ZT-D2SR: dual-domain guided adaptive super-resolution for remote sensing images
TL;DR
A Dual-Domain Guided Adaptive Super-Resolution method that comprehensively outperforms state-of-the-art methods in reconstruction quality but also provides a highly practical solution for large-scale remote sensing image processing with its superior inference efficiency.
Abstract
Remote sensing image super-resolution aims to recover high-frequency details from low-resolution (LR) images. However, when processing complex terrains and tiny objects, existing methods often struggle with insufficient detail recovery, structural distortion, and limited generalization. Although diffusion models possess powerful generative priors, their direct application to the remote sensing domain suffers from high computational costs, false textures, and structural deformations. To address these issues, this paper proposes a dual-domain guided adaptive super-resolution method. Unlike conventional methods that mainly optimize in a single domain, our method builds an efficient adaptive generation framework that improves robustness to complex unknown degradations while reducing the number of reverse-diffusion evaluations. Specifically, we extract a spatial structural feature map from the LR image and use it as geometric guidance for the adaptive initialization. We further integrate a wavelet-domain loss with pixel, perceptual, and adversarial losses to explicitly constrain frequency-specific reconstruction errors. Experiments on multiple benchmark datasets demonstrate a favorable trade-off among structural preservation, perceptual quality, and inference efficiency. The method is competitive on distortion-oriented metrics and shows advantages on selected perception-oriented and no-reference metrics.