This project aims to design and implement a real-time license plate detection system capable of accurately detecting and localizing license plates in images. The system is robust and capable of handling various environmental conditions, including different lighting conditions, vehicle orientations, and background clutter. The goal is to provide reliable results for further processing or use in applications such as traffic monitoring, parking management, or law enforcement.
The dataset used for this project consists of 367 images, each paired with a label file containing bounding box coordinates for the license plates. The dataset is available here, and it serves as the foundation for training and evaluating the license plate detection model. dataset/
├── images/
│ ├── image1.jpg
│ ├── image2.jpg
│ ├── ...
│ └── image367.jpg
└── labels/
├── image1.txt
├── image2.txt
├── ...
└── image367.txt
The project is organized into the following directories and subdirectories:
- Data: Contains the dataset with subdirectories for images and labels.
- Weights: Stores the best-performing model weights obtained during training.
- Results: Holds the end results of the license plate detection system.
- Model-Inference: Provides example code and documentation for model inference.
- Train-Val-Inference: Includes code and instructions for inference during training and validation phases.
- Evaluation Metrics: Contains scripts and documentation for evaluating model performance, including metrics like precision, recall, and mAP.
We used the YOLOv5s architecture for license plate detection. The model was fine-tuned based on a single class (license plates), and the 80 classes from the COCO dataset were removed in the YOLOv5s configuration file (yml).
- Image Size: 512x512 pixels
- Number of Epochs: 10 (adjust based on dataset size and convergence)
- Learning Rate: lr_scheduler
- Optimizer: smart_optimizer
The training process involved iterating through the dataset, feeding images and labels to the model, calculating loss, and optimizing the model using backpropagation. Training metrics were recorded for evaluation.
Model performance was evaluated using various metrics, including object loss, class loss, precision, recall, and mAP (mean Average Precision). The results are shown below:
train/obj_loss train/cls_loss metrics/precision metrics/recall metrics/mAP_0.5 metrics/mAP_0.5:0.95 val/box_loss val/obj_loss val/cls_loss x/lr0 x/lr1 x/lr2
0.014776 0 0.80822 0.67873 0.7727 0.33475 0.041708 0.010181 0 0.00208 0.00208 0.00208
Problem 1: Dataset Preparation
- Problem: Acquiring and annotating the dataset with license plate bounding boxes was time-consuming and challenging.
- Solution: Careful data collection, and quality control processes were implemented to create a reliable dataset.
Problem 2: Model Fine-Tuning
- Problem: Fine-tuning YOLOv5s for a single class, license plates, required selecting suitable hyperparameters and managing overfitting.
- Solution: Extensive experimentation, hyperparameter tuning, and continuous monitoring of the training process helped achieve the desired model accuracy.
Problem 3: Metric Interpretation
- Problem: Interpreting and optimizing evaluation metrics for object detection tasks, such as precision and recall, was complex.
- Solution: In-depth understanding of evaluation metrics and iterative model improvements were employed to balance precision and recall for real-world scenarios.
These solutions allowed for the successful development of an accurate and robust license plate detection system despite the encountered challenges.
This project successfully fine-tuned the YOLOv8 model for license plate detection. The combination of a well-structured dataset, appropriate training parameters, and evaluation metrics has resulted in a robust license plate detection system.