ED-YOLOv8s: An Enhanced Approach for Passion Fruit Maturity Detection Based on YOLOv8s

Accurate assessment of the ripeness of passion fruits, which grow on trellises in orchards, is essential for efficient harvesting, grading, transportation, and storage processes, especially considering the potential occlusion and overlap among the fruits. This study introduces an enhanced maturity detection method, based on YOLOv8s. We first present the DyHead dynamic detection head, which significantly augments the expressive power of the target detection head without adding computational complexity. We then propose the EM_C-C2f lightweight module to optimize the backbone network of YOLOv8s. Experimental results from three custom datasets show considerable improvements achieved by the ED-YOLOv8s model across various evaluation metrics. Compared to other YOLO series object detection networks, ED-YOLOv8s demonstrates superior performance in terms of precision, recall, and mAP. Specifically, on the comprehensive passion fruit dataset, the model achieves a 2.6% increase in precision and a 1.3% increase in recall compared to the original YOLOv8s. Moreover, ED-YOLOv8s attains a mAP@50 score of 98.4%, reflecting a 1.3% improvement, and a mAP@50:95 score of 91.9%, marking a 0.6% improvement over the original YOLOv8s.