Research On Image Super-resolution Algorithms With High Up-scaling Factors

Super-resolution technology(Super-Resolution,SR)is the reconstruction of corresponding high-resolution images from observed low resolution images,which has important application value in illegal license plate shooting,medical image diagnosis and electronic data printing.SR can be divided into two categories:the reconstruction of high resolution images from multiple low resolution images and the reconstruction of high resolution images from a single low resolution image.The main research in this paper is a single image reconstruction method based on low resolution,so it is called a single image super-resolution.However,since the degradation process loses a lot of information,the super-resolution image reconstruction is an ill-posed problem,and is still a challenging job.Image SR reconstruction methods can be generally divided into three categories:image super-resolution based on interpolation,image super-resolution based on reconstruction,and learning.based image super-resolution.The current learning based super-resolution method mainly uses the prior knowledge and spatial similarity between low resolution image(LR)and high resolution image(HR)to establish related mapping relations,and then uses this mapping relation to estimate HR images from LR images.While learning based methods can predict lost information well,they depend largely on the similarity between the images in the training set and the test set.If the training set is not suitable,some unwanted artificial details are generated.Therefore,the prior knowledge of external constraints needs to inhibit the generation of the results.In this paper,we propose three learning based algorithms for image super-resolution,which mainly include the following three parts:Firstly,in our method based on self-learning and sparse representation,the high frequency information(HF)is mainly estimated by the self-learning method and sparse dictionary learning method.The traditional sparse reconstruction method mainly uses the external training dictionary to reconstruct high-frequency information.In this paper,a method of combining the internal samples and the external dictionary training sample is proposed to reconstruct the high frequency information.In the training phase,the method first uses the internal example training method to reconstruct the main high-frequency features(MHF),and then uses the external dictionary training method to train sparse dictionary pairs.In the stage of image reconstruction,we first use the method of internal example training to reconstruct the main high-frequency features(MHF),and then reconstruct the residual high frequency features(RHF)by sparse method.Since the algorithm utilizes the sparse dictionary method to learn and reconstruct images and can better reconstruct high frequency information,better quality of the reconstructed high resolution images can be obtained.Experimental results illustrate that the proposed algorithm performs better than current state-of-the-art super-resolution algorithms.Secondly,in our method on image super-resolution algorithm,it uses iterative back projection method and non-local feature enhancement method.In our method,the algorithm is an effective single image super-resolution(SISR)method based on anchored neighborhood regression.In most regression based image super-resolution methods,the low resolution feature of the linear matching function is directly extracted from the low resolution images after bi-cubic amplification.However,in this paper,we use the iterative backprojection method to enhance the bi-cubic image,and then extract the low resolution feature for the enhanced image.After extracting the low resolution feature,K-means clustering is used to learn M classes,PCA is used to reduce the dimension of each subclass,and the M class projection matrix is learned.With this adaptive selection of PCA and projection matrix,the features of high frequency images with more details are obtained.Finally,the image patch in high resolution images are searched for similar patches inside the high-resolution image,and the enhanced high-resolution images are obtained by using non local similar patch enhancement.Experimental results illustrate that the proposed algorithm performs better than current state-of-the-art super-resolution algorithms.In the third research of learning based anchored neighborhood regression and neighbor embedding for image super-resolution,Based on the assumption that low-resolution(LR)and high-resolution(HR)manifolds are locally isometric,the neighbor embedding super-resolution algorithms try to preserve the geometry of the LR space for the reconstructed HR space,but neglect the geometry of the original HR space.Therefore,our method attempts to reconstruct high resolution images by using the geometric duality between the original high resolution images.The method mentioned in this paper,mainly using two training sets,using the first training set to get the initial high resolution image IOR,The image patch in the initial high-resolution image IOR are searched for similar high resolution image patches in second training set,and using geometric duality between high resolution image patches to get the second high resolution image ImmidSR.Finally,the iterative projection method is used to enhance the coherence between the high-resolution image IOR and ImidSR,and the final high-resolution image ISR is obtained.