The Research Of Super Resolution Based On Learning In Image Reconstruction

Super-Resolution is an effective way to improve the quality of the image which has great significance on image processing, and it's widely used in medicine, surveillance, military, digital television and so on. The traditional reconstruction methods are based on the interpolation and the results are always have block artifacts and not good for edge retention. Super-Resolution image reconstruction algorithm based on learning is a hot topic in recent years, especially the method based on locally linear embedding in manifold learning made the image reconstruction come to a new age, and then with the development of the sparse representation, the image reconstruction technical has made great progress.In this paper, based on the relevant theory of manifold learning and sparse representation, analysis the image reconstruction methods which are based on the manifold learning and the sparse learning, and give a new feature extraction method and the training set generation process, at the same time, as to the over redundancy of the over complete dictionary introduce a new image reconstruction algorithm based on multi-class dictionary. Specific works are as follows:(1) Analysis the image degradation model which is the expression of image degradation process in mathematics and the reverse process of the super resolution, so that we can get the priori information of the image based on this model. Then study the traditional image reconstruction algorithm based on the interpolation such as nearest interpolation, bilinear interpolation, bicubic interpolation, at last, experiments have been done to analysis the strengths and weaknesses of the algorithm.(2) Analysis the relevant theory of manifold learning and especially focus on the image reconstruction based on the locally linear embedding, and as to the feature extraction problem introduce a new method that is a joint of first order gradient and norm luminance. In addition, as to the training set, on the one hand, using the classification method to improve the efficiency of neighbor selection, on the other hand, rotating the image blocks which have rich edge information, then we can get more meaningful training set image blocks without increase the training set images. Experiments show that the method presents better result than others not only in visual aspect but also in peak signal to noise ratio and structural similarity index measurement.(3) Learning the relevant theory of sparse representation, and combined with image resolution based on sparse representation introduce a multi-class dictionary based algorithm to solve the problem of neighbor selection. The method avoids the over redundancy of only one dictionary, and for each category training a sub dictionary so that the sub dictionary may be more representative. Then, for each image block introduce a new distance calculation method to replace the traditional Euclidean distance to find the optimal sub dictionary, thereby improving the quality of the entire image. In addition, as to the ill-posed problem of super resolution image reconstruction introduce a non-local self-similar regularization construction to revise the result. Experiments show that the method presents better result than others not only in visual aspect but also in peak signal to noise ratio and structural similarity index measurement, and the method with non-local self-similar regularization construction can get even better image.