An interactively constrained discriminative dictionary learning algorithm for image classification

Abstract Researches demonstrate that profiles (row vectors of coding coefficient matrix) can be used to select and update atoms. However, the profiles are seldom used to construct discriminative terms in dictionary learning. In this paper, we propose an interactively constrained discriminative dictionary learning (IC-DDL) algorithm for image classification. First, we give a Lemma of the relation between the profiles and atoms. That is, similar profiles can lead to the corresponding atoms which are also similar, and vice verse. Then, we construct a profile constrained term by using the profiles and Laplacian graph of the atoms. Third, we explore the atoms and the Laplacian graph of the profiles to construct an atom constrained term. By alternatively and interactively updating the profiles and atoms, the two proposed constrained terms not only can inherit the structure information of the training samples, but also preserve the structure information of the atoms and profiles simultaneously. Moreover, the atom constrained model also can minimize the incoherence of the atoms. Experiment results demonstrate that the IC-DDL algorithm can achieve better performance than some state-of-the-art dictionary learning algorithms on the six image databases.