Two dimensional K-SVD for the analysis sparse dictionary

Abstract

Analysis sparse model has been successfully used for a variety of tasks such as image denoising, deblurring, and most recently compressed sensing, so it arouses much attention. K-SVD is a mature dictionary learning approach for the analysis sparse model. However, it represents images as one dimension signals, which results in mistakes of spatial correlations. In this paper, we propose a novel analysis sparse model, where analysis dictionary derived from two analysis operators which act on an image, leading to a sparse outcome. And a two dimensional K-SVD (2D-KSVD) is proposed to train the analysis sparse dictionaries. Experiments on image denoising validate that the proposed analysis dictionary can express more image spatial and frequency characteristics and by using the dictionary, the two dimension analysis sparse model outperforms the traditional analysis model in terms of PSNR. © 2012 Springer-Verlag.