Image Classification Based on Discriminative Dictionary Pair Learning

Shuai Yuan, Huicheng Zheng, Dajun Lin

Abstract: Dictionary learning plays an increasingly important role in image classification in recent years. Most of existing dictionary learning methods aim to enhance discrimination of the learned dictionaries. Recently, learning a pair of dictionaries shows effectiveness and efficiency in image classification. Such a pair consists of a synthesis dictionary and a projective analysis dictionary. Different from traditional sparse representation, such a model enforces group sparsity based on structured representation of the pair of dictionaries, which consists with the objective of classification. In this paper, we propose to enhance the discrimination of coding coefficients to further improve the structure of the dictionary pair. More specifically, a regularization term on the coding coefficients is introduced to push pattern representations of the same class closer and those of different classes further away. At the classification stage, we use the learned dictionaries to improve image classification. The experimental results on several representative benchmark image databases demonstrate the effectiveness of the proposed method.