Single Image Super-Resolution Research Via Self-Similarity And Regression Model

Image is the objective and true reflection of all natural objects.Compared with the traditional media such as text and voice,the image communication of information is rich,intuitive and fast.With the rapid development of computer hardware performance and software technology,image applications have gradually penetrated into all aspects of human life,which constantly improves living standards and happiness index.However,due to the limitation of imaging equipment and the acquisition environment,a large number of images exist in the form of low resolution.Single image super-resolution(SISR)is devoted to generating a high-resolution(HR)image from a low-resolution(LR)one,which has been widely used in many fields such as medical treatment,monitoring,remote sensing and so on.Therefore,SISR has been one of the important research hotspots in image processing.All research methods have the same reconstruction goal:generating as much as accurate high-frequency information as possible.At present,the sample-based methods are the mainstream,and the key is to learn reliable mapping relationships from the set of high and low resolution images.However,the dependence on training database and the uncertainty of recovery results in methods based on external database and the time-consuming out of online reconstruction in methods based on internal database are challenges.Considering the above background and analysis,a single image super-resolution method that is based on self-similarity and regression model is proposed in this paper.This method does not use any external information,and completely relies on the inp ut image itself to complete the whole reconstruction process.Firstly,a local-feature based interpolation(LFI)where both edge pixel property and location information are taken into consideration is presented to obtain a better initialization.Then,with the aid of our in-depth study on self-similarity,a training library is constructed and updated by the two nearest down-sampled image layers of the image layer to be constructed.Thus a dynamic and lightweight database composed of its own samples is established,which greatly reduces the time consumption of online searching for similar neighborhood sets.Secondly,in the process of searching similar samples,each low-resolution input image patch is regarded as an anchor point to obtain more accurate and flexible adaptive linear regression matrix,and then the low-resolution image patch is mapped directly to its high-resolution version.In addition,in each step of the step-by-step enlargement strategy,iterative back-projection optimization method is used to improve the content consistency between the final reconstruction result and the input image.In the comparative analysis of the experimental results,firstly,the performance of the local feature interpolation proposed in this paper and the traditional bicubic interpolation are compared.The former can more accurately restore the edge information,which is conducive to improving the quality of the sample training database.Then,the self-similarity of image is explored in depth,and the design motivation of the dynamic light database in this paper is expounded from three aspects:search time,memory occupancy and reconstruction quality.Finally,the reconstruction performance ofthis method is compared with that of other methods.Set5 and Set14 are selected as test libraries.A large number of experimental results prove the competitiveness of the proposed scheme both in visual effects and objective evaluation criteria.This algorithm can reconstruct more accurate edge,texture,details and other information,which has a certain practical significance in real life.