High Quality Image Restoration Based On GAN

High-quality images refer to the images with rich and high-frequency information in them,such as clear textures,bright colors,etc.This high-frequency information can not only improve the accuracy of computer vision tasks,such as face recognition,target detection and so on,but also give people a good visual enjoyment.In classic visual tasks,the ways to obtain high-quality images are to recover them from the low-quality images,usually by using single image super-resolution and image deblurring.Single image super-resolution aims to restore images to high-resolution from the ones in low-resolution.It uses mathematical models to calculate and complete the lost pixels of high-resolution images,which express abundant details.For image deblurring,it uses appropriate mathematical model to restore the blurry images into the clear ones.Due to the rapid development of deep learning in recent years,deep neural network has made great progress in the field of image restoration.Compared to traditional methods,the advantages of deep neural network are that it can restore the low-quality images in an end-to-end manner and achieve better results.Currently,single image super-resolution and image deblurring based on deep learning can achieve a higher PSNR and SSIM score compared to traditional methods.In reality,however,the low-quality images are not limited to just one case and can be more complex.For example,a face or car captured on a roadside camera will suffer from low-resolution and motion blurred.In this case,it is difficult to obtain a high-quality image with a image super-resolution model or image deblurring model.In addition,PSNR and SSIM are mostly used evaluation indexes in image restoration,which can guide the model to generate less distorted images.However,many high-frequency details of images will lose,which not only cause the images look unreal,but also affect their performance in computer vision task.To this end,we propose an end-to-end generative adversarial network which can restore a clear image in high-resolution from a low-resolution with motion blurred or Gaussian blurred,denoted as P~2GAN.The generator of P~2GAN contains deblurring and super-resolution modules.In the deblurring module,we introduce an asymmetric residual encoder-decoder architecture(A-RED),which enlarges the receptive field by stacking residual in residual(RIR)structure with different numbers of convolutional filters,in order to address different sizes of motion blurry kernels.In particular,we extend RIR with channel attention block(CAB)to re-weight the deblurring channel features.Similarly,CAB is introduced into residual in residual dense(RIRD),denoted as RIRD-CA,to re-weight the channel-wise features from input images and deblurring modules in the super-resolution module.RIRD-CA can establish a continuous-memory mechanism which can preserve the features extracted from earlier layers along with the network.In particular,pixel-wise loss,adversarial loss,and contextual loss have been incorporated into P~2GAN from pixel-level and perceptual-level perspectives.Adversarial loss benefits to generating rich textures and high-frequency details,while contextual loss eliminates the unrealistic textures.Extensive experiments conducted on a public dataset show abundant high-frequency details and natural textures of the proposed method,outperforming the state-of-the-art approaches.