Aiming at the problem that the fruit of Camellia oleifera is small and occluding each other in UAV aerial images, an improved YOLOv5s model is proposed. Firstly, SPD-Conv is used to replace the pooling operation in the YOLOv5s model, so that the model can retain more fine-grained information during the down-sampling operation. Then, Coordinate Attention (CA) is introduced at the end of the neck network of YOLOv5s model to improve the robustness of the model to occluding targets. Additionally, the improved YOLOv5s model replaces the YOLOv5s CIOU bounding box loss function with the NWD (Normalized Gaussian Wasserstein) bounding box loss function to improve its ability to detect small Camellia oleifera fruits in drone aerial images. The precision, recall, F1 score, and mean average precision of the improved YOLOv5s model are 93.1%, 90.5%, 91.78% and 91.2%, respectively. Compared to the YOLOv5s model, the improved YOLOv5s model's mean average precision has increased by 3.6 percentage points. The experiments indicate that the improved YOLOv5s has stronger detection capabilities for smaller and occluded Camellia oleifera fruits in aerial images. This research can provide a reference for the estimation of Camellia oleifera fruit yield by using drones.

针对无人机航拍图像中油茶果实小且互相遮挡的问题，提出改进YOLOv5s模型。首先，使用SPD-Conv代替YOLOv5s模型中池化操作，使模型在执行下采样操作时能够保留更多细粒度信息。然后，在YOLOv5s模型的颈部网络末端引入坐标注意力机制CA，提高模型对遮挡目标的鲁棒性。另外，改进YOLOv5s模型使用NWD边界框损失函数替换YOLOv5s中的CIOU边界框损失函数，以提升模型对无人机航拍图像中小油茶果实的检测能力。改进YOLOv5s模型的精确率、召回率、F1分数和平均精度均值分别达到93.1%、90.5%、91.78%和91.2%，与YOLOv5s模型相比，平均精度均值提升3.6个百分点。试验表明，改进YOLOv5s对航拍图像中较小的油茶果实和遮挡果实有更强的检测能力。可为利用无人机进行油茶果实的产量估计研究提供参考。