Research On Super Resolution Reconstruction Algorithm Of Remote Sensing Image Based On GAN

The demand for satellite remote sensing data analysis is increasing in various sectors of society.Due to the complex imaging process of remote sensing image and many uncontrollable factors,the resolution of directly collected remote sensing image may not meet the demand in practical application.It is of great scientific significance and research value to optimize remote sensing image resolution by means of data information processing.According to the algorithm of image super resolution based on generative adversarial network,this thesis analyzes the edge information,non-local features and artifact discrimination,and proposes a super resolution algorithm of remote sensing image combining edge enhancement and non-local modules.Aiming at the problems of serious noise pollution and edge blurring in remote sensing image imaging,a super-resolution algorithm of remote sensing image Edgeenhanced Generative Adversarial Network(EGAN)based on edge enhancement network is proposed.Taking Deep Blind Super-Resolution GAN(BSRGAN)as baseline model,an edge enhancement module is introduced.In order to enlarge the receptive field of the model,the Mask branch in the edge enhancement structure was further optimized.Image consistency loss is introduced to guide edge reconstruction,and subpixel convolution is used to replace nearest neighbor upsampling to obtain clearer edge contour and more consistent style results.Aiming at the problems of low global information utilization rate of remote sensing image in feature extraction stage and artifacts in reconstructed image,A super resolution reconstruction method Non-local Module and Artifact Discrimination EGAN(ENDGAN)for remote sensing images based on non-local module is proposed.Based on EGAN,ENDGAN introduces non-local modules in the feature extraction stage of the model to better utilize the intrinsic correlation of remote sensing images and enhance the ability of the algorithm to extract global target features.At the same time,an artifact discrimination method is used to distinguish artifacts and real details of reconstructed images,and the artifact discrimination loss is introduced on the basis of the original loss to optimize the model update strategy.In this thesis,multiple algorithms were compared on two remote sensing image data sets,NWPU VHR-10 and UCAS-AOD,and the experimental results showed that multiple evaluation indexes of the proposed method were improved.Taking degenerate type Ⅳ as an example,compared with the baseline model,the SSIM,PSNR and RMSE of quadruple super resolution images are increased by 0.074 d B,increased by 1.807 d B and decreased by 13.02%.In addition,the remote sensing image target detection results are used to further evaluate the effectiveness of the proposed method.The experimental results show that the reconstructed remote sensing image is easier to detect ground targets than the original image.