Discriminative Analysis Dictionary Learning With Adaptive Graph Constraint for Image Classification

Abstract

Discrimination of coefficient matrix plays an important role in discriminative analysis dictionary learning (ADL) model. However, the local geometric structure of the profiles(i.e., row vector of coefficient matrix) is seldom exploited to design discriminative terms in the discriminative ADL algorithms. In this paper, we proposed a discriminative ADL algorithm with adaptive graph constrained (DADL-AGC)model, which can adaptively preserve the local geometric structure information of profiles. First, we construct an adaptive graph constrained model by maximizing the information entropy of the similarity matrix of profiles. In this way, the coefficient matrix can preserve and inherit the local geometric information of analysis atoms and training samples by using the K-means method to initialize the analysis dictionary. Moreover, a robust linear classifier is simultaneously learned to improve the classification performance of our DADL-AGC algorithm. On the four deep features and hand-crafted features databases, experimental results demonstrate that our DADL-AGC algorithm can achieve better performance than seven ADL and synthesis dictionary learning algorithms.