This project implements logic for ALPR training and primitive inference using a novel object detection model, D-FINE. The algorithm operates in the following steps:
- Detect all possible license plates in the image
- Extract the license plates
- Run symbol recognition. Symbols are located and classified in one-shot using modified D-Fine architecture
- Enjoy the results!
To get started with ISS, follow these steps:
-
Clone the repository including submodules:
git clone --recurse-submodules https://github.com/ATMI/Plates-Recognition.git -
Navigate to the submodule directory:
cd Plates-Recognition/model/DFine -
Install required packages:
pip install -r requirements.txt -
Download model checkpoints: Google Drive Link
├── data ├── model ├── script ├── src ├── notebook.ipynb └── README.md
Place where all datasets and images are stored. Contains too many files to be published on GitHub/GitLib, provided upon a request.
data ├── demo - images used in demos ├── yolo - original datasets in YOLO format ├── coco - the same YOLO datasets but in CoCo format ├── autocrop - license plate images cropped by the model └── autorecognize - license plate symbols extracted by the model
Directory with different models used in the project. Currently only D-FINE is present.
DFINE ├── configs │ ├── dataset - dataset descriptions │ └── dfine │ └── custom │ ├── plate_detection_n.yml - config for plate detection model (nano) │ └── plate_recognition_n.yml - config for plate recognition model (nano) └── output - checkpoints, training logs
Various utility scripts used during training/inference.
data ├── cocosplit.py - splits coco dataset into train and test parts ├── detect.py - detects the license plate ├── main.py - combines detect.py and recognize.py ├── recognize.py - recognizes symbols in the extracted license plate ├── recognize_correct.py - selects correct annotations └── yolo2coco.py - converts YOLO dataset description to CoCo
Source code for image loading, processing, geometry, dataset handling, etc.
data ├── coco ├── util └── box.py