Image inpainting with group based sparse representation using self adaptive dictionary learning

Abstract

Inpainting is an imaging technique that modifying an image in an undetectable form, is as ancient as art itself. The goals and applications of inpainting are numerous, from the restoration of damaged paintings and photographs to the removal / replacement of selected objects. The popularity of sparse representation and compressed sensing makes the sparse priors to be considered for solving inpainting problems. In earlier works, the patch of an image is taken to be sparse in a particular basis, which is called Patch based sparse representation. The patch based modelling suffers from two severe problems. In our work we exploit the concept of Group based sparse representation (GSR), which takes group (composed of nonlocal patches with similar structures) as the basic unit instead of patch. The GSR sparsely represents natural images in the domain of group, which results intrinsic local sparsity and nonlocal self-similarity of images simultaneously in a unified framework Under the same roof, an efficient self-adaptive dictionary learning method is designed for each group with low complexity, rather than learning the dictionary from natural images.