Fully automatic inpainting method for complex image content

Abstract

A novel, fully automatic framework for restoration of unknown or damaged picture areas is presented. Diverse causes as an accident, manual removal, or transmission loss may have lead to the missing visual information. The challenge then consists in repairing the occluded or missing image regions in an undetectable way. Here, assumption is made that dominant structures are of salient relevance to the human perception. Hence, they are accounted for in the filling process by using tensor voting, which is a structure inference approach based on the Gestalt laws of proximity and good continuation. In fact, based on a new segmentation-based inference mechanism presented in this paper, missing textures crossing dominant structures are robustly recovered. An efficient post-processing step based on cloning via covariant derivatives improves the visual quality of the inpainted textures. The proposed method yields significantly better results than previous approaches.