DecMoiré: Decoupled Illumination Modulation and Moiré Removal for Robust Image Demoiréing

The moiré effect, a common artifact when photographing digital screens, degrades images with a complex mixture of geometric aliasing patterns and severe photometric distortions. These degradations collectively impair visual quality and hinder downstream computer vision tasks. While recent learning-based demoiréing methods have achieved promising results, they typically employ a coupled network to address all degradations simultaneously. They thus often fail to disentangle the intricate interplay between geometric artifacts and photometric aberrations, resulting in suboptimal generalization. To address this, we introduce DecMoiré, a novel retinex-inspired framework for image demoiréing. Inspired by Retinex theory, we explicitly decouple the degraded image into two components: an illumination map that captures the spatially varying photometric deviations caused by screen light interference, and a reflectance map containing both the underlying moiré patterns and texture. We then process each component via specialized modules. For robust photometric restoration, we propose an uncertainty-aware illumination modulation network that predicts adaptive kernels to refocus the scattered screen light, while simultaneously leveraging uncertainty estimation to improve the robustness of illumination reconstruction. For moiré removal, a joint spatial-frequency module operates on the reflectance map, synergizing spatial convolutions for local texture refinement with frequency-domain modulation to suppress large-scale periodic patterns. Extensive experiments on datasets demonstrate that our framework surpasses state-of-the-art methods, producing images with superior illumination fidelity, finer details, and exceptional robustness in challenging scenarios.