Efficient High-Frequency Texture Recovery Diffusion Model for Remote Sensing Image Super-Resolution
Wu-Ding Weng, Chao-Wei Zheng, Jian-Nan Su, Guangyong Chen, Min Gan
Abstract
Remote sensing super-resolution (SR), which aims to reconstruct high-resolution (HR) images with rich spatial details from low-resolution (LR) remote sensing images predominantly composed of low-frequency components, presents a challenging yet practical task. Existing diffusion model (DM)-based methods for remote sensing SR are inefficient, requiring extensive iterations and often failing to recover high-frequency details adequately due to a lack of targeted processing for high-frequency components. To mitigate these challenges, this article introduces an efficient DM for remote sensing image SR, termed image reconstruction representation-diffusion model for super-resolution (IRR-DiffSR). IRR-DiffSR employs a feature extraction encoder to extract the image reconstruction representation (IRR) from ground-truth (GT) images, which makes the reconstruction network focus more on recovering high-frequency textures. Unlike traditional DM-based methods that learn the direct mapping from LR to HR images, IRR-DiffSR employs a pre-trained encoder to guide the DM in extracting consistent IRR directly from LR images. This auxiliary information aids in the efficient and effective reconstruction of high-frequency textures. By serving as an implicit reconstruction prior, this enables the DM to achieve accurate estimations with fewer iterations, thus assisting IRR-DiffSR in recovering high-frequency information more efficiently and effectively. Extensive experiments on four remote sensing datasets demonstrate that IRR-DiffSR achieves state-of-the-art reconstruction results in both real and synthetic scenarios. Specifically, in real scenarios, IRR-DiffSR outperforms the next best method by 0.766 and 0.69 in the naturalness image quality evaluator (NIQE), while in synthetic scenarios, it achieves peak signal-to-noise ratio (PSNR) improvements of 1.07 and 0.51. These results highlight the effectiveness and efficiency of IRR-DiffSR in recovering high-frequency details. Our code and pre-trained models have been uploaded to GitHub (<uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/55Dupup/IRR-DiffSR</uri>) for validation.