Research map / Image generation / Image restoration
Descanning: From Scanned to the Original Images with a Color Correction Diffusion Model
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.
Defines descanning (restoring original digital images from scanned copies), releases the DESCAN-18K dataset, and proposes DescanDiffusion, a color-encoder plus conditional DDPM model with synthetic degradation augmentation.
Problem
Restoring original digital images from scanned copies, which suffer from mixed color-related and non-color-related degradations (noise, halftone, bleed-through, texture distortion, color transition), with no public dataset available.
Approach
A ResNet-34 color encoder predicts per-channel color statistics to globally re-normalize the scanned image; a conditional DDPM starting from the color-corrected image (conditioned on it and the color vector) removes local degradations, with synthetic degradation augmentation for unseen scanners.
Key results
- DescanDiffusion reaches 23.40 dB PSNR, 0.8717 SSIM, 0.042 LPIPS and 13.51 FID on the DESCAN-18K unseen-scanner test set, beating NAFNet (22.03 dB) and commercial apps such as Clear Scan (21.46 dB).
- DescanDiffusion+ with synthetic augmentation gets the best PSNR/SSIM (23.43 dB / 0.8736).
- Compared with DDRM at 256x256, DescanDiffusion gets 25.72 dB PSNR vs 24.14 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add color-corrected image conditioning (CIC) to DDPM (ablation) | DescanDiffusion · DESCAN-18K test | PSNR (dB): 22.72 -> 23.18 | Helped |
| Add color-correction vector conditioning (CVC) (ablation) | DescanDiffusion with CIC · DESCAN-18K test | PSNR (dB): 23.18 -> 23.40 | Helped |
| Add synthetic data generation (SDG) (ablation) | DescanDiffusion with CIC and CVC · DESCAN-18K test | PSNR (dB): 23.40 -> 23.43 | Helped |
| Add synthetic data generation (SDG) to vanilla DDPM (ablation) | Vanilla DDPM · DESCAN-18K test | PSNR (dB): 22.72 -> 23.09 | Helped |
| Use synthetic data generation (ablation) | DescanDiffusion · DESCAN-18K test, unseen scanners | LPIPS: 0.042 -> 0.044 | Hurt |
| Use synthetic data generation (ablation) | DescanDiffusion · DESCAN-18K test, unseen scanners | FID: 13.51 -> 14.60 | Hurt |
| Use synthetic data generation (ablation) | DescanDiffusion · DESCAN-18K test, unseen scanners | PSNR (dB): 23.40 -> 23.43 | Helped |
| Use synthetic data generation (ablation) | DescanDiffusion · DESCAN-18K test, unseen scanners | SSIM: 0.8717 -> 0.8736 | Helped |
Surprising findings
- Applying histogram-matching global color correction improves most baselines, indicating color-related degradations dominate scanned images.
- Adding the synthetic-data scheme to a vanilla DDPM without the color-correction modules yields a performance drop compared with the full model, so global color correction is still needed.
Limitations
- Diffusion inference is much slower (2.58 s) than CNN or transformer baselines.
- Descanning is evaluated mainly on magazine pages from a single publisher source, so diversity of content is limited.
- The paper states that the dataset is licensed CC BY-NC-SA 3.0 (non-commercial terms).
Experimental setup
Datasets: DESCAN-18K DPS dataset OPR dataset DIV2K
Benchmarks: DESCAN-18K DPS dataset OPR dataset
Metrics: PSNR SSIM LPIPS FID NRQM NIQE PI
Baselines: Pix2PixHD CycleGAN HDRUNet Restormer ESDNet NAFNet OPR DPS Clear Scan Adobe Scan Microsoft Lens DDRM
Builds on: DDPM SR3 (conditional DDPM) ResNet UNet
Closest papers by meaning
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.87)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.86)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.85)
- Improving Image Restoration through Removing Degradations in Textual Representations (CVPR 2024 · similarity 0.85)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.85)
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.85)
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.