High-definition processing of remote sensing images based on CUT-CycleGAN

High-definition remote sensing images are more and more widely used in research and life. However, due to hardware conditions and transmission rate limitations, it is too expensive to directly obtain high-definition original images. So, it has become a research hotspot on how to use algorithms to receive high-definition remote sensing images from low-resolution images. In view of the existing super-resolution methods for remote sensing images, the dependence on a large number of matching low-resolution and high-resolution(LR-HR) data sets and the slow network training time. In this paper, contrast learning is used for unpaired image-to-image conversion model (CUT-CycleGAN), which uses cyclic consistency to achieve the purpose of training using unpaired images, and adds a contrast learning framework to effectively shorten CycleGAN's training time and to improve efficiency. The experiment selects SRGAN, CycleGAN, EDSR, and FSRCNN four existing super-resolution methods to compare with the method in this paper. The results show that the training time of CUT-CycleGAN is reduced by nearly 55.7%, and after training with unpaired images, the quality of the generated high-definition images is good enough.