YOLO-Banana-Seg: a lightweight and efficient model for rapid segmentation of banana bunches and stalks in complex orchards
TL;DR
An improved YOLOv11-based segmentation model, YOLO-Banana-Seg, for banana bunches and stalks is proposed, which achieves high segmentation accuracy while reducing parameters and computational complexity, ensuring its applicability in resource-limited scenarios.
Abstract
In banana orchards, accurate segmentation of bunches and stalks is vital for automated harvesting and yield estimation. However, color similarity between green bananas and leaves, irregular shapes, and occlusion makes instance segmentation highly challenging. This study proposes an improved YOLOv11-based segmentation model, YOLO-Banana-Seg, for banana bunches and stalks. We compiled a multi-condition banana dataset encompassing varying illumination intensities and occlusion levels, while the proposed model integrates multi-scale feature fusion and adaptive attention mechanisms to significantly improve occluded target segmentation accuracy. The results show that YOLO-Banana-Seg achieves high segmentation accuracy while reducing parameters and computational complexity, ensuring its applicability in resource-limited scenarios. Experiments demonstrate 97.4% accuracy and 82.9% recall with 16.7% fewer parameters than YOLOv11n, balancing precision and efficiency. The model's high precision and efficiency directly could support the development of harvesting robots and yield estimation systems for agricultural applications.