Research map / Image generation / Image restoration
Learning Image Demoiréing from Unpaired Real Data
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.
Learns image demoiréing from unpaired real moiré and clean images by synthesizing complexity-grouped pseudo-moiré patches with a GAN-based generator and filtering low-quality pairs via adaptive edge-based denoising.
Problem
Image demoiréing normally needs paired real moiré and moiré-free images, which are costly to capture; shooting-simulation synthetic pairs have large domain gaps, so the paper asks how to get useful training pairs from unpaired real data.
Approach
Split real moiré images into patches, group them by moiré complexity (Laplacian frequency and colorfulness), train a per-group adversarial synthesis network that transfers moiré features onto unpaired clean patches while keeping content via a content encoder, then discard pseudo pairs whose edge-structure difference falls below a per-group percentile threshold before training off-the-shelf demoiré
Key results
- On FHDMi with MBCNN at 384 crop, UnDeM reaches 19.89 dB PSNR versus 19.36 dB for cyclic learning and 9.32 dB for shooting simulation.
- On UHDM with ESDNet-L at 768 crop, UnDeM reaches 18.40 dB PSNR versus 18.00 dB for cyclic learning; paired real data reaches 22.12 dB.
- Ablation on MBCNN (FHDMi, crop 192): removing image preprocessing drops PSNR from 19.45 to 16.62 dB, and removing adaptive denoise drops it to 18.96 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on UnDeM-generated pairs instead of cyclic-learning data | MBCNN · FHDMi, crop 384 | PSNR: 19.36 -> 19.89 | Helped |
| Train on UnDeM-generated pairs instead of cyclic-learning data | ESDNet-L · FHDMi, crop 384 | PSNR: 19.05 -> 19.66 | Helped |
| Train on UnDeM-generated pairs instead of cyclic-learning data | MBCNN · UHDM, crop 192, 384, and 768 | PSNR: +0.54 dB, 0.10 dB, and 0.15 dB | Helped |
| Train on UnDeM-generated pairs instead of cyclic-learning data | ESDNet-L · UHDM, crop 192, 384, and 768 | PSNR: 0.28 dB, 0.43 dB, and 0.40 | Helped |
| add image preprocessing (ablation) | UnDeM + MBCNN · FHDMi, crop 192 | PSNR: 16.62 -> 19.45; 2.83 dB | Helped |
| add adaptive denoise (ablation) | UnDeM + MBCNN · FHDMi, crop 192 | PSNR: 18.96 -> 19.45; 0.49 dB | Helped |
| add adversarial loss (ablation) | UnDeM + MBCNN · FHDMi, crop 192 | PSNR: 7.22 -> 19.45; 12.23 dB | Helped |
| add feature loss (ablation) | UnDeM + MBCNN · FHDMi, crop 192 | PSNR: 18.38 -> 19.45; 1.07 dB | Helped |
| add content loss (ablation) | UnDeM + MBCNN · FHDMi, crop 192 | PSNR: 17.90 -> 19.45; 1.55 dB | Helped |
Surprising findings
- Shooting-simulated pseudo moiré data yields very poor demoiréing (around 9-10 dB PSNR), far below cyclic learning and UnDeM.
- Removing the adversarial discriminator loss causes a catastrophic drop (about 12 dB PSNR on FHDMi), showing the adversarial component is critical.
Limitations
- Performance still falls well short of models trained on real paired data (e.g., 19.89 vs 22.73 dB on FHDMi with MBCNN at 384).
- Pseudo-moiré quality improvement remains an open question; the paper suggests better preprocessing, dedicated synthesis networks, and special losses as future directions.
- Adaptive denoise thresholds (gamma_i) were tuned empirically on FHDMi with MBCNN and reused across networks and crop sizes.
Experimental setup
Datasets: FHDMi UHDM
Benchmarks: FHDMi UHDM
Metrics: PSNR SSIM LPIPS
Baselines: Paired training Shooting simulation Cyclic moiré learning MBCNN ESDNet-L
Builds on: cyclic moiré learning PatchGAN least squares GAN MBCNN ESDNet
Closest papers by meaning
- UniDemoiré: Towards Universal Image Demoiréing with Data Generation and Synthesis (AAAI 2025 · similarity 0.91)
- QuantDemoire: Quantization with Outlier Aware for Image Demoiréing (NeurIPS 2026 · similarity 0.85)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.85)
- Slight Corruption in Pre-training Data Makes Better Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.84)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · 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.