Realistic License Plate Restoration and Recognition Dataset (RLPR)

[Dataset Description]
This dataset was constructed to facilitate the development of methods aimed at improving image quality and recognition accuracy of license plates captured from dashcam videos. It contains paired low-quality license plate video sequences and corresponding high-quality target images, enabling quality enhancement model training and evaluation.
Each sample includes a sequence of 31 low-quality license plate frames alongside a single high-quality license plate image that serves as the target (pseudo-ground truth) for quality enhancement. This dataset can be used for training models to enhance image quality and includes recognition labels for assessing recognition accuracy. Additionally, the dataset provides outputs obtained using the MF-LPR² model proposed in our paper, offering a benchmark for performance comparison.

[Dataset Contents]
1) Low-Quality (LQ) License Plate Image Sequences
- 31 frames per sequence.
2) High-Quality (HQ) License Plate Images (Pseudo-Ground Truth)
- Includes both the original images and versions cropped to the Region of Interest (ROI).
- ROI coordinates provided.
3) License Plate Recognition Label
- For privacy reasons, only the numerical characters of the license plate are provided.
4) MF-LPR² Model Super-Resolution (SR) Results
- Outputs generated by the proposed MF-LPR² model
5) Homography Transformation Coordinates
- Coordinates for aligning low- and high-quality images
- Utility tool is also provided

[Citation]
When using this dataset, please cite the following paper:
Na, K., Oh, J., Cho, Y., Kim, B., Cho, S., Choi, J., & Kim, I. (2025). MF-LPR2: Multi-frame license plate image restoration and recognition using optical flow. Computer Vision and Image Understanding, 256, 104361.

[Disclaimer]
All Korean Hangul characters in the license plate regions have been pseudonymized using targeted blurring techniques. In the low-resolution (LR) image sequences, the frames have been cropped to the license plate region to minimize the exposure of the surrounding vehicle or environment. In some cases, small portions of the vehicle may remain visible due to cropping margins. The recognition labels contain only numerical values. No GPS metadata, timestamps, vehicle make/model information, or environmental background data is included in this dataset.

This dataset is released strictly for non-commercial, academic research purposes. By accessing or using this dataset, you agree to the following terms:
* You will not attempt any form of re-identification or misuse of the data.
* You will not use the dataset for surveillance, law enforcement, or any commercial applications.
* You will cite the original paper when publishing any results that use this dataset.
* You will comply with all applicable data protection and privacy laws.
The authors and their affiliated institutions assume no legal responsibility for any use of this dataset that extends beyond its intended academic research purpose.

[数据集描述]
本数据集旨在助力面向提升行车记录仪视频采集的车牌图像质量与识别准确率的方法研发。其包含成对的低质量（LQ）车牌视频序列与对应的高质量（HQ）目标图像，可用于质量增强模型的训练与评估。
每个样本包含31帧低质量车牌帧序列，以及单张用作质量增强目标（伪真值）的高质量车牌图像。本数据集可用于训练图像质量增强模型，并附带用于评估识别准确率的识别标签。此外，本数据集提供了本文提出的MF-LPR²模型生成的输出结果，可为性能对比提供基准测试集。

[数据集内容]
1) 低质量（LQ）车牌图像序列
- 每个序列含31帧图像。
2) 高质量（HQ）车牌图像（伪真值）
- 包含原始图像与裁剪至感兴趣区域（ROI）的版本。
- 附带感兴趣区域坐标。
3) 车牌识别标签
- 出于隐私保护考虑，仅提供车牌的数字字符。
4) MF-LPR²模型超分辨率（SR）结果
- 本文提出的MF-LPR²模型生成的输出结果
5) 单应变换坐标
- 用于对齐低质量与高质量图像的坐标
- 附带配套实用工具

[引用说明]
使用本数据集时，请引用以下论文：
Na, K., Oh, J., Cho, Y., Kim, B., Cho, S., Choi, J., & Kim, I. (2025). MF-LPR2：基于光流的多帧车牌图像复原与识别. 《计算机视觉与图像理解》, 256, 104361.

[免责声明]
车牌区域内的所有韩文字符均通过针对性模糊技术进行了匿名化处理。在低分辨率（LR）图像序列中，帧已被裁剪至车牌区域，以最大程度减少周边车辆或环境的暴露。部分情况下，由于裁剪边界的存在，可能会残留少量车辆区域。识别标签仅包含数字值。本数据集未包含GPS元数据、时间戳、车辆品牌/型号信息或环境背景数据。
本数据集仅用于非商业性学术研究目的。访问或使用本数据集即表示您同意以下条款：
* 不得尝试对数据进行任何形式的重新识别或滥用。
* 不得将本数据集用于监控、执法或任何商业用途。
* 若使用本数据集发表研究成果，请引用原论文。
* 遵守所有适用的数据保护与隐私法律法规。
若超出本数据集原定的学术研究用途进行使用，作者及其所属机构不承担任何法律责任。