Research On Deep-Learning Based Optical Remote Sensing Image Denoising And Super-Resolution Reconstructing Algorithm

The progress of remote sensing technology has greatly promoted the acquisition of optical remote sensing image.Optical remote sensing image has the very high research and the practical value,especially the high quality of optical remote sensing image,has the characteristics of high resolution,low noise,and can be widely used in agriculture,forestry and monitoring,urban planning,military reconnaissance,etc,thus improving the quality of optical remote sensing image method has a high research value and application prospect.However,the method of directly improving remote sensing image quality by upgrading hardware requires high cost,high technology level and long time cycle.Therefore,how to improve the quality of optical remote sensing image economically,conveniently and efficiently has become a major challenge in the field of remote sensing.To address this challenge,deep learning-based image processing algorithm is adopted in this dissertation to improve the quality of a single low-quality image by denoising and super-resolution reconstruction.In this dissertation,a number of studies on optical remote sensing image denoising and super-resolution reconstruction based on deep learning are carried out step by step,aiming to lay a theoretical and technical foundation for subsequent studies and accelerate the engineering application process of optical remote sensing image denoising and super-resolution algorithm based on deep learning technology.The main research results of this dissertation are as follows:(1)A space object image denoising and super-resolution reconstruction method based on deep convolutional neural network is proposed.This method mainly uses the idea of residual network and combines the local residual network and the global residual network to realize the reconstruction of the spatial target image with the super-resolution of 2×,3× and 4×,and at the same time can remove the cosmic ray noise.The experimental data show that this method improves the evaluation standards of Peak Signal-to-Noise Ratio(PSNR)and Human Visual System(HVS)by 0.08?3.1d B and 3.1?7.1d B,respectively,compared with other comparison methods.(2)Aiming at the problem that the denoising and super-resolution reconstruction of optical remote sensing image based on non-Generated Adversiral Netword(GAN)will lead to the smoothing of image details,an algorithm of denoising and super-resolution reconstruction of optical remote sensing image based on generated advertizing network in the wavelet transform domain was proposed.Firstly,in order to reconstruct more detailed information,the architecture of generating antagonistic network is adopted.Then,in order to make the final reconstructed image have good performance in objective evaluation indexes such as PSNR and structural similarity(SSIM),the residual network(local residual network and global residual network)is combined in its generating part.In terms of loss functions,the generated partial loss functions combined with Total Variation(TV)losses to further enhance the detailed information of reconstruction.The discriminant part uses the relative loss to replace the traditional discriminant loss calculation method,in order to make the whole network better convergence.Finally,the algorithm is implemented in the wavelet transform domain.Experimental data show that this method can reconstruct optical remote sensing image with 4× super-resolution and remove gaussian noise or salt noise at the same time.In terms of PSNR and SSIM,it is 0.01?0.56 d B and 0.04?0.07 d B higher than the Ga-based SRGAN,ESRGAN and other algorithms.From the perspective of visual effect,among all experimental comparison algorithms,the image details processed by this method are more abundant,and the evaluation criteria such as Mean Opinion Score(MOS)and Perception Index(PI)are 0.2?2 and 0.07?4.7 higher than other comparison methods.(3)In order to further obtain optical remote sensing images with higher quality than the data set itself,an unsupervised learning optical remote sensing image denoising and super-resolution reconstruction algorithm based on unpaired images was proposed.Firstly,in order to obtain higher quality images,the cyclic generation confrontation architecture is adopted,and the natural image data set with high quality itself is taken as part of the input to jointly train the network.Secondly,relative loss and perception loss are used respectively to calculate counter loss and cyclic consistency loss,so that the whole network can converge better.Experimental data show that this method can reconstruct optical remote sensing images with unsupervised 4× super-resolution and remove Gaussian noise at the same time.In terms of PSNR and structural SSIM,RRDGAN and Cin CGAN were 0.92?2.18 d B and 0.07?0.14 d B higher than gan-based algorithms.In terms of visual effects,the image generated by this method is more detailed,and in terms of evaluation criteria of PI,it is higher than all the algorithms compared in the experiment by 0.01?4.86.To sum up,based on deep learning technology,this dissertation proposes three new optical remote sensing image denoising and super-resolution reconstruction algorithms,which can improve the resolution of optical remote sensing image while removing the corresponding noise,effectively solving some shortcomings of existing methods and achieving ideal results.It provides a new way to obtain higher quality optical remote sensing image.