FPG-YOLO: A detection method for pollenable stamen in 'Yuluxiang' pear under non-structural environments

•A model based on PFG-YOLO has been proposed for the detection of pollenable stamen of 'Yuluxiang' pear.•Designed the C3F module, which enhances feature fusion capability and model lightweightness.•The PfAAMC3 module was constructed to improve the feature extraction capability of the model.•The introduction of GhostConv reduces model complexity and improves detection accuracy. As one of the key technologies of intelligent pollination robots, pollenable stamen detection is of great significance for improving the setting rate of pear fruit, pollination efficiency and labor liberation. In this study, a novel FPG-YOLO model based on YOLOv5n is developed to facilitate the realization of low-cost, high-precision automated pear pollination. The method was based on FasterNet structure to construct C3F module substitute for C3 module of the neck network to improve the fusion capability of features and achieve lightweight design of the model. Meanwhile, the PfAAMC3 module, which is the bottleneck of the C3 module, is constructed by PfAAM, which replaces the C3 module of the backbone network and improves the model's ability to extract features from stamens. Additionally, the improved model uses GhostConv instead of the traditional Conv except for the first Conv, which improves model recognition accuracy while reducing model parameters and dimensions. The results showed that the average precision (AP) of the FPG-YOLO model pollenable stamen and non-pollinated stamen were 96.60 % and 91.56 %, respectively, and the mean average precision (mAP) reached 94.08 %, and the model size and floating-point operations per second (FLOPs) were only 3.09 MB and 3.6 G, respectively, which is 1.2 % improvement in mAP, 0.65 MB reduction in model size, and 0.53 G reduction in FLOPs compared to YOLOv5n. Compared with the commonly used target detection algorithms YOLOv3, YOLOv3-Tiny, YOLOv4, YOLOv4-Tiny, YOLOv7, YOLOV7-Tiny, and YOLOV8n models, the FPS of PFG-YOLO is 90.91 f·s − 1. The model has the highest F1, accuracy, and mAP and smallest model size. FPG-YOLO showed good performance in detecting pollenable stamen of pear in non-structural environments, which could provide technical support for intelligent pollination of 'Yuluxiang' pear pollination robot.