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

Remote sensing images have a wide range of applications in various scenarios such as ground object detection,disaster warning,environmental change detection,and military strikes due to their rich information.For remote sensing images,the amount of information carried by the image depends on the image resolution.However,due to hardware manufacturing process limitations of imaging sensors,optical components etc.and signal transmission bandwidth limitations,it is difficult to directly capture high-resolution remote sensing images.Unknown losses such as blur or noise may occur during the imaging and transmission process of remote sensing images.Improving the resolution of remote sensing images by hardware is costly and infeasible,therefore,improving the resolution of remote sensing images by image processing has gradually become the mainstream.Thanks to the advancement of deep learning technology,image super-resolution reconstruction methods based on deep learning have gradually replaced the previous pixel interpolation-based methods.Currently,research on super-resolution reconstruction of remote sensing images mainly focuses on Generative Adversarial Network(GAN)-based methods.Most deep learning-based image super-resolution reconstruction methods require datasets containing a large number of high-resolution and their corresponding low-resolution remote sensing images.Nevertheless,it is always difficult to obtain highresolution and low-resolution images of exactly the same scene.Aiming at the above problems,a high-order degradation model suitable for simulating the degradation process of remote sensing images in real scenarios is proposed in this thesis,and optimizations on the generator network and discriminator network in the GAN-based super-resolution reconstruction network of remote sensing images are performed.The main research content of this thesis is as follows:(1)A high-order degradation model is constructed to simulate the degradation process of remote sensing image in real scenes,which solves the problem that most existing algorithms using a single double cubic downsampling model to simulate the loss cannot simulate the degradation process of remote sensing image in real scenes and lead to the poor reconstructed image quality.The use of high-order degradation model can not only significantly improve the reconstructed high-resolution remote sensing image quality,but also effectively enhance the generalization ability of the network model.(2)A generator network based on Recurrent Residual Super Dense Block(RRSDB)with recursive residual connections is proposed to address the issue of image blur caused by the generator network of the existing GAN-based image super-resolution reconstruction algorithms,which cannot fully utilize the image feature information in the network due to the network depth limitation.The improved generator can effectively improve the reconstructed image quality without deepening the network depth.(3)A U-shaped discriminator network containing both recursive residual structure and attention mechanism is presented to solve the problem of blurs and artifacts in the details of the generated images brought by the prior art that most discriminator networks can only globally distinguish images.The improved discriminator network can perform more precise discrimination between real and generated images,provide more detailed feedback to the generator,and obtain superior reconstructed high-resolution images with finer texture details and better subjective perceptions.The super-resolution reconstruction network of remote sensing images RAESRGAN proposed in this thesis is composed of a high-order degradation model,a generator network optimized by RRSDB and a U-shaped discriminator network optimized by the recursive residual structure and attention mechanism.Experimental results show that compared with the previous image super-resolution reconstruction algorithms,the reconstructed images by RAESRGAN perform better in texture details and subjective visual effects,and the model has stronger generalization ability.