Minoufiya University Faculty of Electronic Engineering Department of Electronics and Electrical Communications Engineering
Classification of Buried Objects Using Acoustic Waves
A Thesis Submitted for the Degree of M. Sc. in Electronic Engineering, Communications Engineering, Department of Electronics and Communications Engineering
By
Eng. Emad Abd Elhalim Elsayed Elshazly B. Sc. in Electronics and Electrical Communications Engineering, Faculty of Electronic Engineering, Minoufiya University, Menouf 2004
Supervisors
Prof. Mohamed F. El-Kordy Prof. in Electronics and Electrical Communications Engineering Department, Faculty of Electronic Engineering, Minoufiya University
Prof. Sayed M. El-Araby Chairman of Atomic Energy Authority
Dr. Osama F. Zahran Lecturer in Electronics and Electrical Communications Engineering Department, Faculty of Electronic Engineering, Minoufiya University
2012
Minoufiya University Faculty of Electronic Engineering Department of Electronics and Electrical Communications Engineering
Classification of Buried Objects Using Acoustic Waves
A Thesis Submitted for the Degree of M. Sc. in Electronic Engineering, Communications Engineering, Department of Electronics and Communications Engineering
By
Eng. Emad Abd Elhalim Elsayed Elshazly B. Sc. in Electronics and Electrical Communications Engineering, Faculty of Electronic Engineering, Minoufiya University, Menouf, 2004
Supervisors
Prof. Mohamed F. El-Kordy ( ) Department of Electronics and Electrical Communications Engineering, Faculty of Electronic Engineering, Minoufiya University
Prof. Sayed M. El-Araby ( ) Chairman of Atomic Energy Authority
Dr. Osama F. Zahran ( ) Department of Electronics and Electrical Communications Engineering, Faculty of Electronic Engineering, Minoufiya University
2012
Minoufiya University Faculty of Electronic Engineering Department of Electronics and Electrical Communications Engineering
Classification of Buried Objects Using Acoustic Waves
A Thesis Submitted for the Degree of M. Sc. in Electronic Engineering, Communications Engineering, Department of Electronics and Communications Engineering
By
Eng. Emad Abd Elhalim Elsayed Elshazly B. Sc. in Electronics and Electrical Communications Engineering, Faculty of Electronic Engineering, Minoufiya University, Menouf, 2004
Approved by
Prof. Ali Hassan Moustafa ( ) Emeritus Professor Faculty of Electronic Engineering Minoufiya University
Prof. Hesham Fathy Aly Hamed ( ) Professor Faculty of Engineering Minia University
Prof. Mohamed F. El-Kordy ( ) Professor Faculty of Electronic Engineering Minoufiya University
2012
ﺑِﺴﻢِ ﺍﻟﻠﹼﻪِ ﺍﻟﺮﺣﻤـَﻦِ ﺍﻟﺮﺣِﻴﻢِ
((( ﻳﺮﻓﹶﻊِ ﺍﻟﻠﱠﻪ ﺍﻟﱠﺬِﻳﻦ ﺁﻣﻨﻮﺍ ﻣِﻨﻜﹸﻢ ﻭﺍﻟﱠﺬِﻳﻦ
ﺃﹸﻭﺗﻮﺍ ﺍﻟﹾﻌِﻠﹾﻢ ﺩﺭﺟﺎﺕٍ ﻭﺍﻟﻠﱠﻪ ﺑِﻤﺎ ﺗﻌﻤﻠﹸﻮﻥﹶ ﺧﺒِﻴﺮ )))
ﺻﺪﻕ ﺍﷲ ﺍﻟﻌﻈﻴﻢ
ﺳﻮﺭﺓ ﺍﻟﻤﺠﺎﺩﻟﺔ : ﺍﻵﻳﺔ ١١١١١١
ACKNOWLEDGEMENTS
First and foremost, I thank ALLAH , the most gracious, the ever merciful for helping me finishing this work. I want to thank all those, who helped me by their knowledge and experience. I will always appreciate their efforts. I wish to express my sincere thanks to my supervisors Prof. Mohamed El-Kordy, Prof. Sayed El-Araby and Dr. Osama Zahran. I am deeply indebted to them for valuable supervision, continuous encouragement, useful suggestions, and active help during the course of this work. Special thanks to Dr. Fathi Abd El-Samie for his valuable suggestions and continuous support. My sincere appreciation and gratitude to my parents, my brothers, my father in law, my mother in law, my brothers in law, my wife, and my daughter Mariam for their help and patience during the preparation of this work.
I LIST OF PUBLICATIONS
[1] E. A. El shazly, O. Zahran, Sayed M. S. Elaraby, M. El Kordy and F. E. Abd El Samie, “ Automatic Detection of Buried Landmines using Cepstral Approach ”, 1st International Conference on Electrical and Computer Systems Engineering (ECSE 2010), 6th October City, Egypt, 1 3 November, 2010. [2] E. A. El shazly, O. Zahran, Sayed M. S. Elaraby, M. El Kordy and F. E. Abd El Samie, “ Cepstral Detection of Buried Landmines from Acoustic Images with a Spiral Scan ”, 6th International Computer Engineering Conference (ICENCO 2010), IEEE, pp. 97 102, Cairo, Egypt, 28 29 December, 2010. [3] E. A. El shazly, O. Zahran, Sayed M. S. Elaraby, M. El Kordy, S. El Rabie and F. E. Abd El Samie, “ Identification of Buried Landmines Using Mel Frequency Cepstral Coefficients and Support Vector Machines ”, 8th International Conference on Informatics and Systems (INFOS 2012), IEEE, pp. MM60 MM66, Cairo, Egypt, 14 16 May 2012.
II ABSTRACT
There is no doubt that the problem of landmines is one of the most important problems that concerns the whole world. Egypt is one of the countries that suffer from this problem. There are more than 23 million landmines that are subject to explode at any time and these landmines occupy a large area estimated by 3910 square kilometers. The source of these landmines is the wars; World War II and the Arab Israeli wars of 1956, 1967, and 1973 in the eastern desert and Sinai. The problem of landmines is not human casualties only, but there are also serious economic losses. There are several obstacles that are faced in removing the landmines such as the loss or absence of maps as well as the high costs needed to remove them. The work presented in t his thesis provides an introduction to land mines ; definition, their components and types, a summary of the techniques and the different methods used for detecting and clearing them, and the operating principles of each method.
This thesis proposes efficient landmine identification techniques, which help in identifying the several types of landmines with different dimensions, shapes, and types. These techniques use Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) and they are based on the MFCCs to identify the different types of landmines. The performance of these techniques is evaluated in the presence of different types of noise with and without blurring. The thesis also proposes a classification technique using ANNs to classify the different types of landmines into different categories based on MFCCs features.
III TABLE OF CONTENTS
List of Abbreviations ...... …… ...... … ...... …. VIII List of Symbols ...... ……...... …. . . X List of Figures . . . …...... ……...... …. . . . XII List of Tables ...... ……….. XVII
CHAPTER (1) INTRODUCTION ……………………………...…………… 1 1.1 Introduction………………………………...…………...…..…… 1 1.2 Thesis Objectives and Contributions ….………..………….……. 3 1.3 Thesis Outlines……………………….………...…………....…... 4
CHAPTER (2) A SURVEY STUDY ON LANDMINES PROBLEM ….….. 5 2.1 Introduction………..………...….………….……………..…...… 5 2.2 Short Overview of Buried Landmines.....…….…….…...... …..… 5 2.3 Landmine Components …...……..…..…….…………..………… 6 2.3.1 Firing Mechanism ……….………...………...……...….. 7 2.3.2 Detonator or Igniter……….…………………..…...…… 7 2.3.3 The Booster Charge……….………………….…..….…. 7 2.3.4 The Main Charge……….……………...……...…..……. 7 2.3.5 The Case……….……………………..………...... …….. 7 2.4 Types of Landmines ………………….…………..……...….…... 8 2.4.1 AT Landmines…………………...……..……….……… 10 2.4.2 AP Landmines…..…………………….....……………... 11 2.5 Countries and Regions Affected by Landmines……..…………... 13 2.6 Landmines in Egypt………..………………………………...…... 15 2.7 Demining Approaches and Technologies……………..…………. 18 2.7.1 Manual Demining…………..……………………...…. 18 2.7.2 The Use of Animals, Insects and Bacteria...…..….…….. 18 2.7.3 Mechanical Demining….……….………..………..……. 20 2.7.4 Robots and Humanitarian Demining..……….…………. 21
IV 2.7.5 Landmine Detection and Sensing Technologies…….….. 22 2.8 Summary....…...……………………………………….....………. 26
CHAPTER (3) REVIEW STUDY ON LANDMINE DETECTION
SYSTEMS AND ARTIFICIAL NEURAL NETWORKS …….…………… 27
3.1 Introduction....………………………………….………..………. 27 3.2 Literature Review on LDV Based A/S Landmine Detection....…. 28 3.3 Principles of A/S Coupling ……………………………………... 31 3.3.1 Body Waves…………..…………………...…...... 31 3.3.1.1 Primary Waves……...…….……...………….. 32 3.3.1.2 Secondary Waves…...……………………….. 32 3.3.2 Surface Waves………………………………………….. 32 3.3.2.1 Loove Waves……………………………….. 32 3.3.2.2 Rayleigh Waves……………………………... 32 3.4 A/S Transmission System with Acoustic Loudspeakers...………. 33 3.5 A/S Receiving System using LDV………………………………. 33 3.6 Segmentation of Acoustic Landmine Images……………………. 35 3.6.1 Erosion…………………………………………………. 36 3.6.2 Dilation…………………………………………………. 36 3.6.3 Opening…………………………………………………. 37 3.6.4 Closing………………………………………………….. 37 3.7 Artificial Neural Networks (ANNs).………...………………..…. 38 3.7.1 Neural Network Concepts………………….…………… 39 3.7.1.1 Cells………………………...……………….. 39 3.7.1.2 Layers………………………………………... 39 3.7.1.3 Arcs……………………………………...…... 40 3.7.1.4 Weights………………………………...……. 40 3.7.2 The Learning Rules of Neural Networks…………….…. 40 3.7.2.1 Supervised Learning………………...……….. 40 3.7.2.2 Unsupervised Learning…..………………….. 41 3.7.2.3 Reinforcement Learning……...……………… 41
V 3.7.3 Neuron Model…………….…………………………….. 41 3.7.3.1 Simple Neuron………………...…………….. 41 3.7.3.2 Neuron with Vector Input……………...…..... 42 3.7.4 Activation Rules………………………………………... 43 3.7.4.1 Identity Function…...………...……………… 43 3.7.4.2 Step Function..…………………………...….. 43 3.7.4.3 Sigmoid Function……………………………. 43 3.7.5 Network Architectures………………………………….. 44 3.7.5.1 Single Layer Network……………………….. 44 3.7.5.2 Multi Layer Network………………………… 45 3.7.6 Back propagation Neural Network…………..…………. 46 3.8 Types of Noise……………………...……………………………. 49 3.8.1 The AWGN……………………………………………... 49 3.8.2 Impulsive Noise……………………………………….... 50 3.8.3 Speckle Noise…………………………………………... 50 3.8.4 Poisson Noise…...…………………………………….… 50 3.8.5 Image Blurring...………………………………………... 51 3.8.6 Signal to Noise Ratio…………………………………… 51
CHAPTER (4) LANDMINE IDENTIFICATION SYSTEMS USING
ARTIFICIAL NEURAL NETWORKS ……………...……………………… 52
4.1 Introduction……………………………………………………… 52 st 4.2 The 1 Cepstral Landmine Identification Approach...... …… 52 4.2.1 Transformation of 2 D Image to a 1 D Signal……….…. 54 4.2.2 Discrete Transform Domains……...……………………. 55 4.2.2.1 The DCT…………………………...………... 56 4.2.2.2 The DST…………………………...………... 56 4.2.2.3 The DWT…...……………………...………... 57 4.2.3 Feature Extraction………………………………………. 61 4.2.3.1 Extraction of MFCCs………………………... 62 4.2.3.2 Extraction of Polynomial Coefficients……… 65
VI 4.2.4 Feature Matching Technique……..…………………….. 66 st 4.3 Experiments and Results of 1 Approach………..………………. 67 nd 4.4 The 2 Cepstral Landmine Identification Approach.…………… 79 4.4.1 The 2 D DCT…………………..……………...………... 79 nd 4.5 Experiments and Results of 2 Approach……..………………… 80 4.6 Summary....…...……………………………………….....………. 92
CHAPTER (5) LANDMINE IDENTIFICATION SYSTEMS USING
SUPPORT VECTOR MACHINES ….………………...…………………….……… 93 5.1 Introduction……………………………………………………… 93 5.2 Landmines Identification using SVMs…………………………... 94 5.2.1 Support Vector Machines………………………………. 94 5.2.1.1 Overview of SVM…………………………… 94 5.2.1.2 The Separable Case………………………..… 95 5.2.1.3 The Karush Kuhn Tucker Conditions……..… 98 5.2.1.4 Test Phase…………………………………… 99 5.2.1.5 The Non Separable Case……………………. 99 5.2.1.6 Kernel Selection of SVM………….....……… 102 5.3 Proposed Image Identification Approach……………………….. 102 5.4 Experiments and Results.………….…………………………….. 103 5.5 Summary....…...……………………………………….....………. 117
CHAPTER (6) LANDMINE CLASSIFICATION USING ARTIFICIAL
NEURAL NETWORKS ………..………………………………………..…… 118 6.1 Introduction……………………………………………………… 118 6.2 Proposed Landmines Classification Approach………………….. 118 6.3 Experiments and Results………………………………………… 120
CHAPTER (7) CONCLUSIONS ………………………………………..…… 126 7.1 Conclusions……………………………………………………… 126 7.2 Future Work…...…………………………...……..……………… 127
REFERENCES …………………..…………………………………………… 129
VII LIST OF ABBREVIATIONS
1-D One Dimensional 1-NN 1 Nearest Neighbour 2-D Two Dimensional A/D Analog/ Digital A/S Acoustic to Seismic ALD Acoustic Landmine Detection ANN Artificial Neural Network AP Anti Personnel AT Anti Tank AWGN Additive White Gaussian Noise BPNN Back Propagation Neural Network CWT Continuous Wavelet Transform D/A Digital / Analog dB Decibel DCT Discrete Cosine Transform DFT Discrete Fourier Transform DST Discrete Sine Transform DWT Discrete Wavelet Transform ERW Explosive Remnants of War FFT Fast Fourier Transform FM Frequency Modulated GICHD Geneva International Center for Humanitarian Demining GMM Gaussian Mixture Model GPR Ground Penetration Radar HMM Hidden Markov Model ICBL International Campaign to Ban Landmines
VIII IR IR Infrared KKT Karush Kuhn Tucker LDV Laser Doppler Vibrometry LPC Linear Prediction Coefficient LPCC Linear Prediction Cepstral Coefficient MFCC Mel Frequency Cepstral Coefficient MLP Multilayer Perceptron MMC Maximum Margin Classifiers MRF Markov Random Field NDRE National Defense Research Establishment ORNL Oak Ridge National Laboratories PC Personal Computer PR Perfect Reconstruction RBF Radial Basis Function RDV Radar Doppler Vibrometry RDX Royal Demolition Explosive ROI Region of Interest SLDV Scanning Laser Doppler Vibrometer SNR Signal to Noise Ratio SVM Support Vector Machine TNT Trinitrotoluene TRA Time Reversal Acoustic UV Ultraviolet UXO Unexploded Ordnance VQ Vector Quantization
IX LIST OF SYMBOLS
b The scalar bias of ANN θ The threshold th ej(n) Error signal produced at the output neuron j for the n sample th dj(n) The desired response from neuron j for the n sample The learning rate parameter of the back propagation algorithm th δj(n) The local gradient of neuron j for the n sample W The weight matrix of ANN w(n) The Hamming window k The number of neurons in the previous layer σ Variance wAWGN (t) Additive White Gaussian Noise
R (t 1, t 2) The autocorrelation function
Gsn (i, j) The observed image of the speckle noise
Fsn (i, j) The original image of the speckle noise
Usn (i, j) Multiplicative component of the speckle noise
ξsn (i, j) Additive component of the speckle noise w(k) A multiplication factor
H0 The low pass filter
H1 The high pass filter P(z) A zero phase polynomial
αi Lagrange multipliers
ξi Slack variables Θ Erosion morphological operation
⊕ Dilation morphological operation
O Opening morphological operation ● Closing morphological operation
X P Primary Waves S Secondary waves
Se Structuring element F Image of an object z Displacement of the structuring element
XI LIST OF FIGURES
Figure (1 1) An identification system……….……...………...………… 3 Figure (2 1) Landmine components…….……….……………………… 6 Figure (2 2) Assortment of the most common landmines.……………… 8 Figure (2 3) Blast landmine………………..……………………………. 9 Figure (2 4) Fragmentation landmine…………………..……………….. 9 Figure (2 5) Two typical AT landmines; (a) TM 62M and (b) TMA 2… 10 Figure (2 6) Typical AP landmines; (a) PRB M35, (b) PMN, (c) VALMARA 69, and (d) MON 100……..………………… 12 Figure (2 7) Countries around the world affected by mines……………. 13 Figure (2 8) Location map for landmines distributions in Egypt…….…. 17 Figure (2 9) Types of landmines in the western desert, Egypt…………. 17 Figure (3 1) A/S coupling Principle…………………………………….. 31 Figure (3 2) LDV based A/S landmine detection system……………….. 34 Figure (3 3) Samples of acoustic landmine images……………………... 35 Figure (3 4) Block diagram of the wavelet based segmentation method.. 37 Figure (3 5) The Neural Network structure……………………………... 39 Figure (3 6) Simple neuron structure…………………………………… 41 Figure (3 7) Neuron with vector input structure………………………... 43 Figure (3 8) Activation Functions (a) Identity function, (b) Threshold function, and (c) Sigmoid function………………………... 44 Figure (3 9) Single Layer Neural Network……………………………... 44 Figure (3 10) Multi Layer neutral network………………………………. 46 Figure (4 1) Schematic diagram of the proposed landmine identification system (1st approach): (a) Training phase, (b) Testing phase……………………………………………………….. 54 Figure (4 2) The block by block scan…………………………………… 55 Figure (4 3) Spiral scan in a single block……………………………….. 55 Figure (4 4) Spiral scan in a group of neighboring blocks……………… 55 Figure (4 5) Spiral scan in the whole image…………………………….. 56 Figure (4 6) Two level wavelet decomposition tree……………………. 57 Figure (4 7) The two band decomposition reconstruction wavelet filter bank………………………………………………………... 58 Figure (4 8) Some wavelets families (a) Haar, (b) Daubechies, (c) Coiflets, (d) Symlets, (e) Meyer, (f) Morlet and (g) Mexican……………………………………………………. 60 Figure (4 9) Extraction of the MFCCs………………………………….. 62 Figure (4 10) Hamming window (a) Time domain and (b) The frequency response……………………...…………………………….. 63 Figure (4 11) The Mel scale as a function of the linear frequency….…… 64 Figure (4 12) Mel spaced filter bank…………………………………….. 65 Figure (4 13) Recognition rate vs. SNR for the different feature extraction methods from images contaminated by AWGN.. 68 Figure (4 14) Recognition rate vs. SNR for the different feature
XII extraction methods from blurred images contaminated by AWGN…………………………………………………….. 69 Figure (4 15) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by impulsive noise…………………………………………….. 69 Figure (4 16) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by impulsive noise………………………….. 70 Figure (4 17) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by speckle……………………………………………………... 70 Figure (4 18) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by speckle noise……………………………. 71 Figure (4 19) Recognition rate vs. SNR for the different feature extraction methods from landmine images contaminated by AWGN…………………………………………………….. 72 Figure (4 20) Recognition rate vs. SNR for the different feature extraction methods from blurred landmine images contaminated by AWGN…………………………………... 72 Figure (4 21) Recognition rate vs. the percentage error for the different feature extraction methods from landmine images contaminated by impulsive noise………………………….. 73 Figure (4 22) Recognition rate vs. the percentage error for the different feature extraction methods from blurred landmine images contaminated by impulsive noise………………………….. 73 Figure (4 23) Recognition rate vs. the percentage error for the different feature extraction methods from landmine images contaminated by speckle noise…………………………….. 74 Figure (4 24) Recognition rate vs. the percentage error for the different feature extraction methods from blurred landmine images contaminated by speckle noise…………………………….. 74 Figure (4 25) Recognition rate vs. SNR for the different feature extraction methods from landmine images contaminated by AWGN…………………………………………………….. 75 Figure (4 26) Recognition rate vs. SNR for the different feature extraction methods from blurred landmine images contaminated by AWGN………………………………….. 76 Figure (4 27) Recognition rate vs. the percentage error for the different feature extraction methods from landmine images contaminated by impulsive noise…………………….……. 76 Figure (4 28) Recognition rate vs. the percentage error for the different feature extraction methods from blurred landmine images contaminated by impulsive noise…………………….…..... 77 Figure (4 29) Recognition rate vs. the percentage error for the different feature extraction methods from landmine images
XIII contaminated by speckle noise……………………….…… 77 Figure (4 30) Recognition rate vs. the percentage error for the different feature extraction methods from blurred landmine images contaminated by speckle noise……………………….…… 78 Figure (4 31) Schematic diagram of the proposed landmine identification system (2nd Approach): (a) Training phase, (b) Testing phase……………………………………………………….. 80 Figure (4 32) Recognition rate vs. SNR for the different feature extraction methods from images contaminated by AWGN…………………………………………………….. 81 Figure (4 33) Recognition rate vs. SNR for the different feature extraction methods from blurred images contaminated by AWGN…………………………………………………….. 82 Figure (4 34) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by impulsive noise……………………………….………….… 82 Figure (4 35) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by impulsive noise…………………..……… 83 Figure (4 36) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by speckle noise……………………………………………… 83 Figure (4 37) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by speckle noise…………………………… 84 Figure (4 38) Recognition rate vs. SNR for the different feature extraction methods from images contaminated by AWGN……………………………………………………. 85 Figure (4 39) Recognition rate vs. SNR for the different feature extraction methods from blurred images contaminated by AWGN…………………………………………………….. 85 Figure (4 40) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by impulsive noise……………………………………………. 86 Figure (4 41) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by impulsive noise…………………………. 86 Figure (4 42) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by speckle noise……………………………………………… 87 Figure (4 43) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by speckle noise…………………………… 87 Figure (4 44) Recognition rate vs. SNR for the different feature extraction methods from images contaminated by AWGN…………………………………………………….. 88
XIV Figure (4 45) Recognition rate vs. SNR for the different feature extraction methods from blurred images contaminated by AWGN…………………………………………………….. 89 Figure (4 46) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by impulsive noise…………………………………………… 89 Figure (4 47) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by impulsive noise…………………………. 90 Figure (4 48) Recognition rate vs. the percentage error for the different feature extraction methods from images contaminated by speckle noise……………………………………………… 90 Figure (4 49) Recognition rate vs. the percentage error for the different feature extraction methods from blurred images contaminated by speckle noise……………………………. 91 Figure (5 1) Linear separating hyperplanes for the separable case…..…. 96 Figure (5 2) Linear separating hyperplanes for the non separable case… 102 Figure (5 3) Schematic diagram of the proposed image identification approach system, (a) Training Phase and (b) Testing Phase. 103 Figure (5 4) Recognition rate vs. SNR for different MFCC feature extraction methods from images contaminated by AWGN.. 106 Figure (5 5) Recognition rate vs. SNR for different MFCC feature extraction methods from blurred images contaminated by AWGN…………………………………………………..… 107 Figure (5 6) Recognition rate vs. the percentage error for different MFCC feature extraction methods from images contaminated by impulsive noise………………………….. 107 Figure (5 7) Recognition rate vs. the percentage error for different MFCC feature extraction methods from blurred images contaminated by impulsive noise………………………….. 108 Figure (5 8) Recognition rate vs. the noise variance for different MFCC feature extraction methods from images contaminated by speckle noise………………………………………………. 108 Figure (5 9) Recognition rate vs. the noise variance for different MFCC feature extraction methods from blurred images contaminated by speckle noise……………………………. 109 Figure (5 10) Recognition rate vs. SNR for different MFCC feature extraction methods from images contaminated by AWGN.. 110 Figure (5 11) Recognition rate vs. SNR for different MFCC feature extraction methods from blurred images contaminated by AWGN…………………………………………………….. 110 Figure (5 12) Recognition rate vs. the percentage error for different MFCC feature extraction methods from images contaminated by impulsive noise………………………….. 111 Figure (5 13) Recognition rate vs. the percentage error for different MFCC feature extraction methods from blurred images
XV contaminated by impulsive noise………………………….. 111 Figure (5 14) Recognition rate vs. the noise variance for different MFCC feature extraction methods from images contaminated by speckle noise………………………………………………. 112 Figure (5 15) Recognition rate vs. the noise variance for different MFCC feature extraction methods from blurred images contaminated by speckle noise……………………………. 112 Figure (5 16) Recognition rate vs. SNR for different MFCC feature extraction methods from images contaminated by AWGN.. 113 Figure (5 17) Recognition rate vs. SNR for different MFCC feature extraction methods from blurred images contaminated by AWGN…………………………………………………….. 114 Figure (5 18) Recognition rate vs. the percentage error for different MFCC feature extraction methods from images contaminated by impulsive noise………………………….. 114 Figure (5 19) Recognition rate vs. the percentage error for different MFCC feature extraction methods from blurred images contaminated by impulsive noise………………………….. 115 Figure (5 20) Recognition rate vs. the noise variance for different MFCC feature extraction methods from images contaminated by speckle noise………………………………………………. 115 Figure (5 21) Recognition rate vs. the noise variance for different MFCC feature extraction methods from blurred images contaminated by speckle noise….…………………………. 116 Figure (6 1) The schematic diagram of the steps of the proposed cepstral classification approach (a) Training Phase and (b) Testing Phase……………………………………………… 119
XVI LIST OF TABLES
Table (2 1) Typical Specifications of Three Different Types of Mines….. 10 Table (2 2) Worldwide Landmine Distribution and Clearance Status….… 14 Table (2 3) Countries and regions affected by mines around the world….. 15 Table (2 4) Summary of the Detection Technologies Reviewed…………. 24 Table (4 1) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring………………………………………………. 71 Table (4 2) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring……………………………...……………….. 75 Table (4 3) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring………………………...…………………….. 78 Table (4 4) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring…………………...………………………….. 84 Table (4 5) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring……………...……………………………….. 88 Table (4 6) Recognition rates for the different feature extraction methods for landmine images contaminated by Poisson noise with and without blurring………...…………………………………….. 91 Table (5 1) Comparison between CPU processing time using SVM and ANN classifiers for lexicographic ordering scan method….… 109 Table (5 2) Comparison between CPU processing time using SVM and ANN classifiers for block by block scan method……………. 113 Table (5 3) Comparison between CPU processing time using SVM and ANN classifiers for spiral scan method……………………… 116 Table (6 1) Comparison between classification rates (%) for different feature extraction methods for AWGN at different SNRs…… 121 Table (6 2) Comparison between classification rates (%) for different feature extraction methods for AWGN plus image blurring at different SNRs……………………………………………...… 122 Table (6 3) Comparison between classification rates (%) for different feature extraction methods for impulsive noise at different SNRs…………………………………………………………. 123 Table (6 4) Comparison between classification rates (%) for different feature extraction methods for impulsive noise plus image blurring at different SNRs……………………………………. 123 Table (6 5) Comparison between classification rates (%) for different feature extraction methods for speckle noise at different SNRs…………………………………………………………. 124
XVII Table (6 6) Comparison between classification rates (%) for different feature extraction methods for speckle noise plus image blurring at different SNRs……………………………………. 125
XVIII Chapter 1 Introduction
CHAPTER 1
INTRODUCTION
1.1 Introduction Landmines are usually victim triggered explosive devices, which are placed on or near the ground until a person or animal triggers their detonating mechanism. They are intended to damage their target via blast and/or fragments. The use of landmines is controversial, because they are indiscriminate weapons, harming soldiers and civilians alike. Landmines have an indefinite life time, and may still cause horrific personal injuries and economic dislocation for decades after a war has finished [1]. They also remain dangerous after the conflict, in which they were deployed, has ended killing and injuring civilians and rendering land impassable and unusable for years and even decades . To make matters worse, many factions have not kept accurate records of the exact locations of their minefields, making removal efforts painstakingly slow. These facts pose serious difficulties in many developing nations, where the presence of landmines hampers resettlement, agriculture, and tourism. Landmines are made of plastic, metal or other materials. Some landmines contain explosives and some contain pieces of shrapnel. They are classified into Anti Personnel (AP) landmines, which are designed to kill, maim, and injure the human (approximately 26,000 people annually), and Anti Tank (AT) landmines, which are designed to destroy or damage the vehicles [2]. AT landmines are typically larger than AP landmines, and they contain more explosives than AP landmines and often require more pressure or weight on top of them to detonate. The high trigger pressure (normally 100 kg) prevents them from being set off by
1 Chapter 1 Introduction infantry or smaller vehicles of less importance. More modern AT landmines use shaped charges to focus and increase the armor penetration of the explosives [3]. Landmines move as weather conditions change, for example, due to heavy rain, hurricanes or earthquakes. Areas previously considered clear and safe can become mine contaminated, threatening displaced or returning populations. Recent tragic incidents show once again that the only real way to fully ensure safety is by completing clearance of all mined areas as soon as possible, no matter how remote. About 110 million landmines are planted in about seventy six countries and territories in all regions of the world. Egypt has 23 millions, i.e; 20% of the total number of landmines, spreading horror and imminent threats of loss of life and property in the eastern and western parts of the country. Over the past 25 years, about 7,923 people were victimized by landmines, including 3,200 dead and 4,723 injured, according to local and international statistics [4]. Landmine Monitor recorded at least 190 mine/ERW casualties (55 killed and 135 injured) in Egypt between 1999 and 2008. In 2009, 41 mines and explosive remnants of war (ERW) casualties were recorded in 12 incidents in Egypt. The majority of casualties were men, at least 11 were children (nine killed and two injured), and at least one was a woman [4, 5]. There are many obstacles in removing the buried landmines such as the loss or absence of maps or information about these mines or even the areas, where they are laid in, the change of mine locations due to climatic and physical factors, the large variety of types of AP and AT landmines, and the high costs needed to remove mines. It is known that the production cost of landmines is very low (may be $3 per mine), but the detection and removal cost is still high (more than $1000 per mine). The identification system of buried landmine objects consists of two stages to perform both the training of the input image models and the evaluation of the testing image sets. These are feature extraction and identification stages as shown in Figure (1 1). If we consider firstly, the feature extraction stage, we can find that this stage is concerned with converting an input training image into a
2 Chapter 1 Introduction series of vectors, which contain the discriminative information for the main image features. This is achieved by framing and windowing the input images after the transformation of the signals from a 2 D matrix into a 1 D vector, then applying the feature extraction algorithm to each frame. The output of the feature extraction stage is the feature matrix, which combines the feature vectors produced by all frames. Rows of this matrix correspond to the frame number, while columns correspond to the feature vector coefficients [6, 7]. The identification process employs this feature matrix as an input for two processes. For model training, the classifier uses a collection of feature matrices from different training images to construct some type of model. After that, testing of this model should be performed using a test image set to verify the efficiency of the constructed model and attain the optimum model allowing identification to take place.
Training set of Feature matrix Testing samples coefficients decision Feature Identification or extraction Testing set of samples
Figure (1 1): An identification system. 1.2 Thesis Objectives and Contributions
This thesis proposes new approaches for identification and classification of buried landmines from acoustic images. These approaches are cepstral approaches, which are based on generating a database of landmine features using MFCCs and polynomial coefficients extracted from different landmine images. The MFCCs technique can be considered as one of the most effective techniques that are used for feature extraction from speech signals. In this thesis, it is extended to the case of landmine images. It deals with the frames of data in the 1 D signals obtained from the 2 D images. The MFCCs are derived from the Fast Fourier Transform (FFT) magnitude spectrum of each frame by applying a filter
3 Chapter 1 Introduction bank, which has filters evenly spaced on a warped frequency scale. The logarithm of the energy in each filter is calculated and accumulated before a DCT is applied [6, 7]. The MFCCs and shape features extracted from the available training set are used to create a database. The matching process can be performed for any new image to identify it as a landmine or not using ANNs and Support Vector Machines (SVMs) [8, 9, 10]. The proposed approach for feature extraction is based on the MFCCs of the 1 D signals extracted from landmine images. The obtained features are robust to noise and insensitive to time shifts in signals. So, there is no need for registration of images. The database in the training phase of the proposed identification approaches is generated from 20 different landmine images with different dimensions, shapes and types, while the database for classification is generated from 50 images.
1.3 Thesis Outlines
Chapter Two gives a survey on landmines and a literature review of landmines problems and also points to the different types and construction of landmines. Also, it presents some of the demining technologies of buried landmines.
Chapter Three explains the landmine detection techniques and the ANNs.
Chapter Four proposes the feature extraction methods from the different discrete transform domains. It also explains the proposed cepstral identification approaches of buried landmines from acoustic images using ANNs. These approaches are based on generating a database of landmine features using the MFCCs and polynomial coefficients extracted from different landmine images.
Chapter Five proposes a landmine identification technique using an SVM based on MFCCs.
Chapter Six proposes a landmine classification technique using ANNs and MFCCs .
Chapter Seven presents the conclusions and the future work .
4 Chapter 2 A Survey Study on Landmines Problem
CHAPTER 2
A SURVEY STUDY ON LANDMINES PROBLEM
2.1 Introduction
This chapter gives the nature of landmines problem that will be considered in this research. It also gives an overview of buried landmines and landmine detection, identification, and classification techniques. The chapter is organized as follows. Section (2 2) gives a short overview of buried landmines. Section (2 3) explains the main components of a landmine. Section (2 4) discusses the different types of buried landmines. Section (2 5) presents the countries and regions affected by mines. Section (2 6) presents the landmines in Egypt. Finally, section (2 7) presents the demining approaches and technologies of buried landmines.
2.2 Short Overview of Buried Landmines
A landmine is an explosive material with a firing mechanism often laid in groups, called mine fields, and is designed to prevent the enemy from passing through a certain area, or sometimes to force the enemy through a particular area. An army also uses landmines to slow an enemy until reinforcements can arrive. The other sad feature of this weapon is that it fights on long after hostilities have ceased. Many potentially rich agricultural regions are threatened by landmines, and the danger grows as a population increase necessitates the cultivation of new
5 Chapter 2 A Survey Study on Landmines Problem land. The permanent nature of the threat, and the fact that landmines, once laid, are out of anyone's control, distinguishes them from all other weapons. Each day, around the world, tens of people die as a result of landmines. More than 26,000 people are killed or maimed by landmines and unexploded ordnance (UXO) every year, which is equivalent to one victim, every 20 min . Worldwide; landmines are a problem of epidemic proportions. While landmines are produced as "weapons of war," in fact, only 10% of landmine victims are soldiers and 90% of the victims are women and children. Unfortunately, landmine technology is quite simple, and its price is very low as most weapons cost in the range of ($3 $15) to make, while the cost to remove them is between $300 and $1000 [1, 4].
2.3 Landmine Components
A typical landmine includes the following components [11]:
• Firing mechanism or other device (including anti handling devices).
• Detonator or igniter (sets off the booster charge).
• Booster charge (may be attached to the fuse, or the igniter, or be part of the main charge).
• Main charge (in a container, usually forms the body of the landmines).
• Casing (contains all of the above parts).
Figure (2 1): Landmine components.
6 Chapter 2 A Survey Study on Landmines Problem
2.3.1 Firing Mechanism
A landmine can be triggered by a number of things including pressure, movement, sound, magnetism and vibration. AP landmines commonly use the pressure of a person's foot as a trigger.
2.3.2 Detonator or Igniter
It is responsible of the explosion of the mine. It may be a pressure sensor, an electronic sensor or any other type of sensors.
2.3.3 The Booster Charge
The booster charge is a highly sensitive explosive that will explode easily, when subjected to the shock of the detonator. Typically, a pea sized pellet of RDX is used. The purpose of the booster is to amplify the shock of the detonator and initiate the main explosive charge.
2.3.4 The Main Charge
The main charge consists of a stable explosive that is detonated by the booster charge. This is necessary, because making a mine out of a highly sensitive detonator or booster explosive would be more expensive, and makes the device more sensitive and thereby susceptible to accidental detonation. In most AP blast landmines TNT, Composition B or phlegmatised RDX, are used. For example, in a U.S. M14 mine, 29 grams of tetryl are used, while 240 grams of TNT are used in Russian PMN landmines.
2.3.5 The Case
The mine casing houses the components of the landmine and protects it from its environment. Early landmines, such as the ones used in the World War II, had casings made of steel or aluminium. However, by the middle of World War II, the British Army was using the first, practical, portable metal detectors the polish landmines. The Germans responded with landmines that had a wooden or glass casing to make detection harder.
7 Chapter 2 A Survey Study on Landmines Problem
Wooden mines had been used by the Russians in 1939, before the appearance of metal detectors, in order to save steel. Some, like the PP Mi D landmines, continued to be used into the 1980s as they were easy to make and hard to detect. Wood has the disadvantage of rotting and splitting, rendering the landmine non functional after a comparatively short time in the ground. Landmines manufactured after the 1950s generally use plastic casings to hinder detection by electronic landmine detectors. Some, referred to as minimum metal landmines, are constructed with as little metal as possible (often around 1 gram, 0.04 oz) to make them difficult to be detected. Landmines containing absolutely no metal have been produced, but are uncommon. By its nature, a landmine without any metal component in it cannot be found using a metal detector.
2.4 Types of Landmines
Various types of landmines have been manufactured and laid. There exist about 2000 different types of landmines for which catalogs exist [3, 4]. Typical landmines are shown in Figure (2 2).
Figure (2 2): Assortment of the most common landmines.
Although hundreds of landmine varieties exist, landmines generally can be classified as either “ blast ” or “ fragmentation ”. Blast landmines (see Figure (2 3)) are buried at shallow depths. They are triggered by pressure, such as from a person stepping on the mine. The weight needed to activate a blast landmine
8 Chapter 2 A Survey Study on Landmines Problem typically ranges from 5 to 24 lb, meaning that the landmines are easily triggered by a small child’s weight. They cause the affected object to blast into fragments, which blast upwards and often are the major cause of damage [3, 4]. Blast landmines typically are cylindrical in shape, 2 4 inches in diameter, and 1.5 3.0 inches in height. Generally, they contain 30 200 g of explosives. The casing may be made of plastic, wood, or sheet metal. Plastic encased blast landmines are sometimes referred to as “nonmetallic landmines”, but nearly all of them contain some metal parts, usually the firing pin and a spring/washer mechanism, weighing a gram [3, 4].
Figure (2 3): Blast landmine.
Fragmentation landmines (see Figure (2 4)) throw fragments radially outwards at high speeds. Most are lethal and can cause multiple casualties at distances of up to 100 m. One type of fragmentation landmine, known as the “bounding” landmine, is buried underground, but is propelled upwards, when activated and explodes a meter above ground, sending lethal fragments in a wide radius. Other types of fragmentation landmines are mounted on stakes in the ground or on tree trunks [3, 4].
Figure (2 4): Fragmentation landmine.
9 Chapter 2 A Survey Study on Landmines Problem
According to the potential target, landmines can be classified into AT and AP. Typical specifications for the two different types of landmines together with unexploded ordinance (UXO) are summarized in Table (2 1). Generally, UXO represents misfired shells or unexploded bombs that still remain for some reason. UXOs are usually found beneath the former battlefields. UXO has a collective meaning including various types of landmines [3, 4].
Table (2 1): Typical specifications of three different types of landmines [3, 4]
Type UXO AT AP Unspecified, Target Vehicle Human general Weight Various Heavy (6 to 11 kg) Light (0.1 to 4 kg) Size (in diameter) Various Large (13 to 40 cm) Small (6 to 15 cm) Plastic, metal or Case material Mostly metal Metal, plastic wood Detonation pressure Unpredictable 120 kg 0.5 kg
2.4.1 AT Landmines
Most AT landmines are made of metallic materials, and their sizes are bigger than those of AP landmines as indicated in Table (2 1). Since they have been designed to destroy vehicles, their detonation pressure is very high and they generate large metallic splinters after explosion. Two typical AT landmines are shown in Figure (2 5). The TM-62M is a larger size metallic case landmine with a diameter of 31 cm [3, 4]. This device detonator is so insensitive that a human can approach without explosion. The TMA-2 is a different type of AT landmines built in a plastic case.
Figure (2 5): Two typical AT landmines; (a) TM 62M and (b) TMA 2.
10 Chapter 2 A Survey Study on Landmines Problem
2.4.2 AP Landmines
AP landmines are the most difficult type of landmines to find and remove, and most civilian victims have been injured by this type of landmines. Most AP landmines are made of nonmetallic materials, and they are much smaller than AT landmines. AP landmines detonators are so sensitive that less than 10 kg of pressure can make them explode. AP landmines can be laid anywhere and can be set off in a number of ways stepping on them, pulling on a wire or simply shaking them. AP landmines may also explode, when an object placed over them is removed [3, 4]. AP landmines can be divided into three types; (i) blasting, (ii) bounding fragmentation, and (iii) directional fragmentation. The blasting type mines are the most common targets for humanitarian demining work [3, 4, 12, 13]. A blasting landmine is relatively smaller and lighter than other types of landmines. Blasting landmines are usually buried underground, but some models can be scattered by an airplane or floated on a river. For this reason, they can be found on the surface, underground, and at the riverside. Because of its simple mechanism and low material cost, small military groups can easily manufacture this type of mine. Such haphazard manufacturing and deployment of the blasting type AP landmines has resulted in serious mine problems, especially for poorer countries that cannot afford to invest in demining work. The bounding fragmentation type landmines are relatively larger than the blasting type. This type of landmines can destroy a larger area, while the blasting type landmines can damage only a target within a limited distance. Bounding fragment landmines are either buried underground or deployed on the surface. Direct pressure or a trip wire activates their detonators. Once the trigger is activated, they bounce up to a given altitude and explode with their lethal fragments spreading into an area of up to 30 m radius. Most directional fragmentation type landmines are deployed on the surface, and during explosion, they spread their fragments in a specific direction. Some models lethal range reaches over 200 m. Since they are detonated by manual operation as well as a trip wire, sometimes this type of landmines is
11 Chapter 2 A Survey Study on Landmines Problem considered as an active weapon. Some notable AP landmines are shown in Figure (2 6).
(a) (b)
(c) (d) Figure (2 6): Typical AP landmines; (a) PRB M35, (b) PMN, (c) VALMARA 69, and (d) MON 100. Both the PRB-M35 and PMN fall within the realm of blasting type landmines, which can be detonated by 8 kg of pressure. The PRB-M35 is one of the smallest mines with diameter of approximately 6 cm, which is as small as the diameter of a Coke can. If these landmines are buried or scattered on the ground covered with vegetation, they are very difficult to find and eliminate. Even lighter landmines can be spread by floating on water, and their distribution is unpredictable after heavy rains or flooding. The PMN is another example of a cheap, nonmetallic landmine with a cover made of rubber plate. The Valmara-69
12 Chapter 2 A Survey Study on Landmines Problem is a bounding fragmentation type landmine. Once detonated, the device propels upward and explodes with over 2,000 fragments spread over an area of 27 meters in radius. The MON-100 is a directional fragmentation type landmine. Its lethal range reaches over 100 m covering a 9.5 m arc [3, 4, 12, 13]. There are around 110 million landmines in the ground and another 100 million stockpiled around the world. More than 350 different types of AP landmines exist. Even if no more landmines are ever laid, they will continue to maim and kill for many years to come. Bold steps must be taken now to save future generations of innocent civilians. If sufficient funds are provided, deminers from the International Campaign to Ban Landmines (ICBL) say that landmine clearance necessary to restore daily life to near normal levels may be achieved in years, and not the decades once predicted.
2.5 Countries and Regions Affected by Landmines
As shown in the map in Figure (2 7), there is still a long way to go before the world is free of AP landmines. The countries worst affected are those, which have experienced wars in the past 20 years. Some armed groups continue to use them. Worldwide landmine distribution and its clearance status are summarized in Table (2 2) [3, 12].
Figure (2 7): Countries around the world affected by landmines.
13 Chapter 2 A Survey Study on Landmines Problem
Table (2 2): Worldwide landmine distribution and clearance status [3, 4].
Mined Cleared Mines Cleared Countries area area Casualties (million) mines (km2) (km2) Afghanistan 10 158,000 550~780 202 300~360/month Angola 15 10,000 Unknown 2.4 120~200/month Bosnia 3 49.01 300 84 50/month Herzegovina 38,786 or Cambodia 6 83,000 3,000 73.3 100/month Croatia 3 8,000 11,910 30 677 Egypt 23 11,000,000 3,910 942 8301 Eritrea 1 Unknown Unknown 2.48 2000 Iran 16 200,000 40,000 0 6000 Iraq 20 37,000 Unknown 1.25 6715 10,649 or Laos Unknown 251 43,098 Unknown 16~18/month Mozambique 3 58,000 Unknown 28 1759 Somalia 1 32,511 Unknown 127 4500 Sudan 1 Unknown 800,000 0 700,000 Vietnam 3.5 58,747 Unknown 65 180/month
14 Chapter 2 A Survey Study on Landmines Problem
Table (2 3) Countries and regions affected by landmines around the world [3, 4].
South Europe Middle East Africa Asia America Albania Algeria Angola Afghanistan Chile Abkhazia Egypt Burundi Bhutan Colombia Armenia Iran Chad Burma Cuba Bosnia Iraq Congo Cambodia Ecuador Herzegovina Israel/Palestinian Croatia DR Congo China Falklands Territories Cyprus Lebanon Eritrea India Nicaragua Denmark Morocco Ethiopia Laos Peru Guinea Georgia Oman Nepal Venezuela Bissau Greece Syria Libya North Korea Montenegro Tunisia Malawi Pakistan Moldova Yemen Mauritania South Korea Nagorno Mozambique Sri Lanka Karabakh Russia Namibia Taiwan Serbia Niger Tajikistan Turkey Rwanda Thailand Senegal Vietnam Somaliland Swaziland Sudan Uganda Western
Sahara Zambia Zimbabwe
2.6 Landmines in Egypt Due to the central geographical location between Africa, Asia and Europe, Egypt was a location for many battles. During World War II, the most known El Alameen battle, Western Desert, see Figure (2 8), was between the British and German troops. As a result of this fighting, a large number of AT and AP landmines has been left. The 23 million landmines and UXO buried in the Egyptian land are considered to be about 21% of the total number of landmines buried in the whole world [12, 13].
15 Chapter 2 A Survey Study on Landmines Problem
The Egyptian contaminated area is believed to be very rich in natural resources and development potential. Official Egyptian sources have estimated that 16.7 million landmines affect 3910 million square meters in the Western Desert area (from Alexandria to the Libyan border and 30 kilometers wide from the Mediterranean coastline) and 5.1 million landmines affect 200 million square meters in eastern areas (Sinai Peninsula and Red Sea coast). Other Egyptian officials have stated that only 20% ~ 25% of these landmines are really landmines, the remainder being other types of UXO. AP mines believed to be in the Western Desert include German S type bounding fragmentation mines and British Mk 2 mines. Anti vehicle mines are thought to include German Riegalmine 43, Tellermine 35, Tellermine 42 and Tellermine 43 mines, Italian B 2 and V 3 mines, and British Mk 5 and Mk 7 mines [13]. Figure (2 9) illustrate some models of landmines buried in the Egyptian territories. According to the NGO Protection, very few mined areas are marked or mapped, and Egyptian civilians continue to use the mine affected areas for cultivation, grazing, infrastructure projects, and housing. In addition, mines and UXO in the Western Desert deny access to the reserves of an estimated 4.8 billion barrels of oil and 13.4 trillion cubic feet of natural gas [13].
16 Chapter 2 A Survey Study on Landmines Problem
Figure (2 8): Location map for landmine distributions in Egypt.
Figure (2 9): Types of landmines in the Western Desert, Egypt.
17 Chapter 2 A Survey Study on Landmines Problem
2.7 Demining Approaches and Technologies The process of detecting and removing landmines, is called demining . The demining methods currently being used are not safe for clearing landmines. The methods are neither cost effective nor efficient. Mine clearance itself can be accomplished through different methods with varying levels of technology, but the most laborious way is still the most reliable. We should favor technologies that can be manufactured in mined countries, that are transferable, and that provide employment and economic infrastructure, where it is most urgently required.
2.7.1 Manual Demining Manual demining is extremely dangerous; one deminer has been killed for every 2,000 mines removed, with even more civilian victims. Manual demining is a procedure in which landmines are manually detected and neutralized by a human deminer [14 17]. The deminer first scans the ground with a metal detector. Then, the deminer uses a prodder in order to feel, locate, and identify the object causing the signal, after which the deminer carefully uncovers it. When operating in this way, the detection phase still relies heavily on metal detectors, whereby each alarm needs to be carefully checked until it has been fully understood and/or its source removed. Most landmines contain enough metal to be detected by this method; however, only about one in every 1000 signals detected belongs to a landmine or UXO. In general, the ground being cleared is often saturated with metal fragments, shrapnel and cans, making manual demining methods slow, extremely dangerous, and expensive. Complicating matters more, about one third of all AP landmines currently deployed are metal free. The accuracy of metal detection depends heavily on the level of mineralization and debris in the ground, the type of landmine used, and the time needed to clear land varies enormously, depending on local conditions.
2.7.2 The Use of Animals, Insects and Bacteria So far, dogs are considered the best detectors of explosives. Their sensitivity to this kind of substance is estimated to be 10,000 times higher than
18 Chapter 2 A Survey Study on Landmines Problem that of a man made detector [18, 19]. Specially trained dogs are used to detect the characteristic smell of explosive residue that emanates from mines regardless of their composition or how long they have been implanted. This enables the dogs to detect mines with low metal content that are undetectable by metal detectors. In addition, because dogs do not respond to metal, soil or non explosive objects, they eliminate much of the time consuming shortcomings of manual detection techniques. Mine detection dogs can work in almost all types of terrain. They are also easy to transport and highly reliable, and they can screen land up to five times faster than manual deminers. South Africa and Afghanistan have reported success, but it was more in locating the edges of mine fields than in finding individual mines. Dogs can be overwhelmed in areas with dense landmine contamination. Moreover, they can only work for short periods each day (about a couple of hours a day). Dogs can become confused if they can smell explosives coming from several sources at once. The effectiveness of the dogs depends entirely on their level of training, the skill of their handlers, and their correct use. Trained rats may be the best and cheapest form of landmine detector. Rats have certain advantages over dogs. They have a better sense of smell, are cheaper to keep and maintain, and are more resistant to tropical diseases. Since they are smaller, they can be transported even more, easily. In addition, they are very suitable for repetitive tasks. African pouched rats in particular have sensitive noses and can be trained like dogs to detect explosive vapors. The Geneva International Center for Humanitarian Demining (GICHD) is examining the use of rodent detection as part of a dog study [20]. Besides dogs and rats, other animals are being considered for their possible use as mine detectors. Researchers at Sandia National Laboratories and the University of Montana are trying to determine whether foraging bees can reliably and inexpensively detect buried landmines. They are trying to see if bees can be trained to find residues of TNT (the primary ingredient of most landmines) and bring the evidence home [21]. Also, pigs are thought to be better at "sniffing" than dogs and might be better at finding mines. So far, no open literature has been seen describing any tests or trials. An additional technology
19 Chapter 2 A Survey Study on Landmines Problem for getting rid of landmines and UXO that is now under study at Oak Ridge National Laboratories (ORNL) could take advantage of the same microscopic, genetically engineered bacteria that are also being used in waste management technologies. These bacteria can be genetically engineered to glow in the presence of certain compounds, including explosives [19].
2.7.3 Mechanical Demining
Mechanical approaches rely on the use of motorized mine clearers whose design is influenced by military demining requirements. Military devices are designed to clear only a navigable path through a field rather than remove all the mines in the area. A number of mechanical mine clearing machines have been constructed or adapted from military vehicles or armored vehicles of the same or similar type, with the same or reduced size [22]. Mechanical mine clearance systems (such as armored vehicles, plows and flails) unearth mines or force them to explode under the pressure of heavy machinery. Mechanical clearance may be used on large areas (agricultural areas, for instance) and favorable terrain such as flat, sandy areas with no dense vegetation. In small paths or thick bush, such machines simply cannot maneuver. These systems are employed for mine verification and area reduction tasks as well as actual mine field clearance. Large mechanical systems do require substantial investments, not only for machine costs, but also for logistics and maintenance, and they can only be employed on a fraction of the total mined areas. The mechanical approach is fast, but it cannot achieve the humanitarian demining accuracy and safety standards, nor will it in the near future. With this technique, machines often do not destroy all mines in a contaminated area, and AP mines may be pushed to the side or buried deeper or partly damaged, making them more dangerous. However, mechanical clearance in support of manual clearance can be cheaper and significantly safer for deminers. In some terrains and circumstances, it is difficult to imagine mechanical methods being applicable (e.g., in defensive ditches, around large trees, inside residential areas, on soft terrain, etc.) However, machines can speed the clearance process when used in
20 Chapter 2 A Survey Study on Landmines Problem combination with manual clearers, and they may also be useful for quickly verifying that an area is clear of landmines so that manual clearers can concentrate on those areas that are most likely to be infested [22].
2.7.4 Robots and Humanitarian Demining
Most people in the mine clearance community would be delighted if their work could be done remotely or, even better, robotically. The benefits of mounting a mine detector on a remotely controlled vehicle must be balanced against the added cost and possible reduction in efficiency. A cost analysis should be conducted to determine to what extent remotely controlled vehicles are justified [19 24]. Properly sized robotic solutions with a suitable modularized mechanized structure that are well adapted to local conditions of mine fields can greatly improve the safety of personnel as well as the efficiency, productivity, and flexibility of the work. Solving this problem presents challenges in robotic mechanics and mobility, sensors and sensor fusion, autonomous or semi autonomous navigation, and machine intelligence. Furthermore, the use of many robots coordinating their movements will improve the productivity of overall mine detection processes through the use of team cooperation and coordination [19 24]. One benefit would be increased safety by removing the operator from the hazardous area. There are still some doubts whether such equipment will operate effectively when the operator is at a distance or has been removed altogether. There is a little value in a system that makes life safer for the operator, but that is less effective at clearing the ground. Accordingly, a serious evaluation and analysis should be conducted, and efficient designs and techniques should be developed. A reasonably cheap but reliable robot platform is required as the ultimate solution. The target robot should have the capability to operate in different control modes, including the tele operated and semi autonomous modes. The robot should have reliable navigation capabilities over an area to be cleared with efficient and flexible locomotion capability. It will have to be designed to not
21 Chapter 2 A Survey Study on Landmines Problem exceed the threshold that sets off the mines in question. Lastly, it should be easy to use; even someone with only basic training should be able to operate the system. The possible introduction of robots into the demining process can be done through surface preparation and marking, verification, sped up detection and mapping, and mine removal or neutralization. Clearly, it is difficult to design a universal robot/machine that is applicable to different terrains and works under different environmental conditions to meet demining requirements. The high cost and sophisticated technology used in robots that require highly trained personnel to operate and maintain them are additional factors limiting the possibilities of using robots for humanitarian demining. In spite of this, many efforts have been made to develop effective robots for cheap and fast solutions [19 24].
2.7.5 Landmine Detection and Sensing Technologies
Landmine detection represents the slowest yet most important step of the demining process, and the quality of mine detectors affects the efficiency and safety of this process. Landmine detection targets need to achieve a high probability of detection, while maintaining a low probability of false alarms. It is important to develop effective detection technologies that speed up the detection process, maximize detection reliability and accuracy, reduce the false alarm rate, improve the ability to positively discriminate landmines from other buried objects and metallic debris, and enhance the safety and protection of deminers. In addition, there is a need to have simple, flexible, and user friendly interaction that allows safe operation without the need for extensive training. Furthermore, careful study of the limitations of any tool with regard to the location, environment, and soil composition is critical. Knowing the required technical operation and maintenance skills is as important as remembering that not all high tech solutions may be workable in different soil and environmental conditions. The development phase of such new technologies requires a well established set of testing facilities that simulates conditions closely resembling those of the mine affected area. The testing phase should be followed by
22 Chapter 2 A Survey Study on Landmines Problem extensive field trials in real scenarios to validate the new technologies under actual field conditions to specify benefits and limitations of different methods. The work must be performed in close cooperation with end users of the equipment, and real deminers should carry out the test at a real site. This will ensure that the developments are consistent with practical operational procedures in the context of humanitarian demining and that the technology is fulfilling user requirements. Also, there is a need to have a reliable set of global standards for assessing the availability, suitability, and affordability of technology with common information tools that allow for these assessments and evaluations. This can be enhanced by benchmarking the performance levels to develop equipment, systems, and algorithms [18, 25]. The idea of developing multi sensor solutions involving two or more sensors linked to computer based decision support systems with advanced signal processing techniques is attractive and is advocated by many researchers as a fruitful line of development. Because of this, there is a need to use complementary sensor technologies and to have an appropriate sensor data fusion. A critical need is the ability to distinguish fragments or stones from the target material in real time [26, 27]. Finally, Table (2 4) gives a summary of the detection technologies reviewed.
23 Chapter 2 A Survey Study on Landmines Problem
Table (2 4): Summary of the detection technologies reviewed [18 25].
Potential for Humanitarian Technology Operating Principle Strengths Limitations Mine Detection Electromagnetic Electromagnetic induction Induces electric Performs in a range Metal clutter; low Established technology currents in metal of environments metal mines components of mine Ground penetrating radar Reflects radio waves Detects all anomalies, Roots, rocks, water Established technology off mine/soil interface even if nonmetal pockets, other natural clutter; extremely moist or dry environments Electrical impedance Determines electrical Detects all anomalies, Dry environments; can Unlikely to yield major gains tomography conductivity even if nonmetal detonate mine distribution X ray backscatter Images buried objects Advanced imaging Slow; emits radiation Unlikely to yield major gains with x rays ability Infrared/hyperspectral Assesses temperature, Operates from safe Cannot locate Not suitable for close in light reflectance standoff distances individual mines Detection differences and scans wide areas quickly
Acoustic/Seismic Reflects sound or Low false alarm rate; Deep mines; vegetation Promising seismic waves off not reliant on cover; frozen ground mines electromagnetic properties Explosive Vapor Biological (dogs, bees, Living organisms Confirms presence of Dry environments Basic research needed to bacteria) detect explosive vapors explosives determine potential (though dogs are widely used) Fluorescent Measures changes in Confirms presence of Dry environments Basic research needed to
24 Chapter 2 A Survey Study on Landmines Problem
polymer fluorescence explosives determine operational potential in presence of explosive vapors Electrochemical Measures changes in Confirms presence of Dry environments Basic research needed to polymer electrical explosives determine whether detection resistance upon limit can be reduced exposure to explosive vapors Piezoelectric Measures shift in Confirms presence of Dry environments Basic research needed to resonant frequency of explosives determine whether detection various materials upon limit can be reduced exposure to explosive vapors Spectroscopic Analyzes spectral Confirms presence of Dry environments Basic research needed to response of sample explosives determine whether detection limit can be reduced Bulk Explosives Nuclear quadrupole Induces radio Identifies bulk TNT; liquid explosives; Promising resonance frequency pulse that explosives radio frequency causes the chemical interference; quartz bonds in explosives to bearing resonate and magnetic soils Neutron Induces radiation Identifies the Not specific to Unlikely to yield major gains emissions from the elemental content of explosives molecule; atomic nuclei in bulk explosives moist soil; ground explosives surface fluctuations
Advanced Prodders/ Provide feedback about Could deploy almost Hard ground, roots, Promising Probes nature of probed object any type of detection rocks; requires physical and amount of force method contact with mine applied by probe
25 Chapter 2 A Survey Study on Landmines Problem
2.8 Summary
This chapter presented a short survey on landmine problems, their definitions, components, types, the different approaches and technologies used in demining (detection and clearance), and a comparison between them in terms of principles of operation, strengths, limitations, and potential for humanitarian mine detection.
26 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
CHAPTER 3
REVIEW STUDY ON LANDMINE DETECTION SYSTEMS AND ARTIFICIAL NEURAL NETWORKS
3.1 Introduction There are many techniques that can be used for detection of landmines. This chapter will give an overview of one of these techniques in more details, which is the Acoustic to Seismic (A/S) landmine detection technique. The main idea of this technique in order to detect the landmine objects is to vibrate them with acoustic or seismic waves that are generated and received by acoustic and seismic transducers, respectively. The A/S landmine detection technique consists of a transmitter, which generates the acoustic waves and a receiver, which measures the vibrations. In this technique, the transmitting system is composed of acoustic loudspeakers, and the receiving system is composed of a Laser Doppler Vibrometer (LDV). This chapter presents also the segmentation of acoustic landmine images and the ANNs; concepts, learning rules, neuron models, activation rules, network architectures, and back propagation neural networks. This chapter is organized as follows. Section (3 2) presents a literature review on LDV based A/S landmine detection. Section (3 3) presents the principles of A/S coupling. Section (3 4) explains the A/S transmission system with acoustic loudspeakers. Section (3 5) explains the A/S receiving system
27 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks using the LDV. Section (3 6) explains the segmentation of acoustic landmine images based on thresholding, morphology and wavelet transform. Section (3 7) gives the main concepts of ANNs. Finally, Section (3 8) presents the types of noise that might affect the system.
3.2 Literature Review on LDV Based A/S Landmine Detection
There are several studies on the LDV based A/S landmine detection, since the 1950, especially in the last ten years. In 2000, Valeau, Sabatier and Xiang have provided results for different mine depths, types of soil, and sweeping velocities [28]. Also, Sabatier and Xiang presented some measurements using a Scanning Laser Doppler Vibrometer (SLDV) and data analysis techniques for mine detection to increase the speed of scanning [29]. In 2001, Sabatier and Xiang investigated the A/S coupling measurements using the LDV technique for AT and AP landmine detection [2, 30]. In 2002, Johan et al . used two LDV systems with two different wavelengths for the different approaches; one based on a He Ne laser at 0.633 m with acoustic excitation and the other based on an erbium fiber laser at 1.54 m for laser excitation. The acoustic excitation gives a good contrast between the buried mine and the surrounding soil at certain frequencies. Laser excitation gives a pulse response that is more difficult to interpret, but it is potentially a faster technique. In both cases, buried mines could be detected [31]. In 2003, Burgett et al . designed a moving platform, which was tested on different types of mines at various depths and different speeds [32]. Also, in 2003, Amit et al . designed a multiple beam (16 beams) LDV system to locate buried landmines with the laser acoustic technique. The LDV device increases the speed of landmine detection by simultaneously probing 16 positions on the ground over a span of 1 meter and measuring the ground velocity at each of these positions [33]. In 2004, Wagstaff et al . applied the Hilbert and the Fast Fourier Transforms on the multi beam LDV data for demodulating the data. The ground velocity can be obtained from these data to identify a mine presence or absence
28 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
[34]. Also, in 2004, Sabatier et al . presented a sensor fusion technique for the Acoustic Landmine Detection (ALD) data for the non metal buried landmine detection with a Ground Penetration Radar (GPR) technique for the metal buried landmine detection. The experimental results showed the potential and the inherent limitations of the technique [35]. In 2005, Muir made an experiment in a clay soil by using two types of electromagnetic shaker sources. These sources were in co located arrays; one vertically oriented to generate Rayleigh waves and the other is transversely oriented to generate Loove waves. Two different cylindrical landmine targets were tested. The responses of the mine cases to Rayleigh and Loove wave excitation were measured and compared to data from a reference seismometer deployed nearby, on natural, undisturbed soil [36]. In 2006, Sutin et al . suggested a Time Reversal Acoustic (TRA) technique to implement multiple transmitters, simultaneously to increase the scanning speed. In this technique, the TRA signal is transmitted with an opposite sign, and the two received signals are added in a post processing step. The summed signal contains mainly the results of nonlinear wave interaction and tends to cancel the linear response [37]. Also, in 2006, Anthony et al . used the LDV based A/S system to determine the directions of the objects under water [38]. In October 2006, Aranchuk et al . presented a method for buried landmine detection in the field based on using elastic waves in the ground and an LDV as a vibration sensor. This method has shown an excellent performance in field tests [39]. In 2007, Changqing et al . presented a method based on a YAG laser at 1.06 m with acoustic excitation and on an erbium fiber laser at 1.54 m. They reduced the false alarm in plastic landmine detection [40]. In December 2007, Attenborough et al . presented a model for an A/S landmine detection system using a large loudspeaker and accelerometers [41]. In 2008, Kasban et al . presented two innovative techniques for the automation of the landmine detection in a data scanned by the LDV based A/S landmine detection system. These techniques are based on the intensity
29 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks component of the color landmine images or on grayscale versions of these images [42]. In 2009, Kasban et al . proposed some techniques for the automatic detection of objects from the acoustic images, which are obtained from the LDV based A/S landmine detection system. These techniques are based on color image transformations, morphological image processing, and the wavelet transform. The proposed techniques are compared considering the probability of detection, the false alarm rate, the accuracy and the processing speed [43]. Also in 2009, Abd El samie introduced a cepstral approach for the automatic detection of landmines from acoustic images. Cepstral features are extracted from a group of landmine images, which are transformed first to 1 D signals by lexicographic ordering. MFCCs and polynomial coefficients are extracted from these 1 D signals to form a database of features, which can be used to train a neural network with the landmine features. The landmine detection can be performed by extracting features from any new image with the same method used in the training phase. These features are tested with the neural network to decide whether a landmine exists or not [44]. In 2010, Kasban et al . presented a new technique for false alarm rate reduction to improve the performance of automatic object detection systems that operate on digital images. This technique is applied to the images obtained by the LDV based A/S system and based also on morphological image processing and the DWT of the image data. It begins with a color image transformation to obtain a grayscale image, and then a closing operation that enhances the larger objects in the image. After that, the DWT is applied to the closed image. The decimation effect of this DWT helps in the removal of small clutter objects in the image. The approximation component resulting from the DWT is used in the detection process [45].
30 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.3 Principles of A/S Coupling
When sound waves above the ground penetrate the ground surface, the sound energy coupled into the ground causes seismic motion of the ground surface and sub surface as shown in Figure (3 1).
Figure (3 1): A/S coupling principle.
Some of the incident (Ai) acoustic waves are reflected (Ar), and the rest are coupled into the ground resulting in seismic waves. Seismic waves are the waves of energy resulting from the sudden breaking of rock within the earth or an explosion. There are several kinds of seismic waves, which move in different ways. The two main types of waves are body waves and surface waves. Body waves can travel through the earth inner layers, while surface waves can only move along the surface of the planet like ripples on water. The main types of seismic waves are explained below.
3.3.1 Body Waves
Body waves travel through the interior of the earth. These waves have frequencies higher than surface waves. They can be classified into primary waves and secondary waves.
31 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.3.1.1 Primary Waves
The primary waves (P waves) are the first kind of body waves. These waves are the fastest kind of seismic waves, and consequently the first to 'arrive' at a seismic station.
3.3.1.2 Secondary Waves
The secondary waves (S waves) are the second type of body waves. S waves are slower than P waves and can only move through solid rock and not in any liquid medium.
3.3.2 Surface Waves
Surface waves travel only through the crust. They have frequencies lower than body waves and are easily distinguished on a seismogram using the arrival time as they arrive after body waves. These waves include Loove waves and Rayleigh waves.
3.3.2.1 Loove Waves
The first kind of surface waves is called Loove waves, named for A. E. H. Loove, a British mathematician, who worked out the mathematical model for this kind of waves in 1911. They are the fastest surface waves and move in the ground from side to side confined to the surface of the crust.
3.3.2.2 Rayleigh Waves
The other kind of surface waves is the Rayleigh waves, named for Lord Rayleigh, who mathematically predicted the existence of this kind of waves in 1885. Rayleigh waves roll along the ground just like waves rolling across a lake or an ocean. Because they move in the ground up and down and side to side in the same direction that the original acoustic wave is moving. The Rayleigh waves cause the vibrations of landmines. Buried landmines in the subsurface will resonate and induce distinct changes in the seismic motion due to scatter and reflection of the sound energy coupled into the ground. Solid borne (seismic)
32 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks excitation is also possible. The soil surface vibrations are sensed remotely using some sensors. The A/S landmine detection technique consists of a transmission system, which generates acoustic or seismic waves into the area under test and a receiving system, which senses the changes in the mechanical properties of the area under test. The transmission system may be composed of acoustic loudspeakers or electrodynamic shakers. The receiving system may be composed of microphones, geophones / accelerometers, Ultrasonic Doppler Vibrometry (UDV), Radar Doppler Vibrometry (RDV) or LDV [46]. In this thesis, we focus on the acoustic loudspeaker as a transmission system and the LDV as a receiving system.
3.4 A/S Transmission System with Acoustic Loudspeakers
Loudspeakers are electro acoustic transducers, which convert the electrical energy into an acoustic energy. In A/S landmine detection systems, the transmission system is made of two or three subwoofer loudspeakers of 100 200 W, as a sound source with a frequency band 100 – 600 Hertz depending on the depth of the landmine. The loudspeakers are decoupled from the soil using tripods on which speakers are mounted. The sound pressure level in air ranges between 90 dB and 120 dB, and typically a linear power amplifier with output power from 100 to 200W is used [47].
3.5 A/S Receiving System Using LDV
All images used in this thesis are obtained using an LDV system. The vibrational velocity on the ground surface is sensed using the LDV. A laser beam is emitted from the LDV onto a vibrating surface of an object under test as shown in Figure (3 2). The surface vibrational velocity causes a Doppler frequency shift of the laser light. The backscattered light from the measured object takes the opposite path back into the interferometer and is sensed by its photo detector. As a result, a Frequency Modulated (FM) signal from the photo detector carries the information about the surface velocity. After FM demodulation of the detector
33 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks output, the signal is proportional to the surface velocity under test. For the purposes of the present discussion, it is sufficient to mention that the output of an LDV system is a voltage proportional to the instantaneous velocity of a particular spot on the object under test. A monitor of a Personal Computer (PC) displays a 2 D image of the ground surface being scanned by the XY mirrors. Prior to scanning, a measurement grid is defined and superimposed on the image of the ground surface. For more details about the LDV, the reader should see [48, 49, 50].
Figure (3 2): LDV based A/S landmine detection system.
The LDVs are particularly well suited to this measurement application because of their high sensitivity to detect AP and AT mines, excellent spatial resolution and long working distances. The probability that this system introduces some clutter as landmines in the identification process should be small
34 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
(low false alarm rate). Since the removal of a landmine is very expensive, the false alarm should be as low as possible. In our case, the false alarm probability is about 0.03/square meters. Examples for landmine images obtained by this technique are shown in Figure (3 3). These are samples of the images used in the training and testing phases.
(a) AP landmine images.
(b) AT landmine images. Figure (3 3): Samples of acoustic landmine images. 3.6 Segmentation of Acoustic Landmine Images
The segmentation process is very important for the elimination of irrelevant information, removing the background and clutters, and leaving only the objects of interest in the image. Here, segmentation is based on thresholding, morphology, the wavelet transform for increasing the detection probability and decreasing the rate of positive false alarm. The Region of Interest (ROI) is determined by thresholding the landmine image with a threshold obtained automatically to record the
35 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks dimensions of the object area. Then, all clutters are removed by an area thresholding process [51]. Morphology has become a popular tool in many image processing applications such as noise removal, image enhancement, and image segmentation [51]. Morphology is a means of structuring and reshaping a binary or a gray scale image. The tool used in reshaping the image is called the morphological structuring element. The structuring element is a simple matrix or a small window that reshapes the image. There are many shapes for the structuring element such as the rectangle, the square, the octagon, the periodic line and the flat disk shapes. There are four basic types of morphological operations, which will be described briefly below:
3.6.1 Erosion
Each object pixel that is touching a background pixel is changed into a background pixel. This operation makes the objects smaller, and can break a single object into multiple objects. Consider an image (F) and a structuring element (S e), the eroded image of the image F by the structuring element S e can be obtained using the following equation [52]:
F Θ S e = {z (| S e ) z ⊆ F } (3 1) where z is a displacement of the structuring element.
3.6.2 Dilation
Each background pixel that is touching an object pixel is changed into an object pixel. This operation makes objects larger, and can merge multiple objects into one. The dilated image F by a structuring element S e can be obtained using the following equation [52]: ∧ F ⊕ Se = z [(| S e ) z ∩ F] ⊆ F (3 2)
36 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.6.3 Opening
Erosion followed by dilation called opening operation. It removes small islands and thin filaments of object pixels. It is obtained using the following equation [52]:
F o S e = ( F Θ S e ) ⊕ S e (3 3)
3.6.4 Closing
Dilation followed by erosion is called closing operation. It removes islands and thin filaments of background pixels. It is obtained using the following equation [52]:
F • Se = (F ⊕ Se ) Θ Se (3 4) The wavelet transform is a mathematical operation used to divide a given image into different sub bands of different scales to study each sub band, separately [50, 53]. The wavelet transform is used to improve the performance of the segmentation process. Figure (3 4) shows the block diagram of the wavelet based segmentation method, where the variance image is calculated from a grayscale image, then the closing morphological operation is applied to the grayscale variance image using a flat disk structuring element, and the 2 D Haar DWT is applied to the closed image leading to a decomposition of this image into four components; the approximation component and three detail components. Lastly, the variance based segmentation method is applied to the approximation component.
Se
Variance Closed image image Transformation to Image 2 D Haar Landmine image grayscale variance closing DWT Approximation component Segmented
image Automatic segmention
Figure (3 4) Block diagram of the wavelet based segmentation method.
37 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.7 Artificial Neural Networks (ANNs)
An ANN is a powerful data modeling tool that is able to capture and represent complex input/output relationships. The motivation for the development of neural network technology stemmed from the desire to develop an artificial system that could perform “intelligent” tasks similar to those performed by the human brain. Neural networks resemble the human brain in two aspects. A neural network acquires knowledge through learning. The neural network knowledge is stored within inter neuron connection strengths known as synaptic weights. The true power and advantage of neural networks lies in their ability to represent both linear and non linear relationships and in their ability to learn these relationships directly from the data being modeled [54]. The ANNs are widely used for feature matching because of their high efficiency and throughput . The main advantages of the neural network approach are as follow [55]: • Adaptive learning: An ability to learn how to do tasks based on the data given for training or initial experience. • Self Organization: An ANN can create its own organization or representation of the information it receives during the learning time. • Real Time Operation: ANN computations may be carried out in parallel, and special hardware devices are being designed and manufactured taking advantage of this capability. • Fault Tolerance via Redundant Information Coding: Partial destruction of a network leads to a corresponding degradation of performance.
Artificial neurons work in union to solve specific problems. ANNs, like people, learn by examples. An ANN is configured for a specific application, such as pattern recognition or data classification through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurons. This is true for ANNs as well. An artificial neuron (also called a “node”) is a basic unit in an ANN. Artificial neurons are simulations of biological neurons. They receive one or more inputs and sum them to produce an
38 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks output. Usually, the sums of each node are weighted, and the sum is passed through a non linear function. A known form of the transfer function is the sigmoid, but it may also take other nonlinear forms like the piecewise linear form, or the step form.
3.7.1 Neural Network Concepts
Concepts related to neural networks give enough details to provide some understanding of what can be accomplished with neural network models and how these models are developed. The basic concepts of a neural network will be defined in this section. The neural network structure is shown in Figure (3 5).
3.7.1.1 Cells
A cell (or unit) is an autonomous processing element that models a neuron. The cell can be thought of as a very simple computer. The purpose of each cell is to receive information from other cells, and then it performs relatively simple processing tasks of the combined information, and sends the results to one or more other cells.
Input Hidden Layer Output Layer Layer
Figure (3 5): The neural network structure.
3.7.1.2 Layers
A layer is a collection of cells that can be thought of as performing a common function. It is generally assumed that no cell is connected to another cell
39 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks in the same layer. All neural networks have an input layer, and an output layer to interface with the external environment. Each input and output layer has at least one cell. Any cell that is in between the input layer and the output layer is said to be in a hidden layer. Neural networks are often classified as single layer or multi layer networks. 3.7.1.3 Arcs An arc (or connection) is a one way communication link between two cells. A feed forward network is one in which the information flows from the input layer through some hidden layers to the output layer. A feedback network, by contrast, also permits “backward” communication.
3.7.1.4 Weights
A weight w ij is a real number that indicates the influence that a cell u j has on cell u i. The weights are often combined into a weight matrix W. These weights may be initialized as zeros, or initialized as random numbers, but they can be altered during the training phase.
3.7.2 The Learning Rules of Neural Networks
There are three main types of learning in neural networks [56, 57]:
3.7.2.1 Supervised Learning
With this type of learning, we provide the network with input data and the correct answer i.e. what output we wish to receive given that input data. The input data is propagated forward through the network till activation reaches the output neurons. We can then compare the answer, which the network has calculated with that which we wished to get. If the answers agree, we need to make no change to the network; if the answer which the network is giving is different from that which we wished, then we adjust the weights to ensure that the network is more likely to give the correct answer in future if it is again presented with the same (or similar) input data. This weight adjustment scheme is known as supervised learning or learning with a teacher.
40 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.7.2.2 Unsupervised Learning
With this type of learning, we only provide the network with the input data. The network is required to self organize (i.e. to teach itself) depending on some structure in the input data. Typically, this structure may need some form of redundancy in the input data or clusters in the data.
3.7.2.3 Reinforcement Learning
It is a half way house between the above mentioned learning algorithms. In this type of learning, we provide the network with the input data and propagate the activation forward, but only tell the network has produced a right or a wrong answer. If it has produced a wrong answer, some adjustment of the weights is done so that a right answer is more likely in future presentations of that particular piece of input data.
3.7.3 Neuron Model
3.7.3.1 Simple Neuron
Input Neuron without bias Input Neuron with bias
x w v x w v y ƒ y F ƒ
b y=f(wx+b) y=f(wx+b) Figure (3 6): Simple neuron structure.
The scalar input x is transmitted through a connection that multiplies its strength by the scalar weight w, to form the product W.x, again a scalar. Here the weighted input W.x is the only argument of the transfer function f, which produces the scalar output y. The neuron on the right has a scalar bias b. You may view the bias as simply being added to the product W.x as shown by the summing junction or as shifting the function f to the left by an amount b. The bias is much like a weight, except that it has a constant input of 1.
41 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
The transfer function net input v, again a scalar, is the sum of the weighted input W.x and the bias b. This sum is the argument of the transfer function f. Here, f is a transfer function, typically a step function or a sigmoid function, which takes the argument v and produces the output y. w and b are both adjusted parameters of the neuron. The central idea of neural networks is that such parameter can be adjusted so that the network exhibits some desired or interesting behavior. Thus, we can train the network to do a particular job by adjusting the weight or bias parameters, or perhaps the network itself will adjust these parameters to achieve some desired end.
3.7.3.2 Neuron with Vector Input
A neuron with a single m element input vector is shown in the Figure (3
7). Here the individual element inputs x1, x 2 ,……..,x m are multiplied by weights w 1,1 , w 2,1 ,...... w ,1 m and the weighted values are fed to the summing junction. Their sum is simply W.x, the dot product of the single row matrix W and the vector x. The neuron has a bias b, which is summed with the weighted inputs to form the net input v. This sum, v, is the argument of the transfer function f. In mathematical terms, we specify a neuron k by the following equations:
m u k = ∑ w kj x j (3 5) j = 1
vk= u k + b k (3 6)
yk= f( v k ) (3 7) where x 1, x 2,…….x m are the input signals; w k1, w k2, …….,w km are the synaptic weights of neuron k; u k is the linear combiner output due to the weight signals; b k is the bias, which has the effect of applying affine transformation to the output u k in Equation (3 6); f is the activation function; y k is the output signal of the neuron
42 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.7.4 Activation Rules
The network rule is often given by an activation function f(x) to produce the neuron output signal. Most frequently, the same function is used for all of the cells. Several different functions have been used in neural network simulations.
Input Neuron with bias
x1 y w1,1 x2 F ƒ x3
w1,m b xm y=f(wx+b) Figure (3 7): Neuron with vector input structure.
3.7.4.1 Identity Function
f(x) =x for all x (3 8) This activation rule is just the value of the combined input as shown in Figure (3 8a)
3.7.4.2 Step Function
1 for x ≥ θ f (x) = (3 9) 0 for x < θ The output is zero until the activation reaches a threshold θ; then it jumps up by the amount shown in Figure (3 8b).
3.7.4.3 Sigmoid Function
1 f (x) = (3 10) 1+ e −x The sigmoid, (meaning S shaped) function, is bounded within a specific range [0, 1]. It is often used as an activation function for neural networks, in
43 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks which the desired output values are either binary or in the interval between 0 and 1. This is shown in Figure (3 8c).
Figure (3 8): Activation functions. (a) Identity function. (b) Threshold function. (c) Sigmoid function. 3.7.5 Network Architectures
3.7.5.1 Single-Layer Network:
This network has an input layer of source nodes that projects onto an output layer of neuron. With the designation “single layer" referring to the output layer of computation nodes (neurons). We don't count the input layer of source nodes, because no computation is performed there [56].
Input layer Output layer
Figure (3 9): Single layer neural network.
44 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
3.7.5.2 Multi-Layer Network:
This class of neural network distinguishes itself by the presence of one or more hidden layers, whose corresponding computation nodes are called hidden neurons or hidden units. The function of hidden neurons is to intervene between the external input and the network output in some useful manner. ANNs are characterized in principle by a network topology, a connection pattern, neural activation properties, a train strategy and the ability to process data. The most common neural network model is the multilayer perceptron (MLP) [8, 9, 57]. This type of neural network is known as a supervised network, because it requires a desired output in order to learn. The goal of this type of network is to create a model that correctly maps the input to the output using historical data so that the model can then be used to produce the output, when the desired output is unknown. Figure (3 10) shows the block diagram of the MLP. The inputs are fed into the input layer and get multiplied by interconnection weights as they are passed from the input layer to the hidden layer , then they get summed, then processed by a nonlinear function. Finally, the data is multiplied by interconnection weights, then processed one last time within the output layer to produce the neural network output. Mapping is needed to train the neural network. The MLPs have been applied successfully to solve some difficult and diverse problems by training them in a supervised manner with a highly popular algorithm known as the error back propagation algorithm.
45 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
Hidden Layers Connections
Input Layer Output Layer
Figure (3 10): Multi layer neural network. 3.7.6 Back-Propagation Neural Network
The back propagation algorithm trains a given feed forward multilayer neural network for a given set of input patterns with known classifications. When each entry of the sample set is presented to the network, the network examines its output response to the sample input pattern. The output response is then compared to the known, and desired output and the error value is calculated. Based on the error, the connection weights are adjusted. The back propagation algorithm is based on the error correction learning rule. Basically, error back propagation learning consists of two passes through the different layers of the network: a forward pass and a backward pass. In the forward pass, an activity pattern (input vector) is applied to the sensory nodes of the network, and its effect propagates through the network layer by layer. Finally, a set of outputs is produced as the actual response of the network. During the forward pass, the synaptic weights of the network are all fixed. During the backward pass, on the other hand, the synaptic weights are all adjusted in accordance with an error correction rule. The actual response of the network is subtracted from a desired (target) response to produce an error signal. This error signal is then propagated
46 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks backward through the network, against the direction of synaptic connection hence the name ' error back propagation '. The synaptic weights are adjusted to make the actual response of the network move closer to the desired response in a statistical sense. The learning process performed with the algorithm is called back propagation learning. The mean square error is usually used as a measure of the error, which can be defined as [57]:
1 2 E(n) = ∑e j (n) (3 11) 2 j∈c
th where ej(n) is an error signal produced at the output neuron j for the n sample by simply subtracting the actual response of the network from the desired response, the set C consists of all the neurons in the output layer of the network and the error ej (n) is calculated using the formula
e j (n) = d j (n) − y j (n) (3 12)
th where dj(n) refers to the desired response from neuron j for the n sample and th yj(n) is the generated response from neuron j for the n sample. As the change in weights is evaluated by the delta rule:
w ji = δ j (n)y j (n) (3 13) where is the learning rate parameter of the back propagation algorithm, th and δ j (n) is the local gradient of neuron j for the n sample. It is given by [57]:
e() nf' ( v ()) n Neuron j is in output layer j j δ = (3 14) j( n ) fvn' (())δ () nwn () jj∑ k kj Neuron j is in hidden layer k The performance of the back propagation algorithm in the MLP is governed by numerous factors involved in the design.
1. Maximizing information content . Every training sample presented to the back propagation algorithm should be chosen on the basis that its information content is the largest possible for the task at hand.
47 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
2. Choice of activation function . The activation function f(x) defines the output of a neuron. The activation function usually used in the MLP is a sigmoidal nonlinearity; which comes in two basic forms: a logistic function and a hyperbolic tangent function.
3. Normalizing the inputs. Each input variable should be preprocessed so that its mean value, averaged over the entire training set, is either close to zero, or small enough compared to its standard deviation.
4. Initialization . A good choice for the initial values of the synaptic weights of the network has a great impact on the successful network design. When the synaptic weights are assigned large initial values, it is highly likely that the neurons in the network will be driven into saturation, producing a constant activation, which will cause the training of the network to become stuck near the starting point. However, if the synaptic weights are assigned very small initial values, the back propagation algorithm may operate on a very flat area around the origin of the error surface. The back propagation algorithm uses a gradient descent criterion to find the global minimum. Due to the flat area around the origin of the error surface, the performance improvement of the classification will drop to zero, and hence the learning process terminates. For these reasons, the use of both large and very small values for initializing the synaptic weights should be avoided.
5. Setting the learning rate. All neurons in the MLP should ideally learn at the same rate. Assume that for a given neuron, the learning rate should be inversely proportional to the square root of synaptic connections made to that neuron. If k is the number of neurons in the previous layer, then is given by:
1 α (3 15) k 6. Setting the number of hidden neurons . Deciding the number of hidden neurons is considerably difficult. The number of hidden neurons should
48 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks
never exceed twice the number of input layer units, but this number may not be known in advance. The number of hidden neurons controls the degree of generalization in the network. As the number of hidden neurons is increased, the accuracy of input recognition increases, but the capacity for generalization decreases. When the number of hidden units approaches the number of samples, the network can recognize every different sample exactly, but has no ability to generalize. In other words, a large number of hidden neurons can lead to over fitting of the training data.
7. Setting the number of hidden layers. The number of hidden layers required depends on the complexity of the relationship between the inputs and the outputs. If the input/output relationship is linear (can be approximated by a straight line graph), the network does not need a hidden layer at all. It is unlikely that any practical problem will require more than two hidden layers. In theory, an MLP with one hidden layer is sufficient to approximate any continuous function.
3.8 Types of Noise
To evaluate the performance of a system in the presence of noise, four different types of noise are typically considered; Additive White Gaussian Noise (AWGN), impulsive noise, speckle noise, and Poisson noise [53].
3.8.1 The AWGN
AWGN is a stationary random process exhibiting a continuous and uniform frequency spectrum extending across a specified frequency band. AWGN is the most common type of noise resulting from the contributions of many independent signals. AWGN has a constant power spectral density and utilizes a Gaussian density function to describe its amplitude. Mathematically, an AWGN is described with a mean of zero, an autocorrelation in the form of a delta dirac function and a power spectral density equal to its variance. u(t) = E{w(t)} = 0 (3.16) 2 Rtt(,)12= Ewtwt {()()} 12 =σ δ ( tt 21 − )
49 Chapter 3 Review Study on Landmine Detection Systems and Artificial Neural Networks where w P (ω ) = σ 2
50