Remote Image Super-Resolution Using Enhanced Residual Channel Attention Networks

Remote sensing image super-resolution is the process of generating high-resolution remote sensing images from one or more lower-resolution remote sensing images.It is of great use in military and civilian fields such as environmental monitoring,land survey,vegetation monitoring and target detection.The methods based on deep learning can make full use of the information of the input images.Therefore,the methods provides a new technical method for image super-resolution.However,due to the limitations of the sampling locations of convolution kernel and degradation types of the dataset,it is difficult to reconstruct high-resolution remote sensing images.This paper works based on the Residual Channel Attention Network,and the specific improvements are as follows:(1)Considering that the deep convolutional neural network generally adopts regular convolutional kernel,the ability of the network to extract image feature information is limited.In this paper,deformable convolutional modules are introduced and the network structure based on RCAN for remote sensing image super-resolution is redesigned.The deformable convolution modules can dynamically update the convolutional kernel parameters when sampling in different regions,and can shift the sampling positions of rules through deformable convolution.So the network can adaptively extend the receptive field,and make the details of the generated remote sensing images clearer.(2)The existing datasets used for remote sensing image super-resolution are often obtained through bicubic downsampling,which is an ideal situation and far from reality.Therefore,this paper uses realistic datasets for experiments.By utilizing Kernel GAN to estimate the filtering kernel of remote dataset,a dataset with complex degradation type is synthesized.Obviously,networks trained on such datasets have stronger applicability.In this paper,the experiments are conducted on two datasets of differernt types of degration.Numerical results such as PSNR and SSIM are improved,and details such as image texture or edge are also clearer.