A Research Of Image Super-resolution Algorithm Based On Adversarial Neural Network

With the development of technologies and the improvement of people's living stan-dards,people have higher requirements for the visual quality of digital streaming videos and images.The resolution of digital TVs has been upgraded to 4K or even 8K from 720P and 1080P.The low-resolution images in real life are far from satisfying people's needs.Although great progresses have been made in super-resolution reconstruction technology of images in recent years,it can basically resemble images reconstructed in super-resolution,the quality is still not good enough.There will be artifacts in images and the smudging sense of images is serious.The details are distorted and visual effects are not clear enough.To solve the problems above,this paper introduces an adversarial neural network into the super-resolution reconstruction task of images,and designs a new image super-resolution reconstruction model using a new value function.The main content is as fol-lows:(1)This paper first analyzes and implements several typical image super-resolution reconstruction algorithms in detail for comparisons.At the same time,the paper introduces the model structure and principle of the adversarial neural network.The principles and network structures of generative network and adversarial network are also explained in detail.This paper not only verifies the principle of the generative adversarial network is proved based on the elaboration of the adversarial neural network,bus also analyzes the defects existing in the value function of the original GAN network.Then,the paper proves that the value function in WGAN has been improved compared with the value function of the original GAN.(2)This paper explores the super-resolution reconstruction tasks in the frequency domain based on the adversarial neural network.As for the data set,this paper first uses original data set image segmentations of 2K resolutions to make about 60,000 high-and low-resolution images as well as images in the corresponding frequency domain,which are used to train the super-resolution reconstruction model in frequency domain.Although the experimental results of the frequency domain model are not good,the experiment proves that each point contains frequency domain images with strict image information,which cannot fit well with the spatial domain generative model based on pixels.(3)To solve the problems of blurred visual effects of super-resolution reconstructed images at the present stage and the relationship between image visual effects and image sharpening,this paper proposes a Dual-SRGAN model.The model uses the residual dense module and the skip residual connection in the generative network.Because most super-resolution reconstruction models mainly pursue shapes similar to reconstructed images,they lack control over the detailed information of images.In this paper,a siamese adver-sarial neural network is used in the Dual-SRGAN model.On one hand,it can perform the conventional super-resolution reconstruction of images;on the other hand,the network focuses on controlling the details of images and uses the idea similar to contrast mask-ing USM sharpening to fuse ordinary and detail images at the output port,thus achieving better visual effects of super-resolution reconstructed images.In order to test the image reconstruction performance of the Dual-SRGAN model,three commonly used data sets in the field of super-resolution reconstruction are used in this article:Set5,Set14 and BSD100 to verify the model.In the Set5 data set,the PSNR/SSIM of the model in this paper reached 32.24/0.8960,higher than the ESPCN model of 30.76/0.8784 and higher than the SRGAN model of 29.40/0.8472.In the BSD100 data set,the PSNR/SSIM of the model in this paper reaches 28.13/0.7827,which is higher than the 27.02/0.7442 of the ESPCN model and 25.16/0.6688 of the SRGAN model.The experimental results of the model can be seen that the peak signal-to-noise ratio and structural similarity of the reconstructed image have a good improvement.Compared with the rest of the super-resolution algorithms,the images reconstructed by the super-resolution model in this paper are more detailed,the visual contrast is improved better,and the visual effect is better.