A Novel Approach using Degradation Representation for Remote Sensing Image Super-resolution in Real-world Scenarios

Real-world image degradations differ from ideal conditions, where common deep learning models often rely on bicubic interpolation to synthesize low-resolution counterparts. However, the generalizability of these trained models across diverse image datasets with varying distributions is still being determined. To address this challenge, we propose DRSR, Degradation Representation for unsupervised Super-Resolution specifically designed for real-world remote sensing images with unknown arbitrary degradations. The network demonstrates robust generalization capabilities across additional datasets, encompassing both "Ideal"and "Non-Ideal"scenarios. It particularly targets image datasets facing two key limitations: the absence of ground-truth high-resolution images and the presence of arbitrary degradations.