Research map / Image generation / Image restoration
CMA: A Chromaticity Map Adapter for Robust Detection of Screen-Recapture Document Images
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 a chromaticity map adapter (CMA) that extracts RGB chromaticity maps from document images and feeds them as prompt tokens into a frozen ViT (VPT-pretrained) to detect screen-recaptured documents, plus a new ROD dataset.
Problem
Detecting screen-recaptured document images that have been degraded by online social network transmission (compression, resizing, blur), where moiré cues are weakened, and generalizing across document content and acquisition devices.
Approach
A pixel-level sub-pixel resampling distortion model shows that recaptured dark edges produce color imbalance; a normalized-RGB chromaticity map captures these artifacts, which an adapter converts to prompt tokens fed to a frozen ViT alongside RGB tokens.
Key results
- On ROD HQ→ROD LQ, CMA reaches AUC 0.8991 and EER 20.79% versus 35.64% EER for RGB-only ViT (average EER reduced from 26.82% to 16.78% per abstract).
- Under JPEG compression (QF=70), CMA attains AUC 0.8688 versus 0.7402 for RGB-only VPT-style baseline in the robustness table.
- On ROD HQ→ROD M&F, CMA achieves AUC 0.9475 and EER 12.77%, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the CMA adapter to feed chromaticity features as prompts alongside RGB | ViT · ROD HQ→ROD M&F | AUC: 0.8094 -> 0.9475 | Helped |
| Use the CMA adapter to feed chromaticity features as prompts alongside RGB | ViT · ROD HQ→DLC2021 | AUC: 0.7836 -> 0.8489 | Helped |
| Use the CMA adapter instead of directly inputting RGB and chromaticity tokens | ViT · ROD HQ→ROD LQ, Origin | AUC: 0.7714 -> 0.8991 | Helped |
| Add chromaticity map extraction to the CMA method (ablation) | CMA · ROD HQ→ROD LQ | AUC: a reduction of 0.0952 in AUC | Helped |
| Add chromaticity map extraction to the CMA method (ablation) | CMA · ROD HQ→ROD LQ | EER: an increment of 12.61 p.p. in EER | Helped |
| Use the CMA adapter rather than directly inputting multimodal data to ViT (ablation) | ViT with RGB and chromaticity · ROD HQ→ROD LQ | EER: an increment of 13.84 p.p. | Helped |
| Apply Gaussian blur (3) | CMA · ROD HQ→ROD LQ | AUC: 0.8991 -> 0.6282 | Hurt |
Surprising findings
- Chromaticity-only models underperform on ROD LQ, so neither RGB nor chromaticity alone is reliable, while combining them helps.
- A plain multi-modal ViT with chromaticity linear embeddings degrades to 34.63% EER, worse than the adapter-based design.
Limitations
- Performance drops notably under Gaussian blur, where chromaticity maps become less effective.
- Distortion model is one-dimensional, horizontal only; 2D and perspective effects are not modeled.
- Cross-dataset generalization to DLC2021 remains weaker (AUC 0.8489) and the approach depends on high-contrast dark-on-light content.
Experimental setup
Datasets: ROD HQ ROD LQ ROD M&F DLC2021 MP-DocVQA Find it again!
Benchmarks: ROD LQ ROD M&F DLC2021
Metrics: AUC EER
Baselines: ViT VPT MobileNetV2 ResNet50 ResNeXt101 LTC-PE CNN + ViT CMA (w/o Ext.)
Builds on: Visual Prompt Tuning ViT-adapter Cross-modal focal loss Moiré pattern augmentation
Closest papers by meaning
- Efficient and Robust Color Consistency for Community Photo Collections (CVPR 2016 · similarity 0.82)
- RECOVER: Reliable Detection of Unauthorized Data Usage in Text-to-Image Diffusion Models via Inversion Robustness (ICML 2026 · similarity 0.82)
- Descanning: From Scanned to the Original Images with a Color Correction Diffusion Model (AAAI 2024 · similarity 0.82)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.80)
- DCTMamba: Advancing JPEG Image Restoration Through Long-Sequence Modeling and Adaptive Frequency Strategy (AAAI 2025 · similarity 0.79)
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.79)
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.