Super-resolution Reconstruction Of Images Based On Generative Adversarial Networks

Image super-resolution reconstruction is an important research direction in the field of image processing,with the aim of reconstructing clear images from blurred inputs.With the increasing importance of high-resolution images in people’s lives and studies,the demand for image resolution has become higher and higher in various fields.Currently,it is possible to improve the resolution of images through software or hardware methods,but hardware methods bring about high costs,technical risks,and other issues,and software methods still have problems such as insufficient improvement in image details and instability of models.Therefore,research on image super-resolution reconstruction still has great significance.This thesis mainly proposes an improved image super-resolution reconstruction algorithm that addresses the problems of insufficient image detail reconstruction and training stability in existing algorithms.The main optimization of the algorithm is conducted through two parts:The first part optimizes the generation model of the generative adversarial network(GAN).(1)Using the idea of dense network,dense blocks are used instead of residual blocks to obtain more image information,improving the reconstruction quality of images;(2)Replacing the PRelu activation function in the GAN with the Relu activation function;(3)Removing all batch normalization layers from the generation network to reduce computational complexity and improve learning efficiency.The second part optimizes the distinguisher model of the GAN.To address the problem of unstable training of the GAN,this thesisr removes the sigmoid layer from the distinguisher network and introduces the WGAN idea.Using Wasserstein distance,the generated samples and real sample distributions can be well reflected,alleviate the problem of gradient disappearance during training,and improve the stability of model training.Finally,the trained model is tested on the Set5,Set14,BSD100,Siri-WHU-PARK,and DIV-2K-512 data sets for image super-resolution reconstruction,and the experimental results are compared with the SRGAN algorithm and other classic algorithms.The experimental results show that the proposed algorithm can generate clearer reconstruction images,improving the visual effect of images.In objective evaluation indicators,the average peak signal-to-noise ratio(PSNR)of the images tested by the proposed algorithm is higher than that of the SRGAN algorithm,and the average structural similarity index(SSIM)is also improved.