A Universal Framework for Remote Sensing Image Color Correction via Style Transfer

Le résumé fourni par la source

Color inconsistency in optical remote sensing imagery, caused by variations in imaging platforms, illumination, and atmospheric conditions, can reduce the reliability of multi-temporal and multi-source downstream applications. Existing deep learning–based color correction approaches, largely adapted from natural-image processing, often struggle with two intrinsic challenges in remote sensing scenarios: (i) preserving fine geometric structures during appearance distribution alignment, and (ii) scaling to ultra-high-resolution imagery without compromising global color coherence. In this paper, we propose a photorealistic style transfer framework for reference-guided remote sensing image color correction. The core idea is to inject target appearance statistics via a lightweight, learnable mapping of Gram representations into the content feature space dynamically, aiming to improve color consistency while mitigating structural distortions. To address high-resolution processing challenges, we introduce a Pixel Unshuffle/Shuffle strategy, which avoids stitching artifacts and ensures global consistency. Additionally, we design a style fusion strategy that adaptively merges the stylized output with the original content image to restore fine structural details. Extensive experiments on diverse remote sensing datasets demonstrate the stability and consistency of the proposed approach. Further evaluations on change detection, image matching, and 3D reconstruction show that our method offers potential improvements in real-world remote sensing workflows.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.