Research On Image Super-resolution Reconstruction Algorithm Based On Generative Adversarial Networks

Super Resolution Image Reconstruction is a hot issue in the field of image research in recent years.From some traditional algorithms based on interpolation to algorithms based on deep learning,the quality of generated SR images are getting better and better,but the texture details are not rich enough.The algorithm of SRGAN applies the Generative Adversarial Nets to image super-resolution reconstruction for the first time,and generates the SR images with better quality and richer texture details.ESRGAN improves SRGAN by introducing RRDB network architecture units without the batch normalization layers,so that the generated images have better visual quality and more realistic and natural textures.However,the training process of ESRGAN is difficult to supervise,and the edges of the generated images will be geometrically deformed.This article is based on the algorithm of SRGAN and ESRGAN to improve.First,we conduct research on deep convolutional neural networks and generative adversarial networks.Then we studied the basic principles and network architecture of the SRGAN.We also introduced the evaluation index of image super-resolution reconstruction;Then we studied the channel attention mechanism network.We also added it to the RRDB basic block to get a new module CA-RRDB.Its advantage is that it allows us to pay more attention to important features.So we can extract more useful information and features.Regarding the problem that the generated image will be geometrically deformed,we introduce a gradient guidance branch to generate a super-resolution gradient map.This can make the generated SR image geometrically consistent with the real picture.Besides,we give the loss based on WGAN,which can better supervise the training process.The experimental results show that the PSNR and SSIM of the image generated by Algorithm 1 are higher than those of the ESRGAN.Finally we improve the Algorithm 1 by adding a multi-scale receptive field module to the generative network.It can extract different texture details through multi-scale convolution.In order to eliminate the impact of Lipschitz condition restrictions in WGAN,we use the loss function based on WGAN-GP.After that,we use fusion of network parameters to eliminate the unpleasant noise generated by GAN.It also maintains good visual perception quality.The experimental results show that Algorithm 2 have higher PSNR and SSIM for the generated images than Algorithm 1.