Weighted constraint based dictionary learning for image classification

Abstract

Dictionary learning (DL) is a popular approach of image classification. Most DL methods ignore the information hidden in training samples or atoms, and thus cannot enhance the discrimination performance of a dictionary learning algorithm effectively. In addition, the training samples are prone to a wide range of variances such as sample noise and illumination change, which results in the degraded classification performance. Hence, in this paper, we propose a weighted constraint based dictionary learning algorithm to improve the classification performance of dictionary learning. More specifically, the proposed algorithm uses a diagonal weighted matrix to construct a constraint item for reducing the auto-correlation between atoms. Meanwhile, the training samples of the same class enjoy similar coding coefficients such that the reconfiguration and discrimination performance of dictionary is enhanced. Furthermore, in order to avoid over-fitting, we convert a strict two valued label matrix into a flexible matrix in the classification procedure allowing more degrees of freedom to fit the class labels. Experimental results show that the proposed algorithm outperforms massive state-of-the-art dictionary learning and sparse representation algorithms in image classification.