Research On Digital Image Inpainting Algorithm Based On Texture Synthesis

In this thesis, firstly, we present an improved algorithm for solving the problemof error propagation caused by the Criminisi algorithm, a famous method in thetexture synthesis. This improved method takes the boundary information into accountwhen computing the priority values based on the Criminisi algorithm which onlyconsiders the confidence term and the data term, and therefore offers an effectiveand stable treatment of image information for damaged area.Secondly, to accelerate the speed of the algorithm above, we give a fasteralgorithm which only finds the samples in the neighborhood based on the somesimilarity of the damaged region. Experiments show this method is much fasterthan the original one.Lastly, as the Criminisi algorithm can not restore the image which containscurve structural information, we present two methods trying to solve this problem.One is to obtain a sample list by rotating the original image successively, and wecan find the samples from this list during the inpainting process. The other isrotating the patch during the match process to find the best restoration effect. Bothmethods can keep the curve structural information of the original image.