【摘要】 为实现设施环境中樱桃花目标的准确检测，本文在YOLO v8的基础上，提出一种适用于设施环境樱桃花目标检测模型CEG-YOLO。引入跨阶段部分网络与部分卷积结合(Integration of cross-stage partial networks and partial convolution, CSPPC)结构，替换Backbone和Neck网络中的C2f模块，以增强网络的特征提取能力，同时显著降低了模型的浮点运算量和参数量；在SPPF模块前添加轻量化的高效通道注意力(Efficient channel attention, ECA),进一步增强网络的特征提取能力，提高模型的检测精度；将Neck中的Conv模块替换为GSConv(Grouped spatial convolution)模块，进一步降低模型的参数量和浮点运算量。结果显示，CEG-YOLO模型的精确率为89.8%、召回率为91.3%、mAP为94.8%、参数量为2.03×10~6、浮点运算量为5.8×10~9。与原始YOLO v8模型相比，该模型的精确率、召回率和mAP分别提高1.6、1.9、1.8个百分点，参数量和浮点运算量分别减少32.56%和28.40%。与YOLO v3、YOLO v5、YOLO v7、YOLO v9、YOLO v10和YOLO 11模型相比，CEG-YOLO模型mAP分别提高7.7、0.7、1.8、2.1、3.1、1.7个百分点，参数量分别减少76.59%、71.53%、32.56%、55.09%、24.81%和21.31%,浮点运算量分别减少55.04%、45.28%、28.40%、32.56%、29.27%和7.9%。因此，本文提出的CEG-YOLO模型不仅能够准确地检测设施环境中的樱桃花，还具有较低的计算成本，适合部署在樱桃授粉机器人的嵌入式设备上，为设施樱桃的智能授粉提供可行的视觉支持。更多还原

【Abstract】 In order to achieve accurate detection of cherry blossom targets in the facility environment, a model CEG-YOLO for cherry blossom target detection in the facility environment based on YOLO v8 was proposed. Firstly, the structure of integration of cross-stage partial networks and partial convolution(CSPPC) structure was introduced to replace the C2f module in the Backbone and Neck networks to enhance the feature extraction capability of the network, and at the same time, the computational and parametric quantities of the model were reduced significantly. Secondly, a lightweight efficient channel attention(ECA) was added before the SPPF module, which enabled the network to further enhance the characteristics extraction and improve the detection accuracy of the model. Finally, the Conv module in Neck was replaced with the grouped spatial convolution(GSConv) module, which made the number of parameters and computation of the model further reduce. The experimental results showed that the accuracy of the CEG-YOLO model was 89.8%, the recall was 91.3%, the mAP was 94.8%, the number of parameters was 2.03×10~6, and the computational cost was 5.8×10~9. Compared with the original YOLO v8 model, the accuracy, recall, and mAP of the model was improved by 1.6, 1.9, and 1.8 percentage points, respectively, and the number of parameters and computational were decreased by 32.56% and 28.40%. Compared with the YOLO v3, YOLO v5, YOLO v7, YOLO v9, YOLO v10, and YOLO 11 models, the mAP of the CEG-YOLO model was increased by 7.7, 0.7, 1.8, 2.1, 3.1, and 1.7 percentage points, and the number of parameters and computation was reduced significantly. The CEG-YOLO model proposed was not only able to accurately detect cherry blossoms in facility environments, but also possessed low computational overheads, which was suitable for deployment on the embedded devices of cherry pollination robots, which provided feasible visual support for realizing intelligent pollination of facility cherries.更多还原