Research And Implementation Of Image Super-resolution Algorithm Based On Multi-dictionary Learning

Image super-resolution reconstruction refers to the technique of reconstructing highresolution images of the same scene from one or more low-resolution images by using only reasonable prior knowledge and mathematical models without improving hardware.Recently,this technology has been widely used in many fields,including the public security monitoring,remote sensing satellite imaging,medical imaging and high-definition television signals.This paper focuses on the super-resolution algorithm of single image based on the sparse representation.Aiming at the problems of the missing information exists in highfrequency components and the poor classification or clustering of image blocks in multidictionary learning,the corresponding improved algorithms are proposed.The specific work is as follows.(1)An image super-resolution algorithm based on image block multi-level classification is proposed.The algorithm uses the classification strategy to divide the image blocks into multiple categories,therefore,performs dictionary learning for different categories of image blocks.Firstly,the image blocks are divided into smooth blocks and detail blocks by multiple iterations,where the smooth blocks contain less image information and the detail block containing rich edge and texture information.Then the detail block with large amount of information is divided into edge blocks with different angles and irregular blocks according to the gradient axial angle.Secondly,Singular Value Decomposition(KSVD)algorithm is applied to different types of image blocks respectively to realize the dictionary learning,and then obtain a classification dictionary of the image blocks.Thirdly,for the input low-resolution images,the same image block classification strategy is adopted.Some larger smooth blocks directly apply bi-cubic interpolation method,and the detail blocks of different structural features are reconstructed by the corresponding dictionary.Compared with SISR algorithm,the proposed algorithm is improved by 0.4dB and 0.005 respectively on PSNR and SSIM.(2)This paper improves the super-resolution algorithm based on dual-dictionary learning by introducing the intermediate frequency information.The medium frequency information replaces the low frequency information.The main dictionary and the residual dictionary are learned by using the intermediate frequency information and the high frequency information of the training image.Then the dual-dictionary is used to gradually reconstruct the lost high frequency information from the low resolution image.Finally,the IBP algorithm is used to optimize the reconstruction results with respect to the global reconstruction constraint.Experiments show that compared with the original dictionaries learning algorithm,the proposed algorithm gets a much more clear reconstruction result on details and an average increase of 0.35 dB and 0.002 for PSNR and SSIM respectively.(3)Image super-resolution reconstruction system is designed and implemented by using the MATLAB GUI tool,which has three important modules: dictionary learning,dictionary display,image reconstruction.The dictionary learning module can perform dictionary learning through a variety of different algorithms.The dictionary display module converts the selected dictionary matrix into an image and displays it.The image reconstruction module applies a variety of different algorithms for low resolution image reconstruction.