Research On Super-resolution Reconstruction Of Remote Sensing Image Based On Convolution Neural Network

The spatial resolution of remote sensing image is one of the important indexes in remote sensing imaging technology.High-resolution remote sensing image means higher data volume.High-resolution remote sensing image is very important for image processing,analysis and application.However,in the process of remote sensing image acquisition,in addition to the sensor resolution,the performance of imaging equipment and other hardware equipment,it may also be affected by bad weather,system noise and other environmental factors,making the remote sensing image resolution low.Using super-resolution reconstruction method to reconstruct the low-resolution remote sensing image can reduce the acquisition cost and improve the resolution of the remote sensing image,which is of great significance to the subsequent processing and analysis of the remote sensing image.The target of remote sensing image has the characteristics of diversity.To express the relationship between these targets correctly,it is necessary to extract the texture,contour and other features of the target.However,the existing super-resolution reconstruction method is not effective in processing the relevant features of remote sensing image.Therefore,a method of super-resolution reconstruction of remote sensing image based on convolution neural network is proposed.Firstly,the multi-scale feature of remote sensing image is extracted by multi-scale feature extraction,and then the mapping relationship between low-resolution image and high-resolution image is learned by nonlinear mapping unit.Finally,the super-resolution reconstruction of remote sensing image is completed by reconstruction part.The specific work is as follows:(1)This paper analyzes the influence of many factors on the remote sensing image in the imaging process,and constructs the remote sensing image degradation model.Three kinds of typical super-resolution reconstruction methods are introduced.Combined with the characteristics of remote sensing image,the shortcomings of existing methods are analyzed,and a multi-scale depth neural network structure is designed for super-resolution reconstruction of remote sensing image.(2)To solve the problem that the existing methods can not extract the features of different scales of remote sensing image accurately,a super-resolution reconstruction method based on multi-scale feature unit is proposed.In this paper,multi-scale feature units are used to extract multi-scale features from remote sensing images,and the number of parameters in the feature extraction part is reduced by depth separable convolution.The network improves the reconstruction accuracy on the basis of ensuring the reconstruction speed.(3)In view of the problem that the number of convolution layers in the nonlinear mapping part of the existing methods is too small to make the network accurately learn the mapping relationship between the low-resolution remote sensing image and the high-resolution remote sensing image.In this paper,the nonlinear mapping structure is improved,and the multilayer depth separable convolution layer is designed to replace the original convolution layer,so that the network can learn more fully the mapping relationship between low-resolution image and high-resolution image,so as to improve the final reconstruction quality.(4)In order to solve the problem that the existing methods neglect the shallow feature information of low-resolution remote sensing image,a super-resolution reconstruction method based on global residual structure is proposed.In this paper,the shallow features in the original low-resolution image and the deep features obtained by nonlinear mapping are fused together to improve the reconstruction accuracy of the network.In order to verify the performance of this method,we choose this method and all kinds of classical super-resolution reconstruction methods for comparative test.The experimental results show that the structure and texture of the remote sensing image reconstructed by this method are clearer,and the highest value is achieved in the objective evaluation index.The experimental results show the effectiveness of the proposed method.