A Method for 3D Morphological Measurement and Cutting Point Localization of Green Asparagus Based on Improved YOLOv8n-seg and RGB-D Data

Abstract To address the difficulties in recognition and localization caused by the curved and irregular growth forms of green asparagus in natural environments, with parts of the rootstalk hidden in soil, and the challenge of balancing detection accuracy and speed with traditional instance segmentation algorithms during automated harvesting operations, a lightweight green asparagus instance segmentation model named SES-YOLO based on improved YOLOv8n-seg is proposed. First, the StarNet lightweight backbone network is used to replace the original backbone structure, reducing the computational burden of the model. Second, the C2fDWConv_Strip module is designed by integrating depthwise separable convolution and strip convolution to enhance the feature extraction capability for slender targets. Then, the EGAP module is embedded into the SPPF layer to improve edge detail perception and anti-interference capability, and the MPDIoU loss function is adopted to optimize bounding box regression accuracy. Finally, a method combining image masks with depth information to calculate the length and diameter of green asparagus and locate the cutting point is employed. Experimental results show that SES-YOLO achieves precision, recall, and mAP@50 of 98.6%, 97.3%, and 99.3%, respectively. Compared to the baseline model, the parameter count is reduced by 38.1%, and the computational load is decreased by 22.3%. The average relative measurement errors for asparagus length and diameter are 1.97% and 2.07%, respectively, and the average cutting point localization error is 0.34 cm. The model's inference speed on the Jetson Orin NX edge device is 38.69 FPS, which is 58.4% higher than the baseline model. This model maintains high accuracy while possessing good lightweight and real-time performance, making it suitable for embedded deployment scenarios such as green asparagus harvesting robots, providing reliable technical support for achieving automated selective harvesting.