The Research Of Fast Inpainting Algorithm Based On Image Decomposition

Image inpainting is the process of filling missing information of the picture withcertain algorithm on the computer, and requires the inpainting result people canacecept.So far, inpainting technology is mainly divided into inpainting of structurebased on partial differential equation and of texture based on texture synthesis. Forinpainting large area which contains both structure and texture,a good result may not beobtained when using methods above individually. To solve this problem, Criminisialgorithm based on exemplar and layered inpainting methods based on imagedecomposition may be used. These methods are introduced in this paper and layeredinpainting methods based on image decomposition are emphatically studied because ofits complex process and low efficiency leading to less application. For the sake ofimproving the efficiency and effectiveness, from the angles of structure and textureinpainting,the following work is made:For structure inpainting, traditional models based on partial differential equationand interpolation according to the neighbor area are introduced in this paper. The formercan obtain better effectiveness but lead to poor efficiency because of its solutionneeding much time to iterative. Instead, the latter can process very fast but gets worseeffectiveness. Aiming at meeting the requirements of both efficiency and effectivenessfor structure inpainting, a fast method based on gray theory is proposed. It adopts greyforecasting model to mine the tendency of grayscale sequences in different direction andto predict the grayscale value of single unknown pixel using the known informationaround the area to be restored. The inpainting process is same as the silkworms eatmulberry leaves. It replaces the complex priority calculation and guarantees the qualityof edge restoration. The experimental results show that the algorithm not only extendstopology of edge but also meets the real-time requirements.For texture inpainting, texture synthesis based on exemplar is introduced in thispaper. To find the closest sample, in the searching stage, a way of local searchinginstead of global searching is integrated into this paper and a method which uses pixellevel of template center in structure image to filter some source templates is proposed.Then Euclidean Distance between pixels of target template and source template will becalculated to search the most similar sample. Also residual test method will beintegrated into distance calculation. Thus, the matching speed is faster and mirror is decreased. Analysis of the redundant calculation of searching for the target templatewith max priority is made in this paper. And an idea that only recalculating the priorityof changed boundary is pointed out. The experimental results show that improvedmethod in this paper has faster speed and higher matching precision compared withtraditional Criminisi algorithm.At last, the image to be inpainted is decomposed into structure layer and texturelayer using image decomposition model in this paper. Then structure is inpainted usingthe interpolation algorithm based on grey system theory and texture is inpainted usingCriminisi algorithm improved in this paper. The last result will be obtained after addingthe inpainting results of two layers. The experimental results show that inpaintingefficiency and effectiveness is better in this paper compared with traditional Criminisialgorithm.