SVMs multi-class loss feedback based discriminative dictionary learning for image classification

Highlights：

• SMLFDL further employ the Fisher discrimination criterion on the coding coefficients under -norm constraint to make the coding coefficients have small intra-class scatter but big inter-class scatter for countering intra-class variability of datasets.

• An efficient and practical SMLFDL optimization algorithm is presented to learn a dictionary while training SVMs. Experimental results on several widely used image databases show that SMLFDL can achieve a competitive performance with other state-of-the-art methods on classification task.

摘要

•Inspired by the feedback mechanism in cybernetics, a novel discriminative dictionary learning framework, named support vector machines (SVMs) multi-class loss feedback based discriminative dictionary learning (SMLFDL) is proposed to learn a dictionary while training SVMs. As far as we know, it is the first time that the feedback mechanism in cybernetics is adopted for constructing dictionary learning model.•SMLFDL further employ the Fisher discrimination criterion on the coding coefficients under -norm constraint to make the coding coefficients have small intra-class scatter but big inter-class scatter for countering intra-class variability of datasets.•An efficient and practical SMLFDL optimization algorithm is presented to learn a dictionary while training SVMs. Experimental results on several widely used image databases show that SMLFDL can achieve a competitive performance with other state-of-the-art methods on classification task.