Multi-feature Joint Dictionary Learning for Face Recognition

Dictionary learning with sparse representation has been widely used for pattern classification tasks, where an input is classified to the category with the minimum reconstruction error. While most methods focus on single-feature recognition problems, recent studies have proved the superiorities of exploiting multi-feature fusion classification. In this paper, we present a new multi-feature joint dictionary learning algorithm which can enhance correlations among different features via our designed class-level similarity regularization. The proposed algorithm can fuse different information and correlate these dictionary atoms within the same pattern category. Besides, the distinctiveness of several features is weighted differently to reflect their discriminative abilities. Furthermore, a dictionary learning algorithm is used to reduce dictionary size. The proposed algorithm achieves comparable experimental results in several face recognition databases.