Exemplar-Based Pixel by Pixel Inpainting Based on Patch Shift

This paper presents a novel exemplar-based image inpainting method. It performs patch-shift based information statistics and incorporates the result into the data filling process. The central idea to ensure the confidence pixel value to fill is based on patch shift that captures the statistic information from the expanded filling patches' most similar patches. Our approach outperforms the original exemplar-based inpainting in a large number of examples on removing large occluding objects and thin scratches in real and synthetic images, at the cost of some additional computation. It inherits well-known advantages of the exemplar-based inpainting method, such as the capability of maintaining the integrity and consistency of image structure information, and the adaptability to many application scenarios. Experiments on synthetic and natural images show the advantages of our approach, especially, in the big target removing task our method yields generally better results.