Constrained patchmatch for image completion
Abstract
We propose a quick and automatic patch-based image completion method, which uses the PatchMatch framework. We show that PatchMatch can be improved by constraining its random search step, in order to propagate the geometric structures of the image better. Instead of randomly initializing the algorithm and randomly searching, we guide the search by constraining it among only the most likely offsets. Moreover we modify the PatchMatch cost function to ensure a coherence of offset directions. The method is tested with real data, and is compared with state-of-the-art methods