Multi-level K-means clustering and group sparse coding with quasi-SIFT feature for image classification

Abstract

Traditional sparse coding treats local descriptors separately, leading to random locations of non-zero coefficients. To overcome this problem, group sparse coding was proposed. By leveraging mixed-norm regularisation, the locations of non-zero coefficients tend to cluster into groups. However, how to rationally constitute 'a group' is still a problem, since forcing local descriptors in a group to have uniform sparse pattern may be too rigid and bring about larger reconstruction error and information loss, especially when descriptors in a group are not very similar. In this paper, we propose multi-level K-means clustering and group sparse coding to address this issue, which is called MK-GSC. Moreover, we present a novel local descriptor based on SIFT to further promote the power of representation for images. Our method is tested on two public benchmarks, and achieves competitive or better results than the state-of-the-art methods.