Variable Patch Size Sparse Representation over Learned Dictionaries

Abstract

This paper addresses the patch size issue in sparse representation over learned dictionaries. A strategy for selecting the best patch size is proposed. It is empirically shown that the representation quality of natural image patches depends on the patch size considered. The proposed strategy selectively chooses the most appropriate patch size based on the resulting sparse representation error. The sparse representation of each small-sized image region is taken by selecting the most suitable patch size for the patch containing this region. The proposed strategy is shown able to improve the sparse representation quality as seen in numerical experiments, both quantitatively and qualitatively. As tested over a set of benchmark images, the proposed strategy has an average PSNR improvement of 0.99 dB over the standard case of using a fixed patch size. Visual comparison results come inline with the PSNR improvement.