nI=

Pakistan Research Repository Home
 

Title of Thesis

Improved Face Recognition using Image Resolution Reduction and Optimization of Feature Vector

Author(s)

MUHAMMAD ALMAS ANJUM

Institute/University/Department Details
Department of Computer Engineering, College of Electrical and Mechanical Engineering / National University of Sciences and Technology (NUST), Rawalpindi
Session
2008
Subject
Computer Engineering
Number of Pages
209
Keywords (Extracted from title, table of contents and abstract of thesis)
Improved, Face Recognition, Image Resolution, Reduction, Optimization, Feature, Vector, dimension, normalization, computational

Abstract
Face recognition is a difficult problem that involves automated matching of a given face image with corresponding person’s image(s) in a database. Face recognition finds application in areas like surveillance & security, digital libraries and human computer interactions. Successful, speedy and practically feasible face recognition method depends heavily on the choice of feature vector used for classification and addressing the curse of image dimension. The dimension reduction and the skill to acquire minimum size of feature vector required for face recognition for diverse facial expressions is a challenging task in face recognition. Dimension reduction results in removal of irrelevant variables along with noise therein and a lower computation complexity of subsequent processing.
This dissertation addresses the challenges of dimension reduction, choice of minimum size feature vector for face recognition and minimization of adverse effects of varying facial expressions on the recognition through reduction in image resolution. In preprocessing of face images, scale normalization is carried out through a novel scale normalization algorithm to retain only the facial part of images. This helps in reducing computational complexity by restricting dimensions of image to face region only. Tilt of face images is removed by calculating the gradient between the two eyes and applying the reverse rotation. The issue of dimensionality is addressed first by gradually reducing image resolution through spatial domain low pass filtering followed by decimation. The second method involves novel coefficient selection strategies to choose the minimum dimension of feature vector required for recognition with maximum recognition rate and reduced computational complexity. Face images with varying image resolution are obtained by varying the decimation factor. The effects of variation in image resolution on face recognition have been evaluated using template matching and Principle Components Analysis (PCA) based face recognition techniques. Classical PCA technique has been modified into sub-holistic PCA. Better recognition rate is achieved using modified PCA method with reduced image resolution.
Improved recognition rate results are reported using novel coefficients selection and optimization methods in Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT) and Discrete wavelets Transform (DWT) based face recognition methods. The experiments are carried out for various image resolutions using five different datasets. Improved recognition rate of 97.2% (template matching), 87% (PCA), 94% (Sub-holistic PCA), 100% (DFT), 95.75% (DCT) and 99.25% (DWT) is achieved at a specific image resolution for different datasets. The resolution reduction method used with square images is then extended to hexagonal images. A new technique based on Diagonal grow and Butterfly structure methodology has been developed for sampling and indexing hexagonal structure in hexagonal image processing frame work. Proposed strategy offer less pixel redundancy as compared to existing techniques. Reduction in pixel redundancy varies according to size of square image.

Download Full Thesis
3,558 KB
S. No. Chapter Title of the Chapters Page Size (KB)
1 0 CONTENTS

 

ii
250 KB
2

1

INTRODUCTION

1.1 Introduction

1.2 Matching Techniques

1.3 Dimensionality Reduction in Pattern Recognition Models

1.4 Classifiers

1.5 Fundamental Problems in Pattern Recognition System Design

1.6 Biometrics

1.7 Biometrics Qualifying Requirements

1.8 Biometrics Operating Modes

1.9 Applications of Biometrics

1.10 Comparison of Various Biometrics

1.11 Limitations of (Unimodal) Biometric Systems

1.12 Overview of Thesis

4
413 KB
3 2 FACE RECOGNITION FUNDAMENTALS AND CHALLENGES

2.1 Introduction

2.2 Need for Face Recognition

2.3 Five Steps to Facial Recognition

2.4 Human and Technical Difficulties with Facial Recognition

2.5 Typical Face Recognition System

2.6 Brief Overview of Face Recognition Approaches

2.7 Face Recognition Challenges

2.8 Image Illumination Variations

2.9 Facial Tilt and Pose Variations

2.10 Image Resolution Variation

25
404 KB
4 3 FACE IMAGES PRE-PROCESSING

3.1 Introduction

3.2 Color to Grayscale Conversion

3.3 Scale Normalization

3.4 Geometric Normalization

3.5 Varying Image Background Issues in Face Recognition

3.6 Illumination Variation Control

3.7 Databases used to Evaluate the Recognition Models

43
591 KB
5 4 FACE IMAGE DIMENSION REDUCTION THROUGH IMAGE RESOLUTION VARIATION

4.1 Introduction

4.2 Types of Image Resolution

4.3 Image Resolution and Face Recognition

4.4 Image Resolution Variation Through Interpolation

4.5 Image Gaussian Pyramid through Decimation

4.6 Effects of Varying Resolution on Template Matching Face Recognition Model

4.7 Resolution and Dimension Reduction through Decimation Algorithm

4.8 Dissimilarity Space and Matching

4.9 Implementation of Face Recognition Model

4.10 Experiments and Results

4.11 Discussion of Results

4.12 Principal Component Analysis

4.13 Implementation of PCA Based Face Recognition Model

4.14 Experiments and Results

4.15 Sub-Holistic PCA (SHPCA) Recognition Technique and Varying Image Resolution

4.16 Implementation of SHPCA

4.17 Experiments and Results

4.18 Discussion of PCA and SHPCA Results

4.19 Comparison with other PCA Based Face Recognition Techniques

56
1,089 KB
6 5 TRANSFORM DOMAIN FACE RECOGNITION

5.1 Introduction

5.2 Linear Transforms

5.3 Image Resolution Variation in Frequency Domain

5.4 Facial Features Representation through Discrete Fourier Transform (DFT)

5.5 Low Frequency Co efficient Selection Methods

5.6 DFT Based Face Recognition Process

5.7 Implementation

5.8 Experiments and Results

5.9 Discrete Cosine Transform (DCT)

5.10 DCT in Face Recognition

5.11 Implementation of DCT Based Face Recognition Model

5.12 Experiments and Results of DCT Based Recognition Model

5.13 Facial Multiresolution Analysis through Wavelets

5.14 Discrete Wavelet Transform

5.15 Wavelet Application for Multiresolution Analysis and Dimension Reduction

5.16 Classification of Wavelets

5.17 Face Recognition Using Wavelets

5.18 Implementation

5.19 Choice of Wavelet Family for Recognition

5.20 Feature Vector Optimization

5.21 Experiments and Results with Better Choice of Coefficient and Wavelet

5.22 Discussion of Frequency Based Face Recognition Model

98


1,709 KB
7 6 DIMENSION REDUCTION ISSUES IN HEXAGONAL PIXELS

6.1 Introduction

6.2 Hexagonal Structure Sampling and Addressing

6.3 Proposed Hexagonal Processing Methodology

6.4 Alternate Implementation of Proposed Indexing Scheme

6.5 Storage of Hexagonal pixels

6.6 Simulation Results

6.7 Pixel Redundancy Comparison

158


389 KB
8 7 CONCLUSIONS

178


125 KB
9 8 REFERENCES

181


168 KB