| Despite the use of faster and deeper convolutional neural networks,the image super-resolution reconstruction technology has made breakthroughs in the accuracy and speed of single-image super-resolution.But many problems still remain unresolved.Firstly,how to refine the texture problem when performing super-resolution processing under the magnification ratio;Secondly,the existing image super-resolution algorithms of convolutional neural networks are prone to problems of overfitting and insufficient convergence of the loss function.In this context,this paper combines superresolution algorithm and generative adversarial network theory to design an Enhanced Spatial Feature Transformation(ESFT)layer and a superresolution algorithm based on generative adversarial network SRCICGAN,and through PSNR,SSIM.Other indicators are compared with other algorithms.The specific work of this paper is as follows:(1)Aiming at the problem of effective expression of image prior information,an enhanced spatial feature transformation(ESFT)method is proposed.The spatial feature transformation layer is based on the ASSP semantic segmentation probability map.On this basis,the spatial feature transformation layer generates a pair of modulation parameters to spatially transform the feature map on the network.Add this method to SRGAN to change its network structure and conduct a test.The comparison between the experimental results and other methods proves that the method reconstructs the texture of the image and restores the fine grain.It can help to obtain and reconstruct high-quality and high-definition images with more natural and realistic textures.(2)This paper adopts the training method of unsupervised learning,uses the proposed ESFT method,combines Cycle GAN and SRGAN,improves the network structure and optimizes the function,and then proposes the SRCICGAN algorithm,which is used to recover the quadruple down sampled image.The use of ESFT helps to restore the detailed texture of the image;the use of dense residual blocks as the basic structural unit effectively avoids the problem of overfitting;and the improvement of the loss function improves the convergence of the loss function.Experiments prove that SRCICGAN is superior to SRCNN,EDSR,RCAN,Enhance Net,SRGAN,and ESFT-GAN in PSNR and SSIM indicators.In the Flickr2 K data set,the SRCICGAN algorithm is numerically superior to the RCAN algorithm with the highest PSNR and SSIM values among other algorithms.Respectively 1.92% and 5.49% higher,and can also get better visual effects in detail texture.It is verified that the SRCICGAN algorithm can better complete the image super-resolution task.Finally,experiments have proved that the SRCICGAN algorithm has better super-resolution reconstruction effects from the subjective visual evaluation of human eye comfort,and from the objective evaluation of PSNR and SSIM indicators. |
