LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate

Haoyan Gong¹, Zhenrong Zhang¹, Yuzheng Feng¹, Anh Nguyen², Hongbin Liu*^1
1Xi’an Jiaotong-Liverpool University, ²University of Liverpool
Contact: m.g.haoyan@gmail.com

---

📝 Abstract

License plate (LP) recognition is crucial for intelligent traffic management. Real-world LP images are often severely degraded due to distance and camera quality, making restoration extremely challenging.
We introduce the first real-world multi-frame paired LP restoration dataset (MDLP, 11,006 groups) and a diffusion-based restoration model LP-Diff featuring: Inter-frame Cross Attention for multi-frame fusion; Texture Enhancement for recovering fine details; Dual-Pathway Fusion for effective channel/spatial selection Our method outperforms prior SOTA on real LP images, both quantitatively and visually.

---

🔥 Highlights

• [MDLP Dataset]: First real-world, paired, multi-frame LP restoration dataset (11,006 groups).

• [Diffusion-based Model]: Custom architecture tailored for license plate restoration.

• [SOTA Performance]: Best on MDLP for both image quality and LP recognition.

---

🌟 Visual Results

Qualitative comparison on real-world LP images:

These are some confusing letters and complex Chinese characters:

📊 Quantitative Results

Method	PSNR ↑	SSIM ↑	FID ↓	LPIPS ↓	NED ↓	ACC ↑
SRCNN	14.01	0.195	248.3	0.517	0.626	0.041
HAT	14.16	0.250	229.6	0.413	0.613	0.050
Real-ESRGAN	13.93	0.369	31.0	0.176	0.279	0.161
ResDiff	12.00	0.269	35.9	0.277	0.292	0.159
ResShift	12.53	0.321	89.1	0.288	0.332	0.099
LP-Diff	14.40	0.393	22.0	0.159	0.198	0.305

(On MDLP real-world test set. NED: normalized edit distance; ACC: text recognition accuracy)

---

🏗️ Model Overview

• ICAM: Inter-frame Cross Attention Module

• TEM: Texture Enhancement Module

• DFM: Dual-Pathway Fusion Module

• RCDM: Residual Condition Diffusion Module

---

📚 Dataset

The MDLP Dataset consists of 11,006 groups of real-world degraded license plate images. The dataset was collected under diverse real-world conditions, including various distances, illumination changes, and weather conditions. It provides multi-frame degraded images with corresponding clear ground-truth images for robust restoration model training.

Dataset collection pipeline:

Example images:

Detail of one license plate image:

---

🚀 Getting Started

1. Installation

• Install Python and required dependencies.

• Then install remaining Python packages:

  pip install -r requirements.txt

2. Download MDLP Dataset

• Google Drive

• Baidu Netdisk (access code: 1ebm)

3. Training & Evaluation

• Training:

  python run.py -p train -c ./config/LP-Diff.json

• Validation:

  python run.py -p val -c ./config/LP-Diff.json

• Results and checkpoints are saved in ./experiments.

---

📂 Project Structure

LP-Diff/
│
├── config/              # Training and testing config files
├── data/                # Data loading scripts
├── experiments/         # Model checkpoints and logs
├── figs/                # Visualization images for README and paper
├── models/              # Model implementations
├── requirements.txt     # Python dependencies
└── run.py               # Main training/testing script

---

📖 Citation

If you use this work or dataset, please cite:

@inproceedings{gong2025lp,
  title={LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate},
  author={Gong, Haoyan and Zhang, Zhenrong and Feng, Yuzheng and Nguyen, Anh and Liu, Hongbin},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={17831--17840},
  year={2025}
}

---

🤝 Acknowledgements

This project is based on the excellent ResDiff codebase.
We gratefully acknowledge all related open-source works.

---

💬 Contact

For questions, open an issue or email:
m.g.haoyan@gmail.com

---