Research On Image Restoration Algorithm Based On Generative Adversarial Networ

With the continuous progress of information technology,people’s concern and pursuit of image restoration technology is also increasing.In this paper,the digital image restoration algorithm based on generative adversarial network is studied with image restoration technology as the main research object.This paper firstly presents the main research contents and methods of this paper based on the subject of image restoration algorithms based on generative adversarial networks and the status of related research at home and abroad.Secondly,the technologies related to the topic,such as image restoration,deep learning,convolutional neural network,residual network and generative adversarial network,are analyzed respectively to provide theoretical and technical support for the full research.Following that,the ResGAN network is constructed based on the algorithm analysis,and the network is optimized at multiple scales.Finally,experimental design is carried out on the ResGAN network,and experimental simulation is performed based on the experimental data set and parameter settings.The experimental results show that the FID values of the ResGAN algorithm are 270.9and 184.4 in the small batch and 130.6 and 104.6 in the large batch on the Cifar10 and CelebA datasets.The FID values of the ResGAN algorithm are smaller than the other algorithms.In the same dataset,the IS values of the ResGAN algorithm are 1.64 and 1.54 in the small batch and 1.54 and 1.75 in the large batch.The IS values of the ResGAN algorithm are higher than the other algorithms.Meanwhile,in the gradient analysis,the residual generative adversarial network proposed in this paper converges faster in the fifth layer than the DCGAN network;through the back-propagation process of the gradient from the fifth layer to the first layer,it is found that the gradient of the residual generative adversarial network proposed in this paper vanishes more slowly.Therefore,compared with other algorithms,the stability of the residual generative adversarial network algorithm proposed in this paper is better,and a more satisfactory repair effect can be obtained.