基于FGP-YOLO 11n的银杏果柄实时检测方法

Real-time Detection Method of Ginkgo Stems Based on FGP-YOLO 11n

摘要:

银杏果实自动采摘有利于延长水果新鲜期和保证储存期间质量，准确识别银杏果柄对于末端执行器采摘银杏果实至关重要，而复杂的作业环境以及银杏果柄细长等因素会影响银杏果柄识别准确性。为了解决该问题，本文提出了一种基于FGP-YOLO 11n的银杏果柄实时检测方法。在YOLO 11n骨干网络和颈部网络中分别加入FasterNeXt和GELAN网络模块，并将CIoU损失函数替换成PIoU损失函数，以实现银杏果柄实时检测。同时，采用Grad-CAM可视化、消融试验和性能试验等验证FGP-YOLO 11n模型性能，最后分别在不同场景下与经典模型的检测性能进行比较。试验结果表明，FGP-YOLO 11n模型对银杏果柄检测的精确率、召回率和平均精度均值分别为93.80%、90.50%和93.10%;与YOLO 11n模型相比，分别提升5.60、5.60、3.50个百分点，同时检测速度达72.73 f/s;整体性能优于DETR、EfficientDet、RetinaNet和SSD等经典检测模型，满足实际作业情况要求。本研究结果有助于智慧农业的发展，并促进机器视觉在智能采摘领域的实践与应用。

Abstract:

Automated harvesting of ginkgo fruits is beneficial for prolonging post-harvest freshness and maintaining quality during storage. Accurate identification of fruit stems is essential for the precise operation of end-effectors in robotic harvesting systems. However, the complexity of the working environment, coupled with the slender and variable morphology of ginkgo stems, poses significant challenges to reliable detection and recognition. To address this issue, a real-time detection method was proposed based on the FGP-YOLO 11n architecture. The FasterNeXt and GELAN modules were integrated into the backbone and neck networks of YOLO 11n, respectively, while the CIoU loss function was replaced with the PIoU loss function. These enhancements were designed to enable accurate and real-time detection of ginkgo stems. Meanwhile, to comprehensively evaluate the performance of the FGP-YOLO 11n model, a series of validation techniques, including Grad-CAM visualization, ablation experiments, and performance analyses were conducted. Furthermore, the model's detection performance was benchmarked against classical models across various scenarios. Experimental results demonstrated that the FGP-YOLO 11n model achieved a precision of 93.80%, a recall of 90.50%, and a mean average precision (mAP) of 93.10% in detecting ginkgo stems. Compared with the classical YOLO 11n model, the proposed approach achieved improvements of 5.60 percentage points, 5.60 percentage points, and 3.50 percentage points in respective evaluation metrics, while maintaining real-time processing at 72.73 frames per second (FPS). Overall, it outperformed the established object detection frameworks such as DETR, EfficientDet, RetinaNet, and SSD, effectively meeting the demands of real-world agricultural operations. These findings can support the advancement of smart agriculture and facilitate the practical integration of machine vision technologies in intelligent harvesting systems.