Image Super-resolution Based On Double Dictionary Learning And Sparse Representation

Super-resolution(SR)addresses the problem of generating a high-resolution(HR)image from its low-resolution(LR)version,considering about noise,blurry and so on.By super-resolving an LR image,more robust performance can be achieved in many applications such as computer vision,medical imaging,remote sensing,and video surveillance.But in some cases it is difficult to acquire HR images due to the physical limitation of relevant imaging devices,the SR technique can break through the limitation of both imaging equipments and the environment to produce an HR image that traditional digital cameras cannot capture from a real scene.Thus,super-resolution has been an active research topic in the areas of image processing and computer vision these years.This paper research on the existing image super resolution methods based on dictionary learning,and analyse the image degeneration as well as the compressed sensing theory.This paper propose an improved double dictionary learning super-resolution reconstruction method for low-dose X-ray photography and a framework that seamlessly integrates self-similarities and double dictionary for image super-resolution.The main works and innovations are introduced as follows:(1)In order to solve the mismatching exist in medical image super resolution with single high resolution dictionary,this paper presents a new approach to single-image super resolution and apply it on the X-ray photography field to evaluate its feasibility and performance.This method takes the compressed sensing theory to reconstruct the low-resolution images efficiently.Firstly,A novel dictionary learning method is proposed by simultaneous training two dictionaries using the theory of Principal Component Analysis(PCA),in order to enhance the similarity of the sparse representations between the LR and the HR block pairs.Secondly,for each given input LR patch,find the most proper sub-dictionary in the low-resolution dictionary.Thirdly,get the corresponding high-resolution sub-dictionary,solve the ill-posed super-resolution problem by the iterative algorithm and to generate the HR blocks.(2)Recent image super-resolution methods exploit self-similarities and group structural constraints of image patches,not only within the scale but also across scales.It has been widely acknowledged that it is effective to generate interpolated images using this feature.However,for the image without enough repetitive patterns,the methods only use non-local similarity and multi-scale similarities tend to produce sharpened edges rather than fine details.In order to solve this problem,we proposed a framework that seamlessly integrates self-similarities and guided image for image super-resolution.The image nonlocal similarity is introduced as a regularization term into the neighbor embedding methods,which improves the quality of the initial estimation.We incorporate the image multi-scale similarity and the sparse representation of the training set into the reconstruction model as regularization terms and propose a image super-resolution method based on enhanced neighbor embedding and double dictionary learning.The expression ability of prior information is enhanced because of combining the advantages of the two aspects.Experimental results show that the proposed approach can achieve better reconstruction effect in both subjective and objective evaluation criteria.