Discriminative face recognition via kernel sparse representation

Sparse representation (SR) is a popular method in pattern recognition and computer vision, and achieves the noticeable performance for face recognition (FR) task. Nevertheless, the conventional SR algorithm is usually computationally expensive due to the solution of representation coefficients via l1-regularization minimization problem. Besides, the internal relationship of data, such as nonlinear structure is neglected by the classification procedure conducted on the original data space. To solve these problems, this paper proposes a discriminative FR method using kernel sparse representation (KSR) based on the framework of l2-regularization. With the goal of extracting richer information, a kernel function is used to map the original face samples into a high feature space. Then, a new SR method based on the framework of l2-regularization is designed to represent the face samples on this new space. This method can produce a discriminative representation for each face sample. In addition, the proposed method offers a computational efficient algorithm for FR task. Extensive experiments conducted on the face databases show the effectiveness of our method.