An improved scheme for Criminisi's inpainting algorithm

Abstract

Criminisi et al. proposed an exemplar-based inpainting algorithm, which has been one of the most important contributions in this field. In this paper, some heuristics are incorporated into Criminisi's algorithmic framework for performance enhancement. Specifically, (1) variable rather than fixed patch size is used to keep the similarity between the current patch to be inpainted and its exemplar above some prescribed level; (2) the confidence value to be updated of the newly inpainted patch is made explicitly dependent on current searching accuracy. Experimental results confirm the effectiveness of the modifications made.