ABSTRACT
Recently, deep learning techniques have been used significantly for large scale image classification targeting wildlife prediction. This research adopted a deep convolutional neural network (CNN) and proposed a deep scalable CNN. Our research essentially modifies the network layers (scalability) dynamically in a multitasking system and enables real-time operations with minimum performance loss. It suggests a straightforward technique to access the performance gains of the network while enlarging the network layers. This is helpful as it reduces redundancy in network layers and boosts network efficiency. The architecture implementation was done in software using keras framework and tensorflow as the backend on the CPU and to corroborate the universality and robustness of our proposed approach; we train our model on a GPU with a newly created dataset named “Zedataset”, preprocessed for performance evaluation. Results obtained from our experimentations show that our proposed architecture design will perform better with more dataset at the set optimum parameters.
Keywords: GPU, keras, deep CNN, CNN, Scalability, tensorflow, image classification, optimum
TABLE OF CONTENTS
COPYRIGHT ………………………………………………………………………………………………… ii
CERTIFICATION ………………………………………………………………………………………….. iii
SIGNATURE PAGE ……………………………………………………………………………………… iv
ABSTRACT …………………………………………………………………………………………………. v
DEDICATION ………………………………………………………………………………………………. vi
ACKNOWLEDGMENTS ……………………………………………………………………………….. vii
Contents ……………………………………………………………………………………………………. viii
LIST OF TABLES …………………………………………………………………………………………. x
LIST OF FIGURES ……………………………………………………………………………………….. xi
LIST OF ABBREVIATIONS …………………………………………………………………………… xii
Chapter One ………………………………………………………………………………………………… 1
Background of the study …………………………………………………………………………….. 1
1.0 Introduction ………………………………………………………………………………………. 1
1.1 Concept of Deep learning …………………………………………………………………… 2
1.2 Definition of learning ………………………………………………………………………….. 3
1.3 Concept of scalability in machine learning …………………………………………….. 4
1.4 Problem statement …………………………………………………………………………….. 4
1.5 Aim of the research ……………………………………………………………………………. 4
1.6 Objectives of the research ………………………………………………………………….. 5
1.7 Structure of the research ……………………………………………………………………. 5
Chapter Two ………………………………………………………………………………………………… 7
Literature review ……………………………………………………………………………………….. 7
2.0 Introduction ………………………………………………………………………………………. 7
2.1 Basic Concept and Terminology ………………………………………………………….. 7
2.2 Digital image classification ………………………………………………………………….. 8
2.2.1 Supervised learning ……………………………………………………………………… 9
2.2.2 Unsupervised learning ………………………………………………………………… 11
2.3 Neural networks ………………………………………………………………………………. 12
2.3.1 Convolutional Neural Network (CNN) ……………………………………………. 13
2.3.2 Multilayer perceptron (MLP) ………………………………………………………… 14
2.4 Review of similar works ……………………………………………………………………. 15
Chapter three …………………………………………………………………………………………….. 18
Design and Methodology ………………………………………………………………………….. 18
ix
3.0 Introduction …………………………………………………………………………………….. 18
3.1 Concept of Classification Technique …………………………………………………… 18
3.2 Design and requirement phase ………………………………………………………….. 19
3.3 Deep learning network for recognition and identification ……………………….. 19
3.4 Proposed neural network model ………………………………………………………… 20
3.5 Building blocks of Deep CNN…………………………………………………………….. 21
3.5.1 Convolution layer ……………………………………………………………………….. 21
3.5.2 Pooling layer ……………………………………………………………………………… 22
3.5.3 Fully connected layer ………………………………………………………………….. 23
3.5.4 Activation functions …………………………………………………………………….. 24
3.5.5 Last layer activation function ……………………………………………………….. 24
3.6 Training deep CNN ………………………………………………………………………….. 25
3.6.1 Loss function …………………………………………………………………………….. 26
3.6.2 Gradient descent ……………………………………………………………………….. 26
3.7 Dataset description ………………………………………………………………………….. 27
3.8 System architecture of deep CNN for wildlife recognition ………………………. 27
3.9 Implementation details ……………………………………………………………………… 28
3.9.1 Parameters of the network ………………………………………………………….. 29
Chapter four ………………………………………………………………………………………………. 30
Results and discussions …………………………………………………………………………… 30
4.0 Introduction …………………………………………………………………………………….. 30
4.1 Results for 4 convolutions with two output layers …………………………………. 30
4.2 Results for 3 convolution with 3 output layers ………………………………………. 31
4.3 Results summary …………………………………………………………………………….. 32
4.4 Results discussion …………………………………………………………………………… 33
Chapter five ……………………………………………………………………………………………….. 34
Conclusion and future works ……………………………………………………………………… 34
5.0 Introduction …………………………………………………………………………………….. 34
5.1 Conclusions ……………………………………………………………………………………. 34
5.2 Future work …………………………………………………………………………………….. 35
REFERENCES …………………………………………………………………………………………… 36
APPENDIX 1 ……………………………………………………………………………………………… 39
CODES FOR 3 CONVETS, 3 FULLY CONNECTED LAYER …………………………. 39
APPENDIX 2 ……………………………………………………………………………………………… 42
CODES FOR 4 CONVETS, 2 FULLY CONNECTED LAYER …………………………. 42
CHAPTER ONE
Background of the study
1.0 Introduction
The task of identifying and recognition of animals from photos has long been standing as there is no unique method that provides a robust and efficient solution to all situations. Several researchers used long-standing traditional approaches for its implementation with the problem still hanging in limbo as the task hugely involve collecting a large volume of images which predominantly is conducted manually with possibly images having an imperfect quality which sometimes affect the speed of classification, accuracy even for domain experts. More so, processing these image sets is time-consuming, effort demanding, and comes at a very high cost as it is an overwhelming amount of data that is collected.
In recent years, much attention has focused on using deep neural network based techniques in the area of image processing, particularly animal recognition and identification. However, the increase in the performance characteristics of the network depends on how scalable the network is designed. In machine learning, scalability is often defined as the result that even the slightest change in the size of the network parameters such as the network layers, training sets has on the computational performance of an algorithm (accuracy, memory allocation, speed of processing). So the question is to find a balance or in order words getting a suitable solution quick off the mark and most effectively. This is of serious concern as in scathing circumstances where the existence of temporal or contiguous constraints like real-time applications
2
dealing with large datasets, unapproachable computational problems demanding learning or first prototyping needing quickly implemented the result.
To deal with a large dataset, it is expedient to minimize the training time and allotted memory space while preserving accuracy; however, till date, most proposed deep learning algorithms do not proffer a proper trade-off among them. To contain these issues above, we aim to optimize floating points by changing them to fixing points to reduce memory complexity and yield faster processing in the network. In this research, the convolutional neural network framework will be used for animal identification and prediction, while stochastic gradient descent is used to optimize the parameters (i.e., weights, biases) of the network through error backpropagation with momentum and adaptive learning rate. Network layers and nodes in each hidden layer will be added in systematic experimentation and intuition with a robust test to harness.
1.1 Concept of Deep learning
Deep learning is an offshoot of machine learning, which is not new to the field of informatics and predictive analysis. However, recently, it has drawn much attention as neuroscientist, psychologist, engineers, economist, AI workers attempt to explore their learning potential. Deep learning approaches are a set of algorithms that strive to model data with extreme abstractions using a replica architecture with tortuous formation. It is one among the many segments of machine learning techniques based on the concept of learning representations of raw data which could be in a way such as the intensity per pixel value of a data or sections of a specific figure in a more abstract way.
3
There are several numbers of ways the area of deep learning has been represented as it is a subset of machine learning techniques that
i. Uses multiple layers with nonlinear processing units cascaded for feature extraction
ii. Are based on the (unsupervised) learning multiple data representations where hierarchical representation is formed when higher-level features are derived from lower level features.
iii. Learned multiple levels of representations corresponds to different levels of abstraction.
1.2 Definition of learning
One challenging fact when setting up the objectives of deep learning is the definition of learning. Learning is rather conceptual and as to those who have made efforts to give it meaning (psychologists, philosophers, etc.) have only succeeded in uncovering one among the many faces of the complex procedure.
However, there are some views of learning which has been acceding to mostly by those who have made continuous efforts to divulge the concept, and these on many occasion provides reasonable interpretation of the process. Some are the following:
i. There exist a system manipulating information provided by its environment and is capable of improving its self.
ii. The system has numerous ways of altering its current state and information provided can usually take many forms.
iii. The system is capable of remembering and recalling things that it has experienced.
4
1.3 Concept of scalability in machine learning
Scalability has increasingly been integrated over the years as part of deep learning. This is as a result of the likelihood of performance characteristics been affected as recently; most deep neural networks are hugely involved with the overwhelming size of the dataset.
Scalability, as defined in machine learning, is the effect that a change in system parameters has on the performance characteristics of an algorithm. Its methods could be like increase in the number of nodes, network layers, and hidden layers by systematic experimentation and/or intuition. This is done to ensure faster processing with huge dataset while preserving some performance characteristics like (accuracy, memory allocation) and reduction in the network complexity.
1.4 Problem statement
There has been a rise in cases of human-animal attacks and human-vehicle collision with the latter been prevalent in Nigeria. There are about 500-1000 vehicle collisions with large animals each year that result in more than 1 billion Naira in damages. Source (Federal road safety annual report, 2017).
To cope with this problem, machine learning based techniques could be employed, which may be on CCTV cameras connected to the relevant response team for surveillance of animals in both remote and urban places to save lives.
1.5 Aim of the research
The aim of this thesis is to provide a scalable, suitable, more generic and optimized network capable of processing huge amount of dataset even with images having an imperfect quality or varied deformations in real time while preserving better test accuracy.
5
1.6 Objectives of the research
Having at hand the different views of people as regards to what learning seems to be and how to attain it. One can perceive how challenging it is to interpret deep learning and even to set out some clear objectives. Although the concept of learning has cleared the air despite that the approach to deep learning by different people differs. The aim of this research is as follows:
i. Develop an artificial learning system capable of being adaptive and self-improving
ii. Develop a neural network with optimized parameters whose computational performance is unaffected by scalability.
iii. Develop a neural network system architecture with reduced complexity for large scale image classification or prediction.
1.7 Structure of the research
Chapter 1 presents a brief introduction of the research concept primarily deep learning, objectives of the project, and the aims.
Chapter 2 presents supporting theories of the research concept following brief introduction of deep learning concept and learning, forming a link with a classification problem, then give a brief account of the different classification approaches ranging from statistical methods to genetic algorithms. Two best learning approaches will be examined and finally, a brief account of similar works done will follow.
Chapter 3 will presents the theoretical analysis of the adopted algorithm with the proposed layers. The following information is provided:
6
i. A detailed description of the algorithm focusing on its peculiarities
ii. The design of the algorithm with a detailed explanation of its layers.
Chapter 4 will describes the experiments and presents the results which will be statistically analyzed to check for relative performance and the validation of the theoretical estimates presented in the previous chapter.
Chapter 5 summarises the results presented in the thesis and concludes their importance in the context of recognition and identification.
IF YOU CAN'T FIND YOUR TOPIC, CLICK HERE TO HIRE A WRITER»
