platerec

platerec is a lightweight package for reading license plates using an ONNX model. It is designed to be part of a pipeline for detecting, cropping, and reading license plates. The underlying model is a mobilenetv2 as encoder and a light gpt for decoder. The training data comprises primarily Brazilian license plates, sourced from internet images, also synthetic data generated in the same font with transforms. The model repository can be found here.

Now it supports reading from different countries. Currently:

• [br]: Brazil
• [us]: United States
• [ue]: European Union
• [ru]: Russia

The license plate recognition model is continuously being improved. However, accuracy can be significantly enhanced by fine-tuning the model with a dataset specific to your needs. We encourage you to explore the model training repository to learn how to build a customized model with increased accuracy for your format.

The first token after '<' will be the country plate type.

Example: '<[br]ZZZ1Z11>'

Video example

Installation

To install the required dependencies, use the following command:

For cpu

pip install "platerec[cpu]"

For cuda

pip install "platerec[gpu]"

Usage

Command Line Interface

You can use the command line interface to detect license plates in an image:

platerec image_path [--encoder_path ENCODER_PATH] [--decoder_path DECODER_PATH] [--return_types RETURN_TYPE] [--providers PROVIDERS] [--no_platedet]

Arguments

• image_path: Path to the input image. Could be more than one image.
• --encoder_path: Path to the ONNX encoder model (default: artifacts/encoder.onnx).
• --decoder_path: Path to the ONNX decoder model (default: artifacts/decoder.onnx).
• --tokenizer_path: Path to the tokenizer json file (default: artifacts/tokenizer.json).
• --return_type: Output formats (choices: word, char). Word return the plate text and confidence detected, char return the plate chars detected with confidences for each char.
• --providers: ONNX Runtime providers (default: CPUExecutionProvider).
• --no_platedet: Not use platedet to detect plates first.

Example

To just read an already cropped image:

python3 platerec/cli.py examples/1.jpg --return_type word

To detect license plates and read them:

python3 platerec/cli.py examples/1.jpg --return_type word

Using in Code

To just read an already cropped image:

from PIL import Image
from platerec import Platerec

platerec = Platerec()
image = Image.open('examples/1.jpg')
pred = platerec.read(image)

pred will be something like:

{'word': '[br]ZZZ1Z11', 'confidence': 0.98828125}

To detect license plates and read them:

from PIL import Image
from platerec import Platerec

platerec = Platerec()
image = Image.open('examples/1.jpg')
crops = platerec.detect_read(image)
for idx, crop in enumerate(crops['pil']['images']):
    crop.save(f'{idx}.jpg')

pred will be something like:

{'images': [<PIL.Image.Image image mode=RGB size=105x40 at 0x7FEE25B67AD0>], 'confidences': array([0.72949219]), 'words': ['[br]AAA1A11'], 'boxes': array([[ 393, 1188,  498, 1228]], dtype=int32), 'words_confidences': [0.95263671875]}

If you want to use CUDA:

from platerec import Platerec

platerec = Platerec(providers=["CUDAExecutionProvider"])

Check all execution providers here.

---

Extra commands for quick testing:

platerec-video video_path [--font_size FONT_SIZE] [--save_output]

Run platerec on a video file.

platerec-image image_path

Run platerec on a image file.