Improving Indonesian Vehicle License Plate Recognition Using YOLOv8n and Fine-Tuned PaddleOCR

Automatic License Plate Recognition (ALPR) is an important technology in intelligent transportation systems, enabling automatic vehicle identification through license plate detection and character recognition. However, variations in lighting conditions, image quality, viewing angles, and the design characteristics of Indonesian license plates often reduce recognition accuracy. This study proposes an ALPR system that integrates YOLOv8n for license plate detection and Fine-Tuned PaddleOCR for character recognition. A dataset consisting of 900 vehicle images was utilized to train and evaluate the detection model. The YOLOv8n model was trained using a 640 × 640pixel input size and evaluated using Precision, Recall, mAP50, and mAP50-95 metrics. The detected license plate regions were cropped and processed using image preprocessing techniques, including resizing, grayscale conversion, bilateral filtering, and sharpening. Character recognition performance was evaluated by comparing the default PaddleOCR model and a Fine-Tuned PaddleOCR model trained on Indonesian license plate data. Experimental results showed that YOLOv8n achieved a precision of 0.987, recall of 1.000, mAP50 of 0.989, and mAP50-95 of 0.853. Furthermore, the Fine-Tuned PaddleOCR model improved recognition accuracy from 56% to 81%, demonstrating a 25% increase compared to the default PaddleOCR model. These results indicate that the proposed approach effectively improves Indonesian vehicle license plate recognition performance and can be applied to ALPR systems in real-world environments.