The Research Of Image Super-resolution Method Based On Generative Adversarial Network

Single image super-resolution(SISR)is a fundamental low-level vision task,aiming to estimate a high-resolution image from its low-resolution counterpart.Contrast to low-resolution images,high-resolution images provide us with rich details which can not only help understand images thoroughly but also are helpful to other computer vision tasks,such as object detection,object tracking and etc.Therefore,SISR has attracted increasing attention in the research community.With the success of deep convolution neural network(CNN)in computer vision,various CNN-based super-resolution models have been proposed and continuously improved the super-resolution performance.Most current super-resolution models assume that the degradation is an ideal process,thus they can get low-resolution counterparts of high-resolution images in this way,and models are trained on paired low-and high-resolution images in supervised way.However,it's impractical when we want to super-resolve low-resolution images in the real world.On the one hand,the degradation of images in the real world isn't an ideal process and there are many other complicated degradation factors,such as blur,noise and etc.On the other hand,it's hard to get ground truth high-resolution images corresponding to the low-resolution images,thus models mentioned aboved can't be trained in supervised way.To tackle these challeges,we are aiming to accomplish image super-resolution task in the real world.Our works can be summarized as:(1)This paper proposed a Featured-Guided Super-Resolution Generative Adversarial Network(FG-SRGAN).Contrast to current super-resolution models,FG-SRGAN doesn't need paired low-and high-resolution images,because it's trained in unsupervised way.Morever,FG-SRGAN doesn't assume that degradation is an ideal process.To accomplish one-to-one mapping from low-resolution to high-resolution,a guidance module is prposed to guide the super-resolution results of FG-SRGAN.Furthermore,to avoid that guidance module may have a decisive influence on FG-SRGAN's performance,the results of guidance module are treated as fake samples while training,which has been proved effective on FG-SRGAN's performance.(2)This paper proposed a CycleGAN-based super-resolution model,it doesn't assume that degradation is an ideal process,too.There are not only super-resolution network in proposed model,but also degradation network.Super-resolution network learns mapping from lowresolution images to high-resolution images by adversarial training,and degradation network learns mapping from high-resolution images to low-resolution images by adversarial training.It's notable that we can model degradation of images by degradation network.Contrast to ideal degradation process,the proposed model takes various other factors into consideration,such as blur,noise.To accomplish one-to-one mapping from low-resolution to high-resolution,we apply cycle consistence to the proposed model.What's more,the proposed model is trained in unsupervised way.Experiments on real-world face images show that our super-resolution models mentioned above have achieved satisfactory results visually.Furthermore,quantitive comparasions with other relative super-resolution models illustrate improvements obtained by our models.