Single Image Super Resolution Based On Sparse Representation And Local Rank

With the development of society and the increasing of special needs, high resolution images are needed in the social life and the specific situation. However, due to the limit of the hardware, the resolution of actual images, which cannot satisfy the life needs and special requirements. The direct way to get the high resolution images is unable to be extended because of the high cost of updating the hardware level. In recent years, various digital image processing based super resolution algorithms have been proposed to obtain the high resolution images, which thank to the development of the computer technology.Based on the deep research of the sparse representation based single image super resolution method, this paper proposes a new single image super resolution reconstruction method by combining the local rank and the sparse representation. In order to improve the image quality, this paper proposes a non-local and global post processing optimization model. The main research contents of this paper are as follows:① According to the survey and analysis of super resolution, this academic paper proposes a new single image super resolution method based on sparse representation and local rank. The local rank images are generated by processing the high resolution images of the training database. In order to train the corresponding dictionary, the image patches, which includes the high resolution image patches, the low resolution image patches, the positive local rank image patches and the negative local rank image patches, are extracted from the training images. According to the definition of local rank, there are many positive local rank image patches only with zero elements and these zero patches not only cost a lot of training time but also have influence on the accuracy of the dictionary during the dictionary learning process. In order to improve the accuracy of the dictionaries and reduce the size of training samples, the image patches are classified into two patterns.② For every pattern, the corresponding dictionaries are learned. However, the dictionaries are inaccuracy when they are trained separately. In this paper, the multiple dictionary learning models are proposed based the traditional joint dictionary learning model.③ During reconstruct the high resolution image patch, the high resolution image patch is reconstructed by determining which pattern the low resolution image patch belongs to. In order to restrict the edges of the high resolution image, the local rank of the high resolution image is obtained by the local rank dictionary.④ The non-local and global post processing optimization model is proposed to further improve the image quality. During calculating the weights of the non-local term, a local rank based calculation method is proposed.