Research On Image Super-Resolution Reconstruction Based On Deep Learning

Most deep learning-based image super-resolution methods need to be trained on a dataset consisting of a large number of paired low-resolution images(Low-Resolution,LR)and highresolution images(High-Resolution,HR).In practice,most methods use a simple bicubic downsampling strategy to obtain such training images.However,bicubic downsampling strategy will cause the generated images losing a lot of frequency details.Secondly,because the blur kernel and noise of real-world images are more complex or even unknown,these methods are difficult to generalize to real-world image super-resolution reconstruction.In this thesis,we propose a novel deep learning-based SR method.The proposed framework consists of a degradation framework,paired domain correction network and an SR network.The main work and innovations of this thesis are as follows:Firstly,the degradation framework estimates real blur kernels and real noise distributions from real-world images.Based on the degradation framework,the generated realistic LR images can reflect the distribution of real-world images.Secondly,the domain correction network removes noise and adjusts the blur kernel of the inputted LR image.Those generated realistic LR images will be input to the domain correction network and make the output corrected LR images closer to the LR images obtained by bicubic downsampling.Finally,we integrate an existing SR network.The SR network is trained on these corrected LR images and their corresponding HR images in a supervised manner.,which makes it can use the pre-trained model.The result is that we get high-quality reconstructed images and save training time.Finally,the SR network is an independent part,so the existing well-trained existing superresolution networks can be integrated into our proposed method,which can make better use of their research findings.The experimental results on the NTIRE2020 dataset show that the proposed method can not only effectively improve the visual perception quality of the reconstructed image compared with the mainstream models,whether on the traditional synthetic dataset or real-world images.super-resolution reconstruction.Moreover,it also improves the SSIM(Structural Similarity Index,SSIM),PSNR(Peak Signal-to-Noise Ratio,PSNR)and LPIPS(Learned Perceptual Image Patch Similarity,LPIPS),reaching 0.9098,24.97dB and 0.3096 respectively.Especially in the real-world image super-resolution reconstruction task,our method has a large performance advantage.