Improvement Image Super-resolution Method Based On Multi-scale Generative Adversarial Network

The digital age has long arrived,and images,as the main information dissemination medium,have been widely used in various scenarios.In many fields,people have high demand for image quality,such as medical image field,satellite remote sensing field,etc.In the rapidly developing information age,low-resolution images are already difficult to meet the needs of specific scenarios.Therefore,image super-resolution is of great significance in computer vision tasks.The image super-resolution method based on convolutional neural network greatly improves the performance of super-resolution,but many perceptual networks often have the problem of blurred image texture structure and lack of high-frequency information.Therefore,this paper proposes a generative adversarial network structure based on multiscale asynchronous learning,and uses the pyramid structure in the generator model to better integrate high-frequency information of different scales.This paper also adopts the U-net network as the discriminator to consider the global and local context of the input image,and makes some improvements to the loss,and finally obtains better highresolution images.The experimental results of this paper on benchmark datasets such as Set5,Set14,BSD100,and Sun Hays80 prove that the method has a good effect in recovering the detailed texture information of low-resolution images.The specific research contents of this paper are as follows:(1)In order to obtain a more accurate image scale in the field of super-resolution,this paper proposes a super-resolution method based on multi-scale generative adversarial networks.The algorithm combines the structures of many kinds of deep neural networks to obtain an asynchronous network model.(2)Before the data is input into the network,preprocessing training is performed,and in order to ensure the impact of the dataset on the network,relevant ablation experiments are performed to ensure the integrity of the algorithm.In order to enable the network to better extract feature information from low-resolution images,this paper designs a multiscale generative network to capture high-frequency information,optimize the input lowfrequency information,and enhance the network effect.(3)In order to solve the problem that the model level of the existing super-resolution methods cannot well match the feature perception,this paper designs the generative network model level as asynchronous networks of different levels,and adopts a better decoding that can maintain the image dimension.The encoder is used as the discriminator network structure,and finally an improved loss function is designed.(4)This paper discusses the image generation effect of the model through qualitative analysis.Using SSIM,PSNR and LPIPS as the evaluation criteria,the algorithm in this paper is compared with different algorithms on the same test set.Well,the texture and lines of the image are better reconstructed at large magnification factors.