Automatic face recognition is one of the challenging research topics in computer vision, machine learning and biometrics. Although face recognition has been extensively studied in recent decades, various methods have been proposed for facial extraction and classification, among which the representatives include subspace learning based Eigenface, Fisherface, Laplacianfaces, Gabor feature based classification. machine learning capabilities based SVM. However, due to occlusion, posture, lighting and other effects, recognition performance is still severely limited. Therefore, recognition is still necessary in the real situation to improve the robustness.In2008, the sparse representation based classification(SRC) of face images is proposed by John Wright et al., sparse representation is introduced to the face recognition at the first time. This method which the training face images are used as the dictionary, get the image to be recognized in the dictionary sparse representation coefficients by minimizing l1, and solving minimum residuals to be identified. Given the sparse representation classification methods successfully applied in face recognition, sparse representation based face recognition were studied. For illumination and occlusion of the image, focus on how to extract local features and constructe reasonable occlusion dictionary in two parts. The main work and research content are given as follows:1) When using the SRC method for face recognition, feature extraction is not so important, but this is limited to the overall characteristics. In order to make a coefficient vector with more significant sparsity, inspired by the Gabor based sparse representation classification algorithm, Monogenic feature, this local feature is introduced into the sparse representation classification. The article propose Momogenic feature based sparse representation classification algorithm(MSRC). Due to energy feature, structural feature and geometric feature can be extraced by Monogenic feature, and more comprehensive image information extraction, MSRC can improve the recognition rate. In the processing time of the algorithm, with respect to the multi-scale and multi-direction of Gabor feature, and only a single multi-scale of Monogenic feature can be reduced in the processing time.On AR library and Extend Yale B library conducted experiments to verify the effectiveness of the proposed algorithm.2) When the image appears damaged or occlusion in the face recognition, the use of occlusion of the face image is obscured by a original face and superposition error, we research the effect of the expansion of the dictionary with occlusion to the recognition performance focuing on the matrix block dictionary and Gabor occlusion dictionary. Taking into account the complexity of the extension dictionary, using compressed sensing theory to remove the area of occlusion, the use of non-occlusion area as a dictionary to identify. On the AR face database containing blocked conducted experiments to compare the performance of these types of dictionaries identification block image. When the block size is large, the use of non-occluded region as recognition performance superior to the dictionary block matrix and Gabor occlusion dictionary. |