Graph-Regularized Discriminative Analysis-Synthesis Dictionary Pair Learning for Image Classification

Abstract

Analysis-synthesis dictionary pair learning, which can provide a comprehensive view of data representation, has been applied in various computer vision tasks. Although good performance has been reported in image denoising, discriminative dictionary pair learning for image classification remains unsolved. In this paper, we propose a novel model of graph-regularized discriminative analysis-synthesis dictionary pair learning (GDASDL), in which a graph-regularized term and a discriminative term are incorporated into dictionary pair learning. By taking advantage of graph constraints, the proposed GDASDL can preserve the local geometry structure of the data. Global information is introduced by associating label information with dictionary atoms. In this paper, an iteration algorithm is presented to efficiently solve the proposed GDASDL. We extensively conduct experiments on three public image datasets and one face dataset in comparison with the existing dictionary learning approaches, and the experimental results show that the proposed model achieves superior performance using a simple linear classifier.

Keywords

• Representation learning
• dictionary learning
• image classification
• local geometry structure