Freqformer: Image-Demoiréing Transformer via Efficient Frequency Decomposition

TL;DR

Image demoiréing is hindered by the intertwined nature of moiré textures and color distortions. Freqformer addresses this by using a frequency decomposition to separate high-frequency moiré textures from low-frequency color distortions, processed in a dual-branch Transformer with a learnable Frequency Composition Transform to fuse results. A lightweight Spatial-Aware Channel Attention module enhances moiré-sensitive regions while keeping computational costs low. Across high-resolution benchmarks (FHDMi and UHDM), Freqformer delivers state-of-the-art restoration with a compact model, highlighting the practical impact of frequency-aware design for real-world demoiréing.

Abstract

Image demoiréing remains a challenging task due to the complex interplay between texture corruption and color distortions caused by moiré patterns. Existing methods, especially those relying on direct image-to-image restoration, often fail to disentangle these intertwined artifacts effectively. While wavelet-based frequency-aware approaches offer a promising direction, their potential remains underexplored. In this paper, we present Freqformer, a Transformer-based framework specifically designed for image demoiréing through targeted frequency separation. Our method performs an effective frequency decomposition that explicitly splits moiré patterns into high-frequency spatially-localized textures and low-frequency scale-robust color distortions, which are then handled by a dual-branch architecture tailored to their distinct characteristics. We further propose a learnable Frequency Composition Transform (FCT) module to adaptively fuse the frequency-specific outputs, enabling consistent and high-fidelity reconstruction. To better aggregate the spatial dependencies and the inter-channel complementary information, we introduce a Spatial-Aware Channel Attention (SA-CA) module that refines moiré-sensitive regions without incurring high computational cost. Extensive experiments on various demoiréing benchmarks demonstrate that Freqformer achieves state-of-the-art performance with a compact model size. The code is publicly available at https://github.com/xyLiu339/Freqformer.

Figures (5)

• Figure 1: Frequency Decomposition.
• Figure 2: Visualization of the frequency-decomposed images (zoom in for better analysis). The left three columns show the moiré images along with their 2D and 3D visualizations, while the right three columns display the clean images. The first, second, and third rows represent the original images, low- and high-frequency components, respectively. The low-frequency curves of moiré and clean images exhibit similar trends, primarily reflecting color distortions. The high-frequency curves differ significantly in shape, indicating structural differences caused by moiré patterns.
• Figure 3: The whole pipeline of our Freqformer. The figure shows the dual-branch learning with crop/resize strategy and a learnable FCT module.
• Figure 4: Left part is the detailed transformer architecture of the SA-CA module utilized in all the encoders and decoders. The right part shows the hierarchical fusion of the high-frequency branch.
• Figure 5: Visual comparisons of different baselines on FHDMi and UHDM datasets (please zoom in for better details). While OSEDiff, AdaIR, MoCE-IR, and FHDe$^2$Net all struggle with removing moiré artifacts, ESDNet-L also leaves residual traces behind. In addition, these baselines suffer from color distortions and fail to accurately recover the original tones. In contrast, our approach effectively eliminates both moiré and color artifacts, as shown in the last row.