Image Inpainting Based On Weight Variation Of Neighborhood Window

As a typical algorithm of exemplar-based algorithm, Criminisi algorithm has achieved good inpaint result in existing works. However, when it is used to inpaint the damaged area with both structure and rich texture information, since the priority measure and sample block matching of Criminisi algorithm is unreasonable, Criminisi algorithm easily confuses texture for structure and inpaints texture prior to structure, which will result in texture extension and affect the inpaint result. This paper focuses on the structure and texture characteristics, two novel algorithms are proposed based on weight variation of neighborhood window and structure distribution rate:Firstly, an image inpainting method based on weight variation of neighborhood window is proposed, to modify the priority measure in Criminisi’s algorithm, the weight variation is introduced by combining the total variation and intrinsic variation in a neighborhood window. Therefore, the ability of identifying geometric and texture information has been improved and the geometric information can be priority inpainted. Furthermore, through introducing structure difference operator in combination with pixel color comparison, the matching accuracy has been improved. Compared with the original Criminisi and its improved algorithms, the proposed algorithm can achieve the better result to inpaint the damaged region with both geometric structure and rich texture as well as to inpaint some ordinary damaged region so that it has generality.Secondly, an image inpainting method based on structure distribution rate is proposed in this paper. To improve priority algorithm, the similarity distribution rate and the structure distribution rate are constructed by the characteristics of structure and texture of the damaged image, which can distinguish the structure and the texture effectively. Furthermore, the searching and matching algorithm are developed by enlarging the samplers, which can enhance the matching accuracy and efficiency. The experimental results demonstrate that the improved algorithm solves the existing problem of mismatch and texture extension in the original Criminisi algorithm. Further, it can maintain the visual connectivity of the inpainted image, and improve the peak signal-to-noise ratio (PSNR) about 2-3 dB compared with the original algorithm.