| --- |
| license: agpl-3.0 |
| tags: |
| - computer-vision |
| - object-detection |
| - image-to-text |
| - yolov8 |
| - svtr |
| - alpr |
| - tensorrt |
| - automatic-license-plate-recognition |
| - automatic-number-plate-recognition |
| - license-plate-recognition |
| - onnxruntime |
| - plate-recognition |
| - plate-detection |
| - ocr |
| pipeline_tag: image-to-text |
| --- |
| |
| # LiteALPR Models |
|
|
| Pretrained weights for **LiteALPR**, a lightweight open-source library for fast and accurate Automatic License Plate Recognition (ALPR), explicitly optimized for high-throughput execution on both edge and cloud environments. |
|
|
| ## Model Details |
| - **Architecture**: |
| - **Detector**: YOLOv8n-Efficient (utilizing Ghost Convolution via lightweight C3Ghost blocks). |
| - **Recognizer**: SVTR26-Tiny (featuring an Efficient RCTC Decoder to bypass computationally heavy attention mechanisms). |
| - **License**: GNU Affero General Public License v3.0 (AGPL-3.0) |
| - **Repository**: [LiteALPR GitHub](https://github.com/vn-anhnth/LiteALPR) |
|
|
| ## Training Data |
| To ensure robust, multi-national generalization, the models were trained and validated on a large-scale aggregated dataset of 52,595 images from diverse geographic regions: |
| - **Brazil**: [RodoSol-ALPR Dataset](https://github.com/laroca/RodoSol-ALPR) |
| - **China**: [CBLPRD Dataset](https://github.com/detectRecog/CBLPRD) |
| - **Vietnam**: [Kaggle License Plates Dataset](https://www.kaggle.com/datasets/duydieunguyen/licenseplates) alongside proprietary self-collected real-world traffic footage. |
|
|
| Half of the recognition crops were synthetically degraded (e.g., motion blur, Gaussian noise, perspective skew) during training to ensure high resilience against challenging real-world scenarios. |
|
|
| ## Evaluation Results |
| - **Detection (YOLOv8n-Efficient)**: 99.45% mAP@50, 5.69 GFLOPs, 2.00M parameters. |
| - **Recognition (SVTR26-Tiny)**: 89.15% Sequence Accuracy, 3.28% Character Error Rate (CER), 4.22M parameters. |
| - **End-to-End Speed**: 66.5 FPS on GPU (NVIDIA RTX 3060) and 23.4 FPS on CPU (AMD Ryzen 5 4600G) using ONNX FP32 execution. |
|
|
| ## Known Limitations |
| - The model currently provides highly accurate out-of-the-box support for **single-line alphanumeric license plates** (A–Z, Đ, 0–9) typical of many global regions. |
| - Recognition accuracy on heavily curved multi-line plates or non-Latin characters (e.g., Cyrillic, Arabic) may require custom fine-tuning using the provided training scripts. |
|
|
| ## How to Cite |
| If you use these models in your research, please cite our paper: |
|
|
| ```bibtex |
| @article{nguyen2026litealpr, |
| title={LiteALPR: A lightweight open-source library for fast and accurate license plate recognition}, |
| author={Nguyen, Anh and Hoang, Vinh Truong}, |
| journal={SoftwareX}, |
| year={2026} |
| } |
