ODD: An Algorithm of Online Directional Dictionary Learning for Sparse Representation

Abstract

Recently, some sparse representation based image reconstruction methods have demonstrated with a learnt dictionary. In this paper, we propose a block-based image sparse representation approach with an online directional dictionary (ODD). Unlike the conventional dictionary learning approaches for image sparse representation aims at learning some signal patterns from a large set of training image patches, the proposed joint dictionary for each patch is composed by an original offline or online trained sub-dictionary from a training set and an novel adaptive directional sub-dictionary estimated from the reconstructed nearby pixels of the patch itself. A joint dictionary with ODD has two main advantages compared with the conventional dictionaries. First, for each patch to be sparse represented, not only the most general contents, but also the most possible directional textures of the image patch are considered to improve the reconstruction performance. Second, in order to save storage costs, only the original trained sub-dictionary should be stored, the proposed ODD can be obtained consistently. Experimental results show that the reconstruction performance of the proposed approach exceeds other competitive dictionary learning based image sparse representation methods, validating the superiority of our approach.