Research map / Image generation / Image restoration
UniDemoiré: Towards Universal Image Demoiréing with Data Generation and Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UniDemoiré, which collects a real moiré pattern dataset on white backgrounds, generates more patterns with an LDM, and synthesizes moiré images with a blending module plus a tone refinement network to train universal demoiréing models.
Problem
Image demoiréing models overfit to a single moiré domain and generalize poorly to new moiré patterns and natural images, due to scarce, non-diverse, and unrealistic training data.
Approach
Decouple moiré patterns from image content by capturing pattern-only data on a white screen, expand pattern diversity with a latent diffusion generator, and synthesize realistic moiré images by handcrafted blending refined with a learned tone network using feature-statistic fusion and color/perceptual/TV losses.
Key results
- Trained only on synthesized data, the method reaches 17.92 dB PSNR on UHDM with MBCNN, more than 3.2 dB above the next best synthesis method.
- In cross-dataset evaluation (source UHDM, target FHDMi) with ESDNet-L, PSNR rises from 18.00 dB baseline to 18.94 dB with synthesis, and 20.75 dB with the full method (ESDNet-L).
- Ablation shows removing the Moiré Pattern Generator drops PSNR from 20.75 to 20.16 dB on FHDMi.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with UniDemoiré synthesized data instead of the compared synthesis approach | MBCNN and ESDNet-L · UHDM and FHDMi zero-shot | PSNR: MBCNN UHDM: 14.7826 -> 17.9162; MBCNN FHDMi: 18.5523 -> 19.0094; ESDNet-L UHDM: 15.2269 -> 17.2524; ESDNet-L FHDMi: 18.5 | Helped |
| Train with UniDemoiré synthesized data alongside source-domain real data | MBCNN and ESDNet-L · Cross-dataset evaluation | PSNR: UHDM → FHDMi: 19.4676 -> 19.8625; 20.4014 -> 20.7543; UHDM → TIP: 18.9214 -> 19.3922; 19.3964 -> 19.5009 | Helped |
| Add Moiré Pattern Generator (MPG) (ablation) | UniDemoiré · UHDM → FHDMi | PSNR: 20.1607 -> 20.7543 | Helped |
| Add Tone Refinement Network (TRN) (ablation) | UniDemoiré · UHDM → FHDMi | PSNR: 20.1691 -> 20.7543 | Helped |
| Add perceptual loss (ablation) | TRN in UniDemoiré · UHDM → FHDMi | PSNR: 20.3076 -> 20.7543 | Helped |
| Add color loss (ablation) | TRN in UniDemoiré · UHDM → FHDMi | PSNR: 20.2692 -> 20.7543 | Helped |
| Add total variation loss (ablation) | TRN in UniDemoiré · UHDM → FHDMi | PSNR: 20.3961 -> 20.7543 | Helped |
| Add feature fusion block (ablation) | TRN in UniDemoiré · UHDM → FHDMi | PSNR: 20.2868 -> 20.7543 | Helped |
Surprising findings
- UnDeM using the authors' generated moiré patterns outperforms UnDeM using real TIP moiré patterns in all experiments.
Limitations
- Limitations are stated only in the appendix, which is not fully available in the extract.
- Synthesis depends on a handcrafted blending module plus a TRN trained on real moiré images from TIP, FHDMi, or UHDM.
Experimental setup
Datasets: Moiré Pattern Dataset (new) MHRNID TIP FHDMi UHDM MoireSpace
Benchmarks: UHDM FHDMi TIP
Metrics: PSNR SSIM LPIPS
Baselines: Shooting UnDeM MoireSpace MBCNN baseline without synthesis
Builds on: Latent Diffusion Model Uformer HistoGAN MoireSpace UnDeM
Closest papers by meaning
- Learning Image Demoiréing from Unpaired Real Data (AAAI 2024 · similarity 0.91)
- Direction-Aware Video Demoiréing with Temporal-Guided Bilateral Learning (AAAI 2024 · similarity 0.85)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.85)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.85)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.85)
- IR-CM: The Fast and General-purpose Image Restoration Method Based on Consistency Model (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.