Image Inpainting Based on Non-local Learned Dictionary

A novel learning-based image inpainting method is presented. As a further development of classical sparse representation model, the non-local self-similar patches are unified for joint sparse representation and learning dictionary, in which each element of the self-similar patches has the same sparse pattern. The method assures the self-similar patches possess similarity when projected on the sparse space, and efficiently builds the sparse association among them. This association is next taken as a priori knowledge for image inpainting. The paper uses numerous samples and non-local patches of input image to train overcomplete dictionary. The method not only takes into account the priori knowledge of samples, but also considers the non-local self-similar information of input image. Large and small region inpainting experiments and text removing experiments on natural images show the good performance of the method.