Share
Export Citation
Comparative Analysis of Series YOLO Models for Loose Fruits Palm Oil Detection: A Comprehensive Performance Evaluation
Warni E.
International Conference on Wireless Networks and Mobile Communications Wincom
Abstract
Automated detection of loose fruits has become a critical factor in minimizing harvest losses in oil palm plantations. This study aims to evaluate and compare the performance of four generations of YOLO models (YOLOv5, YOLOv8, YOLOv9, and YOLOv11) for detecting loose fruits palm oil, focusing primarily on accuracy and inference speed. A total of 20 different model scales were tested using a carefully annotated plantation image dataset resized to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$640 \times$</tex> 640 pixels. Model performance was evaluated based on precision, recall, mean average precision (mAP), and inference time (milliseconds) on both GPU based systems and edge devices. Results showed that lightweight models such as YOLOv5s and YOLOv8s achieved mAP scores around 0.83 with inference times below 340 ms, making them suitable for real time applications. Conversely, YOLOv9m and YOLOv11m achieved the highest accuracy with mAP scores exceeding 0.86, but their inference times were longer, surpassing 800 ms, making these models more appropriate for server based inspection applications. The largest models in each generation provided only slight improvements in accuracy but were accompanied by significant increases in inference time, highlighting a saturation point in the trade off between model complexity and performance. Consequently, this study offers a comprehensive performance map, serving as a practical guide for selecting appropriate YOLO based loose fruit detection models tailored to specific application requirements. The key contribution of this research lies in providing a scientific basis for choosing detection models that effectively balance accuracy and computational efficiency, thereby accelerating the integration of computer vision technologies into smart agricultural practices.