Study Of Digital Image Inpainting Algorithms

With the development of computer’s storage and computing speed, digitalimage processing technology has been improved greatly. As an important branch indigital image processing field, digital image inpainting technique is playing anincreasingly important role in various fields, becoming the focus of research.Digital image completion method mainly comprise two categories: one is theinpainting technique for restoring small-scale damaged area, such as the algorithmbased on partial differential equations; another one is the texture completion techniquefor large-area damaged image, such as the exemplar-based texture synthesis algorithm.Currently, the inpainting algorithm for small-scale damaged area is relatively mature,so this thesis mainly discusses the texture synthesis technology which focuses onrepairing the image with large-scale damaged region.Firstly, this thesis presented the research background and significance of digitalimage completion algorithm, and summarized the research status of differentcompletion algorithms. Then the thesis deeply discussed the exemplar-based texturesynthesis algorithm, and on this basis proposed an adaptive method to determinepatch size, the method can automatically adjust the patch size according to differentstructure region. The simulation results show that this method can realize perfectcompletion effect on images with complex structure.Secondly, the thesis discussed two improved completion methods based on theimage’s structure features. One is a fast texture image completion algorithm based ondirectional search. It utilizes Hessian matrix to judge the structure direction ofrepairing patch, decides the fan shaped searching areas, and then the search region islimited to the regions which are similar to the structure of repairing patch.Experimental results show that the algorithm can improve completion speedeffectively and reduce the mismatch. Another one is an improved priority algorithmbased on structure sparsity, which make use of the structure sparsity of repairing patchto search the filled frontier of structure, instead of traditional structure term based onisophote. The simulation results proved that this algorithm can achieve better effectfor images with obvious structure and maintain the structure continuity.In the end, the thesis proposed a completion algorithm based on structure propagation. Firstly, connected the damaged edge according to the extracted edgeinformation, and then the connected edge was utilized as guidance to repair thedamaged structure with the structure information. In the end, the remaining texturesub-regions were filled separately with the partition result of image. Besides that, bysolving the Poisson equation, we employed the photographic correction for thestructure which human eyes are sensitive to. Experimental results proved the validityof the algorithm.