Research On Image Super-resolution Method Based On Generative Adversarial Networ

Image is one of the important ways of modern human information,its resolution size are the important factors influencing the image with the amount of information,the resolution of the image size specific said how many pixels in the image,high resolution image can carry more useful information,and the low resolution image increasingly unable to meet the needs of people for image information,Therefore,image super-resolution reconstruction has always been a hot topic in contemporary research,and it plays an important role in medical images,remote sensing images and other fields.The research on image super-resolution reconstruction first began in the 1960 s.With the development of science and technology,interpolation,reconstruction and learning-based methods have emerged successively.This paper makes an improvement on the super-resolution method of generative adversarial network based on deep learning,and mainly does the following work:1.Based on generative adversace network,corresponding improvement schemes are proposed to solve the problems of model collapse and insufficient feature extraction.First,combined with the Wasserstein Generative Adversarial Networks proposed by WGAN,Jen-shannon divergence is used to improve the network training process instead of generating adversarial network to measure data distribution.Second,referring to the dense connection network theory,a residual dense connection module is designed to improve the network’s ability to extract features,speed up the information flow between deep networks,and make the network have more resources to learn features.Through simulation and experiment,it is proved that the peak signal-to-noise ratio and structure similarity of objective evaluation index are improved to different degrees.2.Integrate the Attention mechanism into the field of super-resolution reconstruction.Firstly,aiming at the problems of insufficient high frequency information of reconstructed images and insufficient details,the Residual Attention Block(RAB)extracted high frequency information of reconstructed images was proposed to realize the right allocation of network resources to high and low frequency information of reconstructed images.Enhance the learning ability of network for image high frequency information.Secondly,L1 loss is used as network loss function to make the reconstructed image more realistic and soft.The experimental results on the test set show that the improved method not only improves the objective evaluation index peak signal-to-noise ratio and structure similarity,but also improves the subjective visual effect of reconstructed images,which makes the image more detailed and achieves better reconstruction effect.3.In view of the problems of complex practical operation and insufficient practical significance of the algorithm model,this paper designs and implements an image super-resolution system with simple operation,which mainly includes user operation module,model call module and data management module.Through the Web visual interface,users can easily and intuitively complete the super-resolution reconstruction of low-resolution images.