Novel Dimension Reduction Method of Gabor Feature and Its Application to Face Recognition

In order to improve the face recognition performance using Gabor wavelet, a multi-channel dimension reduction scheme (MDRS) employing PCA algorithm is proposed. In addition, the selection of scales and orientations in Gabor transformation is investigated. Different from existing dimension reduction methods which employed ensemble dimension reduction scheme (EDRS), in MDRS model, PCA algorithm is performed on total convolution results related to a certain Gabor filter to get feature vector whose dimensions are reduced. The process is repeated according to the number of other different Gabor filters. So a given face image sample has the same number of feature vectors as that of Gabor filters, and the final augmented feature vector could be derived by concatenating all these feature vectors. The experiment results in the popular face databases such as YALE and FERET demonstrate not only that the proposed method is effective but also that the traditional selection of scale and orientation is not optimal.