Learning A Deep Residual Network For JPEG Image Super-Resolution

With the development and popularization of information technology,the number of network pictures has gradually increased.The demand for image quality is getting higher and higher.Due to the limitation of network bandwidth and storage space,low resolution and lossy compression images are often used in network transmission.This requires the restoration of high resolution clear images from the low resolution compressed images.Because most of the image super-resolution methods are based on the assumption that low resolution images are lossless compression,the results cause the JPEG compression artifacts to deteriorate during super-resolution,resulting in bad visual effects.This paper firstly analyzes and summarizes the image super-resolution methods and image quality evaluation standards at home and abroad.Then,on the basis of this,the method of this paper and the improved method are put forward.(1)A common assumption for JPEG degraded images is that HR images are sampled first and then JPEG compression with a certain quality factor is carried out.Therefore,an intuitionistic solution is to first restore the clear LR image by JPEG artifact removal algorithm,and then carry out super-resolution processing on the restored image.Through experimental comparison,it finds that two cascaded networks can achieve the effect of super-resolution of JPEG images,and the effect of TNRD-VDSR is better than that of AR-SRCNN,but there is still a subtle artifact in the near look.In the analysis of the reasons,we remove the compression artifacts from the JPEG image by cascading network,and then amplify,although it has achieved good results,but because of the lack of end to end training,the connected pipeline will introduce cumulative error,so people can't be satisfied with this result.(2)Aiming at the problems of cascaded networks,this article design a deep residual convolution neural network.The network consists of two subnets,forming a cascade structure.They are used to remove JPEG compression artifacts,image super resolution and information integration respectively.The network shares the convolution feature by using the skip connection between the two modules to improve the network performance.The network can achieve end to end joint learning to avoid pipeline errors.Finally,we implement a robust quality factor model for different quality factor training models.Experiments show that our method can achieve satisfactory results in both the artifact removal task and the JPEG image super-resolution task.