Learning structured sparse representation for single sample face recognition

In this paper, we propose a robust sparse representation method to address single sample per person problem by simultaneously exploiting the local and global structure of data. Considering the fact that most sparse representation methods use each testing sample separately and ignore the prior information from testing data, we seek the sparse representation of all testing samples together to capture the global structure of data. Moreover, we adopt an intra-class variance dictionary to describe various facial changes that can not be captured by the single training sample. To make use of local structure, we divide each face image into some blocks consisting of overlapped patches and assume the overlapped patches in a local block are different samples from the same class, which makes their coefficients have row-wise sparse structure. Finally, by imposing group sparsity constraint and sparsity constraint respectively on the coefficients corresponding to the training patches dictionary and variance dictionary, we obtain more discriminative sparse representation, whose coefficients can be directly utilized for classification. Experimental results on three public databases not only demonstrate effectiveness of the proposed approach but also show robustness to various facial variation.