SlideShare a Scribd company logo
1 of 8
Download to read offline
Gender and Ethnic Classification From Frontal Face
Images
An Investigation Into Methods of Feature Extraction and Classification for Estimating
Gender and Ethnicity From Face Images
Presented by:
David Beretta
Prepared for:
Dr. F. Nicolls
Department of Electrical Engineering
University of Cape Town
Submitted to the Department of Electrical Engineering at the University of Cape Town in
partial fulfilment of the academic requirements for a Bachelor of Science degree in Electrical
Engineering.
October 16, 2012
Keywords: Machine Learning, Facial Recognition, Computer Vision, Eigenface, Fisherface,
PCA, LDA, QDA
Abstract
This work involves investigating various methods of feature extraction and classification tech-
niques for the particular problem of determining a person’s gender and ethnicity from an image
of their face. Different methods of feature extraction and classification are tested and compared
in constrained and unconstrained environments to determine the best algorithms for gender and
ethnic classification. Ethnic classification is simplified to a two class problem that only involves
differentiating between African and Caucasian faces. Various parameters such as image reso-
lution, number of training images and number of features are varied to gauge their effect on
performance. The most effective classifier for gender is found to be a Quadratic Discriminant
classifier using PCA with 100 eigenfaces, achieving an accuracy of 97.6 ± 1.6%. The best clas-
sifier for estimating 2-class ethnicity is found to be a Quadratic Discriminant classifier using
50 eigenfaces and achieving an accuracy of 98.2 ± 1.8%. A real-time face classification system
is also implemented using the Viola-Jones face detector, a Quadratic Discriminant classifier
with PCA and a webcam. It is found that while illumination variations impede the success of
real-time face classification, gender and 2-class ethnic classification can be achieved with high
success in conditions of controlled illumination.
This work presents results using the MUCT face database which contains an ethnically diverse
set of faces and has not been used in other literature on gender and ethnic classification.
iv
Contents
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Plan of Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Literature Review 4
2.1 Facial Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Facial Recognition in Humans . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Facial Recognition in Machines . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.3 Face Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.4 Feature Extraction for Facial Recognition . . . . . . . . . . . . . . . . . . 7
2.1.5 Classifiers for Facial Recognition . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.6 Limitations in Machine Facial Recognition . . . . . . . . . . . . . . . . . . 10
2.2 Methods of Gender and Ethnic Classification . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Appearance-Based Methods using Holistic Features . . . . . . . . . . . . . 10
v
2.2.2 Geometric and Local Feature Methods . . . . . . . . . . . . . . . . . . . . 12
3 Face Detection 13
3.1 Introduction to the Viola-Jones Face Detector . . . . . . . . . . . . . . . . . . . . 13
3.2 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Learning Best Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Cascaded Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.5 Final Detector Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 Feature Extraction 18
4.1 Image Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Feature Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Principal Component Analysis (PCA) . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3.1 Finding Principal Components . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3.2 Eigenfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Fisher’s Linear Discriminant (FLD) . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4.1 Finding Features with FLD . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.4.2 Problems with FLD in Facial Recognition . . . . . . . . . . . . . . . . . . 27
4.4.3 Fisherfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5 Face Classification 29
5.1 Bayes Decision Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2 Quadratic Discriminant Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
vi
5.3 Mahalanobis Distance/Linear Discriminant Classifier . . . . . . . . . . . . . . . . 33
5.4 Euclidean Distance or Nearest Mean Classifier . . . . . . . . . . . . . . . . . . . . 34
5.5 Na¨ıve Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.6 k-Nearest Neighbour Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.7 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.7.1 Linear SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.7.2 Kernel SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6 Implementation and Experimental Philosophy 39
6.1 Testing Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.1 Hardware and Operating System . . . . . . . . . . . . . . . . . . . . . . . 39
6.1.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2 Software Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2.2 Classifier Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.3 Performance Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.1 Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3.2 Processing Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4 Cross-validation of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4.1 k-Fold Cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5 Gender and Ethnic Categorisation . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.5.1 Classing Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
vii
6.5.2 Classing Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.6 Experimental Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.6.1 Training and Testing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.6.2 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.6.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.6.4 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
7 Results 48
7.1 Extracting Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.1.1 Using Eigenfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.1.2 Using Fisherfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.2 Classification Under Controlled Conditions . . . . . . . . . . . . . . . . . . . . . 58
7.2.1 Comparing Feature Extraction Methods and Classifiers . . . . . . . . . . 58
7.2.2 Effect of Varying Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.3 Classification Under Uncontrolled Conditions . . . . . . . . . . . . . . . . . . . . 71
7.3.1 Static Image Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.3.2 Real-time Video Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
8 Conclusions 77
8.1 Conclusions from Testing in Controlled Conditions . . . . . . . . . . . . . . . . . 77
8.1.1 Eigenfaces and Fisherfaces Useful Features for Gender and Ethnic Clas-
sification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.1.2 Quadratic Classifier Superior for Gender and Ethnic Classification . . . . 78
viii
8.1.3 The Effects of Varying Different Parameters . . . . . . . . . . . . . . . . . 78
8.2 Conclusions from Testing in Uncontrolled Conditions . . . . . . . . . . . . . . . . 79
8.2.1 Extreme Lighting Variations and Non-constant Backgrounds Greatly Af-
fect Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.3 Conclusion on Classifying Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.4 Conclusion on Classifying Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . 80
9 Recommendations and Future Research 81
9.1 Recommendations to Improve Feature Extraction . . . . . . . . . . . . . . . . . . 81
9.1.1 Use Feature Subset Selection with PCA . . . . . . . . . . . . . . . . . . . 81
9.1.2 Use Null-space Methods for Computing Better MDF’s with FLD . . . . . 82
9.1.3 Investigate Other Feature Extraction Methods . . . . . . . . . . . . . . . 82
9.2 Recommendations to Improve Classification . . . . . . . . . . . . . . . . . . . . . 82
9.2.1 Use a Combination of Different Training Databases . . . . . . . . . . . . . 82
9.2.2 Discard Bad Training Images . . . . . . . . . . . . . . . . . . . . . . . . . 82
9.2.3 Investigate SVM with Different Kernels . . . . . . . . . . . . . . . . . . . 83
9.3 Suggestions for Mitigating Effects of Varying Illumination . . . . . . . . . . . . . 83
9.3.1 Train Classifiers for Illumination Invariance . . . . . . . . . . . . . . . . . 83
9.3.2 Use Illumination Compensation Methods . . . . . . . . . . . . . . . . . . 83
9.4 Suggestions for Mitigating Effects of Inconsistent Backgrounds . . . . . . . . . . 83
Bibliography 86
A Extra resources 87
ix

More Related Content

What's hot

Protege owl tutorialp3_v1_0
Protege owl tutorialp3_v1_0Protege owl tutorialp3_v1_0
Protege owl tutorialp3_v1_0Patricia Silva
 
Multimodal Image Processing in Cytology
Multimodal Image Processing in CytologyMultimodal Image Processing in Cytology
Multimodal Image Processing in CytologyUniversity of Zurich
 
Keraudren-K-2015-PhD-Thesis
Keraudren-K-2015-PhD-ThesisKeraudren-K-2015-PhD-Thesis
Keraudren-K-2015-PhD-ThesisKevin Keraudren
 
Group Exercise Instructor Certification Manual
Group Exercise Instructor Certification ManualGroup Exercise Instructor Certification Manual
Group Exercise Instructor Certification ManualJimmyGFlores
 
From sound to grammar: theory, representations and a computational model
From sound to grammar: theory, representations and a computational modelFrom sound to grammar: theory, representations and a computational model
From sound to grammar: theory, representations and a computational modelMarco Piccolino
 
7 introplasma
7 introplasma7 introplasma
7 introplasmaYu Chow
 
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.FrdricRoux5
 
Mansour_Rami_20166_MASc_thesis
Mansour_Rami_20166_MASc_thesisMansour_Rami_20166_MASc_thesis
Mansour_Rami_20166_MASc_thesisRami Mansour
 
Trade-off between recognition an reconstruction: Application of Robotics Visi...
Trade-off between recognition an reconstruction: Application of Robotics Visi...Trade-off between recognition an reconstruction: Application of Robotics Visi...
Trade-off between recognition an reconstruction: Application of Robotics Visi...stainvai
 
A Bilevel Optimization Approach to Machine Learning
A Bilevel Optimization Approach to Machine LearningA Bilevel Optimization Approach to Machine Learning
A Bilevel Optimization Approach to Machine Learningbutest
 
biometry MTH 201
biometry MTH 201 biometry MTH 201
biometry MTH 201 musadoto
 

What's hot (17)

Protege owl tutorialp3_v1_0
Protege owl tutorialp3_v1_0Protege owl tutorialp3_v1_0
Protege owl tutorialp3_v1_0
 
Multimodal Image Processing in Cytology
Multimodal Image Processing in CytologyMultimodal Image Processing in Cytology
Multimodal Image Processing in Cytology
 
phd-thesis
phd-thesisphd-thesis
phd-thesis
 
Keraudren-K-2015-PhD-Thesis
Keraudren-K-2015-PhD-ThesisKeraudren-K-2015-PhD-Thesis
Keraudren-K-2015-PhD-Thesis
 
Group Exercise Instructor Certification Manual
Group Exercise Instructor Certification ManualGroup Exercise Instructor Certification Manual
Group Exercise Instructor Certification Manual
 
Data structures
Data structuresData structures
Data structures
 
From sound to grammar: theory, representations and a computational model
From sound to grammar: theory, representations and a computational modelFrom sound to grammar: theory, representations and a computational model
From sound to grammar: theory, representations and a computational model
 
Javanotes5 linked
Javanotes5 linkedJavanotes5 linked
Javanotes5 linked
 
Scala reference
Scala referenceScala reference
Scala reference
 
thesis
thesisthesis
thesis
 
7 introplasma
7 introplasma7 introplasma
7 introplasma
 
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.
Alpha and Gamma Oscillations in MEG-Data: Networks, Function and Development.
 
Chem sep book2
Chem sep book2Chem sep book2
Chem sep book2
 
Mansour_Rami_20166_MASc_thesis
Mansour_Rami_20166_MASc_thesisMansour_Rami_20166_MASc_thesis
Mansour_Rami_20166_MASc_thesis
 
Trade-off between recognition an reconstruction: Application of Robotics Visi...
Trade-off between recognition an reconstruction: Application of Robotics Visi...Trade-off between recognition an reconstruction: Application of Robotics Visi...
Trade-off between recognition an reconstruction: Application of Robotics Visi...
 
A Bilevel Optimization Approach to Machine Learning
A Bilevel Optimization Approach to Machine LearningA Bilevel Optimization Approach to Machine Learning
A Bilevel Optimization Approach to Machine Learning
 
biometry MTH 201
biometry MTH 201 biometry MTH 201
biometry MTH 201
 

Similar to Thesis Abstract

Final Report - Major Project - MAP
Final Report - Major Project - MAPFinal Report - Major Project - MAP
Final Report - Major Project - MAPArjun Aravind
 
High Performance Traffic Sign Detection
High Performance Traffic Sign DetectionHigh Performance Traffic Sign Detection
High Performance Traffic Sign DetectionCraig Ferguson
 
Classification System for Impedance Spectra
Classification System for Impedance SpectraClassification System for Impedance Spectra
Classification System for Impedance SpectraCarl Sapp
 
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...Automatic Detection of Performance Design and Deployment Antipatterns in Comp...
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...Trevor Parsons
 
Stochastic Processes and Simulations – A Machine Learning Perspective
Stochastic Processes and Simulations – A Machine Learning PerspectiveStochastic Processes and Simulations – A Machine Learning Perspective
Stochastic Processes and Simulations – A Machine Learning Perspectivee2wi67sy4816pahn
 
Aspect_Category_Detection_Using_SVM
Aspect_Category_Detection_Using_SVMAspect_Category_Detection_Using_SVM
Aspect_Category_Detection_Using_SVMAndrew Hagens
 
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...Artur Filipowicz
 
Dragos Datcu_PhD_Thesis
Dragos Datcu_PhD_ThesisDragos Datcu_PhD_Thesis
Dragos Datcu_PhD_Thesisdragos80
 
An Optical Character Recognition Engine For Graphical Processing Units
An Optical Character Recognition Engine For Graphical Processing UnitsAn Optical Character Recognition Engine For Graphical Processing Units
An Optical Character Recognition Engine For Graphical Processing UnitsKelly Lipiec
 
Au anthea-ws-201011-ma sc-thesis
Au anthea-ws-201011-ma sc-thesisAu anthea-ws-201011-ma sc-thesis
Au anthea-ws-201011-ma sc-thesisevegod
 

Similar to Thesis Abstract (20)

Final Report - Major Project - MAP
Final Report - Major Project - MAPFinal Report - Major Project - MAP
Final Report - Major Project - MAP
 
High Performance Traffic Sign Detection
High Performance Traffic Sign DetectionHigh Performance Traffic Sign Detection
High Performance Traffic Sign Detection
 
Thesis
ThesisThesis
Thesis
 
Mak ms
Mak msMak ms
Mak ms
 
Classification System for Impedance Spectra
Classification System for Impedance SpectraClassification System for Impedance Spectra
Classification System for Impedance Spectra
 
Thesis_Prakash
Thesis_PrakashThesis_Prakash
Thesis_Prakash
 
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...Automatic Detection of Performance Design and Deployment Antipatterns in Comp...
Automatic Detection of Performance Design and Deployment Antipatterns in Comp...
 
Stochastic Processes and Simulations – A Machine Learning Perspective
Stochastic Processes and Simulations – A Machine Learning PerspectiveStochastic Processes and Simulations – A Machine Learning Perspective
Stochastic Processes and Simulations – A Machine Learning Perspective
 
M.Sc thesis
M.Sc thesisM.Sc thesis
M.Sc thesis
 
Intro photo
Intro photoIntro photo
Intro photo
 
btpreport
btpreportbtpreport
btpreport
 
Aspect_Category_Detection_Using_SVM
Aspect_Category_Detection_Using_SVMAspect_Category_Detection_Using_SVM
Aspect_Category_Detection_Using_SVM
 
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...
Virtual Environments as Driving Schools for Deep Learning Vision-Based Sensor...
 
Dragos Datcu_PhD_Thesis
Dragos Datcu_PhD_ThesisDragos Datcu_PhD_Thesis
Dragos Datcu_PhD_Thesis
 
Analytical-Chemistry
Analytical-ChemistryAnalytical-Chemistry
Analytical-Chemistry
 
Neural networks and deep learning
Neural networks and deep learningNeural networks and deep learning
Neural networks and deep learning
 
thesis
thesisthesis
thesis
 
Thesis
ThesisThesis
Thesis
 
An Optical Character Recognition Engine For Graphical Processing Units
An Optical Character Recognition Engine For Graphical Processing UnitsAn Optical Character Recognition Engine For Graphical Processing Units
An Optical Character Recognition Engine For Graphical Processing Units
 
Au anthea-ws-201011-ma sc-thesis
Au anthea-ws-201011-ma sc-thesisAu anthea-ws-201011-ma sc-thesis
Au anthea-ws-201011-ma sc-thesis
 

Thesis Abstract

  • 1. Gender and Ethnic Classification From Frontal Face Images An Investigation Into Methods of Feature Extraction and Classification for Estimating Gender and Ethnicity From Face Images Presented by: David Beretta Prepared for: Dr. F. Nicolls Department of Electrical Engineering University of Cape Town Submitted to the Department of Electrical Engineering at the University of Cape Town in partial fulfilment of the academic requirements for a Bachelor of Science degree in Electrical Engineering. October 16, 2012 Keywords: Machine Learning, Facial Recognition, Computer Vision, Eigenface, Fisherface, PCA, LDA, QDA
  • 2.
  • 3. Abstract This work involves investigating various methods of feature extraction and classification tech- niques for the particular problem of determining a person’s gender and ethnicity from an image of their face. Different methods of feature extraction and classification are tested and compared in constrained and unconstrained environments to determine the best algorithms for gender and ethnic classification. Ethnic classification is simplified to a two class problem that only involves differentiating between African and Caucasian faces. Various parameters such as image reso- lution, number of training images and number of features are varied to gauge their effect on performance. The most effective classifier for gender is found to be a Quadratic Discriminant classifier using PCA with 100 eigenfaces, achieving an accuracy of 97.6 ± 1.6%. The best clas- sifier for estimating 2-class ethnicity is found to be a Quadratic Discriminant classifier using 50 eigenfaces and achieving an accuracy of 98.2 ± 1.8%. A real-time face classification system is also implemented using the Viola-Jones face detector, a Quadratic Discriminant classifier with PCA and a webcam. It is found that while illumination variations impede the success of real-time face classification, gender and 2-class ethnic classification can be achieved with high success in conditions of controlled illumination. This work presents results using the MUCT face database which contains an ethnically diverse set of faces and has not been used in other literature on gender and ethnic classification. iv
  • 4. Contents 1 Introduction 1 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.3 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.4 Plan of Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2 Literature Review 4 2.1 Facial Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 2.1.1 Facial Recognition in Humans . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.1.2 Facial Recognition in Machines . . . . . . . . . . . . . . . . . . . . . . . . 6 2.1.3 Face Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.1.4 Feature Extraction for Facial Recognition . . . . . . . . . . . . . . . . . . 7 2.1.5 Classifiers for Facial Recognition . . . . . . . . . . . . . . . . . . . . . . . 8 2.1.6 Limitations in Machine Facial Recognition . . . . . . . . . . . . . . . . . . 10 2.2 Methods of Gender and Ethnic Classification . . . . . . . . . . . . . . . . . . . . 10 2.2.1 Appearance-Based Methods using Holistic Features . . . . . . . . . . . . . 10 v
  • 5. 2.2.2 Geometric and Local Feature Methods . . . . . . . . . . . . . . . . . . . . 12 3 Face Detection 13 3.1 Introduction to the Viola-Jones Face Detector . . . . . . . . . . . . . . . . . . . . 13 3.2 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.3 Learning Best Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3.4 Cascaded Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3.5 Final Detector Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4 Feature Extraction 18 4.1 Image Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 4.2 Feature Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.3 Principal Component Analysis (PCA) . . . . . . . . . . . . . . . . . . . . . . . . 20 4.3.1 Finding Principal Components . . . . . . . . . . . . . . . . . . . . . . . . 20 4.3.2 Eigenfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 4.4 Fisher’s Linear Discriminant (FLD) . . . . . . . . . . . . . . . . . . . . . . . . . . 24 4.4.1 Finding Features with FLD . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.4.2 Problems with FLD in Facial Recognition . . . . . . . . . . . . . . . . . . 27 4.4.3 Fisherfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 5 Face Classification 29 5.1 Bayes Decision Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 5.2 Quadratic Discriminant Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 vi
  • 6. 5.3 Mahalanobis Distance/Linear Discriminant Classifier . . . . . . . . . . . . . . . . 33 5.4 Euclidean Distance or Nearest Mean Classifier . . . . . . . . . . . . . . . . . . . . 34 5.5 Na¨ıve Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 5.6 k-Nearest Neighbour Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 5.7 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 5.7.1 Linear SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 5.7.2 Kernel SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 6 Implementation and Experimental Philosophy 39 6.1 Testing Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 6.1.1 Hardware and Operating System . . . . . . . . . . . . . . . . . . . . . . . 39 6.1.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 6.2 Software Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 6.2.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 6.2.2 Classifier Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 6.3 Performance Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 6.3.1 Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 6.3.2 Processing Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 6.4 Cross-validation of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 6.4.1 k-Fold Cross-validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 6.5 Gender and Ethnic Categorisation . . . . . . . . . . . . . . . . . . . . . . . . . . 43 6.5.1 Classing Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 vii
  • 7. 6.5.2 Classing Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 6.6 Experimental Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 6.6.1 Training and Testing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 6.6.2 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 6.6.3 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 6.6.4 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 7 Results 48 7.1 Extracting Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 7.1.1 Using Eigenfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . 50 7.1.2 Using Fisherfaces as Features . . . . . . . . . . . . . . . . . . . . . . . . . 55 7.2 Classification Under Controlled Conditions . . . . . . . . . . . . . . . . . . . . . 58 7.2.1 Comparing Feature Extraction Methods and Classifiers . . . . . . . . . . 58 7.2.2 Effect of Varying Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 61 7.3 Classification Under Uncontrolled Conditions . . . . . . . . . . . . . . . . . . . . 71 7.3.1 Static Image Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 7.3.2 Real-time Video Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 8 Conclusions 77 8.1 Conclusions from Testing in Controlled Conditions . . . . . . . . . . . . . . . . . 77 8.1.1 Eigenfaces and Fisherfaces Useful Features for Gender and Ethnic Clas- sification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 8.1.2 Quadratic Classifier Superior for Gender and Ethnic Classification . . . . 78 viii
  • 8. 8.1.3 The Effects of Varying Different Parameters . . . . . . . . . . . . . . . . . 78 8.2 Conclusions from Testing in Uncontrolled Conditions . . . . . . . . . . . . . . . . 79 8.2.1 Extreme Lighting Variations and Non-constant Backgrounds Greatly Af- fect Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 80 8.3 Conclusion on Classifying Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 8.4 Conclusion on Classifying Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . 80 9 Recommendations and Future Research 81 9.1 Recommendations to Improve Feature Extraction . . . . . . . . . . . . . . . . . . 81 9.1.1 Use Feature Subset Selection with PCA . . . . . . . . . . . . . . . . . . . 81 9.1.2 Use Null-space Methods for Computing Better MDF’s with FLD . . . . . 82 9.1.3 Investigate Other Feature Extraction Methods . . . . . . . . . . . . . . . 82 9.2 Recommendations to Improve Classification . . . . . . . . . . . . . . . . . . . . . 82 9.2.1 Use a Combination of Different Training Databases . . . . . . . . . . . . . 82 9.2.2 Discard Bad Training Images . . . . . . . . . . . . . . . . . . . . . . . . . 82 9.2.3 Investigate SVM with Different Kernels . . . . . . . . . . . . . . . . . . . 83 9.3 Suggestions for Mitigating Effects of Varying Illumination . . . . . . . . . . . . . 83 9.3.1 Train Classifiers for Illumination Invariance . . . . . . . . . . . . . . . . . 83 9.3.2 Use Illumination Compensation Methods . . . . . . . . . . . . . . . . . . 83 9.4 Suggestions for Mitigating Effects of Inconsistent Backgrounds . . . . . . . . . . 83 Bibliography 86 A Extra resources 87 ix