Structured supervised dictionary learning based on class-specific and shared sub-dictionaries

In recent years, sparse representation and dictionary learning has been widely used in signal processing tasks, especially for classification aim. For the aim of high accuracy classification, sparse coefficients obtained based on learned dictionary should have high discrimination power, which common sparse representation techniques cannot well satisfy that because most techniques ignore the underlying structural information of the data. Instead, structured sparsity coding techniques capture the structural information of data and improve the classification accuracy accordingly. On the other hand, one should notice that data samples of different classes might share some similarities which can decrease the discrimination ability and the classification accuracy. This implies to use a shared dictionary among all classes to capture the similarities while class-specific sub-dictionaries describe the intra-class features properly. In this paper, inspired by DL-COPAR method, a structured sparse coding technique is proposed that learns discriminative class-specific sub-dictionaries and a shared dictionary which its atoms are shared among all classes and have no discrimination capability. Also, the proposed method is based on l2,1 norm to use the structure of the data too. The optimization function of the proposed method is solved by an efficient alternating iterative scheme. The proposed method is evaluated by conducting experiments on four datasets and the experimental results demonstrate the effectiveness of the proposed method.