Research On Blind Super-Resolution Algorithm Based On Generative Adversarial Network

When it comes to Single Image Super-resolution(SISR)methods,most convolutional neural networks adopt synthesized Low-Resolution(LR)and High-Resolution(HR)image pairs to train the super-resolution network.However,in practical applications,there is a lack of paired HR and LR images for training.Traditional super-resolution networks’ performance will significantly decrease when there is a domain difference between the test and training images.To solve this problem,many unsupervised training methods have emerged,among which Generative Adversarial Networks(GAN)are widely used to address the problem of unpaired images through domain transfer techniques.However,directly transforming one domain to another remains very challenging.If the transformed domain has a significant difference from the target domain,the super-resolution network trained using pseudo paired images’ performance is also unsatisfactory.To address this problem,this paper proposes a new unsupervised blind super-resolution framework based on bidirectional domain transformation.This method reduces the difficulty of domain transformation by making two domains approach each other and converting their images to an adaptable intermediate domain as a transition for domain conversion,and builds a new training framework around the intermediate domain.The transformed intermediate domain images are then used as pseudo paired images with HR images to train the super-resolution network.The contributions of this paper include:(1)Propose the idea of exploring an intermediate domain between two domains,using two GAN networks to process images from both domains simultaneously.The two GAN networks use each other’s output image domain as their target domain,allowing the two domains to be transformed into an intermediate domain in a mutually close way.Then,use the paired images of the intermediate domain and the high-resolution(HR)images to train the super-resolution(SR)network.(2)A new blind super-resolution framework is designed around the intermediate domain.By utilizing the intermediate domain transition with the Cycle Generative Adversarial Network(CycleGAN)structure through cycle learning,the effectiveness of CycleGAN can be better exerted.The three networks can precisely support the use of the CycleGAN structure,without the need for an additional restoration network,The obtained cycle-consistency loss can also effectively constrain the three networks to stably train together.(3)Conducting additional ablation experiments to demonstrate the effectiveness of this paper’s method,and achieving good results on various competition datasets.This paper’s method uses an additional discriminator to demonstrate that the intermediate domain is between the HR and degraded domains,reducing the domain distance.