This project presents a License Plate Recognition (LPR) system using YOLO models for object detection and EasyOCR for text recognition, designed to achieve real-time, accurate license plate detection. Trained on a diverse, hybrid dataset of over 52,000 labeled images, the system demonstrates strong performance with high precision and recall, making it suitable for applications like traffic monitoring, automated tolling, and parking management.
- YOLO-based Object Detection: Uses YOLO models for detecting license plates in real-time.
- EasyOCR for Text Recognition: Extracts alphanumeric characters from detected license plates.
- Hybrid Dataset: Combines three different datasets containing a variety of challenging scenarios such as occlusions, lighting inconsistencies, and pose variations.
- Data Augmentation: Includes transformations like flipping, cropping, rotation, and color adjustments to simulate real-world conditions and improve model robustness.
The dataset consists of images sourced from multiple projects:
- Dataset 1: 21,175 images from a license plate recognition computer vision project.
- Dataset 2: 6,784 images from the license plates of vehicles in Turkey project.
- Dataset 3: 24,242 images from a vehicle registration plates project.
This hybrid dataset contains a total of 52,201 labeled images, split into training, validation, and test sets:
- Training: 46,278 images (87%)
- Validation: 3,957 images (8%)
- Test: 1,966 images (4%)
- Resize: All images were resized to 640x640 pixels.
- Data Augmentation: Includes horizontal flipping, rotation, cropping, grayscale conversion, and adjustments in brightness, hue, saturation, and contrast.
We evaluated several YOLO models, including:
- YOLOv8 and YOLOv11 (with various configurations: n, s, m, l, x).
- EasyOCR for optical character recognition from detected license plates.
- YOLOv11x achieved the highest mean Average Precision (mAP) of 0.98466 at IoU 50% and 0.71605 at IoU 50-95% after 20 epochs.
- YOLOv11n, a lightweight model, achieved high performance with similar precision and recall values after 100 epochs.
- The YOLOv11m model achieved the best balance of performance and accuracy, with the highest mAP@50-95 of 0.71743.
The system was evaluated using Precision, Recall, and mean Average Precision (mAP) at both IoU 50% and IoU 50-95%. The results showed improvements in recall and mAP with extended training, particularly in YOLOv8 and YOLOv11 models trained for 100 epochs.
-
YOLOv8 (10 epochs) achieved:
- Precision (P): 0.974
- Recall (R): 0.954
- mAP@50: 0.978
- mAP@50-95: 0.682
-
YOLOv11 (10 epochs) achieved:
- Precision (P): 0.981
- Recall (R): 0.951
- mAP@50: 0.981
- mAP@50-95: 0.682
Training for 100 epochs significantly enhanced detection capabilities, improving mAP@50-95 by up to 4.25%.
An Ablation Study was conducted to evaluate the performance of different variants of YOLOv11 models trained on the hybrid dataset. The study aimed to assess the trade-offs between model precision, recall, and computational efficiency.
The following YOLOv11 variants were evaluated:
- YOLOv11n (lightweight)
- YOLOv11s
- YOLOv11m
- YOLOv11l
- YOLOv11x (largest model)
The results of the ablation study are summarized in the table below:
| Model | Precision (B) | Recall (B) | mAP@50 (B) | mAP@50-95 (B) |
|---|---|---|---|---|
| YOLOv11x | 0.97384 | 0.95932 | 0.98387 | 0.71605 |
| YOLOv11l | 0.96733 | 0.96591 | 0.98497 | 0.71454 |
| YOLOv11m | 0.97035 | 0.96761 | 0.98582 | 0.71743 |
| YOLOv11s | 0.97312 | 0.96170 | 0.98477 | 0.71344 |
| YOLOv11n | 0.97125 | 0.95849 | 0.98235 | 0.71064 |
- YOLOv11m achieved the highest mAP@50-95 (0.71743), making it the best choice for applications requiring high accuracy.
- YOLOv11x achieved the highest Precision (B) (0.97384), suitable for applications prioritizing detection accuracy.
- YOLOv11l achieved the highest Recall (B) (0.96591), indicating its capability to minimize missed detections.
- YOLOv11n and YOLOv11s are lightweight models that balance computational efficiency with satisfactory detection metrics, making them ideal for deployment on edge devices.
These results provide insights into the trade-offs between precision, recall, and computational requirements for each variant, allowing tailored model selection based on application-specific needs.