Research map / Image generation / Image restoration
DocNLC: A Document Image Enhancement Framework with Normalized and Latent Contrastive Representation for Multiple Degradations
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 DocNLC, a unified document image enhancement framework that handles multiple degradations via degradation normalization/compensation and a Barlow Twins-style latent contrastive loss, pre-trained on synthetic degraded DIBCO data and fine-tuned on real DIBCO data.
Problem
Document image enhancement is hard because camera-captured document images suffer multiple simultaneous degradations (watermarks, blur, noise, background, shadow), while most existing methods target only one degradation type.
Approach
Instead of pushing apart different degradation types as in SimCLR-style contrastive learning, it normalizes degradation features with instance normalization, adds a compensation branch to retain information, and applies an asymmetric latent contrastive loss (Barlow Twins-style invariance and redundancy reduction) plus a direct pixel-level contrastive loss across degradation types of the same docum
Key results
- Pre-trained DocNLC averages 17.17 PSNR / 0.90 SSIM across four DIBCO test sets, versus 16.03 for BCDUnet and 15.16 for DocEnTr, with 29.84M parameters.
- After fine-tuning, DocNLC reaches 22.15 PSNR on DIBCO2011 and 23.91 on DIBCO2012, outperforming Text-DIAE's 21.29 and 23.66 on those sets, but trails it on DIBCO2018 (18.23 vs 19.95).
- Ablation shows that using background as the primary modality gives the best result (16.31 PSNR), while blur as primary gives the worst (13.08).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune DocNLC on real DIBCO data after pre-training | DocNLC · DIBCO11 | PSNR: 20.30 -> 22.15 | Helped |
| Fine-tune DocNLC on real DIBCO data after pre-training | DocNLC · DIBCO12 | PSNR: 21.91 -> 23.91 | Helped |
| Fine-tune DocNLC on real DIBCO data after pre-training | DocNLC · DIBCO17 | PSNR: 19.64 -> 20.24 | Helped |
| Fine-tune DocNLC on real DIBCO data after pre-training | DocNLC · DIBCO18 | PSNR: 20.18 -> 18.23 | Hurt |
| Add LBCE to LMSE (ablation) | DocNLC · DIBCO11–18 average | SSIM: 0.84 -> 0.87 | Helped |
| Add LLcon to LMSE, LBCE, and LDcon (ablation) | DocNLC · DIBCO11–18 average | PSNR: 16.29 -> 17.17 | Helped |
| Add LDcon to LMSE, LBCE, and LLcon (ablation) | DocNLC · DIBCO11–18 average | PSNR: 16.75 -> 17.17 | Helped |
| Add LMSE to LBCE, LDcon, and LLcon (ablation) | DocNLC · DIBCO11–18 average | PSNR: 14.57 -> 17.17 | Helped |
| Add LBCE to LMSE, LDcon, and LLcon (ablation) | DocNLC · DIBCO11–18 average | PSNR: 16.81 -> 17.17 | Helped |
| Use background rather than blur as the primary modality (ablation) | DocNLC · DIBCO2018 | PSNR: 13.08 -> 16.31 | Helped |
Surprising findings
- Using blur as the primary modality performs worst (13.08 PSNR), which the authors link to low mutual information with the ground truth.
- DocNLC underperforms on DIBCO2018 after fine-tuning, which the authors attribute to text colour being similar to background colour.
Limitations
- Performs poorly on heavily blurred document images.
- Struggles with multi-layered shadows, producing inconsistent character colours.
- Underperforms some baselines on DIBCO2018 after fine-tuning.
Experimental setup
Datasets: DIBCO 2008-2019 DIBCO2011 DIBCO2012 DIBCO2017 DIBCO2018
Benchmarks: DIBCO2011 DIBCO2012 DIBCO2017 DIBCO2018
Metrics: PSNR SSIM
Baselines: U-Net BCDU-Net DE-GAN DocEnTr Otsu Sauvola Cascaded cGAN CMU-Net Text-DIAE DeepOtsu
Builds on: Barlow Twins U-Net instance normalization DeepOtsu
Closest papers by meaning
- DocRes: A Generalist Model Toward Unifying Document Image Restoration Tasks (CVPR 2024 · similarity 0.84)
- Enhancing Visual Document Understanding with Contrastive Learning in Large Visual-Language Models (CVPR 2024 · similarity 0.84)
- Predicting the Original Appearance of Damaged Historical Documents (AAAI 2025 · similarity 0.84)
- Improving Image Restoration through Removing Degradations in Textual Representations (CVPR 2024 · similarity 0.84)
- Descanning: From Scanned to the Original Images with a Color Correction Diffusion Model (AAAI 2024 · similarity 0.83)
- Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration (CVPR 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.