Abstract:

This study delves into the practical application and fine-tuning of YOLOv8 models for real-time tomato recognition using drone imagery within greenhouse environments. We evaluated YOLO’s speed, robustness, and adaptability, finding that varying batch sizes and epochs had minimal impact on performance. Notably, YOLOv8n performed on par with the extra-large YOLOv8x model, offering a significant advantage: training time was up to 60 times shorter. Further tuning and innovative training strategies revealed that the Final Learning Rate (lrf) and dataset annotation quality were the most influential factors for model performance. Fine-tuning the lrf and re-annotating datasets markedly improved accuracy, underscoring the importance of optimizing learning rates and maintaining high-quality annotations for effective YOLO models. Our results also demonstrated the superiority of YOLOv8 over YOLOv5. The optimized YOLOv8n model is well-prepared for deployment in upcoming tomato recognition tasks, paving the way for more efficient agricultural monitoring. This work also provides valuable insights into the broader field of object recognition and offers practical guidance for researchers tackling similar challenges.