A Statistical Evaluation Framework with Application to YOLO-Based Models in Cross-Language License Plate Recognition: Arabic and Latin Alphabets

Abstract

This study presents a statistical evaluation framework with application to YOLO-based models in cross-language License Plate (LP) recognition. Rather than relying solely on conventional accuracy metrics, the proposed framework integrates statistical performance analysis with deep learning techniques to provide a comprehensive and objective comparison of object detection models. The framework employs multiple statistical evaluation measures, including precision, recall, mean Average Precision (mAP), confusion matrices, and confidence-based performance curves, to quantify detection accuracy, classification reliability, and model robustness. Experiments were conducted using an Iraqi license plate dataset comprising 1,834 annotated images containing Arabic and Latin license plate characters. Two state-of-the-art object detection models, YOLOv5 and YOLOv8, were trained and evaluated under identical experimental settings. The results demonstrate that YOLOv8 consistently outperformed YOLOv5 across all statistical evaluation measures, achieving higher detection accuracy and improved classification performance. The proposed statistical evaluation framework provides a robust and objective approach for assessing and comparing deep learning models. By integrating statistical evaluation principles with advanced object detection techniques, this framework offers an effective methodology for model evaluation and selection in cross-language license plate recognition and can be extended to various computer vision applications.