Korean License Plate Recognition

This repository is no longer being updated. For a more readable version of License Plate Recognition using Jax, please refer here.

This repository is based on Self-supervised Implicit Glyph Attention for Text Recognition. It is a Tensorflow implementation of Korean license plate recognition.

Demo

Requirements

• Python 3.11.6
• Tensorflow 2.15.0
• numba
• numpy
• opencv-python
• Pillow

Dataset

• Real Korean License Plate Dataset
• Synthetic Korean License Plate Dataset

Training

python train.py

Evaluation

python model/eval.py

Model Architecture

Model Performance

The best model saved in ./checkpoints/backup/best.keras.

task	accuracy
LPR w/ RE	100.00 %
LPR w/o RE	99.15 %
Character Recognition	99.89 %

* All accuracy is calculated on unquantized model

Computational Cost

task	parameters	FLOPs	size
LPR deployment	32,207	9.58 M	83 KB

Speed of Inference

task	platform	quantization	time
LPR w/o RE	Apple M2	uint8	0.14 ms
LPR w/o RE	Intel i9-10900K	uint8	- ms
LPR w/o RE	AMD EPYC 7V12	uint8	0.42 ms
LPR w/o RE	Coral Edge TPU	uint8	- ms

One More Thing

The jax implementation of this repository is available at here.