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

Because hardware devices such as cameras are not easy to update,software upgrades are relatively easy and portable.Therefore,it is a research hotspot to use the Super-Resolution technology to improve the resolution of the original image through the software method.Image super-resolution reconstruction refers to adding a high-frequency detail information to a given low-resolution image through a specific algorithm,improving spatial resolution and clarity,and generating a corresponding high-resolution image.The image super-resolution algorithm based on deep learning can solve the problems of traditional algorithms such as large calculation amount,poor robustness,edge blur,loss of details,and low quality of generated images.However,there are still situations such as unstable network training and gradient disappearance.Images with rich and complex textures are prone to blurry and distorted images when over-division.In view of the above problems,this paper studies the application of deep learning methods such as convolutional neural networks and generative confrontation networks in the field of image super-resolution.First,a super-resolution reconstruction model based on dense residual jumper network is proposed,and then combined with generative confrontation Network,a new image super-resolution algorithm based on generative adversarial network is proposed.The main work of this thesis are as follows:(1)Using the characteristics of residual network and dense network,a super-resolution algorithm based on Super-Resolution via Residual Dense connected network(SRRDCN)is proposed.The residual dense unit of the network is formed by cascading multiple residual dense blocks,extracting high-level abstract features with richer semantic information;introducing global shortcut connections,fusing shallow and deep features together,from the original LR Obtain global dense features in the image.The local fast connection of the dense residual block combines local low-level features and local high-level features to increase the flow of information and further improve the network's presentation ability.The experimental results show that the SRRDCN algorithm has a clearer visual effect than the images reconstructed by some common algorithms on the three magnification factors of the four test sets,and has a greater improvement in the objective evaluation index.(2)Combining the advantages of dense residual jumper network and relative discriminator architecture,a new image super-resolution algorithm based on generative adversarial network(Novel Super Resolution Generative Adversarial Network,NSRGAN)is proposed based on generative adversarial network.In order to solve the problem that the image super-resolution reconstruction algorithm based on generative adversarial network can generate realistic textures and produce unpleasant artifacts at the same time,the algorithm has been improved in three aspects.The improvement of the generated network is to remove the batch normalization layer and combine the dense residual network to improve the expression ability of the network.Obtain the perceptual loss before activating the function to optimize the perceptual loss and make the reconstructed picture similar to the real picture in brightness.Experimental results show that the algorithm produces clearer image details,more realistic textures,and no artifacts.