Research On Video Super-Resolution Technology Based On Sparse Representation

With the progress of information age,images and video have become important carriers of multimedia information.Perfect visual perception and high-quality images are greatly pursued today;however,the bad imaging equipment and shooting environment make it very difficult obtain high-resolution data sources.Therefore,researches on super-resolution reconstruction based on low-resolution images or video becomes increasingly crucial.The focus of research in the field of super-resolution is the image super-resolution based on sparse representation incorporating the concept of learning.Firstly,a large number of high-and low-resolution image pairs are trained and learned,and a relation model between sparse representations of high-and low-resolution image blocks is constructed,which is a high-resolution dictionary and a low-resolution dictionary.Then a sparse solution is applied to the waiting super-resolution block based on the dictionary to achieve the optimization of image super-resolution model in the entire sample space.Based on the sparse representation of single image super resolution,the paper proposes an improved sparse representation super resolution model,and extends this model to the application of video super resolution.The main research focuses include:(1)The traditional sparse representation image super-resolution is sparsely solved based on a single image block.The sparse reconstruction of each image block is independent.In this paper,the sparse solution of a single image block is replaced by the joint sparse solution of simultaneous orthogonal matching pursuit(SOMP)based on similar block sets,and the association constraint between image blocks is added to the sparsely solved model,making sparse solution more precise.SOMP joint sparse super resolution model has better effect on visual and evaluation indicators.(2)The accuracy of the set of similar blocks has an influence on the precision of the sparse solution of SOMP.In view of the characteristics of non-local means(NLM)and local control kernel regression(SKR),similar criteria for non-local kernel regression(NL_KR)and improved adaptive NL_KR similarity criteria have been proposed.Since there are fewer blocks with higher similarity in complex corners of the image,using adaptive NL_KR similarity criteria can obtain more accurate similar block sets.In this paper,NL_KR similarity criterion is used at non-corner point and adaptive NL_KR similarity criterion used at corners to determine the similar block set of each point,and the simulation verifies that the similar block set obtained by this model is more accurate;(3)Expand the single image super resolution method to the video super resolution field.A 3-D NL_KR similarity criterion is proposed for the video sequences with time dimension,and the 3-D similarity criterion is used to determine the similar block set of the current frame image block in the 3-D search domain.The process includes no explicit motion estimation.SOMP sparse solution is performed based on similar block sets and then super-resolution reconstruction of image block is conducted on the current frame.Simulations show that the video super resolution method is superior to the general video super resolution algorithm in the evaluation indicators of PSNR and SSIM.