Face recognition is a technology using of computers to extract effective information from the face image for authentication. As face recognition technology is friendly and non-intrusive and has a bright future in criminal cases detecting, document authentication and other fields, researching on the face recognition technology has great theoretical and practical values. Therefore, how to improve the recognition rate, reduce the time consumption, and improve the robust on illumination, pose and expression of face recognition algorithms become the focus of researchers. This thesis studied the problem of feature extraction of face image and the selection of kernel function and two methods were proposed.Based on the analysis of results of statistical learning methods for feature extraction on face image, the limitations of current linear learning method in the extraction of non-linear characteristics of face image, the advantages of kernel method on non-linear feature extraction and good reflect of Gabor on the local features of face images, full using of the advantages of signal processing and kernel methods on image feature extraction, a supervised Gabor-based kernel locality preserving projections (SGKLPP) algorithm and a Gabor-based kernel Independent component analysis (GKICA) algorithm are proposed in this thesis. SGKLPP algorithm first extract the feature of face image using Gabor transform, then the extracted feature vectors are mapped to the kernel space and the supervised adjacency matrix is constructed in kernel space, finally extract feature of the feather vector under kernel space and classify them by Locality Preserving Projections algorithm; as the same to SGKLPP, GKICA algorithm first extract the feature of face image using Gabor transform, then the extracted feature vectors are mapped to the kernel space by Kernel Principal Component Analysis algorithm, finally extract feature of the feather vector under kernel space and classify them by Independent Component Analysis algorithm.In this thesis, we do experiments on ORL and Yale dataset about recognition rate and time costs under different feature vector dimensions, the selection of kernel function and different number of training samples, and give the analysis of results. |