Research On Non-Local Image Inpainting Algorithm Based On Exemplar-Learning Sparse Representaion

Digital image inpainting is a process of filling in missing data in a damaged region of an image in a visually plausible way. Image inpainting has broad applications in the error concealment of video, special effects in film making, artwork restoration, etc. It's becoming one of hot topics of image processing. The selection of an image model is critical to solving the image inpainting problem. In the foundation of the deeper research on the existing image inpainting methods, this thesis carries out some researches on image inpainting.Firstly, in order to promote the exploitation of image local neighborhood information and guarantee the inpainting direction keeps consistent with the structure information in the image, an iterative image inpainting using sparse constrain with local adaptive learned dictionary and informational priority selected diffusion is proposed in this paper. On the basis of modeling the local neighborhood of the lost area by learning, the algorithm reconstruct the missing data along the isophote direction with sparse representation which would realize the effective restoration of different structural content in the lost region.Secondly, in order to obtain the optimal solution of image inpainting problem according to the image self-similarity property, progressive non-local mean based global image inpainting optimization with informational priority selected diffusion is presented. A global inpainting energy function is built on the lost area by taking advantage of image self-similarity, and the function is optimized using non-local mean filter based expectation maximization algorithm. This algorithm further improve the quality of structure restoration by constraining the inpainting is performed alone the isophote. The experimental results show the effectiveness of the method.Finally, in order to strengthen the sparse constrain according the image self-similarity property, and construct an image content adaptive dictionary, non-local sparsity representation based image inpainting with multi-region learning dictionary is proposed. The algorithm models every region which contains similar geometric structure by a best learned dictionary. The learned dictionary is then employed to recover the damaged pixel iteratively by using a non-local sparse expectation-maximization representation modal which exploits the self-similarity dependency in natural image to reveal the organizational structure between image blocks. The result of experiments shows that our method effectively improves the quality of the recovered image.