Analisis Kinerja Transfer Learning Pada Model Yolo Dan Paddleocr Dalam Deteksi Dan Pengenalan Plat Nomor Kendaraan

Abstract

Automatic License Plate Recognition (ALPR) systems are crucial in intelligent transportation and security. Traditional ALPR challenges, such as number plate variations and environmental conditions, can be addressed with deep learning. This research analyzes the performance of transfer learning on the YOLOv8 model for number plate detection and the PaddleOCR model for character recognition, utilizing an Indonesian vehicle number plate dataset. The methodology involves collecting and preprocessing the Indonesian number plate dataset, followed by fine-tuning the YOLOv8 and PaddleOCR models using transfer learning. The YOLOv8 model is responsible for detecting number plate locations, and the detected areas are then processed by PaddleOCR for character recognition. Performance evaluation is conducted using Precision, Recall, F1-score, and mAP for detection, as well as Character Recognition Rate (CRR), License Plate Recognition Rate (LPRR), and Edit Distance for character recognition. The results show that fine-tuning the YOLOv8 model is highly effective, achieving an mAP50 of 99.31%, Precision of 98.06%, Recall of 97.77%, and an F1-score of 97.91%, indicating excellent detection capabilities and rapid training convergence (15-20 epochs). For character recognition, PaddleOCR achieved a CRR of 84.20% and an LPRR of 71.40%, with most recognition errors being minor (Edit Distance predominantly at 0 and 1). Overall, the transfer learning approach proved to be highly efficient and effective in building an accurate and adaptive ALPR system for the Indonesian number plate context. The advantages of this approach include the highest detection accuracy, computational efficiency, high adaptability to Indonesian number plate characteristics, optimal Precision-Recall balance, and robustness to various visual conditions.