Blind Super-Resolution on Remote Sensing Images with Blur Kernel Prediction
Runmin Dong, Lixian Zhang, Haohuan Fu · 2021
Single image super-resolution (SISR) is essential in many remote sensing applications. Most of the existing SISR methods on remote sensing images assume that the low resolution (LR) images are synthesized from high-resolution (HR) images by bicubic downscaling. However, the performance of those methods is limited in the real-world remote sensing scenario as the actual degradation is sometimes different from the assumption. Therefore, we introduce the blind super-resolution (SR) concept and propose a super-resolution method with blur kernel prediction (BKPSR). BKPSR first predicts the blur kernel code for an image and then utilizes the blur kernel code to assist the image super-resolution. Experimental results indicate that our method outperforms existing SISR methods on real-world remote sensing images.