| The digital image inpainting technology is an important branch in digital image processing, it iswidely used in domains such as restoration of damaged image and video, inpainting of ancientcultural relic, removal of occlusions, special effects of films, image transmission and compressionand so on. The goal of the image inpainting is to repair damaged area with minimizing artificialmarks. There are two primary image inpainting methods: image inpainting based on partialdifferential equations and image inpainting based on texture synthesis. The former is based on PDEmodel, which establishes partial differential equation on the boundary of the damaged area, thenfills the areas to be inpainted by smoothly propagating information from the surrounding areasalong the the normal of direction isophote. The latter makes use of the effective synthesis of thetexture to complete the inpainting, it is mainly suitable for the image with rich texture and bigdamaged area.Criminisi algorithm is a classical image inpainting method which is based on sample. Firstlyfind the the boundary of the damaged area and select the block with biggest priority, then search inthe source region for seeking the most similar template for the current block, following find thesubstitute for the current block, finally update boundary. Repeat the above steps until the inpaintingis completed. Criminisi algorithm can inpaint image with higher speed, less garbage growing andaccurate propagation of linear structures, especially the image with quite big damaged area.However, Criminisi algorithm still has shortcomings such as the inpainting order is not fullyreasonable, the choice of the similar patches is easy to make mistake and so on. This article mainlystudies Criminisi algorithm, and has made two improvements to its existed insufficiencies:Frist, modify the repair sequence, by introducing dynamic factors, the weight coefficients ofconfidence and data in the priority calculation formula dynamically adjusted during the repairprocess, at the same time the adjustment is different due to the inpaint image, which makes theinpainting order more reasonable.Second, modify the matching criteria, when the SSD of several source blocks are the same,which has the smallest physical distance away from the target block, is the optimal matching. In thisway, the optimal matching with the object is not only simila r in color but also in field to ensure themore accurate selection of the best matching block.At the same time, on the base of a full theoretical analysis, this article carries on massive simulating experiment about the two proposed improvements, which co nfirms the feasibility andthe validity of the improved methods. |