ALPR

An Automatic License Plate Recognition Algorithm using YOLOv5 and EasyOCR.

Sample Output

The algorithm is divided into 2 stages, first locating license plates (bounding boxes) then readout the characters.

Output: readout characters, license plate confidence level

Introduction

This project is aimed at US License Plate Recognition for parking management. The YOLOv5 framework is adopted from https://github.com/ultralytics/yolov5 and trained using images from OpenALPR benchmark dataset https://github.com/openalpr/benchmarks.

For EasyOCR https://github.com/JaidedAI/EasyOCR, default settings and weights are used.

Environment Setup

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.6. To install run:

$ git clone https://github.com/YuTingChow/ALPR.git
$ cd ALPR
$ pip install -r ./requirements.txt

Optional: Docker build

docker build -t alpr:v1 - < Dockerfile
docker run --gpus all -it alpr:v1

Tests

To test if the environment is setup correctly, you may run

$ python ./src/detect.py

The expected output is as follows:

Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.5, device='', img_size=640, iou_thres=0.5, output='inference/output', save_txt=False, source='tests/images', update=False, view_img=False, weights='best.pt')
Using CUDA device0 _CudaDeviceProperties(name='Your GPU Name', total_memory=XXXXMB)

Fusing layers... 
Model Summary: 191 layers, 7.25509e+06 parameters, 0 gradients
image 1/3 ~/alpr-insight/ALPR/tests/images/car7.jpg: 0   I0DTM59
448x640 1 license_plates, Done. (0.069s)
image 2/3 ~/alpr-insight/ALPR/tests/images/car8.jpg: 0   FZRULZ
640x512 1 license_plates, Done. (0.042s)
image 3/3 ~/alpr-insight/ALPR/tests/images/car9-2.jpg: 0   4639520
384x640 1 license_plates, Done. (0.043s)
Results saved to inference/output
Done.

Inference

Inference can be run on images or videos. Results are saved to './inference/output'

$ python ./src/detect.py --source  file.jpg  # image 
                            file.mp4  # video
                            path/  # directory
                            path/*.jpg  # glob

To run inference on images in the './tests/images' folder:

$ python ./src/detect.py --source ./tests/images/

Custom Training

You can train the network from scratch with your own data as follows:

$ python ./src/train.py --data ./data/licensePlate.yaml --cfg models/yolov5s.yaml --weights '' --batch-size 16

You just need to replace the data path to the training and validation set in licensePlate.yaml.

For the requirements of data labelling, please refer to this wiki.

pubpy2015/ALPR