Face Recognition Based On Group Sparse And Weighted Sparse Representation

Intelligent analysis of video surveillance images for maintaining public order, crime prevention, criminal detection and other applications has a crucial role. Although after30years of research it presents a variety of face recognition methods, the recognition success in numerous practical applications is limited by many environmental conditions, for example different image resolutions, local non-uniform illumination, occlusion, rotation, noise, facial expression, age and other factors. In other words, there is still a far distance between the effectiveness of the human cognitive system and the current machine recognition system.The new sparse representation proposed in recent years has shown superior performance and robustness to noise and occlusion for face recognition, and therefore receives wide attention of researchers. Classification based on Sparse Representation (the SRC method) will model the classification problem as looking for an optimal linear combination of atoms in the dictionary problem by minimizing Lo norm or L1norm of sparse coefficients, this method will introduce sparse representation to pattern classification, making the emergence of new groundbreaking direction. This paper explores various optimal linear combination methods in framework of SRC, including robust face recognition with individual and group sparsity constraints; weighted sparse representation based cascading Gabor feature, the WGSRC method; weighted sparse representation based multidirectional Gabor feature, the multidirectional WGSRC method. Experiments of these methods on Extended Yale B face database show excellent classification performance.The main content and innovation of this paper are given as follows:1) This paper first introduced the current status of relevant technologies. By presenting the real-time video recognition system our team realizes in the8th UTMVP face information recognition contest in2013, the paper outlines the framework for face recognition flow and the algorithms of the various steps. I have a more detailed description of the age recognition task I undertake which provided technical background of this thesis.2) This thesis proposed a new robust face recognition method based on sparse representation with individual and group sparsity constraints. Taking it into account that all the pictures of everyone make up different groups in the training dictionary, this paper discussed the L1norm individual sparse precision problem in SRC method. This thesis introduced the model which minimizes the number of non-zero reconstructed vector with group sparsity into SRC framework.Combining the individual and group sparsity constraints can improve the solution precision, and experiments in the common face database Extended Yale B demonstrated good classification performance of the method, especially in the low dimension cases.3) This thesis proposed the Classification method based on Weighted Sparse Representation via cascading Gabor feature(WGSRC) and further proposed the Classification method based on Weighted Sparse Representation via cascading Gabor feature(multidirectional WGSRC). Recent studies indicate that a weighted sparse representation based on the classification (WSRC) makes full use of data locality to improve the classification performance. Due to global features used in the WSRC method, in order to further improve the recognition rate, we used local Gabor features showing excellent test results in the8th UTMVP face information recognition contest in WSRC and further studied multidirectional Gabor feature. Experiments in AR face database demonstrated good classification performance of the methods, especially in the low dimension cases.