This project focuses on Automatic Number Plate Recognition (ANPR) and vehicle classification based on the PLUS highway toll classification system in Malaysia. The model is trained using YOLOv8 for object detection and classification.
- Dataset includes images of Malaysian number plates and various vehicle classes.
- Number plates are detected and classified separately.
- Vehicles are categorized based on PLUS toll classification.
- Dataset augmentation was performed to improve generalization.
- Model: YOLOv8
- Training Images: 44
- Validation Images: 10
- Epochs: 300
- Batch Size: 4
- Image Size (imgsz): 640
- Best F1 Score: 0.58 at Confidence 0.615
- Class-wise Performance:
- Class 1: High precision and recall.
- Number Plate: Moderate precision but lower recall.
- Class 2: Minimal detections, requires further training.
- Box Loss (Bounding Box Regression): Decreasing over time, but validation loss fluctuates.
- Class Loss (Classification Accuracy): Training improving, validation remains unstable.
- Distribution Focal Loss (DFL): Gradually decreasing.
- Precision: Fluctuating, showing inconsistent detections.
- Recall: Unstable, indicating potential false negatives.
- mAP50: Peaks at 0.65 but varies across training.
- mAP50-95: Lower performance (~0.40), meaning detections across IoU thresholds are not strong.
- Overfitting: Training loss decreases, but validation loss remains unstable.
- Class Imbalance: Some classes have very few examples.
- Dataset Size: More training data needed for better generalization.
- Number Plate Variations: Need to train on both white and black plates to generalize across different Malaysian plates.
- Increase dataset size and include more black and white number plates.
- Improve augmentation strategies (color variations, rotations, brightness changes).
- Tune hyperparameters (batch size, learning rate, optimizer selection).
- Try alternative models such as YOLOv8x or fine-tuning on a pre-trained OCR model for number plate reading.