A Modular Kernel Linear Discriminant Analysis of Gabor Features for Improved Face Recognition

Abstract

Automatic face recognition is one of the major challenges in computer vision and pattern analysis. This thesis presents an efficient face recognition system that is robust with regards to changes in illumination, facial expressions and partial occlusions. Modular Kernel Linear Discriminant Analysis performed on Gabor Features obtained from the face images is employed for improving face recognition accuracy. A face image is pre-processed using the 2D Gabor wavelet transform to achieve invariance to illumination in images. Modular approaches that divide the pre-processed images into smaller sub-images provide improved accuracy, as the facial variations in an image are confined to local regions. Kernel methods are applied to these images in order to extract the most discriminating features, thus improving the classification accuracy of the system. Dimensionality reduction of these generated higher dimensional features is obtained by applying Linear Discriminant Analysis, thus improving the computational speed. Performance of the proposed technique is tested and evaluated for face recognition accuracy including factors like changes in illumination conditions, facial expressions and partial occlusions. The AR and FERET databases are used for training and testing processes. Results indicate that the proposed technique has better face recognition accuracy when compared to state of the art techniques like Principal Component Analysis (PCA), Modular Principal Component Analysis (MPCA), Linear Discriminant Analysis (LDA), and Kernel Principal Component Analysis (KPCA). Research is continuing for pose invariant face recognition by considering multiple face recognition modules, trained for different facial views.