Patch-based inpainting is widely used in interactive photo editing scenarios. It iteratively fills the target region by searching the candidate from the source region. However, the high computational cost is a long-lasting concern that prevents users from seeing instant results, bringing limitation to its further application. In this paper, we present a novel instant inpainting technique based on a multi-scale framework, pushing the speed to an unprecedented interactive level (around seconds). Our key insights are that filling process becomes very efficient at low scale. Also, scale changes do not significantly affect the match correspondence, allowing the source-target match correspondence to be quickly collected at low scale, and to be delivered as a priori to higher scales. At high scale, thanks to the built-in coherence in natural images, we propose a mechanism named propagation optimization to further fine-tune the match correspondence based on prior, eliminating the side effects caused by scale recovery. Experiments demonstrate that our method is 10-300 times faster while keeping the same image quality as previous works did. We believe our work contributes a new strategy to real-time application of image inpainting techniques.
|