الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 137 |  |  |
| | | from | 137 | | |
Abstract
Face Recognition is an important pattern recognition problem, compared to other biometrics because it is non-intrusive, non-invasive and requires no participation from the subject. However, the face is highly deformable and its appearance alters significantly due to pose, illumination or expression changes. These changes in appearance are most notable for texture images or two dimensional (2D) face data where intra-class variations due to these factors are usually greater than inter-class variations. However, the underlying structure of the face, or three dimensional (3D) face data, is not changing by pose or illumination variations. 3D face data has been used to deal with the unsolved issues in 2D face recognition. However, 3D scanners are more expensive than 2D cameras and the significant computational cost of 3D face recognition has made large scale deployment of 3D face recognition impractical.
This research is intended to enhance and improve the performance of automatic 2D and 3D face recognition systems. Evaluation of current systems concluded that facial images need to be normalized and correctly aligned for better feature extraction process and high recognition rate. Also, the size of extracted features needs to be minimized for better system performance. In this context, the thesis presents the following approach to enhance the accuracy and system performance;
A proposed approach to improve 2D face recognition system performance by enhancing the 2D image through a preprocessing technique to align and normalize all images in the database based on eyes centers localization using 2D Normalized Cross Correlation (2DNCC) template matching. The detected eyes centers have been used to align all images and crop face Region of Interest (ROI). The segmented ROI of the 2D is compensated for illumination changes using Adaptive Scale Retinex (ASR) Technique to reduce effects of illumination changes. The proposed approach extracted 2D face features using a set of selected orthogonal Gabor filters. This approach minimizes the feature vector extracted compared to full Gabor
filters bank. A further compression to the extracted features has been accomplished using Linear Discriminant Analysis (LDA) before the classification stage .
The proposed 3D face recognition approach exploited the proposed 2D frame work to preprocess 3D face data using depth map representation of the 3D data. The eye localization technique has been applied to depth map and results have been used to segment and align 3D data. The proposed approach extracted 3D face features from normal map representation using a set of selected orthogonal Gabor filters followed by LDA to compress the features.
Finally, the performance of the proposed system has been enhanced by integrating 2D and 3D extracted features to form a robust Multi-Modal Face Recognition system.
The system has been tested on CASIA, ORL, Cropped YaleB and Gavab DB face images Databases and achieved high recognition rate compared to current systems.
Experimental results show that the proposed system is effective for both dimensions reduction and recognition performance when compared to the complete Gabor filter bank based current systems