Group lasso based collaborative representation for face recognition

Abstract

Based on the idea of collaborative representation, a novel approach CRC-GLasso is proposed for face recognition. Our main contributions lie in two aspects: 1) Instead of sparse representation, collaborative representation is employed to compute sparse representations of face images to solve the `lack of samples' problem. The reason is that face images of different classes share similarities, and some face images from one class may be very helpful to represent those from another class. 2) As the regularization term of collaborative representation, group lasso can be used to construct our objective function, which can make collaborative representation well-structured according to two physical meanings of group lasso: 1) The coefficients of training samples from certain class can be enhanced. 2) The coefficients of most classes can be alleviated. Our proposed method is applied to the well-known public face databases, AR database, and the experimental results show that CRC-GLasso outperforms other state-of-the-art algorithms for face recognition, such as SRC, CRC, KSVD, D-KSVD and LC-KSVD.