Research On Remote Sensing Image Super-Resolution Based On Deep Learning

With the continuous commercialization and miniaturization of China’s remote sensing satellites,the use of existing data and information processing methods to improve the spatial resolution of remote sensing images has very important scientific significance and research value.This article focuses on the need for remote sensing satellites to obtain high-resolution remote sensing data.Based on the existing deep learning network theory and experiment,it focuses on the research of adaptive remote sensing image super-resolution reconstruction technology suitable for various remote sensing bands.Under the requirement of retaining the overall geometric structure of the image,it focuses on the spatial resolution enhancement and detailed feature reconstruction of each spectrum segment of the remote sensing image.Aiming at the problem that the existing super-resolution reconstruction network lacks effective model design for feature learning between bands of hyperspectral images,this paper designs algorithms from three perspectives: image feature extraction,feature nonlinear mapping,and image reconstruction,and proposes a combination of multiple Convolutional neural network super-resolution reconstruction algorithm based on scale feature extraction and multi-level feature fusion structure.The innovative work of this article mainly lies in the following three aspects:(1)Aiming at the problem of nonlinear learning between features of different bands of remote sensing images,firstly,a basic network module based on multi-scale feature extraction is designed for the fusion expression of features of different sizes of receptive fields.At the same time,in order to prevent the explosion of the overall network parameters,the Depthwise Separable Convolution can be used to performs feature fusion of spatial domain and band dimension on high-dimensional features.Then,in order to better aggregate the weights of the features,this paper introduces the traditional wavelet transform to realize the multi-level feature extraction of the details of the features and the weight distribution of the features based on the spatial self-attention mechanism.(2)In view of the insufficient expression of feature information in the feature extraction stage of image input and the cascaded form of the overall residual network,the model’s nonlinear learning ability of prior knowledge is suppressed.First,the standardization of input data is introduced to eliminate the difference in the distribution of different images and bands in the same feature space reduces the difficulty of the overall nonlinear learning of the network.On the other hand,the pixel domain-based self-attention fusion module is used to fuse the multi-level features of the feature nonlinear mapping stage,and the features of different levels and receptive fields are used to assist the final image reconstruction.(3)In the traditional super-resolution method based on convolutional neural network,the final image reconstruction stage mainly uses transposed convolution for feature fusion to obtain a fixed-magnification super-resolution image.This pure convolutional form the application of reconstruction methods to remote sensing image networks with higherdimensional features will greatly increase the overall parameters of the network and make it difficult to fully express and reconstruct the input abstract prior information.Inspired by the idea of 2D sub-pixel convolution,this paper proposes an adaptive sub-pixel convolution super-resolution reconstruction method for the final image reconstruction,which improves the overall model’s ability to reconstruct the details of each band of remote sensing images.The experimental results show that in the same data set and training environment,compared with the existing algorithms in the reconstruction situations of 2,4,and 8,the algorithm in this paper can improve the PSNR by an average of about 0.3d B,and at the same time,this algorithm is better than existing algorithms in subjective effect.