Sparse Representation And Its Application In Face Recognition

Face recognition is the largest and most important ability that we used in our daily lives, and thus it becomes one of the hottest areas in the current pattern recognition research topics. For now, the face recognition method is generally based on the subspace feature extraction. But the robustness of subspace feature extraction will be unstable in dealing with changes of face expression, posture and illumination, limiting automatic face recognition in practical applications.Face recognition algorithm based on sparse representation can well avoid the dimension disaster and the loss of data structure in subspace feature extraction, largely increasing the robustness of face recognition in practical application. In this paper, the sparse representation theory and its application in face recognition are studied, and sparsity preserving projections (SPP) is improved. According to SPP without supervision, we combined the SPP with the graph embedding dimensional model, proposing a new feature extraction algorithm named Supervised Discriminant Sparsity Preserving Projections (SDSPP), and its extensions named weighted SDSPP (WSDSPP) and two-dimensional SDSPP (2DSDSPP). SPP can effectively improve the discrimination ability of SPP by greatly preserving the sparse reconstructive relationship in samples of same classes and reducing the sparse reconstructive relationship in samples of different classes at the same time.In order to verify the effectiveness of improved algorithms, a series of simulation experiments in some public face database is carried on for the comparison between the proposed method and other methods. Experimental results show that the recognition rate of improved algorithm is better than the original algorithm, demonstrated the feasibility of SDSPP algorithm.