Super-resolution Reconstruction Of Remote Sensing Image Based On Dense Convolutional Neural Network

When using remote sensing images to solve practical problems at work,the spatial resolution is one of the main considerations.However,the development of high-resolution satellites is expensive and time-consuming,so the field of remote sensing has been exploring ways to obtain high-resolution images economically and quickly.In recent years,the deep learning algorithm has been developed rapidly and compared with the traditional machine learning method.The deep neural network can automatically learn the complex abstract features of the object and improve the learning ability.Therefore,this paper proposes a remote sensing image super-resolution reconstruction algorithm based on dense convolutional neural network,which provides a method for generating high-resolution remote sensing images economically and quickly.The main research contents of this paper are as follows:(1)The degradation model of remote sensing image is established based on the analysis of various factors affecting the degradation of image;Then the common super resolution reconstruction algorithm based on single frame image is studied and the convolution neural network based super resolution reconstruction algorithm is selected for remote sensing image processing.(2)Aiming at the large amount of low-frequency information in low-resolution remote sensing images,a deep-dense residual neural network is proposed as a super-resolution reconstruction algorithm for remote sensing images.In the network,the combination of subpixel convolution and ordinary convolution is used instead of the traditional interpolation preprocessing or deconvolution completion feature,which effectively reduces the influence of noise and the possible checkerboard effect in reconstructing images;The attentionconcentration structure can make the network more focused on the high-frequency information in the image,thereby improving the quality of the reconstructed image;using the residual network with deep dense connection and large convolution to increase the network’s receptive field and enhance its characteristics.The ability to extract information,improve the ability to reconstruct details,and solve the problem of gradient explosion,gradient disappearance and long-term dependence caused by the network too deep,improve the network’s ability to acquire prior information,and accelerate the convergence of the training process.(3)Based on the MATLAB platform,the program is used to construct the data set,and the algorithm and other reconstruction algorithms are used to perform super-resolution reconstruction experiments on the constructed data set.In the case of magnifications of 2x,3x and 4x,the average peak signal-to-noise ratio and average structural similarity of the proposed algorithm are the best.In terms of visual visual effect,the algorithm reconstructs the image details more clearly and the Stronger noise suppression.(4)The GF2 image was used as the verification data for super-resolution reconstruction experiments.Under three magnification multiples,the objective evaluation index of the algorithm in this paper is the highest,and the subjective visual perception is the best.The above experimental results verify that the proposed algorithm has better super-resolution reconstruction effect on the high-resolution satellite remote sensing data.The experimental results show that compared with bi-cubic interpolation,SRCNN,RDN and EDSR,the proposed algorithm has better super-resolution reconstruction effect.The superresolution reconstruction algorithm proposed in this paper has practical value in remote sensing image application.