Detection of sugarcane stalk node based on improved YOLOv8 and its deployment on edge device

Accurate sugarcane detection is a critical upstream task for intelligent harvesting, but it faces several challenges, including occlusion, varying lighting conditions, and unclear morphological features in sugarcane field. To address the above issues, a Sugarcane Stalk Node Dataset (SSND) was constructed, which encompasses multi-angle acquisitions, diverse lighting conditions, multi-stage growth variations, and meteorological variations to facilitate deep learning-based sugarcane stalk node detection tasks. An improved YOLOv8 model for sugarcane stalk node was proposed in this study and was subsequently deployed on edge device for practical application. Specifically, the BRA (Bi-level Routing Attention) was integrated into the YOLOv8 to reduce background interference from sugarcane fields. The traditional convolution was then replaced with SPD-Conv to improve small-target detection capabilities. Finally, the ASFF (Adaptive Spatial Feature Fusion) was incorporated to enhance overall detection performance. The experimental results demonstrate that the improved YOLOv8 model achieved a P of 96.8%, R of 94.1%, [email protected] of 98.6%, with 4.08 million parameters, 10.4 GFLOPs, and a model size of 8.1 MB, outperforming other models across these metrics. Compared to the original YOLOv8n, a 1.4% improvement in precision for sugarcane stalk node is observed. To verify its operational feasibility on edge devices, the model was deployed on an NVIDIA Jetson TX 2, where successful detection of sugarcane stalk nodes was achieved. This study will provide valuable technical support for sugarcane harvesting robots.