Image Classification with Group Fusion Sparse Representation

Abstract

In this paper we introduce a novel framework for image classification using local visual descriptors - group fusion sparse representation (GFSR), which casts the classification problem as a linear regression model with sparse constraints of the regression coefficients. Considering the intrinsic discriminative property of prior class label information, and the requirement of local consistency within a class, we add two penalties, one is for sparsity at group level, and the other is for the fusion demand. Experiments on several benchmark image corpora demonstrate that the proposed representation and classification method achieves state-of-the-art accuracy.