Image Inpainting Based On Tensor Structural Sparse Representation

Signal image inpainting is to repair the damaged images using information we have known or to remove the unwanted objects in the images, it is one of the most important contents in the study of image restoration. At present, there are three kinds of methods of inpainting: texture diffusing, sample texture synthesis, and sparse representation. In the third method, we can using the property that the sparse coefficients of the damaged and the undamaged image are the same to inpaint images completely. In this paper, because common image inpainting methods which have to put the image blocks into columns, which damages the structure information and the global and non-global self similarity of the images,which in turn, affects the quality of inpainting, we do the following researches:1、It achieved a method of image restoration based on classification dictionary and sparse representation. In view of the problem that the traditional dictionary learning does not take into account the similar structure information, this paper proposes a sparse representation model of the classification dictionary, makes full use of the structural information of the image block, and then carries on the texture classification, and the image patches. Using the above models, we repair the mask, letters and noise of the damaged natural images. The results show that the numerical results of SSIM and PSNR based on the structural information classification and the visual effects are the best algorithm.2、It achieved a new image inpainting method based on patch sparse representation of image block dual geometry adaptive dictionary.This paper proposes 2D dual geometry structure model taking full use of local self similarity and non local similarity of image blocks, and introduces the model and algorithm of adaptive dictionary learning. According to the different images, a large amount of time is saved. At last, we construct group sparse representation model to improve the quality of image inpainting. Experimental results show that: compared to other similar methods, the numerical results of group based dual geometry adaptive dictionary of image method have a greater improvement.3、A tensor sparse representation algorithm based on structural information is designed. We need take the image blocks into vectors when inpaint a image, which destructs the texture structure of the image blocks. On the basis of 2D model, the spatial structure information of the natural image is considered, and 3D spatial model is built, which is used to improve the efficiency of sparse encoding. The simulation experiments of the mask, letter and noise destructions of the images show that the proposed method can obtain the numerical results and improve the visual effect.