Your page title
Block-Structured Dictionary Learning for Sparse Representation-Based Face Recognition
Yixiong Liang*, Ke Nai*, Gang Hua+, Beiji Zou*
*Central South University +Stevens Institute of Technology
Abstract. Sparse representation-based classification (SRC) for face recognition has gained its popularity due to its simplicity and effectiveness. In SRC face recognition method, the dictionary has a natural block structure, i.e., all the samples from the same person form a block. Instead of using the off-the-shelf dictionary, in this paper we propose a novel method, the Block K-SVD, which generalizes the well-known K-SVD to learn a block-structured dictionary for SRC face recognition. Similar to K-SVD, Block K-SVD also adopts an alternating optimization scheme which involves the iteration between sparse coding and dictionary updating. During the sparse coding stage, we propose a named Simultaneous Block Orthogonal Matching Pursuit (SB-OMP) algorithm to simultaneously perform the structured sparse coding of all samples from the same class, while during the dictionary update stage, we update the dictionary atoms and the corresponding coefficients block-by-block rather than one-by-one. This characteristic gives rise to taking full advantage of the block structure both in the coding step and in the dictionary updating step. Furthermore, the learned structural dictionary is less coherent than the ones learned by K-SVD and its variants. Experiment results on two public face databases demonstrate that our algorithm outperforms many state-of-the-art dictionary learning methods.
Supplement
Source code.