Research On Remote Sensing Image Super Resolution Reconstruction Technology Based On Deep Learning

Super-resolution(SR)is the technic which reconstructs a low-resolution image into a high-resolution one with the help of software or arithmetic operation.In recent years,the reconstruction of image super-resolution based on deep learning has experienced rapid development,meanwhile,with advantages of low economic cost and easy deployment,SR networks based on deep learning have become the effective method to improve image resolution.However,compared with nature images,remote sensing imageries(RSI)contain more information,and the relationship between objects is more complex.Therefore,networks which are specifically designed for SR of nature images also encounters problems involving unstable training,bad convergence,spectral distortion,and edge information loss when they are used in SR of RSI with no change of the structure.To handle these problems,this paper proposes a generative adversarial network with high attention of spectral information and edge information(EAGAN)for SR of RSI.The main researches done in this paper are listed as followings:(1)In this paper,structures of three typical image super-resolution reconstruction based on deep learning are analyzed,and a composite loss function is designed to handle their existing problems involving in training oscillation with non-convergence and the local optimum.The composite loss function contains two main improvements:First,the improved adversarial loss function.The traditional adversarial loss function based on Jensen-Shannon divergence has a severe problem of vanishing gradient,thus,the improved adversarial loss function adopts the modified Wasserstein distance as the new evaluation index of reconstructed results.In that,the max distance between generative network and discriminate network is set to 2 rather than infinity,which help network to reduce training time.Second,the improved spectral loss function.The improved spectral loss function introduces the empirical distribution of pixels as a weighting factor which adaptively weights the reconstructed pixel points according to the pixel occupancy ratio.Which helps network to overcome the problem of generally low saturation of reconstruction results caused by the uneven overall color distribution.,while ensuring that the network can generate reconstructed images with multiple color families and high saturation.(2)In order to solve the problems of spectral distortion and contour information loss that may occur in reconstructed images,the spectral attention focus and edge attention focus mechanism are designed in this paper embedded in the image generation network.First,the spectral attention mechanism generates the spectral weight map for each pixel point through an adaptive low-pass filtering calculation,and then updates values with the convolution operation,so that the network can correctly grasp the overall spectral distribution of the image while removing unnecessary noise information.Second,the edge attention mechanism obtains the boundary and contour information of the feature image by Laplace operation,and normalizes it as the boundary weight value to enhance the boundary on the feature map,thus further enriching the edge feature details of the reconstructed image.Finally,this paper conducts comparison experiments on AID dataset,Kaggle open source dataset,and Jilin-1 optical satellite image with three typical SR networks including fast superresolution convolutional neural network(FSRCNN),VDSR and super-resolution adversarial neural network(SRGAN).The results show that EAGAN is effective to different kinds of RSI,and the super-resolution reconstructed images generated by EAGAN contains more realistic spectral information and richer boundary features,which cannot be achieved by the other counterpointed networks.