PERFORMANCE COMPARISON OF EASYOCR AND PADDLEOCR IN YOLO-BASED INDONESIAN LICENSE PLATE RECOGNITION

Abstract

Vehicle license plate recognition plays an important role in intelligent transportation systems for automated traffic monitoring and management, yet real world deployment remains challenging due to variations in illumination, image quality, and plate condition. This study designs and evaluates an artificial intelligence based license plate recognition system combining object detection and character recognition. The detection model was trained on the publicly available Roboflow “plate-indonesia-uakjc” dataset (v1, CC BY 4.0), comprising 661 training, 108 validation, and 92 test images of Indonesian plates; the 92 image test set combines 44 field collected images from Bangka Belitung Province and 48 images from other Indonesian provinces. The proposed pipeline detects license plates using YOLOv8 Nano, then crops and recognizes characters using EasyOCR and PaddleOCR. Detection performance was evaluated using precision, recall, and mean average precision (mAP), while OCR performance was assessed using plate level accuracy, character level accuracy, and Character Error Rate (CER). YOLOv8 achieved strong detection performance, with precision of 0.9975, recall of 1.0000, [email protected] of 0.9950, and [email protected]:0.95 of 0.8328. For character recognition, EasyOCR outperformed PaddleOCR on the full 92 image test set, achieving 78.26% plate accuracy (CER: 3.26%) versus 72.83% (CER: 7.03%) for PaddleOCR, with character accuracies of 96.74% and 92.97%, respectively. Error analysis shows that most errors arise from confusion between “Q” and “O” and from border artifacts in bounding box crops of older plates with white frames. These findings demonstrate that YOLOv8 combined with EasyOCR provides an effective, reproducible approach for Indonesian license plate recognition under real world conditions.