LAUSR

Abstract Over the past decade, discriminative dictionary learning (DDL) has demonstrated the great success in various pattern classification problems. However, in previous DDL methods, the scheme that how to generate the effective coding coefficients for classification has not been well addressed. This paper proposes a novel DDL method, named CW-DDL, to learn a discriminative dictionary for classification by exploiting class-wise coding coefficients. In the proposed method, a label-aware constraint is first presented to make the coefficient matrix has class-wise approximate sparse structure. Then the graph regularization is further enforced on coding coefficients by utilizing the locality information of dictionary atoms. These two constrained terms reinforce each other in the learning process, resulting in a very robust and discriminative dictionary. Moreover, to obtain class-wise separation of coefficient vectors from different classes, a support vector based classifier is integrated into the objective function. Finally, an iterative algorithm is devised to solve the proposed method efficiently. Experimental results illustrate that the optimal coding coefficients derived by CW-DDL are very effective for pattern classification. The proposed method shows the superior performance to related DDL methods on several benchmark datasets, and coupled with the CNN features, it also leads to the state-of-art performance on the more challenging dataset.