A No-Reference Super Resolution for Satellite Image Quality Enhancement for KOMPSAT-3
Yeonju Choi, Yongwoo Kim · 2020
Recently, a deep learning based super-resolution (SR) technology has been applied to satellite images to improve spatial resolution and sharpness, and to increase extractable information. In this paper, we propose a no-reference single image super-resolution method that improves the image quality by doubling spatial resolution of Korea Multi-Purpose Satellite-3 (K3), achieving a ground sampling distance (GSD) of 0.7 m. When training SR networks, the proposed method generates low-resolution (LR) images by applying the degradation model to K3 images and creates enhanced high-resolution images (HRe) by applying the top and bottom hat transformation to the original high-resolution (HRo) images. As a result of applying SR to the original K3 image, it was possible to obtain an image with improved quality. Additionally, as a result of testing the Baotou area used for satellite image quality evaluation, it was confirmed that the resolution is similar to the spatial resolution of Korea Multi-Purpose Satellite-3A (K3A), which is a GSD of 0.55 m.