Fisher discriminant analysis of Gabor texture features for face recognition

This paper proposes a novel approach, called the Fisher Discriminant Analysis of Gabor Texture Features (FGTF), for robust face recognition. The purpose of FGTF is to reduce the dimension of Gabor features while making full use of the original information. In FGTF, the Gabor-filtered image is partitioned into many non-overlap sub-images (blocks) equally to calculate the statistics such as mean and standard deviation of each block, which are used for local texture representation. These local texture features of all the blocks are concatenated to form a low-dimensional Gabor texture feature vector, which is then normalized by Gaussian normalization. Before being used for face recognition, the Gaussian normalized feature vector is subjected to Fisher Discriminant Analysis (FDA) to obtain enhanced discriminative power. Experiments carried out on two face databases, ORL and UMIST, have shown that our method can effectively reduce the dimension of Gabor features and increase the recognition accuracy when compared with other popular techniques. In addition, our experimental results indicate its robustness to moderate variations in pose, illumination and facial expression. Copyright © 2009 Binary Information Press.