp
10
2 cores
5 4 cores
0 16 cores
1GB 2GB 4GB 8GB
Size of dataset in GB
10 1 GB
3 GB
0 5 GB
1 2 4 8
Number of cores 10 GB
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.
sureshsatapathy@ieee.org
A Framework for Discovering Important Patterns … 405
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)
sureshsatapathy@ieee.org
406 S. Dasgupta and B. Saha
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)
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
408 G. Gugnani et al.
1 Introduction
sureshsatapathy@ieee.org
Implementing DNA Encryption Technique in Web Services … 409
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
sureshsatapathy@ieee.org
410 G. Gugnani et al.
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.
3 Proposed Methodology
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
sureshsatapathy@ieee.org
Implementing DNA Encryption Technique in Web Services … 411
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.
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
sureshsatapathy@ieee.org
412 G. Gugnani et al.
sureshsatapathy@ieee.org
Implementing DNA Encryption Technique in Web Services … 413
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
sureshsatapathy@ieee.org
414 G. Gugnani et al.
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
sureshsatapathy@ieee.org
Implementing DNA Encryption Technique in Web Services … 415
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)
sureshsatapathy@ieee.org
An Efficient Motion Detection Method
Based on Estimation of Initial Motion
Field Using Local Variance
1 Introduction
sureshsatapathy@ieee.org
418 S. Kumar and J.S. Yadav
2 Related Work
sureshsatapathy@ieee.org
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].
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.
sureshsatapathy@ieee.org
420 S. Kumar and J.S. Yadav
sureshsatapathy@ieee.org
An Efficient Motion Detection Method Based on Estimation … 421
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.
4 Experimental Results
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).
sureshsatapathy@ieee.org
422 S. Kumar and J.S. Yadav
Ground
Truths
Proposed
Method
Ground
Truths
Proposed
Method
sureshsatapathy@ieee.org
An Efficient Motion Detection Method Based on Estimation … 423
Ground
Truths
Proposed
Method
The relevant and irrelevant true-positive pixels in the foreground mask are given
through recall and precision, respectively, which are assessed as follows:
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
sureshsatapathy@ieee.org
424 S. Kumar and J.S. Yadav
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.
sureshsatapathy@ieee.org
An Efficient Motion Detection Method Based on Estimation … 425
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
sureshsatapathy@ieee.org
A Survey on Texture Image Retrieval
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.
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,
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
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:
Feature extraction time solely depends on the method used to extract the fea-
tures, while computational time depends upon the distance measure formula.
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.
sureshsatapathy@ieee.org
430 G. Raghuwanshi and V. Tyagi
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
sureshsatapathy@ieee.org
A Survey on Texture Image Retrieval 431
sureshsatapathy@ieee.org
432 G. Raghuwanshi and V. Tyagi
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,
sureshsatapathy@ieee.org
A Survey on Texture Image Retrieval 433
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
sureshsatapathy@ieee.org
434 G. Raghuwanshi and V. Tyagi
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)
sureshsatapathy@ieee.org
A Survey on Texture Image Retrieval 435
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)
sureshsatapathy@ieee.org
Converged OAM
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Converged OAM 439
sureshsatapathy@ieee.org
440 P. Saste and J. Martis
sureshsatapathy@ieee.org
Converged OAM 441
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
sureshsatapathy@ieee.org
442 P. Saste and J. Martis
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.
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.
sureshsatapathy@ieee.org
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
Ethernet IP Y.1731
Header Header OAM PDU
sureshsatapathy@ieee.org
444 P. Saste and J. Martis
7 Conclusion
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
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy
PI Controller with Real-Time
Implementation on a Speed
Control Process
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
sureshsatapathy@ieee.org
446 A. Ghosh et al.
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
sureshsatapathy@ieee.org
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.
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].
sureshsatapathy@ieee.org
448 A. Ghosh et al.
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
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy PI Controller … 449
6 Simulation Results
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy PI Controller … 451
sureshsatapathy@ieee.org
452 A. Ghosh et al.
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy PI Controller … 453
(a) (b)
0.8 0.8
change-in-control-action
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
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.
sureshsatapathy@ieee.org
454 A. Ghosh et al.
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy PI Controller … 455
(a) (b)
0.8 0.8
change-in-control-action
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
-1 -1 change-of-error
change-of-error error error
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.
sureshsatapathy@ieee.org
456 A. Ghosh et al.
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
sureshsatapathy@ieee.org
Rule Reduction of a Neuro-Fuzzy PI Controller … 457
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
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Construction of Binary and Nonbinary
LDPC-like Codes from Kernel Codes
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.
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Construction of Binary and Nonbinary LDPC-like Codes … 461
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.
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}.
sureshsatapathy@ieee.org
462 C.P. Kumar et al.
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.
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
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.
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.
sureshsatapathy@ieee.org
Construction of Binary and Nonbinary LDPC-like Codes … 463
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.
sureshsatapathy@ieee.org
464 C.P. Kumar et al.
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
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.
sureshsatapathy@ieee.org
Construction of Binary and Nonbinary LDPC-like Codes … 465
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)
sureshsatapathy@ieee.org
466 C.P. Kumar et al.
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)
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
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.
2 Proposed Methodology
sureshsatapathy@ieee.org
Simulation-Level Implementation of Face Recognition … 469
sureshsatapathy@ieee.org
470 S.L. Fernandes and G. Josemin Bala
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
472 S.L. Fernandes and G. Josemin Bala
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)
sureshsatapathy@ieee.org
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.
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
sureshsatapathy@ieee.org
474 S.L. Fernandes and G. Josemin Bala
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.
Multimodal fusion techniques are usually classified as: score level fusion [15],
decision level fusion [16], and feature level fusion [17].
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
sureshsatapathy@ieee.org
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.
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.
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
sureshsatapathy@ieee.org
476 S.L. Fernandes and G. Josemin Bala
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
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
478 S.L. Fernandes and G. Josemin Bala
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)
sureshsatapathy@ieee.org
Genetic Algorithmic Approach to Mitigate
Starvation in Wireless Mesh Networks
Keywords Wireless mesh networks Starvation Contention window Genetic
algorithm Multimedia traffic Priority-based scheduling method
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
482 P. Balamuralikrishna et al.
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.
sureshsatapathy@ieee.org
Genetic Algorithmic Approach to Mitigate … 483
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.
sureshsatapathy@ieee.org
484 P. Balamuralikrishna et al.
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.
Is Yes
termination End
met?
No
sureshsatapathy@ieee.org
Genetic Algorithmic Approach to Mitigate … 485
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.
sureshsatapathy@ieee.org
486 P. Balamuralikrishna et al.
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.
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
sureshsatapathy@ieee.org
Genetic Algorithmic Approach to Mitigate … 487
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
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
sureshsatapathy@ieee.org
488 P. Balamuralikrishna et al.
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)
sureshsatapathy@ieee.org
Genetic Algorithmic Approach to Mitigate … 489
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)
sureshsatapathy@ieee.org
Handwritten Indic Script Identification
from Document Images—A Statistical
Comparison of Different Attribute
Selection Techniques in Multi-classifier
Environment
sureshsatapathy@ieee.org
492 S.M. Obaidullah et al.
1 Introduction
sureshsatapathy@ieee.org
Handwritten Indic Script Identification from Document Images … 493
2 Proposed Methodology
2.1 Preprocessing
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.
sureshsatapathy@ieee.org
494 S.M. Obaidullah et al.
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
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.
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.
sureshsatapathy@ieee.org
Handwritten Indic Script Identification from Document Images … 495
sureshsatapathy@ieee.org
496 S.M. Obaidullah et al.
sureshsatapathy@ieee.org
Handwritten Indic Script Identification from Document Images … 497
sureshsatapathy@ieee.org
498 S.M. Obaidullah et al.
sureshsatapathy@ieee.org
Handwritten Indic Script Identification from Document Images … 499
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)
sureshsatapathy@ieee.org
500 S.M. Obaidullah et al.
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)
sureshsatapathy@ieee.org
An FPGA-Based Embedded System
for Real-Time Data Processing
Keywords Coherent averaging Embedded system Field programmable gate
arrays (FPGAs) Random noise
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
An FPGA-Based Embedded System for Real-Time Data Processing 503
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.
The architecture for ping-pong memory-based coherent averaging with trigger unit
is described as follows.
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
An FPGA-Based Embedded System for Real-Time Data Processing 505
Fig. 2 a Architecture of the ping-pong memory-based coherent averaging with trigger unit,
b 16-bit Pipelined adder
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
An FPGA-Based Embedded System for Real-Time Data Processing 507
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.
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
An FPGA-Based Embedded System for Real-Time Data Processing 509
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.
References
sureshsatapathy@ieee.org
Clustering of Noisy Regions (CNR)—A
GIS Anchored Technique for Clustering
on Raster Map
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.
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
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).
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.
sureshsatapathy@ieee.org
514 A. Chakraborty et al.
sureshsatapathy@ieee.org
Clustering of Noisy Regions (CNR) … 515
Fig. 4 Digitization
sureshsatapathy@ieee.org
516 A. Chakraborty et al.
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.
sureshsatapathy@ieee.org
Clustering of Noisy Regions (CNR) … 517
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
Fig. 8 Graphical
representation of comparison
between CNR and K-means
on clustering time
sureshsatapathy@ieee.org
518
sureshsatapathy@ieee.org
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.
Clustering of Noisy Regions (CNR) … 519
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)
sureshsatapathy@ieee.org
520 A. Chakraborty et al.
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
sureshsatapathy@ieee.org
Performance of Brightness Enhancement
Technique for Narrow-Band
and White-Band Images
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
sureshsatapathy@ieee.org
522 S.R. Challa and D.V. Rama Koti Reddy
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
sureshsatapathy@ieee.org
Performance of Brightness Enhancement Technique … 523
YCbCr
BPF
Adaptive illu-
mination adjust-
ment
Color restoration
Input Image
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
sureshsatapathy@ieee.org
524 S.R. Challa and D.V. Rama Koti Reddy
3 Methodology
Fig. 2 stage by stage output images as a input, b luminance, c background, and d output image
sureshsatapathy@ieee.org
Performance of Brightness Enhancement Technique … 525
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
sureshsatapathy@ieee.org
526 S.R. Challa and D.V. Rama Koti Reddy
Fig. 4 NB images at each stage as a input, b luminance, c background, and d output image
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.
sureshsatapathy@ieee.org
Performance of Brightness Enhancement Technique … 527
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)
sureshsatapathy@ieee.org
Z Transformation-Based High Payload
Authentication Technique with Higher
Region of Convergence Value (ZATHRoc)
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
sureshsatapathy@ieee.org
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
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
532 J.K. Mandal and S. Mahapatra
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.
sureshsatapathy@ieee.org
Z Transformation-Based High Payload Authentication Technique … 533
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.
sureshsatapathy@ieee.org
534 J.K. Mandal and S. Mahapatra
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
sureshsatapathy@ieee.org
Z Transformation-Based High Payload Authentication Technique … 535
sureshsatapathy@ieee.org
536 J.K. Mandal and S. Mahapatra
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’
sureshsatapathy@ieee.org
Z Transformation-Based High Payload Authentication Technique … 537
4 Conclusion
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)
sureshsatapathy@ieee.org
Breast Tissue Density Classification Using
Wavelet-Based Texture Descriptors
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
sureshsatapathy@ieee.org
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
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.
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.
sureshsatapathy@ieee.org
Breast Tissue Density Classification … 541
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
sureshsatapathy@ieee.org
542 J. Virmani and Kriti
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
sureshsatapathy@ieee.org
Breast Tissue Density Classification … 543
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
sureshsatapathy@ieee.org
544 J. Virmani and Kriti
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
sureshsatapathy@ieee.org
Breast Tissue Density Classification … 545
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)
sureshsatapathy@ieee.org
546 J. Virmani and Kriti
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)
sureshsatapathy@ieee.org
On New Families Related to Bernoulli
and Euler Polynomials
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.
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!
sureshsatapathy@ieee.org
548 S. Khan and M. Ali
2ext X
1
tn
¼ En ðxÞ ; jtj\p ð3Þ
etþ 1 n¼0 n!
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:
and
^ and P
The operators M ^ also satisfy the commutation relation
^ M
½P; ^ ¼P
^M^ M
^P^ ¼ ^1 ð7Þ
^ Pfp
M ^ n ðxÞg ¼ npn ðxÞ: ð8Þ
sureshsatapathy@ieee.org
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!
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]:
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.
sureshsatapathy@ieee.org
550 S. Khan and M. Ali
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!
X
1
@ m1 @ r1 tn
¼ Bn y þ mD1 þ rz r1 : ð19Þ
n¼0
x
@ym1 @y n!
sureshsatapathy@ieee.org
On New Families Related to Bernoulli and Euler Polynomials 551
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
and
^ LHB ¼ Dy ;
P ð23Þ
respectively.
Proof Consider the identity:
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!
sureshsatapathy@ieee.org
552 S. Khan and M. Ali
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
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
sureshsatapathy@ieee.org
On New Families Related to Bernoulli and Euler Polynomials 553
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
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:
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
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
sureshsatapathy@ieee.org
554
sureshsatapathy@ieee.org
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
On New Families Related to Bernoulli and Euler Polynomials 555
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)
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming
Pile-Ups in Data Dissemination
Using HTPVANET Algorithm
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.
1 Introduction
sureshsatapathy@ieee.org
558 R.V.S. Lalitha and G. Jaya Suma
Fig. 1 MANET
Fig. 2 VANET
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 559
2 Related Work
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:
sureshsatapathy@ieee.org
560 R.V.S. Lalitha and G. Jaya Suma
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 561
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
sureshsatapathy@ieee.org
562 R.V.S. Lalitha and G. Jaya Suma
VANET
formation
Optimized
VANET
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.
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 563
sureshsatapathy@ieee.org
564 R.V.S. Lalitha and G. Jaya Suma
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.
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 565
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).
sureshsatapathy@ieee.org
566 R.V.S. Lalitha and G. Jaya Suma
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.*/
6. return
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 567
sureshsatapathy@ieee.org
568 R.V.S. Lalitha and G. Jaya Suma
B:7
D:1
A:2
C:4
E:4
D:6
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
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 569
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
sureshsatapathy@ieee.org
570 R.V.S. Lalitha and G. Jaya Suma
sureshsatapathy@ieee.org
Vehicular Ad Hoc Networks: Trimming Pile-Ups … 571
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
sureshsatapathy@ieee.org
Information Extraction from Research
Papers Based on Statistical Methods
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.
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Information Extraction from Research Papers … 575
3 Proposed Work
sureshsatapathy@ieee.org
576 S.D. Kavila and D.F. Rani
4 Proposed Architecture
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.
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.
sureshsatapathy@ieee.org
Information Extraction from Research Papers … 577
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.
sureshsatapathy@ieee.org
578 S.D. Kavila and D.F. Rani
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
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
sureshsatapathy@ieee.org
Information Extraction from Research Papers … 579
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.
sureshsatapathy@ieee.org
580 S.D. Kavila and D.F. Rani
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)
sureshsatapathy@ieee.org
Artificial Neural Network Classifier
for Intrusion Detection System
in Computer Network
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)
sureshsatapathy@ieee.org
582 N. Lokeswari and B. Chakradhar Rao
1 Introduction
sureshsatapathy@ieee.org
Artificial Neural Network Classifier for Intrusion Detection … 583
2 Proposed Architecture
sureshsatapathy@ieee.org
584 N. Lokeswari and B. Chakradhar Rao
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.
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).
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.
sureshsatapathy@ieee.org
Artificial Neural Network Classifier for Intrusion Detection … 585
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
4 Normalization
sureshsatapathy@ieee.org
586 N. Lokeswari and B. Chakradhar Rao
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,
5 0 0 0 83 253 0
0.009 0 0 0 0.32 0.99 0
sureshsatapathy@ieee.org
Artificial Neural Network Classifier for Intrusion Detection … 587
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.
sureshsatapathy@ieee.org
588 N. Lokeswari and B. Chakradhar Rao
5.2.1 Preprocessing
Preprocessing, the first stage is shown in figure; here also the same 12 features are
considered in the section.
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).
sureshsatapathy@ieee.org
Artificial Neural Network Classifier for Intrusion Detection … 589
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
sureshsatapathy@ieee.org
590 N. Lokeswari and B. Chakradhar Rao
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).
sureshsatapathy@ieee.org
Artificial Neural Network Classifier for Intrusion Detection … 591
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
sureshsatapathy@ieee.org
Edge Detection on an Image Using Ant
Colony Optimization
1 Introduction
sureshsatapathy@ieee.org
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 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
sureshsatapathy@ieee.org
Edge Detection on an Image Using Ant Colony Optimization 595
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).
3 Optimization Technique
sureshsatapathy@ieee.org
596 P. Hinduja et al.
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.
sureshsatapathy@ieee.org
Edge Detection on an Image Using Ant Colony Optimization 597
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.
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.
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.
sureshsatapathy@ieee.org
598 P. Hinduja et al.
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.
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
sureshsatapathy@ieee.org
Edge Detection on an Image Using Ant Colony Optimization 599
6 Conclusion
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
sureshsatapathy@ieee.org
Cryptography Technique for a Novel Text
Using Mathematical Functions
1 Introduction
sureshsatapathy@ieee.org
602 P.P. Raj and Ch.S. Rao
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
sureshsatapathy@ieee.org
Cryptography Technique for a Novel Text … 603
Encryption
Plain Cipher
algorithm
text text
Key
Show in image
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
sureshsatapathy@ieee.org
604 P.P. Raj and Ch.S. Rao
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
sureshsatapathy@ieee.org
Cryptography Technique for a Novel Text … 605
Plain
text Key
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
sureshsatapathy@ieee.org
606 P.P. Raj and Ch.S. Rao
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
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.
sureshsatapathy@ieee.org
Cryptography Technique for a Novel Text … 607
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)
sureshsatapathy@ieee.org
A New Hybrid Approach for Document
Clustering Using Tabu Search and Particle
Swarm Optimization (TSPSO)
1 Introduction
sureshsatapathy@ieee.org
610 T. Haribabu and S. Jayaprada
2 Partitional Clustering
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.
sureshsatapathy@ieee.org
A New Hybrid Approach for Document Clustering Using Tabu Search … 611
The variance ratio criterion (VRC) is the mostly widely used partitioned clustering
strategy [4]:
Bd n k
VRC ¼ : ð1Þ
Wd k 1
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
sureshsatapathy@ieee.org
612 T. Haribabu and S. Jayaprada
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.
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″.
sureshsatapathy@ieee.org
A New Hybrid Approach for Document Clustering Using Tabu Search … 613
Clusters Objects
1 1,6,8
2 2,5,7
3 3,4,9
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.
sureshsatapathy@ieee.org
614 T. Haribabu and S. Jayaprada
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.
sureshsatapathy@ieee.org
A New Hybrid Approach for Document Clustering Using Tabu Search … 615
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.
4.3 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
sureshsatapathy@ieee.org
616 T. Haribabu and S. Jayaprada
sureshsatapathy@ieee.org
A New Hybrid Approach for Document Clustering Using Tabu Search … 617
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)
sureshsatapathy@ieee.org
An Effective and Efficient Clustering
Based on K-Means Using MapReduce
and TLBO
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
An Effective and Efficient Clustering Based on K-Means … 621
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.
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.
sureshsatapathy@ieee.org
622 P.K. Pedireddla and S.A. Yadwad
• 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.
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.
sureshsatapathy@ieee.org
An Effective and Efficient Clustering Based on K-Means … 623
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 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.
sureshsatapathy@ieee.org
624 P.K. Pedireddla and S.A. Yadwad
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.
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.
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
sureshsatapathy@ieee.org
An Effective and Efficient Clustering Based on K-Means … 625
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.
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.
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
sureshsatapathy@ieee.org
626 P.K. Pedireddla and S.A. Yadwad
5 Experimental Results
sureshsatapathy@ieee.org
An Effective and Efficient Clustering Based on K-Means … 627
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
sureshsatapathy@ieee.org
628 P.K. Pedireddla and S.A. Yadwad
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
sureshsatapathy@ieee.org
Detection of Sinkhole Attack in Wireless
Sensor Network
Keywords Wireless sensor networks Security RSSI LQI Intrusion detection
system Sinkhole attacks
1 Introduction
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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.
sureshsatapathy@ieee.org
632 I. Raju and P. Parwekar
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
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.
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.
sureshsatapathy@ieee.org
Detection of Sinkhole Attack in Wireless Sensor Network 633
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:
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.
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.
sureshsatapathy@ieee.org
634 I. Raju and P. Parwekar
Node Forward
Cost Alert
id Node id
sureshsatapathy@ieee.org
Detection of Sinkhole Attack in Wireless Sensor Network 635
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.
sureshsatapathy@ieee.org
636 I. Raju and P. Parwekar
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
sureshsatapathy@ieee.org
Enhancement of Stream Ciphers Security
Using DNA
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
sureshsatapathy@ieee.org
638 B. Ramesh et al.
2 Existing Work
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
sureshsatapathy@ieee.org
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:
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
sureshsatapathy@ieee.org
640 B. Ramesh et al.
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
sureshsatapathy@ieee.org
Enhancement of Stream Ciphers Security Using DNA 641
Cipher text 2
TCP/IP protocol
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.
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.
sureshsatapathy@ieee.org
642 B. Ramesh et al.
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
sureshsatapathy@ieee.org
Enhancement of Stream Ciphers Security Using DNA 643
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)
sureshsatapathy@ieee.org
Object Recognition with Discriminately
Trained Part-Based Model on HOG
(Histogram of Oriented Gradients)
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
Object Recognition with Discriminately Trained Part-Based Model … 647
3 Proposed Architecture
sureshsatapathy@ieee.org
648 T. Alekhya and S. Ranjan Mishra
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.
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:
sureshsatapathy@ieee.org
Object Recognition with Discriminately Trained Part-Based Model … 649
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
sureshsatapathy@ieee.org
650 T. Alekhya and S. Ranjan Mishra
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).
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.
5 Experimental Results
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
sureshsatapathy@ieee.org
Object Recognition with Discriminately Trained Part-Based Model … 651
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
sureshsatapathy@ieee.org
652 T. Alekhya and S. Ranjan Mishra
sureshsatapathy@ieee.org
Object Recognition with Discriminately Trained Part-Based Model … 653
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
sureshsatapathy@ieee.org
Accuracy Assessment of Images
Classification Using RBF
with Multi-swarm Optimization
Methodology
1 Introduction
sureshsatapathy@ieee.org
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
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].
sureshsatapathy@ieee.org
Accuracy Assessment of Images Classification Using RBF … 657
sureshsatapathy@ieee.org
658 G. Shyama Chandra Prasad et al.
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.
Fig. 2 Comparison of
accuracy values for case 1
sureshsatapathy@ieee.org
Accuracy Assessment of Images Classification Using RBF … 659
Fig. 3 Comparison of
accuracy values for case 2
Fig. 4 Comparison of
accuracy values for case 3
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.
Case 4: Training Data is with 500 images and the other 500 images are con-
sidered as test data Table 4 and Fig. 5 describe the results of test case 4, with the
sureshsatapathy@ieee.org
660 G. Shyama Chandra Prasad et al.
Fig. 5 Comparison of
accuracy values for case 4
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.
Case 5: Training Data is with 400 images and the other 600 images are con-
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.
Case 6: Training Data is with 200 images and the other 800 images are con-
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.
sureshsatapathy@ieee.org
Accuracy Assessment of Images Classification Using RBF … 661
Fig. 6 Comparison of
accuracy values for case 5
Fig. 7 Comparison of
accuracy values for case 6
sureshsatapathy@ieee.org
662 G. Shyama Chandra Prasad et al.
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
4 Conclusions
sureshsatapathy@ieee.org
Accuracy Assessment of Images Classification Using RBF … 663
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)
sureshsatapathy@ieee.org
664 G. Shyama Chandra Prasad et al.
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)
sureshsatapathy@ieee.org
Microstrip Patch Antenna Array
with Metamaterial Ground Plane
for Wi-MAX Applications
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
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
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
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].
sureshsatapathy@ieee.org
668 C. Arora et al.
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
sureshsatapathy@ieee.org
Microstrip Patch Antenna Array with Metamaterial Ground Plane … 669
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
sureshsatapathy@ieee.org
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)
sureshsatapathy@ieee.org
Microstrip Patch Antenna Array with Metamaterial Ground Plane … 671
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)
sureshsatapathy@ieee.org
Tracking and Speed Estimation of Moving
Vehicle for Traffic Surveillance System
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.
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.
sureshsatapathy@ieee.org
674 K. Kamakula et al.
2 Related Work
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.
sureshsatapathy@ieee.org
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.
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.
sureshsatapathy@ieee.org
676 K. Kamakula et al.
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.
Estimation
Feature of vehicle
extraction velocity
sureshsatapathy@ieee.org
Tracking and Speed Estimation of Moving Vehicle … 677
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
sureshsatapathy@ieee.org
678 K. Kamakula et al.
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
sureshsatapathy@ieee.org
Tracking and Speed Estimation of Moving Vehicle … 679
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)
sureshsatapathy@ieee.org
Authentication of Audio Signals
in Frequency Domain (AASF)
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
sureshsatapathy@ieee.org
682 U.Kr. Mondal and J.K. Mandal
2 The Technique
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.
sureshsatapathy@ieee.org
Authentication of Audio Signals in Frequency Domain (AASF) 683
Step 3: Apply step 2 of algorithm 1 for generating lower encoded values of selected
hidden signal.
Step 1: Find sampled values of original song and apply Fast Fourier Transform (FFT)
over the song for getting frequency component.
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.
sureshsatapathy@ieee.org
684 U.Kr. Mondal and J.K. Mandal
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)
over the song for getting frequency component.
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.
sureshsatapathy@ieee.org
Authentication of Audio Signals in Frequency Domain (AASF) 685
2.4 Authentication
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.
3 Experimental Results
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.
sureshsatapathy@ieee.org
686 U.Kr. Mondal and J.K. Mandal
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.
Fig. 2 Modified song after modulating amplitude and adding authenticating code
sureshsatapathy@ieee.org
Authentication of Audio Signals in Frequency Domain (AASF) 687
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.
3.2 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
sureshsatapathy@ieee.org
688 U.Kr. Mondal and J.K. Mandal
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
sureshsatapathy@ieee.org
Authentication of Audio Signals in Frequency Domain (AASF) 689
100 SNR
50 PSNR
0
BER
MOS
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
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.
sureshsatapathy@ieee.org
690 U.Kr. Mondal and J.K. Mandal
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)
sureshsatapathy@ieee.org
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
sureshsatapathy@ieee.org
692 Author Index
sureshsatapathy@ieee.org
Author Index 693
sureshsatapathy@ieee.org