PBL 4: Deep Learning

Topic: License plate recognition with Deep Neural Networks

---

Summary

This topic aims to implement the LPRNet model [1] with the Chinese license plate dataset compiled from the CCPD2020 dataset [2] and the Vietnamese license plate dataset collected through real photos.

Dataset

Chinese license plate dataset: https://github.com/lephamcong/PBL4_LPRNet/tree/main/Dataset/BiensoxeTrungQuoc

Vietnamese license plate database: https://github.com/lephamcong/PBL4_LPRNet/tree/main/Dataset/BiensoxeVietNam

Directory structure

• Dataset/: This folder contains the Chinese and Vietnamese license plate datasets.
• LPRNet_Pytorch_China/: Contains source code for training and evaluation with the Chinese license plate dataset.
• LPRNet_Pytorch_VietNam/: Contains source code for training and evaluation with the Vietnamese license plate dataset.
• readme.md: this file
• requirements.txt: Required library files (instructions for installing the library are below)

LPRNet Architecture

Dataset

Chinese License Plate Dataset: This dataset is curated from the CCPD2020 collection, which contains approximately 11,776 street-captured images of Chinese license plates. Within the scope of this project, only high-quality samples were selected and enhanced through data augmentation techniques, resulting in a final set of approximately 11,000 images. This data is divided into 10,000 images for training and 1,000 images for evaluation.

Vietnamese License Plate Dataset: This dataset was compiled through real-world photography. After selecting high-quality samples and applying various processing and data augmentation techniques, a total of 13,320 Vietnamese license plate images were obtained. The dataset is split into 12,320 images for training and 1,000 images for evaluation purposes.

Model training

The model training process was conducted in the Google Colab environment with the following hardware resource configuration:

• CPU: Intel(R) Xeon(R) CPU @ 2.00GHz
• RAM: 12GB
• GPU: Tesla T4 12GB VRAM

The following table outlines the specific parameters used for training the models on both Chinese and Vietnamese license plate datasets:

Parameter	Chinese License Plate Data	Vietnamese License Plate Data
max epoch	150	150
img size	[94, 24]	[94, 24]
dropout rate	0.5	0.5
learning rate	0.001	0.001
lpr max len	8	9
train batch size	32	32
test batch size	20	20
momentum	0.9	0.9
weight decay	2e-5	2e-5

Results and evaluation

Model	Dataset	Mean Accuracy	Mean Levenshtein Distance	Processing Time Per Sample
LPRNet	Chinese license plates with 10,000 training and 1,000 evaluation images	90%	0.045	9ms
LPRNet	Vietnamese license plates with 12,320 training and 1,000 evaluation images	75%	0.155	10ms
Tesseract	1,000 Vietnamese license plate images	84%	0.22	129.8ms

How to use source code?

1. Clone repo:

https://github.com/lephamcong/PBL4_LPRNet.git

2. Install the necessary libraries.:

pip install -r requirements.txt

3. Training and evaluation with the Chinese license plate dataset.

cd LPRNet_Pytorch_China

Training (adjusting the path to the dataset)

python train_LPRNet.py

Evaluation (Results are displayed as the Notebook file LPRNet_Pytorch_China.ipynb)

python test_LPRNet.py

4. Training and evaluation with the Vietnamese license plate dataset.

cd LPRNet_Pytorch_VietNam

Training (adjusting the path to the dataset)

python train_LPRNet.py

Evaluation (Results are displayed as the Notebook file LPRNet_Pytorch_VietNam.ipynb)

python test_LPRNet.py

References

[1] Zherzdev, Sergey, and Alexey Gruzdev. "Lprnet: License plate recognition via deep neural networks." arXiv preprint arXiv:1806.10447 (2018).

[2] Xu, Zhenbo, et al.“Towards end-to-end license plate detection and recogni- tion: A large dataset and baseline." Proceedings of the European conference on computer vision (ECCV). 2018

[3] https://github.com/lyl8213/Plate_Recognition-LPRnet

[4] https://github.com/mesakarghm/LPRNET

[5] https://github.com/xuexingyu24/License_Plate_Detection_Pytorch.git

Contact information: lephamcong.bk@gmail.com