The Research Of Image Super-resolution Reconstruction Algorithm Based On Adversarial Learning

Image super-resolution technology is an approach that could transform lowresolution images into high-resolution images at a specific magnification factor.Due to its ability to restore true detail information in low resolution images,it has been widely used in lots of fields,including satellite remote sensing,security surveillance video,medical diagnostics,etc.,and this technology can also provide a research base for sophisticated computer vision tasks like image classification,scene detection and segmentation,and recognition detection.However,in current research,many image super-resolution algorithms still have defects such as detail blurring,false artifacts,and are not suitable for complex degraded images in real scenes.The focus of this thesis is to enhance the visual quality of super-resolution reconstructed images by leveraging adversarial learning techniques in classical and blind super-resolution algorithms The main research work of this thesis is as follows:For the classical image super-resolution task,a region adversarial learning and backprojection based image super-resolution reconstruction algorithm is explored to address the issue of detail blurring and artifacts in the output super-resolution images of the generative adversarial network-based reconstruction model.The algorithm distinguishes between smooth regions and texture detail regions in the image,and uses adversarial learning to repair the texture-rich regions,while only using the generator for reconstruction in the smooth regions,thereby alleviating the problem of false artifacts produced by the reconstruction model.In order to achieve better reconstruction results,this thesis utilizes channel attention mechanisms to optimize the network structures of both the generator and discriminator.This approach emphasizes more informative features and,in conjunction with back projection constraints,further refines the loss function in the training process to restore true details in the reconstructed images.Finally,the effectiveness of the proposed algorithm is validated through super-resolution experiments on multiple benchmark image datasets,as conducted in this thesis.To enhance the applicability of image super-resolution reconstruction algorithms in complex real-world scenarios,this thesis explores an image blind super-resolution reconstruction algorithm that is based on implicit modeling techniques.To overcome the problem of insufficient prior information in implicit modeling methods,this thesis designs a VQVAE-based image prior information learning model,which continuously extracts rich detail features of the original high-resolution image as feature vectors in the discrete codebook during the training process,and then trains the decoder to reconstruct the original high-resolution image.The resulting discrete codebook and decoder are used as prior information for the blind super-resolution reconstruction model.To more effectively extract feature information from real degraded images,this thesis uses a more powerful encoder structure and shares the effective prior features in the discrete codebook for shared mapping,then inputs them into the trained decoder and U-Net discriminator for adversarial training to achieve super-resolution reconstruction of degraded images in real scenes.Finally,extensive experiments on multiple benchmark datasets and real scene datasets demonstrate that this method can improve the super-resolution reconstruction effect of degraded images in real scenes.