Robust Remote Sensing Super-Resolution With Frequency Domain Decoupling for Multiscenarios

Super-resolution (SR) for remote sensing has the potential for a huge impact on diverse applications by reconstructing an accurate high-resolution (HR) image from its low-resolution (LR) counterpart. Despite a lot of attention, most recent methods concentrate on the overall modeling of the original inputs and treat various frequency information within the image equally, which can lead to distortion of image details or artifacts. To tackle this issue, we proposed a frequency-domain decoupling approach for image SR. It uses nonparametric clustering to obtain the low-frequency contours, applies a linearized transformer to model the long-range dependency of the mid-frequency details, processes the high-frequency textural with the attention mechanism (AM), and employs the Fourier transform to integrate different frequency information effectively. By processing the different frequencies independently, we achieve a more accurate enhancement of the original scene's visual quality without adding noise. We conducted evaluations of our method across five real-world scenarios with distinct characteristics. The results show that our model has a significant improvement in peak signal-to-noise ratio of 2.4-5.5 dB over the state-of-the-art method.