Exemplar based image inpainting on a projection framework

Abstract

An exemplar based inpainting method on a projection framework is presented. A ker- nel based projection is adopted to detect dominant gradient in the neighbor regions of the restoration target. The priority function which determines the patching order is redefined based on the projection framework. In addition, considering the observation that similar patches are usually found near the reference points, a distance based weighting is added to the similarity metric between two patches. It reduces the artifact effects caused by exhaus-tive patch searching. It improves the plausibility of the restoration especially when global structures are involved in the target region. The experimental results show that the proposed method enhances the plausibility of restored regions, especially when global structures are overlaid on the target region.