Jointly projection and graph-regularization coupled discriminative dictionary learning for image classification

Analysis synthesis dictionary pair learning methods have achieved good performance in image classification. Due to the redundancy contained in the original data, learning a more compact and discriminative analysis synthesis dictionary is still open. In this paper, we propose a jointly projection and graph-regularized coupled discriminative dictionary learning (JPGCDDL) for image classification.Specifically, JPGCDDL obtains feature more suitable for dictionary learning via simultaneously learning the projection matrix and analysis synthesis dictionary pair. Then in the low-dimensional subspace, we consider improving the discriminability of the analysis dictionary by introducing a constraint term on the coding coefficients, which can ensure both the within-class compactness and between-class separation. And the synthesis dictionary atoms are utilized to construct the graph regularization term to obtain a robust and discriminative synthesis dictionary. Moreover, the structural consistency term of the projection samples is introduced to make the low-dimensional data features have a common sparsity structure in each class, so that the final classification is focused on more important low-dimensional features. Finally, an effective iterative algorithm is devised to solve the optimization problem. Experimental results on several benchmark databases show the superior performance of JPGCDDL.