Fabric Image Demoiréing Benchmark from Synthesis to Restoration¶

Conference: ECCV 2026
arXiv: 2606.24072
Project Page: https://weipengchao.top/PRISM-page
Code: TBD
Area: Image Restoration
Keywords: Fabric demoiréing, benchmark dataset, physical simulation, conservative restoration, anisotropic filtering

TL;DR¶

This paper presents the first systematic study on the fabric image demoiréing problem. It introduces PRISM, a physical imaging chain simulation-based residual injection synthesis framework (including the first fabric moiré benchmark with 16,050 image pairs), and FaDeNet, a conservative restoration network specifically designed for the spectrally entangled characteristics of fabrics, significantly outperforming existing screen demoiréing methods in terms of PSNR, SSIM, and LPIPS.

Background & Motivation¶

Moiré patterns are aliasing artifacts generated in the frequency domain when camera sensors under-sample high-frequency scene content that exceeds the Nyquist limit, manifesting as periodic stripes, ripples, and color distortions. Over the past decade, research on screen-to-camera demoiréing has made significant progress—ranging from multi-scale decomposition methods like DMCNN and frequency-domain bandpass filtering like MBCNN, to ESDNet and P-BiC designed for 4K resolutions. These methods largely rely on an implicit assumption: image content and moiré artifacts are spectrally separable. However, fabric moiré is dramatically different. The woven structure of fabrics contains dense, anisotropic, and quasi-periodic micro-patterns (stripes, plaids, herringbone, etc.), whose spectrum is broadband with energy widely distributed and highly overlapping with the fabric's intrinsic texture. When such a complex spectrum is folded by sensor sampling, moiré patterns and original textures become completely entangled in the frequency domain, rendering the inverse restoration inherently more ill-posed. Directly applying screen demoiréing models to fabric scenes inevitably results in two types of failure: either a large amount of residual moiré artifacts remain, or the fine texture details of the fabric are over-smoothed and erased. These failures point to a fundamental conflict—screen moiré is discretely separable in the spectrum, whereas fabric moiré is continuously inseparable.

Conducting research on fabric demoiréing faces two substantial engineering barriers. First, acquiring pixel-aligned real paired data is extremely difficult. Screen demoiréing can easily construct paired data by displaying internet images on a screen and re-photographing them. However, fabrics are flexible, non-rigid surfaces prone to wrinkles, stretching, and pose changes, making global homography alignment nearly impossible and physically unfeasible. The most relevant existing work (the focus/defocus dual-view method by Liu et al.) requires dedicated hardware and cannot be scaled up. Second, the lack of a dedicated benchmark dataset means researchers cannot systematically compare different methods, seriously hindering progress in this direction—while screen demoiréing has multiple mature benchmarks like TIP2018, UHDM, and FHDMi, fabric demoiréing remains completely unexplored.

This paper addresses this gap from both the data and algorithmic perspectives. On the data side, the authors propose PRISM (Physics-based Residual Injection for Synthetic Moiré), a synthesis framework that extracts pure moiré residual fields from physical imaging chain simulations using a round-trip alignment strategy. These are then injected back into the original clean images, achieving perfect pixel-level alignment while fully preserving fabric textures. On the algorithm side, they design FaDeNet, which adopts a base/detail decomposition combined with a mask-gated conservative restoration strategy, optimized specifically for the spectral entanglement of fabrics. Core Idea: Shift demoiréing from "global uniform correction" to "mask-gated conservative restoration"—allowing the network to obtain spatial confidence masks to automatically identify moiré-dominated regions and apply corrections only when necessary, thereby achieving a delicate balance between removing artifacts and preserving natural textures.

Method¶

Overall Architecture¶

This paper's solution revolves around a dual-track pipeline consisting of "synthetic data + restoration network". The PRISM synthesis pipeline performs physical imaging chain simulation—including viewpoint projection, discrete under-sampling, Bayer mosaicing, noise addition, and demosaicing ISP—on each clean fabric image to generate sensor outputs with synthetic moiré. It then extracts pure moiré residuals via a round-trip alignment strategy and injects them back into the original clean image to produce pixel-aligned training pairs. The FaDeNet restoration network first performs content-adaptive base/detail decomposition on the input moiré image to process low-frequency color/illumination variations and high-frequency texture structures separately. The decomposed base image is fed into a 4-scale U-Net backbone (constructed by stacking SAGB modules) to predict base residuals and a spatial confidence mask. This mask simultaneously controls the base residual correction and the amplitude-constrained refinement of the detail branch. Finally, the corrected components are recombined to yield the final restored image.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Moiré Image X"] --> B["Content-Adaptive<br/>Base/Detail Decomposition"]
    B --> C["Low-Frequency Base B"]
    B --> D["High-Frequency Detail D"]
    C --> E["4-Scale U-Net Backbone<br/>(with SAGB Modules)"]
    E --> F["Residual R + Confidence Mask M"]
    D --> G["Lightweight Residual Predictor"]
    F --> H["Mask-Gated Base Correction<br/>B̂ = B − M⊙R"]
    F --> I["Amplitude-Constrained Detail Correction<br/>D̂ = D − α·M⊙tanh(·)"]
    G --> I
    H --> J["Recombination B̂ + D̂"]
    I --> J
    J --> K["Restored Image Ŷ"]

Key Designs¶

1. Content-Adaptive Base/Detail Decomposition: Adaptively separating frequency components based on local texture density

Existing demoiréing methods usually employ fixed-parameter low-pass filters (such as Gaussian blur) to separate an image into base and detail components. However, different regions of fabric textures exhibit huge variations in texture density—the spectral characteristics of dense stripe regions are vastly different from flat backgrounds. This paper proposes a learnable adaptive decomposition: a lightweight convolutional gating sub-network (a two-layer Conv-GELU-Conv stack with only 372 parameters) is used to predict a fusion weight \(W \in (0,1)\) for each pixel location. This weight is then used for pixel-wise linear interpolation between the depthwise separable Gaussian blurred results and the original input. This is equivalent to letting the network learn "whether to smooth the current pixel": structural texture regions yield weights close to 0 (preserving original details), while flat uniform regions yield weights close to 1 (allowing base smoothing). This design endows the decomposition itself with content awareness rather than a one-size-fits-all simple filtering, providing a better starting point for subsequent divide-and-conquer processing. Notably, implementing this module entirely with depthwise separable convolutions keeps the parameters extremely compact—372 parameters are practically negligible for the entire network.

2. Mask-Gated Conservative Restoration: Applying corrections only to moiré-dominated regions

Conventional demoiréing networks apply a uniform residual prediction across the entire image, which is highly risky in fabric scenes—since fabric textures inherently contain abundant high-frequency information, indiscriminate correction can easily erase natural textures. The core protection mechanism in this work is a spatial confidence mask \(M \in (0,1)^{H \times W \times 1}\), generated by applying a sigmoid activation to the outputs of the U-Net backbone, indicating the "confidence level that correction is needed" at each position.

This mask controls the correction intensity of both branches simultaneously. In the low-frequency base branch, the U-Net predicts the base residual \(R\), and the mask restricts the final correction to \(\hat{B} = B - M \odot R\): regions with mask values close to 1 (moiré-dominated) are fully corrected, whereas regions close to 0 (clean texture) remain practically untouched. In the high-frequency detail branch, the correction is similarly gated by the mask, but is further constrained by a \(\tanh\) function to limit the absolute magnitude of a single correction within \(\alpha=0.35\). This is a critical design choice because the detail branch directly handles the high-frequency components of the texture itself, and an excessively large single-step correction could completely destroy the texture structure. This dual conservative mechanism ensures that the network "corrects only what needs to be corrected and leaves the rest untouched." Ablation studies demonstrate that removing the mask gating (forcing \(M\) to be all ones) drops the PSNR by about 0.56 dB and leads to visible over-smoothing of textures.

3. Spectral-Anisotropic Gated Block (SAGB): Joint modeling of spatial directional textures and frequency-domain periodic signals

Fabric moiré exhibits two distinctive features: spatially, it manifests as striped structures extending along specific directions; in the frequency domain, it presents as narrow-band periodic peaks. SAGB handles these two representations in parallel, with a dedicated branch capturing each.

Spatial Anisotropy Branch: Expands the input features by 4 times in channels and splits them into 4 parts, which are fed into a \(3 \times 3\) standard depthwise convolution, a dilated \(3 \times 3\) depthwise convolution with dilation rate of 2, and \(1 \times 5\) and \(5 \times 1\) strip depthwise convolutions, respectively. These four branches capture spatial directional patterns at different scales and aspect ratios. The fused features are then projected back to the original channel count via a \(1 \times 1\) convolution. Designing all branches with depthwise convolutions yields exceptionally high parameter efficiency while maintaining explicit directional selectivity—\(1 \times 5\) captures horizontal stripes, \(5 \times 1\) captures vertical stripes, and \(3 \times 3\) captures isotropic patterns.

Spectral Gating Branch: Computes the amplitude spectrum of the 2D real FFT on the input feature maps, applies global average pooling across the spatial dimensions, and passes it through an MLP to generate channel-wise gating weights \(W_f\). The original features are then modulated channel-by-channel with these weights. The elegance of this lightweight design lies in the fact that the global average amplitude extracts a rough profile of the frequency-domain energy distribution of the entire feature map. Based on this, the MLP determines which channels "currently contain periodic interference components" and subsequently suppresses or enhances those channels. Spectral gating is only enabled at coarse scales (1/4 and 1/8 resolutions, where the effective receptive field is large enough to capture low-frequency periodic patterns) and skipped at fine scales to protect high-frequency texture details.

The outputs of both branches are concatenated and passed through a gated residual update to produce the final output: simultaneously predicting a gate \(G\) and a residual \(\Delta F\), outputting \(O = F + G \odot \Delta F\). This allows stronger corrections in moiré-dominated regions while suppressing unnecessary updates in clean texture areas. Ablation studies show that removing either the spectral gating or the spatial branch causes a drop of approximately 0.6 dB in PSNR, while replacing SAGB with a standard ConvBlock of equivalent parameters drops the performance by 0.81 dB, validating the necessity of their joint design.

Loss & Training¶

The training objective consists of a pixel-level \(\ell_1\) loss and a Laplacian high-frequency constraint loss: \(\mathcal{L} = \|\hat{Y} - Y\|_1 + \lambda_{hp} \cdot \|Lap(\hat{Y}) - Lap(Y)\|_1\), where \(\lambda_{hp}=0.3\). The Laplacian operator acts as a highpass filter, forcing the network to preserve high-frequency details in edges and textures, preventing the \(\ell_1\) loss from smoothing out fine structures. The optimizer is Adam (\(\beta_1=0.9, \beta_2=0.999\)), with an initial learning rate of \(2 \times 10^{-4}\) coupled with a cyclic cosine annealing scheduler. The model is trained on a single RTX 4090 for 200 epochs, with a batch size of 4 and randomly cropped \(384 \times 384\) patches.

Key Experimental Results¶

Main Results¶

Quantitative comparisons with 9 state-of-the-art screen demoiréing methods on the PRISM test set:

Method	PSNR↑	SSIM↑	LPIPS↓	Parameters	FPS
Input (with moiré)	16.998	0.7268	0.3495	—	—
DMCNN (TIP 2018)	24.659	0.9307	0.0772	1.43M	164.9
MDDM (ICCVW 2019)	24.916	0.9260	0.0755	8.01M	17.7
WDNet (ECCV 2020)	26.141	0.9444	0.0524	3.92M	65.0
MBCNN (CVPR 2020)	27.911	0.9617	0.0405	14.19M	68.4
MopNet (ICCV 2019)	28.190	0.9648	0.0362	60.23M	10.1
FHDe2Net (ECCV 2020)	28.243	0.9654	0.0343	13.60M	55.3
P-BiC (ACM MM 2024)	28.590	0.9689	0.0334	4.92M	38.2
ESDNet-L (ECCV 2022)	28.809	0.9711	0.0268	10.62M	19.2
FaDeNet (Ours)	32.159	0.9859	0.0169	7.05M	38.1

FaDeNet significantly leads across all three quality metrics: its PSNR is 3.35 dB higher than the runner-up, ESDNet-L, while its LPIPS is reduced by 37%, all with fewer parameters and nearly double the inference speed of ESDNet-L. This clearly demonstrates that screen demoiréing methods suffer severe performance limitations in fabric scenes, highlighting the massive advantage of FaDeNet's fabric-specific design.

Ablation Study¶

Configuration	PSNR↑	SSIM↑	LPIPS↓	Description
(A) w/o Decomposition	31.560	0.9837	0.0186	Remove base/detail decomposition and the detail branch
(B) w/o Mask Gating	31.599	0.9840	0.0185	Mask \(M\) forced to be all ones
(C) w/o SAGB	31.349	0.9831	0.0191	Replace SAGB with standard ConvBlock of equivalent parameters
(D) SAGB w/o Spectral Gating	31.541	0.9837	0.0189	Keep spatial anisotropy branch only
(E) SAGB w/o Spatial Branch	31.355	0.9829	0.0194	Keep spectral gating branch only
(F) \(\ell_1\) loss only	31.777	0.9843	0.0183	Remove Laplacian high-frequency constraint
Full (FaDeNet)	32.159	0.9859	0.0169	Full model

Key Findings¶

SAGB makes the most prominent contribution: Replacing it with a standard ConvBlock of equivalent parameters (Config C) drops the PSNR by 0.81 dB—the largest degradation among all settings, indicating that joint modeling of anisotropy and spectrum customized for moiré patterns is far superior to general convolutions.
Three key designs are mutually orthogonal and complementary: Removing decomposition (A), removing the mask (B), or removing SAGB (C) each decreases performance by about 0.6–0.8 dB, showing that every component contributes to the final performance without redundancy.
Laplacian loss mainly enhances perceptual quality: Removing the high-frequency constraint (F) drops the PSNR by about 0.38 dB, but increases the LPIPS from 0.0169 to 0.0183 (an 8.3% degradation). This suggests that pixel-level supervision via \(\ell_1\) alone is insufficient to maintain high-frequency texture fidelity, making the high-frequency constraint crucial for perceptual quality.
Zero-shot generalization is statistically significant: In a blind MOS user study on real, unpaired fabric images, FaDeNet achieved a MOS of 3.39/5 and an average rank of 1.97, significantly outperforming comparison methods (\(p < 0.01\) after Holm-Bonferroni correction).
Failure cases indicate performance boundaries: When extremely fine repetitive textile structures are highly entangled with moiré, or when colored moiré overlaps with genuine clothing patterns, the model might over-smooth local textures, indicating room for further improvement in these challenging scenarios.

Highlights & Insights¶

PRISM's round-trip aligned residual injection is an elegant data synthesis design: In conventional synthesis, simulating the entire physical imaging chain with an under-sampling step irreversibly loses high-frequency texture details, meaning clean textures cannot be used as the ground truth. PRISM avoids the typical path of "generating degraded images from scratch" and instead uses "imaging chain simulation \(\rightarrow\) round-trip alignment and pure residual extraction \(\rightarrow\) injection back into original images". This guarantees pixel-level alignment while perfectly preserving the intrinsic clean textures. This paradigm can be extended to other image degradation tasks that require realistic paired data (e.g., defocus deblurring, scanline removal).
The philosophy of "conservative restoration" perfectly fits the essence of the problem: The distinct challenge of fabric demoiréing lies in the thin line between "removing artifacts" and "preserving textures." FaDeNet's confidence mask + amplitude constraint establishes a natural safety boundary: what is not strong is not corrected, and what needs correction is corrected gently. Visualization of the spatial confidence mask (Appendix Fig. 4) reveals that the mask response is not only concentrated in fabric areas but is also positively correlated with moiré severity, showing that the network indeed learns "what actually needs to be corrected."
The dual-branch design of SAGB combining spatial directionality and frequency periodicity is conceptually clear: Each branch targets a physical property of moiré patterns (spatial stripes and frequency periodicity), being non-overlapping yet mutually complementary. The design choice of enabling spectral gating only at coarse scales avoids texture degradation at fine scales, offering generalizable insights into scale-selective processing.
Highly complete experimental framework: Beyond the main synthesis benchmark experiments, the evaluation includes zero-shot transfer to real unpaired scenes, no-reference IQA metrics (ARNIQA, CLIP-IQA+), a 20-person blind MOS user study, paired Wilcoxon signed-rank tests with Holm-Bonferroni multi-comparison correction, and visualizations of failure cases—providing experimental depth that far exceeds standard benchmark papers.

Limitations & Future Work¶

The authors acknowledge that when moiré patterns are highly entangled with fabric textures (such as fine herringbone) or when colored moiré overlaps with genuine printed patterns, the model may still over-suppress texture details. This limitation intrinsically reflects the ill-posed nature of spectral separation, which might be addressed in the future by incorporating stronger texture-aware priors (e.g., diffusion-model-based generative priors).
Although the domain gap between synthetic and real data is validated via zero-shot experiments, the lack of real paired data makes it difficult to precisely measure the quantitative gap. A promising future direction is to introduce a small amount of real paired data for fine-tuning or adopt adversarial learning-based domain adaptation.
The safety of the masking mechanism relies on the U-Net backbone's capability to accurately identify moiré regions. If the mask itself fails (e.g., misidentifying fine grid patterns as moiré), the conservative strategy may instead hinder proper restoration. Exploring uncertainty estimation for masks or integrating multi-scale mask predictions could be valuable avenues.

vs. Screen Demoiréing Methods (DMCNN / MBCNN / WDNet / ESDNet): These methods assume components are spectrally separable and apply global uniform correction strategies. FaDeNet differentiates itself by recognizing the spectral inseparability of fabric scenes, resorting to a mask-gated conservative strategy for spatially adaptive correction, achieving an absolute lead of 3.35 dB in PSNR on PRISM.
vs. PRISM and Existing Moiré Synthesis Methods (UniDemoiré / CycleMoiré): Existing synthesis methods mainly target screen scenes, treating moiré as an independent foreground layer. PRISM's residual injection and round-trip alignment physically guarantee that fabric textures are not lost, producing moiré patterns naturally entangled with fabric textures, which closer matches physical processes.
vs. Liu et al. (Focus/Defocus Demoiréing): This is currently the only prior work dedicated to fabric demoiréing, but it requires specialized hardware to capture focus/defocus image pairs. PRISM's synthesis strategy needs no special hardware and can be scaled extensively with regular datasets.

Rating¶

Novelty: ⭐⭐⭐⭐☆ This work is the first to systematically define the fabric image demoiréing task and construct a dedicated benchmark dataset. The residual injection synthesis strategy in PRISM possesses clear methodological innovation. While individual components of FaDeNet (adaptive decomposition, mask gating, frequency-domain modules) are not entirely new at the molecular level, their specific combination and adaptation to the alignment requirements of the problem show true innovation.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Highly comprehensive coverage including main experiments on the synthetic benchmark, full ablation studies, real-world zero-shot testing, no-reference IQAs, blind MOS user studies, statistical significance tests, failure analysis, and detailed implementation specs in the appendix.
Writing Quality: ⭐⭐⭐⭐⭐ The paper builds the motivation logically from the physical principles of sampling theorems down to the spectral entanglement of fabric moiré. The methodology is naturally aligned with formulas and figures, and results are clearly presented. The appendix provides rich details on hyperparameters, spectral entropy analyses, and mask visualizations to facilitate reproducibility.
Value: ⭐⭐⭐⭐☆ Fills an important yet overlooked research gap in fabric demoiréing, with PRISM offering a standardized benchmark platform and FaDeNet establishing a strong baseline. Clear utility for real-world applications (e-commerce apparel photography, textile quality inspection, virtual try-ons).