Learning a structure adaptive dictionary for sparse representation based classification

Dictionary learning (DL), playing a key role in the success of sparse representation, has led to state-of-the-art results in image classification tasks. Among the existing supervised dictionary learning methods, the label of each dictionary atom is predefined and fixed, i.e., each dictionary atom is either associated to all classes or assigned to a single class. In this paper, we propose a structure adaptive dictionary learning (SADL) method to learn the relationship between dictionary atoms and classes, which is indicated by a binary association matrix and jointly optimized with the dictionary. The binary association matrix can not only represent class-specific dictionary atoms, but also hyper-class dictionary atoms shared by multiple classes. Furthermore, discrimination is explored by introducing Fisher criterion on coding coefficient and reducing between-class dictionary coherence. The extensive experimental results have shown that the proposed SADL can achieve better performance than previous supervised dictionary learning methods on various classification databases.