Exploiting label consistency in structured sparse representation for classification

Abstract

Sparse representation with adaptive dictionaries has emerged as a promising tool in computer vision and pattern analysis. While standard sparsity promoted by $$\ell _0$$ or $$\ell _1$$ regularization has been widely used, recent approaches seek for kinds of structured sparsity to improve the discriminability of sparse codes. For classification, label consistency is one useful concept regarding structured sparsity, which relates class labels to dictionary atoms for generating discriminative sparsity patterns. Motivated by the limitations of existing label-consistent regularization methods, in this paper, we investigate the exploitation of label consistency and propose an effective sparse coding approach. The proposed approach enforces the sparse approximation of a label consistency matrix by sparse code during dictionary learning, which encourages the supports of sparse codes to be consistent for intra-class signals and distinct for inter-class signals. Thus, the learned dictionary can induce discriminative sparsity patterns when used in sparse coding. Moreover, the proposed method is computationally efficient, as the label consistency regularization developed in our method brings very little additional computational cost in solving the related sparse coding problem. The effectiveness of the proposed method is demonstrated with several recognition tasks, and the experimental results show that our method is very competitive with some state-of-the-art approaches.