Research On Image Super-Resolution Based On Generative Adversarial Networks

With the development of image optimization and other technologies,image super-resolution reconstruction has become a hot research direction,which is widely used in medical,military,aerospace and other fields.The traditional super-resolution algorithm can be divided into interpolation,reconstruction and learning,but the traditional method has some disadvantages,such as fuzzy edge,missing details,too long reconstruction time and so on.At present,the mainstream method is super-resolution reconstruction based on deep learning.Its optimization goal is to achieve the balance between performance and reconstruction quality,that is,to reduce processing time and improve reconstruction quality.The effect is much better than the traditional method.In the image super-resolution method based on deep learning,srcnn is a convolutional neural network with three convolution layers,which is also the first work.Fsrcnn is improved on the basis of srcnn,espcn uses sub-pixel convolution layer to realize super-resolution of single image and video,VDSR is a deeper reconstruction method based on residual network.Before theappearance of image super-resolution method srgan against neural network,the follow-up methods are mostly the deepening of network structure,and srgan is to improve the resolution of super-resolution problem to a new height.Under the requirement of 4-power magnification,the image of the traditional method will be smoother and lack of details,while srgan can reconstruct details and make the image more realistic.However,in the network structure of srgan,there are many residual blocks in the generated network,the internal structure of the residual blocks is complex,and the traditional convolution method is used,which results in a large amount of computation,so it is difficult to improve the reconstruction speed under the condition of ensuring the reconstruction quality,especially under the condition of limited computing power on the mobile end,the reconstruction quality is greatly affected by the computing power.In view of the low reconstruction quality of srgan on devices with limited computing power,this paper proposes an improved model,shufflenetsrgan,to change the network structure and reduce the computational complexity.The specific research is as follows:(1)A shufflenetsrgan model is proposed by replacing theRESNET residual block in srgan generation network with the shufflenet cell structure.In the experiment,the training data image is processed by different scale fuzzy preprocessing and then transmitted to the generating network to increase the robustness of the model.Through the experiments in this paper,we know that the shufflenetsrgan with 13 layers can achieve better reconstruction quality in less time.(2)Ressrcnn,a limited model for generating abnormal texture in shufflenetsrgan model,is proposed.The experimental results show that the reconstruction effect is 10% higher than that of srcnn,and the abnormal texture suppression effect on shufflenetsrgan is also well completed.The effect of image reconstruction is improved obviously.obviously.