Research On Image Super-Resolution For Real Scenes

In the digital information age,as a relatively convenient carrier of information dissemination,image is in an irreplaceable position,and high-quality images containing more image details are of great value,especially in the fields of medical diagnosis,intelligent monitoring,remote sensing imaging,etc.Single image super-resolution(SISR)is a ”soft”technology to solve the problem that only blurred low-resolution images can be obtained due to external factors such as hardware constraints and weather noise.Since this technology has flourished,synthetic datasets have been unable to meet the performance requirements because of the difference between their degradation process and the real scenes,so a series of real-world datasets and real-world single image super-resolution(RSISR)have gradually emerged.This thesis aims to design a super-resolution algorithm with better performance in the real-world datasets,and focuses on the problems of insufficient receptive field which hinders the utilization of global information for image restoration and different requirements for super-resolution in different regions in the current super-resolution algorithm.The main research work is as follows.First of all,in order to meet the different reconstruction requirements of flat,edge and corner areas in the image,and taking account of the different reconstruction difficulties of the three components,this thesis implies an image feature area decomposition module,which contains three characteristic-attentive blocks(CABs),and respectively and successively reconstruct the flat,edge and corner areas of the image.In each CAB,the corner detection algorithm is used to extract the corresponding region from the high-resolution image,drive the hourglass model to learn the characteristic-attentive mask,and apply it to the intermediate rough SR image learned by the hourglass model.Secondly,based on the fact that the phase component of the Fourier transform spectrum reflects the shape and contour information of the image,and the amplitude component reflects the color information,this thesis proposes a frequency decompose block(FDB).After Fourier transform of the hourglass output intermediate SR image to the frequency domain,it is decomposed into phase part and amplitude part,and trained separately with the corresponding part of high-resolution image as contrast.The four blocks are cascaded based on the hourglass model,forming a progressive example,and finally aggregated to produce the final superresolution image.This whole network is the divide-and-conquer super-resolution model based on image component decomposition proposed in this thesis.Finally,in order to further improve the effect of supersampling,this thesis explored the influence of the size of receptive field on image-related tasks,and speculated that it would also improve the supersampling task to a certain extent.This thesis proposed a global upsampling method based on linear attention mechanism to replace the traditional upsampling method.It can make every pixel in the image interact with the whole image.Compared with the traditional upsampling method,it can effectively increase the receptive field.The experimental results show that the divide-and-conquer SR model based on image component decomposition presented in this thesis shows excellent performance on the real-world datasets DReal SR and Real SR.The ablation experiment also proved that the improved receptive field brought by the global up-sampling module proposed in this thesis further promoted the improvement of the super resolution performance.