Research On Super-resolution Reconstruction Of Remote Sensing Images Based On Cycle Consistent Adversarial Network

High-resolution remote sensing images have high practical value in the fields of target extraction,ecological environment monitoring,disaster rescue,etc.However,due to the influence of hardware and environmental factors,the resolution of remote sensing images is reduced and detailed information is missing.The research on resolution reconstruction methods is of great significance.Inspired by the idea of dual learning and unsupervised learning image conversion tasks,this paper proposes a method for super-resolution reconstruction of remote sensing images based on a recurrent generation confrontation network.The main contents are as follows:(1)In view of the problem that the cumbersome steps of the reconstruction-based multi-image super-resolution method and the convolutional neural network image super-resolution reconstruction method is not ideal for the reconstruction of remote sensing images,this paper proposes a remote sensing image super-resolution reconstruction method based on CycleConsistent Adversarial Network.Network training does not require paired high resolution and low resolution data of the same scene,the cycle-consistent loss is used to constrain the content of the unpaired data,and the perceptual loss is used to constrain the quality of the reconstructed image from the perspective of high-level semantic features.The subjective and objective evaluation of the reconstructed image shows the feasibility of this method for super-resolution reconstruction of a single remote sensing image.(2)The remote sensing image super-resolution reconstruction method based on CycleConsistent Adversarial Network is composed of a reconstruction network,an image degradation network and two discriminator networks.The dense multi-scale feature module is designed in the degradation network,which can effectively extract the multi-scale features of the image,and realize the conversion mapping from high resolution to low resolution;Then,a combination of dense residual structure,self-attention dilated convolution residual structure and sub-pixel convolution is used in the reconstruction network to realize the conversion mapping from low resolution to high resolution,which can strengthen the local features of the image,and more reconstruct the texture details of the target well and speed up the model convergence At the same time,an improved fully convolutional ResNet50 discriminator network structure is designed,and a two-dimensional probability matrix is output,which can improve the authenticity of the local features of the reconstructed image.The experimental results show that the method in this paper has certain advantages in peak signal-to-noise ratio,structural similarity and noise removal compared with the traditional method and the classic model of deep learning,and it can better reconstruct the texture details of the ground objects.The experimental results also show that the method in this paper has an improved extraction effect on target detection tasks such as road extraction,and the feasibility of superresolution reconstruction of the resource ZY-3 image.The paper has 40 pictures,4 tables,and 70 references.