Fabric Image Demoiréing Benchmark from Synthesis to Restoration

Fabric moiré is a persistent artifact in digital photography, caused by the interference between fine textile patterns and camera sensor grids. Unlike screen-induced moiré, which is relatively periodic and global, fabric moiré is characterized by broadband, semi-periodic, and anisotropic patterns that are deeply entangled with the underlying textile texture. This spectral overlap makes traditional demoiréing methods—often designed for screens—prone to over-smoothing or failing to remove artifacts entirely. Furthermore, the lack of pixel-aligned real-world training pairs has historically hindered the development of robust, learning-based solutions for this specific domain.

To address the data scarcity, the authors present PRISM (Physics-based Residual Injection for Synthetic Moiré). Instead of attempting to capture real-world pairs, which is complicated by non-rigid fabric deformation and alignment errors, PRISM models the physical imaging chain. It uses a round-trip construction process to isolate aliasing residuals and inject them into clean fabric images. This approach generates a large-scale dataset of 16,050 multi-resolution, pixel-aligned pairs, providing a standardized benchmark for training and evaluating demoiréing models.

Building on this benchmark, the authors introduce FaDeNet, a network tailored for fabric restoration. FaDeNet employs a content-adaptive base/detail decomposition to separate low-frequency color/illumination shifts from high-frequency textures. It uses a U-shaped trunk to predict residuals and a spatial confidence mask, ensuring that corrections are applied only in moiré-dominant regions. To prevent the loss of fine textile details, the model incorporates Spectral-Anisotropic Gated Blocks (SAGB) and magnitude-controlled refinement, which allow for targeted artifact suppression without degrading the original fabric structure.