Face Recognition Research Based On Gabor Wavelet And Sparse Representation Algorithm

Face recognition technique is of great value for its wide application and research significance,which has become one of hotspots of research in the field of biological recognition and computervision. Face recognition consists of feature extraction and classification. Up now, the mostoften-used methods for feature extraction are the subspace analysis method and elastic graphmatching method. And for classifications method, there are nearest neighbor method, SVM methodand sparse representation classification method.Based on the2DGabor wavelet theory and sparsepreserving projection algorithm, the focus of this thesis is on the study of face recognition, whichincludes follow sections.(1) Study of sparse preserving projection (SPP) algorithm. The algorithm may reduce thedimension of data in multiple-dimensional space by preserving the global sparse reconstructionrelationship in sample space and as a result, it is available of the method for processing thedatabase of Yale, AR and Extended Yale B. However it is still not good enough to preservingthe local neighbor information in sample space. To solve this problem, an improved method ispresented in paper to reduce the dimensions based on SPP, which is about applying localpreserving projection (LPP) in SPP or called LSPP algorithm. This is also verified in ORLdatabase.(2) Study of GLSPP algorithm. This is proposed by applying the2D Gabor wavelet into LSPP, orcalled GLSPP. Usually in the case of big difference into face images due to light, pose andexpression, the conventional SPP is not so efficient for feature extraction. However, the2DGabor wavelet can extract the constant information of face image and thus overcomes thedisturbance of those factors. Then the improved SPP method is done to reduce the dimensionsof database and in the end, the nearest neighbor method is adopted for classification. Thesimulation shows that this method can improve the face recognition result.(3) Study of comparison between various algorithms. The comparison is also done in this thesisbetween the improved GLSPP and those methods like PCA, SPP and LSPP for processing theORL database. Results show that the improved method has more advantages than othermethods.