Relativistic Super-Resolution Generative Adversarial Network With Stability Optimization

Single image super-resolution has received an increasing research attention within the computer vision community owing to its high practical value in restoring details and textures in image.Deep neural network can learn the complex feature extraction and transformation process from low-resolution image to high-resolution image,so in recent years,deep learning methods have been widely studied and applied to image super-resolution problem.Generative adversarial network is an important branch of deep learning.In recent years,generative adversarial network based super-resolution methods can help to synthesize more high-frequency details in the super-resolved images.This thesis studies the shortcomings of the current super-resolution generative adversarial networks,the main works are as follows:(1)Many researchers have proposed different super-resolution models by combining different network modules.This thesis does not focus on each method,but modularly reviews the framework of existing super-resolution models,and analyzes the effectiveness of each component and its limitations.Based on this,a post-sampling super-resolution model based on compact residual network is proposed and adopted as the generator network.unnecessary modules in the traditional residual block are removed to improve the generalization performance and save computing resources.(2)The training of generative adversarial network is unstable,existing super-resolution generative adversarial network is difficult for training and easily produce less meaningful visible artifacts.Although some studies have been carried out to stabilize GAN training,how to ensure that the mechanisms integrated into SISR models are trained correctly and play an active role remains a problem.This thesis studies the training stability of super-resolution generative adversarial network and propose adopting a gradient regularization term on current fake data to penalize the discriminator,thus overcoming the above problems and ensure the super-resolved images more realistic.(3)A relativistic super-resolution generative adversarial network with simplified gradient penalty(RSRGAN-SGP)is proposed for image super-resolution.The proposed method is modified with a relativistic discriminator,the relativistic discriminator will learn to distinguish “whether real image is more realistic than the generated fake counterpart” rather than “whether one image is real or fake”.The experiments show that relative discriminator can guide generator to produce results with richer details and sharper edges.The proposed method is trained by a fusion of pixel loss,VGG perception loss,and adversarial loss,which ultimately make the super-resolved images have a more natural and realistic visual perception.The proposed method is performed on a series of benchmarks to evaluate each component's effectiveness and compared with existing methods,the experiments show that the proposed method produces the results with richer and more realistic details,outperforming related works in terms of visual perception.Except that,the proposed method is with smaller number of parameters,easy to train and possesses good generalization capability.