Engineering Applications of Neural

Networks 17th International Conference

EANN 2016 Aberdeen UK September 2 5
2016 Proceedings 1st Edition Chrisina
Jayne
Visit to download the full and correct content document:
https://textbookfull.com/product/engineering-applications-of-neural-networks-17th-inte
rnational-conference-eann-2016-aberdeen-uk-september-2-5-2016-proceedings-1st-e
dition-chrisina-jayne/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Engineering Applications of Neural Networks 19th

International Conference EANN 2018 Bristol UK September
3 5 2018 Proceedings Elias Pimenidis

https://textbookfull.com/product/engineering-applications-of-
neural-networks-19th-international-conference-eann-2018-bristol-
uk-september-3-5-2018-proceedings-elias-pimenidis/

Web Reasoning and Rule Systems 10th International

Conference RR 2016 Aberdeen UK September 9 11 2016
Proceedings 1st Edition Magdalena Ortiz

https://textbookfull.com/product/web-reasoning-and-rule-
systems-10th-international-conference-rr-2016-aberdeen-uk-
september-9-11-2016-proceedings-1st-edition-magdalena-ortiz/

Engineering Applications of Neural Networks: 20th

International Conference, EANN 2019, Xersonisos, Crete,
Greece, May 24-26, 2019, Proceedings John Macintyre

https://textbookfull.com/product/engineering-applications-of-
neural-networks-20th-international-conference-
eann-2019-xersonisos-crete-greece-may-24-26-2019-proceedings-
john-macintyre/

Artificial Intelligence Methodology Systems and

Applications 17th International Conference AIMSA 2016
Varna Bulgaria September 7 10 2016 Proceedings 1st
Edition Christo Dichev
https://textbookfull.com/product/artificial-intelligence-
methodology-systems-and-applications-17th-international-
conference-aimsa-2016-varna-bulgaria-
Agile Processes in Software Engineering and Extreme
Programming 17th International Conference XP 2016
Edinburgh UK May 24 27 2016 Proceedings 1st Edition
Helen Sharp
https://textbookfull.com/product/agile-processes-in-software-
engineering-and-extreme-programming-17th-international-
conference-xp-2016-edinburgh-uk-may-24-27-2016-proceedings-1st-
edition-helen-sharp/

Security and Cryptography for Networks 10th

International Conference SCN 2016 Amalfi Italy August
31 September 2 2016 Proceedings 1st Edition Vassilis
Zikas
https://textbookfull.com/product/security-and-cryptography-for-
networks-10th-international-conference-scn-2016-amalfi-italy-
august-31-september-2-2016-proceedings-1st-edition-vassilis-
zikas/

Advances in Neural Networks ISNN 2016 13th

International Symposium on Neural Networks ISNN 2016 St
Petersburg Russia July 6 8 2016 Proceedings 1st Edition
Long Cheng
https://textbookfull.com/product/advances-in-neural-networks-
isnn-2016-13th-international-symposium-on-neural-networks-
isnn-2016-st-petersburg-russia-july-6-8-2016-proceedings-1st-
edition-long-cheng/

E Commerce and Web Technologies 17th International

Conference EC Web 2016 Porto Portugal September 5 8
2016 Revised Selected Papers 1st Edition Derek Bridge

https://textbookfull.com/product/e-commerce-and-web-
technologies-17th-international-conference-ec-web-2016-porto-
portugal-september-5-8-2016-revised-selected-papers-1st-edition-
derek-bridge/

Web Age Information Management 17th International

Conference WAIM 2016 Nanchang China June 3 5 2016
Proceedings Part I 1st Edition Bin Cui

https://textbookfull.com/product/web-age-information-
management-17th-international-conference-waim-2016-nanchang-
china-june-3-5-2016-proceedings-part-i-1st-edition-bin-cui/
Chrisina Jayne
Lazaros Iliadis (Eds.)

Communications in Computer and Information Science 629

Engineering Applications
of Neural Networks
17th International Conference, EANN 2016
Aberdeen, UK, September 2–5, 2016
Proceedings

123
Communications
in Computer and Information Science 629
Commenced Publication in 2007
Founding and Former Series Editors:
Alfredo Cuzzocrea, Dominik Ślęzak, and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Xiaoyong Du
Renmin University of China, Beijing, China
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Orhun Kara
TÜBİTAK BİLGEM and Middle East Technical University, Ankara, Turkey
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Ting Liu
Harbin Institute of Technology (HIT), Harbin, China
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
More information about this series at http://www.springer.com/series/7899
Chrisina Jayne Lazaros Iliadis (Eds.)
•

Engineering Applications
of Neural Networks
17th International Conference, EANN 2016
Aberdeen, UK, September 2–5, 2016
Proceedings

123
Editors
Chrisina Jayne Lazaros Iliadis
Robert Gordon University Lab of Forest Informatics (FiLAB)
Aberdeen Democritus University of Thrace
UK Orestiada
Greece

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-319-44187-0 ISBN 978-3-319-44188-7 (eBook)
DOI 10.1007/978-3-319-44188-7

Library of Congress Control Number: 2016947184

© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
 Preface

The 17th International Conference on Engineering Applications of Neural Networks

(EANN) was held at Robert Gordon University in Aberdeen, UK, during September,
2016. The supporters for the conference were the International Neural Network Society
(INNS), The Scottish Informatics and Computer Science Alliance (SICSA), Visit
Aberdeen, Visit Scotland, and Robert Gordon University in Aberdeen, UK. The 17th
EANN 2016 attracted delegates from 12 countries across the world: Czech Republic,
China, Chile, Colombia, Greece, Italy, Japan, Poland, Portugal, Russia, the UK, and
USA.
The volume includes 22 full papers, three short papers, and two tutorial papers. All
papers were subject to a rigorous peer-review process by at least two independent
academic referees. EANN 2016 accepted approximately 53 % of the submitted papers
as full papers. The authors of the best 10 papers were invited to submit extended
contributions for inclusion in a special issue of Neural Computing and Applications
(Springer). The papers demonstrate a variety of novel neural network and other
computational intelligence approaches applied to challenging real-world problems. The
papers cover topics such as: convolutional neural networks and deep learning appli-
cations, real-time systems, ensemble classification, chaotic neural networks, self-
organizing maps applications, intelligent cyber physical systems, text analysis, emotion
recognition, and optimization problems.
The following keynote speakers were invited and gave lectures on exciting neural
network application topics:
– Professor Nikola Kasabov, Director and Founder, Knowledge Engineering and
Discovery Research Institute (KEDRI), Chair of Knowledge Engineering, Auckland
University of Technology, New Zealand
– Professor Marley Vellasco, Head of the Electrical Engineering Department and the
Applied Computational Intelligence Laboratory (ICA) at PUC-Rio, Brazil
– Professor John MacIntyre, Dean of the Faculty of Appliced Sciences, Pro Vice
Chancellor Director of Research, Innovation and Employer Engagement, University
of Sunderland, UK
On behalf of the conference Organizing Committee we would like to thank all those
who contributed to the organization of this year’s program, and in particular the
Program Committee members.

September 2016 Chrisina Jayne

Lazaros Iliadis
 Organization

General Chair
Chrisina Jayne Robert Gordon University, UK

Advisory Chair
Nikola Kasabov Auckland University of Technology, New Zenland

Program Chairs
Chrisina Jayne Robert Gordon University, UK
Lazaros Iliadis Democritus University of Thrace, Greece

Local Organizing Committee Chair

Michael Heron Robert Gordon University, UK

Program Committee
A. Canuto Federal University of Rio Grande do Norte, Brazil
A. Petrovski Robert Gordon University, UK
B. Beliczynski Institute of Control and Industrial Electronics, Poland
D. Coufal Czech Academy of Sciences
D. Pérez University of Oviedo, Spain
E. Kyriacou Frederick University, Cyprus
H. Leopold Austrian Institute of Technology GmbH, Austria
I. Bukovsky Czech Technical University in Prague, Czech Republic
J.F. De Canete University of Malaga, Spain
Rodriguez
K.L. Kermanidis Ionian University, Greece
K. Margaritis University of Macedonia, Greece
M. Holena Academy of Sciences of the Czech Republic
M. Fiasche Politecnico di Milano, Italy
M. Trovati Derby University, UK
N. Wiratunga Robert Gordon University, UK
P. Hajek University of Pardubice, Czech Republic
P. Kumpulainen Tempere University of Technology, Finland
S. Massie Robert Gordon University, UK
V. Kurkova Czech Academy of Sciences
Z. Ding Hewlett Packard Enterprise, USA
VIII Organization

A. Papaleonidas Democritus University of Thrace, Greece

A. Kalampakas Democritus University of Thrace, Greece
B. Ribeiro University of Coimbra, Portugal
D. Gorse University College London, UK
E. Elyan Robert Gordon University, UK
F. Marcelloni University of Pisa, Italy
I. Bougoudis Democritus University of Thrace, UK
I. Stephanakis Hellenic Telecommunication Organization SA, Greece
K. Demertzis Democritus University of Thrace, Greece
K. Koutroumbas National Observatory of Athens, Greece
M. Gaber Robert Gordon University, UK
M. Kolehmainen Environmental Science University of Eastern, Finland
M. Tauber Austrian Institute of Technology GmbH, Austria
N. Nicolaou Imperial College London, UK
P. Gastaldo Università degli Studi di Genoa, Italy
P. Vidnerov Czech Academy of Sciences
R. Tanscheit PUC-Rio, Brazil
S. Sani Robert Gordon University, UK
Y. Manolopoulos Aristotle University of Thessaloniki, Greece

Supporting Organizations

International Neural Network Society (INNS)

The Scottish Informatics and Computer Science Alliance (SICSA)
Visit Aberdeen
Visit Scotland
Robert Gordon University, Aberdeen, UK
 Contents

Active Learning and Dynamic Environments

Deep Active Learning for Autonomous Navigation . . . . . . . . . . . . . . . . . . . 3

Ahmed Hussein, Mohamed Medhat Gaber, and Eyad Elyan

2D Recurrent Neural Networks for Robust Visual Tracking of Non-Rigid

Bodies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
G.L. Masala, B. Golosio, M. Tistarelli, and E. Grosso

Choice of Best Samples for Building Ensembles in Dynamic Environments . . . 35

Joana Costa, Catarina Silva, Mário Antunes, and Bernardete Ribeiro

Semi-supervised Modeling

Semi-supervised Hybrid Modeling of Atmospheric Pollution in Urban

Centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Ilias Bougoudis, Konstantinos Demertzis, Lazaros Iliadis,
Vardis-Dimitris Anezakis, and Antonios Papaleonidas

Classification Applications

Predicting Abnormal Bank Stock Returns Using Textual Analysis

of Annual Reports – a Neural Network Approach . . . . . . . . . . . . . . . . . . . . 67
Petr Hájek and Jana Boháčová

Emotion Recognition Using Facial Expression Images for a Robotic

Companion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Ariel Ruiz-Garcia, Mark Elshaw, Abdulrahman Altahhan,
and Vasile Palade

Application of Artificial Neural Networks for Analyses of EEG Record

with Semi-Automated Etalons Extraction: A Pilot Study . . . . . . . . . . . . . . . 94
Hana Schaabova, Vladimir Krajca, Vaclava Sedlmajerova,
Olena Bukhtaieva, Lenka Lhotska, Jitka Mohylova,
and Svojmil Petranek

Clustering Applications

Economies Clustering Using SOM-Based Dissimilarity . . . . . . . . . . . . . . . . 111

Adam Chudziak
X Contents

Elastic Net Application: Case Study to Find Solutions for the TSP
in a Beowulf Cluster Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Marcos Lévano and Andrea Albornoz

Comparison of Methods for Automated Feature Selection Using

a Self-organising Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Aliyu Usman Ahmad and Andrew Starkey

EEG-Based Condition Clustering Using Self-Organising Neural

Network Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Hassan Hamdoun and Aliyu Ahmad Usman

Cyber-Physical Systems and Cloud Applications

Intelligent Measurement in Unmanned Aerial Cyber Physical Systems

for Traffic Surveillance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Andrei Petrovski, Prapa Rattadilok, and Sergey Petrovskii

Predictive Model for Detecting MQ2 Gases Using Fuzzy Logic on IoT
Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Catalina Hernández, Sergio Villagrán, and Paulo Gaona

A Multi-commodity Network Flow Model for Cloud Service Environments . . . 186

Ioannis M. Stephanakis, Syed Noor-Ul-Hassan Shirazi,
Antonios Gouglidis, and David Hutchison

Designing a Context-Aware Cyber Physical System for Smart Conditional

Monitoring of Platform Equipment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
Farzan Majdani, Andrei Petrovski, and Daniel Doolan

Time-Series Prediction

Convolutional Radio Modulation Recognition Networks . . . . . . . . . . . . . . . 213

Timothy J. O’Shea, Johnathan Corgan, and T. Charles Clancy

Mutual Information with Parameter Determination Approach for Feature

Selection in Multivariate Time Series Prediction . . . . . . . . . . . . . . . . . . . . . 227
Tianhong Liu, Haikun Wei, Chi Zhang, and Kanjian Zhang

Learning-Algorithms

On Learning Parameters of Incremental Learning in Chaotic Neural

Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Toshinori Deguchi and Naohiro Ishii
 Contents XI

Accelerated Optimal Topology Search for Two-Hidden-Layer Feedforward

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Alan J. Thomas, Simon D. Walters, Miltos Petridis,
Saeed Malekshahi Gheytassi, and Robert E. Morgan

An Outlier Ranking Tree Selection Approach to Extreme Pruning

of Random Forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Khaled Fawagreh, Mohamed Medhat Gaber, and Eyad Elyan

Lower Bounds on Complexity of Shallow Perceptron Networks . . . . . . . . . . 283

Věra Kůrková

Kernel Networks for Function Approximation. . . . . . . . . . . . . . . . . . . . . . . 295

David Coufal

Short Papers

Simple and Stable Internal Representation by Potential Mutual Information

Maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Ryotaro Kamimura

Urdu Speech Corpus and Preliminary Results on Speech Recognition . . . . . . 317

Hazrat Ali, Nasir Ahmad, and Abdul Hafeez

Bio-inspired Audio-Visual Speech Recognition Towards the Zero

Instruction Set Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
Mario Malcangi and Hao Quan

Tutorials

Classification of Unbalanced Datasets and Detection of Rare Events

in Industry: Issues and Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Marco Vannucci and Valentina Colla

Variable Selection for Efficient Design of Machine Learning-Based

Models: Efficient Approaches for Industrial Applications . . . . . . . . . . . . . . . 352
Silvia Cateni and Valentina Colla

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

Active Learning and Dynamic
Environments
 Deep Active Learning for Autonomous
Navigation

Ahmed Hussein(B) , Mohamed Medhat Gaber, and Eyad Elyan

School of Computing, Robert Gordon University,

Garthdee Road, Aberdeen AB10 7QB, UK
a.s.h.a.hussein@rgu.ac.uk

Abstract. Imitation learning refers to an agent’s ability to mimic a

desired behavior by learning from observations. A major challenge fac-
ing learning from demonstrations is to represent the demonstrations in a
manner that is adequate for learning and efficient for real time decisions.
Creating feature representations is especially challenging when extracted
from high dimensional visual data. In this paper, we present a method
for imitation learning from raw visual data. The proposed method is
applied to a popular imitation learning domain that is relevant to a vari-
ety of real life applications; namely navigation. To create a training set,
a teacher uses an optimal policy to perform a navigation task, and the
actions taken are recorded along with visual footage from the first per-
son perspective. Features are automatically extracted and used to learn
a policy that mimics the teacher via a deep convolutional neural net-
work. A trained agent can then predict an action to perform based on
the scene it finds itself in. This method is generic, and the network is
trained without knowledge of the task, targets or environment in which it
is acting. Another common challenge in imitation learning is generalizing
a policy over unseen situation in training data. To address this challenge,
the learned policy is subsequently improved by employing active learn-
ing. While the agent is executing a task, it can query the teacher for
the correct action to take in situations where it has low confidence. The
active samples are added to the training set and used to update the ini-
tial policy. The proposed approach is demonstrated on 4 different tasks
in a 3D simulated environment. The experiments show that an agent can
effectively perform imitation learning from raw visual data for naviga-
tion tasks and that active learning can significantly improve the initial
policy using a small number of samples. The simulated testbed facilitates
reproduction of these results and comparison with other approaches.

1 Introduction
One of the important aspects of artificial intelligence is the ability of autonomous
agents to behave effectively and realistically in a given task. There is a rising
demand for applications in which agents can act and make decisions similar to
human behavior in order to achieve a goal. Imitation learning is a paradigm in
which an agent learns how to behave by observing demonstrations of correct

c Springer International Publishing Switzerland 2016
C. Jayne and L. Iliadis (Eds.): EANN 2016, CCIS 629, pp. 3–17, 2016.
DOI: 10.1007/978-3-319-44188-7 1
4 A. Hussein et al.

behavior provided by a teacher. In contrast to explicit programming, learning

from demonstrations does not require knowledge of the task to be integrated in
the learning process. It favors a generic learning process where the task is learned
completely from observing the demonstrations. Thus, an intelligent agent can be
trained to perform a new task simply by providing examples. Since an agent is
able to learn complex tasks by mimicking a teacher’s behavior, imitation learning
is relevant to many robotic applications [2,4,6,11,13,24,29,36] and is considered
an integral part in the future of intelligent robots [31].
One of the biggest challenges in imitation learning is finding adequate repre-
sentations for the state of the agent in its environment. The agent should be able
to extract meaningful information from sensing of its surroundings, and utilize
this information to perform actions in real time. Deep learning methods have
recently been applied in a wide array of applications and are especially successful
in handling raw data. One of the most popular deep learning techniques is Con-
volutional Neural Networks (CNNs). CNNs are particularly popular in vision
applications due to their ability to extract features from high dimensional visual
data. The ability of deep networks to automatically discover patterns provides
a generic alternative to engineered features which have to be designed for each
specific task. For instance traditional planning approaches that use computer-
vision methods of object recognition and localization need to tailor the methods
for every individual target and task. CNNs achieve results competitive with the
state of the art in many image classification tasks [8,17] and have been recently
used to learn Atari 2600 games from raw visual input [20,21]. These and other
recent attempts have shown that deep learning can be successful in teaching an
agent to perform a task from visual data. However, most studies focus on 2D
environments with stationary views; which does not reflect real world applica-
tions. Moreover, direct imitation is performed without considering refining the
policy based on the agent’s performance. To the best of our knowledge, training
an agent from raw visual input using deep networks and active learning in a 3D
environment has not been done.
In this paper we present a novel method that utilizes deep learning and active
learning to train agents in a 3D setting. The method is demonstrated on sev-
eral navigation tasks in a 3D simulated environment. Navigation is one of the
most explored domains in imitation learning due to its relevance to many robotic
applications, such as flying [1,23,29] and ground vehicles [7,26,27,32]. Naviga-
tion is also an essential base task in high degree of freedom robots (e.g. humanoid
robots) [7,30]. We propose a generic method for learning navigation tasks from
demonstrations that does not require any prior knowledge of the task’s goals,
environment or possible actions. A training set is gathered by having a teacher
control the agent to successfully perform the task. The controlled agent’s view of
the 3D environment is captured along with the actions performed in each frame.
A deep convolution network is used to learn visual representation from the cap-
tured video footage and learn a policy to mimic the teacher’s behavior. We also
employ active learning to improve the agents policy by emphasizing situations in
which it is not confident. We show that active learning can significantly improve
 Deep Active Learning for Autonomous Navigation 5

the policy with a limited number of queried instances. Once trained, the agent is
able to extract features from the scene and predict actions in real time. We con-
duct our experiments on benchmark testbed that makes it seamless to replicate
our results and compare with other approaches.
Benchmark environments are useful tools for evaluating intelligent agents.
A few benchmarks are available for 2D tasks such as [3,15,25] and are being
increasingly employed in the literature. 3D environments however have not been
as widely explored, although they provide a closer simulation to real robotic
applications. We use mash-simulator [19] as our testbed to facilitate the evalua-
tion and comparison of learning methods. It is also convenient for extending the
experiments to different navigation tasks within the same framework.
In the next section we review related work. Section 3 describes the proposed
methods. Section 4 details our experiments and results. Finally we present our
conclusions and discuss future steps in Sect. 5.

2 Related Work
2.1 Navigation
Navigation tasks have been of interest in AI in general and imitation learning
specifically from an early stage. Sammut et al. [29] provides an early exam-
ple of an aircraft learning autonomous flight from demonstrations provided via
remote control. Later research tackle more elaborate navigation problems includ-
ing obstacles and objects of interest. Chernova et al. [7] use Gaussian mixture
models to teach a robot to navigate through a maze. The robot is fitted with
an IR sensor to provide information about the proximity of obstacles. This data
coupled with input from a teacher controlling the robot is used to learn a policy.
The robot is then able to make a decision to execute one of 4 motion primi-
tives(unit actions) based on its sensory readings. In [10] the robot uses a laser
sensor to detect and recognize objects of interest. A policy is learned to predict
subgoals associated with the detected objects rather than directly predicting
the motion primitives. Such sensing methods provide an abstract view of the
environment, but can’t convey visual details that might be needed for intelligent
agents to mimic human behavior. [22] use neural networks to learn a policy for
driving a car in racing game using features extracted from the game engine (such
as position of the car relative to the track). Driving is a complex task compared
to other navigation problems due to the complexity of the possible actions. The
outputs of the neural network in [22] are high DOF low level actions. However,
the features extracted from the game engine to train the policy would be dif-
ficult to extract in the real world. Advances in computational resources have
prompted the use of visual data over simpler sensory data. Visual sensors pro-
vide detailed information about the agents surrounding and are suitable to use in
real world applications. In [28] a policy for a racing game is learned from visual
data. Demonstrations are provided by capturing the games video stream and the
controller input. The raw frames (downsampled) without extracting engineered
features are used as input to train a neural network.
6 A. Hussein et al.

2.2 Deep Learning

Deep learning methods are highly effective in problems that don’t have estab-
lished sets of engineered features. CNNs have been used with great success to
extract features from images. In recent studies [20,21] CNNs are coupled with
reinforcement learning to learn several Atari games. A sequence of raw frames is
used as input to the network and trial and error is used to learn a policy. Trial
and error methods such as reinforcement learning have been extensively used
to learn policies for intelligent agents [16]. However, providing demonstrations
of correct behavior can greatly expedite the learning rate. Moreover, learning
through trial and error can lead the agent to learn a way of performing the
task that doesn’t seem natural or intuitive to a human observer. In [12] learn-
ing from demonstrations is applied on the same Atari benchmark. A supervised
network is used to train a policy using samples from a high performing but non
real time agent. This approach is reported to outperform agents that learn from
scratch through reinforcement learning. Other examples of using deep learning
to play games include learning the game of ‘GO’ using supervised convolution
networks [9] and a combination of supervised and reinforcement learning [33].
These examples all focus on learning 2D games that have a fixed view. However
in real applications, visual sensors would capture 3D scenes, and the sensors
would most likely be mounted on the agent which means it is unrealistic to have
a fixed view of the entire scene at all times.
In [18] a robot is trained to perform a number of object manipulation tasks.
First a trajectory is learned using reinforcement learning with the position of
the objects and targets known to the robot. These trajectories then serve as
demonstrations train a supervised convolutional neural network. In this case no
demonstrations are needed to be provided by a teacher. However, this approach
requires expert knowledge for the initial setup of the reinforcement learning
phase. Compared to related work that employs deep learning to teach an intel-
ligent agent, this is a realistic application implemented with a physical robot.
However, the features are extracted from a set scene with small variations. This
is different from applications where the agent moves and turns around, and with
that completely altering it’s view.

2.3 Active Learning

In many imitation learning applications direct imitation is not sufficient for

robust behavior. One of the common challenges facing direct imitation is that the
training set doesn’t fully represent the desire task. The collected demonstrations
only include optimal actions performed by the teacher. If the agent makes an
error it arrives at a state that was not represented in its learned policy [35].
It is therefore necessary in many cases to provide further training to an agent
based on its own performance of the task. One of the methods to enhance a
trained agent is active learning. Active learning relies on querying a teacher
for the correct decision in cases where the trained model performs poorly. The
teacher’s answers are used to improve the model in its weakest areas. In [7]
 2D Recurrent Neural Networks for Robust Visual
Tracking of Non-Rigid Bodies

G.L. Masala ✉ , B. Golosio, M. Tistarelli, and E. Grosso

( )

Department of Political Science, Communication,

Engineering and Information Technologies - Computer Vision Laboratory, University of Sassari,
Sassari, Italy
{gilmasala,golosio,tista,grosso}@uniss.it

Abstract. The efficient tracking of articulated bodies over time is an essential

element of pattern recognition and dynamic scenes analysis. This paper proposes
a novel method for robust visual tracking, based on the combination of image-
based prediction and weighted correlation. Starting from an initial guess, neural
computation is applied to predict the position of the target in each video frame.
Normalized cross-correlation is then applied to refine the predicted target posi‐
tion.
Image-based prediction relies on a novel architecture, derived from the
Elman’s Recurrent Neural Networks and adopting nearest neighborhood connec‐
tions between the input and context layers in order to store the temporal infor‐
mation content of the video. The proposed architecture, named 2D Recurrent
Neural Network, ensures both a limited complexity and a very fast learning stage.
At the same time, it guarantees fast execution times and excellent accuracy for
the considered tracking task. The effectiveness of the proposed approach is
demonstrated on a very challenging set of dynamic image sequences, extracted
from the final of triple jump at the London 2012 Summer Olympics. The system
shows remarkable performance in all considered cases, characterized by changing
background and a large variety of articulated motions.

Keywords: Recurrent neural network · Tracking · Video analysis

1 Introduction

Motion has been one of the main cues studied in Computer Vision and Pattern Recog‐
nition. As stated by David Marr [1]: “Motion pervades the visual world, a circumstance
that has not failed to influence substantially the process of evolution. The study of visual
motion is the study of how information about only the organization of movement in an
image can be used to make inferences about the structure and the movement of the
outside world”.
In many cases motion enables three-dimensional perception (consider for example
the counter-rotating cylinders effect described by Ullman [2]) and, in its simplest form,
explains the unrivaled ability of humans to perform scene segmentation and object
tracking [3].

© Springer International Publishing Switzerland 2016

C. Jayne and L. Iliadis (Eds.): EANN 2016, CCIS 629, pp. 18–34, 2016.
DOI: 10.1007/978-3-319-44188-7_2
 2D Recurrent Neural Networks for Robust Visual Tracking 19

Object tracking can be defined as the estimation of the trajectory of an object in the
image plane as it moves around a scene. Errors in tracking are often due to abrupt changes
in object motion, changes in the objects appearance, non-rigidity of the object, occlu‐
sions and non-linear camera motion [4]. The robustness of the representation of target
appearance, against these and other unpredictable events, is crucial to successfully track
objects over time. Interestingly, assumptions are often made to constrain the tracking
problem within the context of a particular application.
Recent tracking algorithms are classified into two major categories, based on the
learning strategy adopted: generative and discriminative methods. Generative methods
describe the target appearance by a statistical model estimated from the previous frames.
To maintain the integrity of the target appearance model, various approaches have been
proposed, including sparse representation [5, 8, 9], on-line density estimation [10]. On
the other hand, discriminative methods [11, 13] directly implement classifiers to discrim‐
inate the target from the surrounding background. Several learning algorithms have been
adopted, including on-line boosting [13], multiple instance learning [11], structured
support vector machines [12] and random forests [14, 15]. These approaches are often
limited by the adoption of hand-crafted features for target representation, such as iconic
templates, Haar-like features, histograms and others, which may not generalize well to
handle the challenges arising in video sequences from everyday life scenes.
In this paper, an original method is proposed for robust visual tracking, based on a
combination of image prediction and weighted correlation matching techniques. Image
prediction is based on a novel recurrent neural network which can be easily generalized
to track any visual pattern in dynamic scenes. The proposed approach is derived from
both Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs).
A Recurrent Neural Network (RNN) is an artificial neural network with feedback
connections between nodes, with the capability to model dynamic systems [16]. Elman’s
neural network, also known as Simple Recurrent Network (SRN), is a partially recurrent
neural network first proposed by Elman [17]. Because of the context neurons and local
recurrent connections between the context layer and the hidden layer, the Elman’s neural
network has several dynamic advantages over a static neural network. Training and
convergence of SRNs usually take a long time, which makes them useless in time critical
applications [10] and/or when dealing with high resolution images. Therefore, to effi‐
ciently process high resolution images, a compromise between the representation power
and the dimension of the network must be sought.
A CNN is a feed-forward artificial neural network where individual neurons are
arranged to respond to overlapping regions in the visual field. A CNN consists of
multiple layers of small neuron collections taking as input small overlapping areas of
the image. The outputs of each layer are tiled and overlapping to better represent the
original image. This feature ensures a reasonable invariance to planar translation on the
image plane [18]. Due to their representation power, Convolutional Neural Networks
have recently attracted a considerable attention in the Computer Vision community [7],
particularly for image- and video-based recognition. However, only few attempts can
be found in the literature to employ CNNs for visual tracking. One reason is that off-
line classifiers require a model of the objects class. On the other hand, performing on-
line learning based on CNNs is not straightforward, due to the large network size and
Another random document with
no related content on Scribd:
which a sliding element containing a fulminate strikes. The sliding
block carries a small charge of black powder which is set off by the
fulminate, thus igniting the train which leads to the high explosive
charge detonator. Were this sliding block left free to slide back and
forth at all times it would be unsafe to transport the fuze, as it might
be set off by accident. There must be therefore some means of
holding it safely away from the anvil until it is desired to detonate the
charge. There are thus two conflicting conditions to be met: safety
during transportation and sensitiveness at the point of departure. It
may not be understood at first why sensitiveness at the point of
departure should be a condition to be met. Suffice it to say that all
fuzes are designed to arm at discharge or soon after leaving the
bore for they must be ready to act at any time after leaving the
muzzle. Were they to be safe during flight they might be so safe that
the remaining velocity would not be sufficient to set them off. All
fuzes are designed to arm as we say either during travel through the
bore or immediately after.

Methods of Arming.
Spring method.—Let us suppose that after our projectile has
started on its way the sliding block is free to move within a cavity at
the forward end of which is the anvil. If the projectile comes to a
sudden drop or even sudden reduction of velocity the block if
unrestrained will, according to the principle of inertia, keep on going
till something stops it. The something in this case is the anvil and the
fulminate cap is set off. But it is not so simple. For while the projectile
is in flight it is acted upon by the air resistance and slows down but
the block in the cavity of the head is not subjected to this resistance.
It therefore gains on the projectile or creeps forward in the cavity
unless restrained as it is by a spring. Now one more point and this
type of fuze is complete. We supposed that our block was free to
slide. For safety’s sake it is pinned to the cavity. Again we call upon
inertia to bread the pin so as to leave the block free to slide. The
strength of the pin is calculated so that the force of inertia of the
mass of the block is greater than the resistance of the safety pin and
when the projectile starts the pin breaks and the spring forces the
block to the rear of the cavity until the sudden stop of the projectile
permits the block to slide forward as explained. Such a fuze requires
a comparatively high initial velocity and is not adapted to howitzers
using low muzzle velocities.
There are three other methods in use to arm the fuze. They are
inertia of a sleeve; centrifugal force and powder pellet system,
that is, combustion of a grain of powder holding the sliding block
from the anvil by means of an arm resting against the unburned
powder grain. These are more sensitive than the type described.
In the first system, a sleeve fitting around the plunger carrying
the cap slides to the rear by inertia when the projectile starts and two
clips engage in notches on the plunger body making the sleeve and
plunger thereafter move as one body, they are thus held together by
a plunger spring which before arming held the plunger away from the
anvil. The safety spring held the sleeve and plunger away from the
anvil and after arming prevents forward creeping by the plunger and
sleeve now locked together. Upon striking, the plunger and sleeve
move forward as one body and the cap strikes the anvil.
In centrifugal systems the primer plunger is kept safely away
from the anvil by a lock which is kept in place by springs. When the
rotational velocity reaches a certain point the force of the springs is
overcome by the centrifugal force and the locks are thrown aside or
opened and the plunger is free to move forward on impact.
In the powder pellet system (the one largely used by the
Germans) there is a well or channel filled with compressed powder,
this is set off by a fulminate cap which is fired by inertia, a small
plunger-anvil striking the cap. When the powder is consumed it
leaves a channel into which an arm attached to the sliding block
carrying the igniting fulminate for the charge may slide, thus
permitting the block to slide forward to the anvil fixed in the forward
part of the cavity. It is held from creeping forward after the
compressed powder is burned by a safety spring, thus insuring
sufficiently hard an impact to set off the cap.
 Heretofore in our service the fulminating cap has been fixed and
the plunger carried the anvil or as we call it the firing pin. Such is
now the system in our base detonating fuzes, and in our combination
fuze.
The new point detonating fuzes are patterned after the French and
are practically French fuzes.

Fuzes Classification.
Fuses are classified as:
(a) Percussion if it acts on impact, producing a low order of
explosion.
(b) Time when it acts in the air at a certain point of the trajectory.
(c) Combination if it is able to act in the air or upon impact.
(d) Detonating when it contains a fulminate which will bring about
detonation upon impact.
The detonator may be separate or incorporated in the fuse. For
the 75-mm gun and the 155-mm howitzer it forms a part of the fuze.
Many fuzes are armed on set-back. An exception to this is the long
detonating fuse, MK 111, which is armed by the unrolling of a brass
spiral holding together two half rings made of steel so fitted as to
prevent the anvil and the head of the fuse from getting close
together. The spiral unrolls when the rotational velocity of the
projectile reaches a certain speed, thus drawing away the two steel
rings and arming the fuse.
 DETONATING FUZE—MARK-III.

DETONATING FUZE—MARK-V.

It is of great importance that the spiral spring be not unrolled

during transportation or storage. This is prevented by winding a tape
of tarred canvas around the spirals, the head being covered by a thin
band of tinfoil. Just before loading the projectile the head and tape
are removed by pulling the free end of the tape.
 The following precautions concerning fuses must be rigidly
observed to prevent grave accidents:
1. All detonators and detonating point fuses must be fitted with a
felt washer underneath, thus insuring proper seating in the central
tube.
2. Never disassemble a fuse by unscrewing.
3. Any fuse, the parts of which have become accidentally
unscrewed, must be destroyed at once. If fired it may cause a
premature burst; if handled a grave accident may result.
4. Any fuse or projectile which has been fired is dangerous,
because it may then be able to detonate by a very slight shock. It is
forbidden to touch it.
5. Never remove the tin hood from the long fuse before having
screwed the fuse in the central tube.
6. After having removed the tin hood, be sure that the spiral is in
its proper position. Never use a long fuse without the spiral.
7. Be sure the men understand that this spiral must not be
removed. It has happened that men have removed this spiral,
thinking that it was a device similar to the safety ring in trench mortar
fuzes, MK VII E.
8. See that the ring of the long fuze which connects the powder
train to the fuze body cannot be unscrewed. If it can be unscrewed
the fuze should be sent back to the depot.
9. If it is necessary to remove a shell with a long fuze by means of
the rammer, be sure to have a special rammer cup in the shape of a
hollow cylinder of wood which will fit between the shell and the
rammer.
10. Time and combination fuzes cannot be made absolutely water-
tight; the cover must therefore not be removed until the projectile is
about to be loaded.

Fuse Tables.
 Tables showing American and French fuses to be used by our
Field Artillery, with information concerning markings, color, time of
delay, size of fuse, etc.
DETONATING FUSES.
Size
Corresponding
Time of delay. Color. of Cannon.
to.
Fuse.
2- 3” gun for target
MK I White head. Short. Russian 3GT. practice
100 only.
2-
M II (non delay) 8”, 9.2”,
100
2-
MK II (non delay) White top. Short. 204-m/m.
100
5- Gun and
MK II (short delay) Black top. Short. Modified.
100 Howitzer.
15-
M II (long delay) Black head. Short. Russian.
100
75 G; 3.8”G and
H; 4.7 in. G
MK III
zero No color. Long. French IAL. and H; 6”H;
(Supersensitive)
155H; all
gas shells.
2- French 24/31
MK IV (non-delay) White top. Short. SR (99- Howitzer only.
100 15).
5- French 24/31
MK IV (short delay) Black top. Short. AR (99- Howitzer only.
100 15).
Black top
15- French 24/31
violet
MK IV (long delay) Short. SR (99- Howitzer only.
100 detonator
15).
socket.
2- French 24/31
All guns, but not
MK V (non-delay) White top. Short. SR (99-
100 Howitzers.
08).
MK V (short delay) 5- Black top. Short. French 24/31 All guns, but not
AR (99- Howitzers.
 100 08).
Mark—VII (non 2-
White. Short 6” T. M.
delay) 100
Black top with
20- violet
Mark VII (long delay) Short 6” T. M.
100 detonator
socket

Letter “E” after mark VII indicates safety device.

Note:—All American point detonating fuses are stamped on head cap in letters
and figures, .125 in high, with name of use, amount of delay, initials of loader, lot
and number; thus: PDF. MIV, xx Delay, FA, Lot No. xx.

45-SECOND COMBINATION FUZE MARK I.

 21 SECOND COMBINATION FUZE MODEL OF 1907 M.

COMBINATION FUSES.
Total time By what
Corresponding On what projectile Wt. of
Fuse. burning cannon
French Type. used. fuse.
Sec. fired.
21 s/comb. F. All 3” and
22/31M 1897, Com. Shrapnel.
A., 1907 21 75-mm 1¼ lbs.
24 sec. MKi.
M. guns
21 s/comb. 22/31M 1916, All 3” and
Com. Shrapnel.
F.A., 21 24 sec. 75-mm 1¼ lbs.
MKi.
1915 AA. guns
31 s/comb. F. 30/55M 1889,
31 Com. Shrapnel. 4.7” gun. 2 lbs.
A. 1915 40 sec.
45 s/comb. F.
Same as
A. 1907 45
above.
M.
30/55M 1889, Com. Shrapnel,
155 How.
40 sec. MKi.
 30/55M 1913, C. S. Shell AA 4.7” gun
40 sec. MKiii AA. Anti-
AA. Shrapnel. aircraft.

ACTION OF AMERICAN AND FRENCH DETONATING FUSES.

Time zero 1/ 2/ 5/ 15/
100 100 100 100
No Black with
Color Red. White. Black.
color. violet socket.
MK ii
American MKii None being made. MK i MK ii (LD)
(SD)
MK ii MK ii
Detonating Fuse is considered MK iv (LD)
(ND) (SD)
MK iV MK V
unsafe
(ND) (SD)
MK iV MK V
safe
(ND) (SD)
Will be abandoned MK V
Fuses
by French (ND)
French
detonating iAL. 1 SR. AR. LR.
fuses

Notes on Ammunition Marking.

Marks on H. E. Shell. These are of two kinds.
(a) Stamped marks made with a steel punch on the body of the
projectile just above the rotating band. These refer to the
manufacture of the projectile.
(b) Painted marks or bands which are clearly visible. They refer to
the loading, to the weight of the projectile and to the special
purposes for which the projectile is to be used.
Painted marks referring to loading are found on the ogive.
H. E. shells are usually painted red.
 Marks referring to weight are painted in black just above the
rotating band, as follows:
L.— very light.
+— light.
++— normal.
+++— heavy.
++++— very heavy.
A white cross above these marks means that a plate has been
welded on the base. These marks are also painted on the boxes.
Shells fitted with cartridge cases (fixed ammunition) are not
painted below the rotating bands.
Special Shell.
Incendiary shells.—These incendiary shells are filled with some
flame-producing liquid, alumino thermic charge or incendiary cylinder
composed of slow burning linstock and string strongly impregnated
with saltpeter.
Markings.—Green with red ogive.
All shells containing black powder are more or less incendiary.
Percussion shrapnel is incendiary.
75-mm Ammunition.
 Star Shells.—For 155 howitzer. Upon bursting, they liberate eight
white stars fitted with silken parachutes. The stars are projected
backward through the base of the projectile at the point of burst. The
parachutes open, the stars descending very slowly, illuminating the
surrounding objects for about 45 sec. The best height of burst is
about 300 m.; the burst interval should not be over 300. These shells
are also incendiary. Markings: a blue star and an “E.”
Gas shells are either toxic or tear-producing.
(a) Toxic shells are numbered either 4 or 5. The liquids 4 and 5
volatize, immediately upon contact with the air. The gases are
quickly diffused and easily carried by the wind.
Effect.—Liquid 4 acts immediately and is felt instantly.
Liquid 5, on the contrary, works more slowly and its effects are
apparent only after several hours. Markings: Green with white bands,
and numbers 4 or 5 on the ogive.
(b) Tear shells.—These shells are numbered 11, 12 and 13. They
are filled with two liquids, either mixed or separated, one liquid being
tear producing, the other smoke producing. When the shell bursts, a
greater part of the liquid is volatilized, the remainder being projected
to the ground in small drops which volatize with variable speed.
Markings: Green with numbers 11, 12 or 13 on the ogive.
Tracer shell.—This shell is fitted with a time fuse which ignites the
inside charge, the flames of which pass through the holes in the
ogive thus tracing the trajectory. Tracer shells are used in fire for
adjustment on aircraft. They are also incendiary. Markings: White
with blue ogive. Letter “T” painted on body.

PRECAUTIONS IN SEPARATE LOADING

PROJECTILES.
All projectiles must be seated accurately and carefully in loading,
otherwise not only inaccurate fire will result but also premature
detonations may occur.
 Rotating bands should be smoothed and lightly greased just
before loading. In transport and in storage the bands should be
protected by rope bands, straw tithes, etc., to prevent deformation.
 CHAPTER XIII
CARE AND PRESERVATION.

OILS AND CLEANING MATERIAL, TOOLS AND

ACCESSORIES FOR ARTILLERY MATERIEL
WITH THEIR USE.
In order that all parts of the materiel may function easily, it is
necessary that all the working and bearing surfaces may be properly
cleaned and lubricated with the appropriate lubricant. Where such
surfaces are not directly accessible, oil holes are provided; these
holes should be kept free from grit and dirt. Except during oiling, they
should be kept fully closed by the means provided.
For use in service, for the cleaning and preservation of this
materiel, the ordnance department issues hydroline oil, lubricating
oil, clock oil, vaseline, sperm oil, coal oil, neat’s-foot oil and light
slushing oil. Each of these oils are suited for the particular purpose
for which it is issued, as stated below, and care should be taken that
it is not used for other purposes.
Hydroline oil.—Used in the recoil cylinders of the carriage and for
no other purpose. Never used as a lubricant. It is characterized by its
low freezing point and its non-corrosive action on metals.
Lubricating oil (Engine oil Number 1).—A light petroleum oil used
exclusively in all oil holes of the materiel, and in lubricating such
parts as wheels and axles, guns and cradle slides, cradle pintle and
socket elevating and traversing mechanisms, exterior of cylinders,
brake bearings, hinges, different surfaces of breechblocks, threads,
breech recess, et cetera.
Clock oil.—Used on the spindle and all gearings of the Battery
Commander’s telescope, bearings of the panoramic sight, range
quadrants and fuze setters. In all cases clock oil should be used only
when the instruments mentioned are disassembled for cleaning. It
should be applied by dropping from the end of the dropper attached
to the end of the cork. In case of emergency, use as a substitute
either sperm oil or Engine oil No. 1, in the order mentioned.
Vaseline (Petrolatum).—The heavy petroleum oil free from rosin.
Used on the worm gears and the worm racks of the panoramic sight,
the hand and bracket fuze setter, B. C. telescope, and on the
micrometer screw and bushing of the quadrant. The spare parts of
the breech mechanism should also be coated with vaseline and each
piece then wrapped in paper to prevent the oil from being rubbed off.
Sperm oil.—A lighter lubricant than the lubricating oils, and may
be used on the gears of sights, fuze setters, ranges, quadrants, parts
of revolvers, et cetera; lubricating oil may also be used on such
parts. It is also used as a temporary rust preventive. Its low viscosity
and light body make it unsuitable for this purpose for more than a
few days.
Coal oil.—Used for cleaning purposes. In the field it may be used
for lanterns. Coal oil for general illuminating purposes is furnished by
the quartermaster department.
Neat’s-foot Oil.—An animal oil used for softening and preserving
leather. Applied with a moistened cloth to the flesh side of moistened
leather.
Light slushing oil.—The heavy petroleum oil similar to cosmic.
Used as a rust preventive. Essentially a mineral oil containing a large
per cent of rosin. Prescribed for use in the protection and
preservation of all bright or unpainted of steel or iron on all parts of
the equipment when the materiel is to remain unused for an
appreciable length of time. Its use as a lubricant for mobile artillery is
forbidden. Before applying the slushing oil to any surface, the parts
should be thoroughly cleaned so as to be free from rust, coal oil,
lubricating oil, et cetera, as their presence will cause rusting under
the slushing oil. The slushing oil should then be applied in a thin,
uniform coat, since this is all that is necessary to give good
protection. Except in very cold weather it can be applied by using a
paint brush as when painting, in cold weather it should be applied by
stippling—that is, by holding the brush perpendicular to the surface
to be coated and then tapping the surface with the point of the brush.
It can be applied through the bore of the gun by a slush brush issued
for that purpose. In cold weather it should be warmed before used in
the bore of the gun. It may be readily removed by burlap or waste
dipped coal oil.
Borax.—Issued for use as a flux in welding.
Lavaline.—A metal polish issued interchangeable with Gibson’s
soap polish. Used on bits and collars.
Lye, powdered.—When dissolved in water, one pound to six
quarts with sufficient lime to give a consistence of paint, is used to
remove old and blistered paint.
Napthaline.—A moth preventive, effective only after eggs and
grubs already present have been removed. Used in the storage of
blankets, et cetera.
Polish, Gibson’s Soap.—A metal polish issued interchangeably
with lavaline. Used on bits and collars.
Paint, rubberine.—Used in connection with loading ammunition in
accordance with instructions regarding the same.
Primer, brown enamel.—A hard, quick drying enamel used for
painting parts of horse collars, draft springs, et cetera.
Sal Soda, Bicarbonate of Soda.—A saturated solution of soda
and water makes an alkaline solution that will not rust. The solution
must be saturated, that is, at least 20% or one-fourth pound of soda
(6 heaping spoonfuls to one cup of water). This solution is an
effective solvent of powder fouling and should always be used after
firing, whether metal fouling solution is to be used or not. It reduces
the labor of cleaning with oil alone by more than half. Used also in a
weaker solution (one-half pound to 8 quarts of water) in washing
surfaces to be painted and to remove dirt and grease.
Soap, H. & H.—A neutral naphtha soap used in washing blankets,
web and cloth equipment. Applied in the form of a solution (one cake
to 9 cups of hot water). If for any cause this soap is not obtainable, a
good laundry soap (ivory or equal) may be used, but in no case
should yellow soap containing a large percentage of alkali be used.
Soap, castile.—An alkaline soap used in cleaning leather
equipment. Applied on a moistened sponge.
Soap, saddle (Hollingshead).—A soap used as a dressing for
leather equipment. Applied with a thick lather on a moistened
sponge.
Swabbing solution, contains.—Ammonium persulphate, 60
grains or one half spoonful smoothed off. Ammonia 28%, 6 oz. or ⅜
of a pint or 12 spoonfuls. Water, 4 oz. or ¼ pint or 8 spoonfuls.
Dissolve the ammonium persulphate in the water and add the
ammonia. Keep in a tightly corked bottle. Pour out only what is
necessary at a time and keep the bottle corked.

TOOLS AND ACCESSORIES.

In the repair of all equipment, it is literally true that “a stitch in time
saves nine,” and that a timely repair will save the entire article.
Tool Kits will be kept complete and serviceable; edges of cold
chisels free from nicks; drifts and punches properly shaped
immediately after using; and files kept clean.
To prevent unscrewing, copper wire is used to lash nuts and other
threaded parts which are not secured by split pins.
Contents of Leather Pouch for Spare Parts (carried in Trail
Boxes of 3-inch Guns):—
For Breechblock—
50 Split pins
1 Block latch and spring
1 Firing pin and spring
1 Firing pin sleeve
2 Handy oilers, 5-16 inch
Hinge pin catch
 1
1 Lever latch spring
1 Locking bolt, nut and pin
1 Locking bolt spring
2 Oil hole covers with screws
1 Pallet pin
1 Sear
2 Trigger shaft detent
For Hand Fuze Setter—
2 Corrector scale screws
1 Guide plate lock screw
2 Index bar screws
1 Index plunger and spring
2 Oil hole screws
1 Range index
3 Range ring screws
1 Stop pin screw
For Bracket Fuze Setter—
1 Corrector scale screw
3 Guide screws
4 Housing screws & split washers
1 Knob washer
1 Range worm crank & knob taper pin
1 Range worm crank handle
2 Range ring screws
3 Split pins (0.125)
1 Spring and spring cover with screw
2 Stop pins with rivets
For Cylinder—
1 Drain-plug, cylinder
1 Elevating & traversing lock spring
1 Filling plug (piston rod)
 5 Rings Garlock packing, ¾ in

Special Wrenches, Spanners, other tools and accessories will be

used only for the purposes for which they are intended. This purpose
is usually stamped upon the tool.
In assembling or disassembling parts of the materiel, no part will
be struck directly with a hammer. If force is necessary, a piece of
wood or copper should be interposed between the hammer and the
part struck. All nuts are provided with split pins as keepers.
A pair of wire cutting pliers is provided for use in pulling split
pins, cutting wire lashings, etc. When a nut is assembled the split pin
should always be inserted and properly opened.
Axes, hatchets, picks, pick mattocks and shovels are carried
on the carriage for use in the field and will not be put to other uses.
The working edges will be kept bright and lightly oiled, the edges
being sharpened if intended for cutting, or smooth if intended for
digging. Deformed blades, edges or points should be straightened at
the anvil and forge or in a vise. Shovel points are straightened with a
hammer on a block of wood.
The side edges of shovel blades should not be used as a mattock,
as such treatment will deform the blade. In the field, split handles
should be wrapped with a cord until they can be replaced by new
handles.
Canvas Buckets are used for watering animals, for washing
carriages and equipment. Whenever possible, they should be dried
before folding and replacing in the holders on the carriages.
A rip or hole may be patched and made practically water-tight by a
coat of shellac.
Lanterns are used for illuminating purposes in the field only.
Paulins are used to cover the harness and guns when in the field
or in park. On the march they are carried on the carriages, being
folded to serve as seat cushions. Holes and tears should always be
properly sewed, stitched or darned as soon as practicable.
 Picket Ropes are used in the field as drag ropes for the carriages
or as picket lines for the animals. The ropes must be in a serviceable
condition and free from knots. To keep them in a serviceable
condition, splicing may often be necessary.

CARE AND CLEANING OF THE DIFFERENT

PARTS OF THE CARRIAGES.
To disassemble and to clean the cylinder.—For cleaning, the
cylinder is dismounted and emptied and the cylinder head, counter
recoil buffer, and piston rod removed. The interior of the cylinder, the
piston, the counter recoil buffer and the stuffing box should then be
thoroughly cleaned by the use of cotton waste. The removal of the
packing is not necessary in cleaning. The cylinder bore should be
carefully inspected, and if any rust has formed it should be removed
with coal oil, using if necessary, fine emery cloth. The latter must be
used with great care to prevent any increase in the clearance
between the piston and the cylinder. If rubbing, burring, or scoring of
the parts is noted, the rough spots should be carefully smoothed
down by a skilled workman with a dead smooth file or with emery
cloth, and the cause of the roughness ascertained and removed.
Where unusual rubbing or scoring has occurred, the facts will be
reported to the Officer of the Ordnance Department charged with the
duty of keeping the battery in repair, for his information and action.
The exterior of the cylinder should be kept well oiled and free from
rust and dirt, and an inspection made at least once a month to
ascertain its condition. Where rust has formed it should be removed
with coal oil, and, if necessary, emery cloth. For shipment or storage,
or where the carriage is to stand without firing for extended periods,
the cylinder should be coated with the light slushing oil used for the
bore of guns.
To fill the recoil cylinder.—If the cylinder is not completely filled,
loss of stability will occur and higher stresses than normal will be
developed in the carriage. For this reason the cylinder should be
filled with the greatest care, a commissioned should, himself, verify
that the cylinder is full and that no air is left in it, exception of the void