Research map / Image generation / Image restoration
Towards Automated Chinese Ancient Character Restoration: A Diffusion-Based Method with a New Dataset
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.
Introduces ARMCD, a dataset of ~15.5k real ancient Chinese single-character images with synthetic erosion masks, and DiffACR, a cold-diffusion-based restoration method that uses the eroded image's white pixels as a prior mask.
Problem
Automated Chinese ancient character restoration suffers from small paired datasets and non-professional synthetic masks that do not match real erosion, leading to overfitting and poor restoration of real eroded characters.
Approach
Treat real erosion as a degradation ("erosionfication") built from overlaid local erosion masks and apply cold diffusion to invert it, conditioning restoration on a prior mask extracted directly from the eroded image's white pixels.
Key results
- DiffACR achieves FID 12.87 and SSIM 0.9988 on ARMCD, the best FID and SSIM among compared methods; CharFormer and SwinIR have higher PSNR (24.38, 24.08).
- Erosionfication outperforms Gaussian noise and snowification degradations in ablation (MAE 0.0187 vs 0.0563 and 0.0356).
- In a 2AFC user study, 63% of synthetic erosion was judged real by restoration experts and 55% by graduate students.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use erosionfication instead of Gaussian noise (ablation) | DiffACR · ARMCD | MAE: 0.0563 -> 0.0187 | Helped |
| use erosionfication instead of snowification (ablation) | DiffACR · ARMCD | MAE: 0.0356 -> 0.0187 | Helped |
| add prior mask guidance in all stages (ablation) | DiffACR · ARMCD | MAE: 0.0287 -> 0.0187 | Helped |
| add prior mask guidance in training (ablation) | DiffACR · ARMCD | MAE: 0.0259 -> 0.0187 | Helped |
| add prior mask guidance in sampling (ablation) | DiffACR · ARMCD | MAE: 0.0236 -> 0.0187 | Helped |
Surprising findings
- DiffACR does not have the best PSNR or MAE; SwinIR and CharFormer have higher PSNR and CIDG/IPT-level MAE differences are small, so gains are metric-dependent.
- Restoration performance collapses for eroded masks above 60% coverage (PSNR 8.35 at 81%+).
Limitations
- Only 460 real eroded images are available, so the real-erosion distribution is only partially captured by the synthetic masks.
- Performance degrades sharply at high mask ratios (above 60%).
- Evaluation relies on synthetic masks applied to uneroded images rather than paired real restorations.
Experimental setup
Datasets: ARMCD
Benchmarks: ARMCD
Metrics: MAE PSNR SSIM FID LPIPS
Baselines: DNCNN Cycle-Dehaze VDN CIDG SCCGAN SGGAN SwinIR IPT CharFormer
Builds on: cold diffusion RePaint DDPM snowification
Closest papers by meaning
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.85)
- Component-Level Segmentation for Oracle Bone Inscription Decipherment (AAAI 2025 · similarity 0.84)
- Descanning: From Scanned to the Original Images with a Color Correction Diffusion Model (AAAI 2024 · similarity 0.82)
- A Multi-Style Chinese Characters Writing Intelligent Tool Based on Small-scale Training Data (AAAI 2025 · similarity 0.82)
- Diffusion-based Blind Text Image Super-Resolution (CVPR 2024 · similarity 0.82)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.82)
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.