Research On Super-resolution Reconstruction Of Remote Sensing Images Based On Deep Learning

With the development of science and technology,humans are exploring more and more unknown perspectives.Remote sensing images are becoming an emerging direction for image mission applications.The demand for high-resolution remote sensing images is increasing sharply.Due to the limitation of hardware level and economic cost,and the rapid development of computer vision technology,super-resolution reconstruction technology is becoming an important research field,and super-resolution reconstruction technology based on remote sensing images will also play a huge role in our life.At present,super-resolution reconstruction technology has been developed to some extent,but super-resolution reconstruction technology based on specific scenes is still yet to be developed,especially for super-resolution reconstruction in remote sensing image scenes.Therefore,this paper will mainly focus on the migration of super-resolution reconstruction methods to remote sensing images.By designing the network,super-resolution reconstruction techniques can be better applied to remote sensing images.The main contents and innovations of the work are as follows:According to the needs of the paper,super-resolution reconstruction datasets of remote sensing images were constructed,including a satellite image dataset and an aerial photography dataset.Since the remote sensing images have the characteristics of huge difference in object scale,and the low-resolution and high-resolution image are similar in low frequency,this paper proposes a residual multi-scale convolutional neural network to solve remote sensing image super-resolution reconstruction problem.This paper is based on the RED network.The encoder-decoder structure is used to extract features.In particular,a multi-scale module consisting of serial and parallel dilated convolutions is added between the convolution layer and the deconvolution layer to extract multi-scale features.In addition,this paper also introduces a channel-wise attention mechanism to enhance useful features and suppress useless information,which enhances the ability to express features for long connections.Finally,on the overall network structure,this article will only learn high-frequency residuals,perform sub-pixel convolution operations on low-resolution images and high-frequency residual features,and add them as outputs.After applying the idea of residuals,because the residual image is relatively sparse,many values are small or even zero,which speeds up the training speed and improves the accuracy.In this paper,through the experiments on two self-built data sets,PSNR and SSIM evaluation indicators are applied to prove the effectiveness of the increased module,and the number of parameters is reduced.To a certain extent,it meets the needs of super-resolution reconstruction of remote sensing images.