Research On Key Technologies Of Blurred Image Super-resolution In Real Scenes

Single Image Super-Resolution(SISR)is defined as enlarging the spatial resolution of images,which is one of the classic low-level visual tasks.SISR has a wide range of practical application demands in public safety,medical care,remote sensing,and mobile internet.With the advent of the intelligent age,image data has become an important channel for obtaining information.Hundreds of millions of image data are generated in different devices and different scenarios.Due to the influence of lowresolution cameras,compression for storage,high magnification and other factors,a large number of images have problems of low resolution,blur,and noise.How to enhance the quality of low-resolution photos in different scenes and restore the rich details of natural images has important research and application value.This paper analyzes the advantages and disadvantages of recent representative SR algorithms.The SR problem has two main research difficulties: high-resolution reconstruction and degradation modeling.Most existing SR methods train SR networks under known degradation assumptions,and it is difficult to guarantee the performance on blurred image in real scenes.This paper explores the design of SR reconstruction network structure and the modeling of blurred image degradation in real scenes.The specific work of this paper is as follows:1.Aiming at the problem of poor recovery of low-frequency features in the image super-resolution process,a novel context-aware residual Super-Resolution network(CASR)is proposed.The existing SR methods process all channels and regions indiscriminately when recovering low-frequency pixel features,and do not model the relationship between the local and the context.The CASR combines the in-novatively designed fusion attention residual structure and enhanced gating structure,which can integrate contextual information more efficiently.Among them,the fusion attention residual structure is based on the channel attention and spatial attention mechanisms to enhance the information representation ability,which can efficiently describe the local context and model important low-frequency features.In addition,the enhanced gating structure proposed in this paper is used to modulate the information flow between multi-level features,further filter redundant information and enhance high-contribution features.This paper verifies the rationality and effectiveness of CASR through experiments on five datasets.The CASR network improves the performance of the residual network by restoring natural and clear image details with a lightweight structure.2.To address the problem of image domain shift when the image super-resolution algorithm is applied to real scenes,a novel Frequency Consistent Adaptation method(FCA)is proposed.This paper finds that most of the existing SR methods are based on degraded prior knowledge to construct low-resolution images,and the trained models often have better reconstruction effects on the training data.However,since the image degradation method and prior knowledge in real scenes are not necessarily the same,it is difficult for existing models to achieve the expected results when tested in real scenes.The FCA proposed in this paper estimates the SR degradation kernel from unsupervised source-domain images,and then generates low-resolution images with the same frequency density as the source-domain images,thereby providing training sample pairs consistent in the image domain for the SR model.In FCA,a Frequency Density Comparator(FDC)is proposed,which is used to provide the generator with a frequency consistency loss.The generator receives the gradient information fed back by the FDC,and learns to estimate the degradation kernel in the source domain image.In this paper,experiments on synthetic datasets and real datasets verify the improvement effect of FCA on the SR model.Compared with the latest methods of solving blind super-resolution,FCA shows obvious advantages in both quantitative metrics and visual effects.