Spatial-Frequency Guided Moiré Removal with Multi-Stage Feature Fusion

Moiré patterns severely degrade the visual quality of images by introducing complex, multi-scale, and frequency-entangled artifacts that are often intertwined with real image content. To address this challenge, we propose a novel multi-stage progressive restoration framework, named SFMFNet, which jointly exploits spatial- and frequency-domain representations to enhance moiré removal. The network adopts an encoder-decoder structure with multi-scale token mixing blocks and a dedicated frequency-aware module based on Fourier transforms. To improve information flow across stages, we introduce a cross-scale feature aggregation (CSFA) mechanism that effectively aggregates and redistributes features at different resolutions. Experimental results demonstrate that SFMFNet can effectively suppress moiré artifacts while preserving structural integrity and texture details, confirming the advantage of combining spatial structure and frequency cues for generative image restoration.