Super-resolution Research Based On Generative Adversarial Network

With the rapid development of computer software and hardware technology,image and video processing technology has also made great progress.The problem of super-resolution of images and video has always been a hot research issue in the field of computer vision,in video surveillance,remote sensing technology,military and medicine fields have extensive application prospects.Based on the given low-resolution images and video,algorithms are used to improve the super resolution of images and videos to obtain high-definition images and videos.On the one hand,it can save a lot of bandwidth,and on the other hand,it provides additional experience and better details in special application scenarios Traditional super-resolution algorithms are generally classified into three categories based on interpolation,reconstruction and learning.Although traditional methods have achieved a lot of achievements in this field in recent years,they can basically achieve the purpose of super-resolution quickly,but the image will appear blurred and smooth,resulting in noise and other issues,which will lead to the loss of some details of interest.Compared with the traditional method,the recently popular method based on deep learning has achieved good results,and can partially solve the problems that cannot be solved by traditional super-resolution,and gradually expands the research and application of deep learning in the field of super-resolution.In the field of deep-resolving super-resolution problems,there are known SRCNN,ESPCN,VESPCN,SRGAN and other algorithms.These image and video super-resolution methods have achieved very good results on some data sets.Previous studies have shown that deep-learning-based super-resolution methods have indeed been achieved in general image evaluation indicators PSNR,SSIM and human visual perception.Far beyond the performance of traditional methods.The main research content of this paper is to use GAN network for super-resolution research.We use the generative network to generate the image,and use the adversarial network to judge whether the image is network-generated or original,and through learning and training to make the network-generated image gradually approach the original high-definition image.In the first algorithm SRGAN,which uses GAN for super-resolution research,it is applied to the field of image super-resolution,followed by some related research.However,there are few researches on using GAN to do video super resolution.This paper uses GAN to study video super resolution.First of all,this paper uses GAN to study the super resolution of a single image.From the aspects of the internal structure of the GAN generation network and the loss function,the design of the GAN network is optimized,and better super-resolution results are obtained.Second,we propose a video super-resolution algorithm based on GAN.By introducing timing information and motion information in the video and combining it with the structure of the GAN network,a GAN video super-resolution structure based on motion alignment is designed,and a large number of experiments are performed on the data set.The results show that we propose The algorithm has achieved better results in images and video than previous algorithms.