Wang Zhi-Yong, Zhang Cui-Ping
Shandong Provincial University Laboratory for Protected Horticulture, Weifang University of Science and Technology, Weifang, 262700, Shandong, People's Republic of China.
School of Computer Science, Weifang University of Science and Technology, Weifang, 262700, Shandong, People's Republic of China.
Plant Methods. 2025 May 27;21(1):71. doi: 10.1186/s13007-025-01390-9.
Artificial pollination can considerably improve pollination success and boost chilli pepper fruit set and quality when grown in enclosed environments (e.g., greenhouses). Artificial pollination, on the other hand, raises production costs while also necessitating specific operating abilities. The precise and efficient identification of pepper blossoms is a critical step in the development of robotic pollinators or pollination drones. In this paper, we propose a pepper flower detection method based on YOLOv8 that incorporates multi-scale, attention, and conditional information. To begin, the CBAM structure that incorporates edge information is integrated into Backbone to expand the feature extraction receptive field and facilitate the learning of long-distance dependency. The BERT model is then used to encode conditional information, which is integrated into the backbone via the ELAN layer to assist the training and inference processes. Finally, an improved MPDIoU is applied to increase detection accuracy while increasing flexibility. The experimental results show that the modification enhances the network depth and reduces the number of parameters from 4M to 2.85M, while improving the mean average accuracy (mAP) by 3.1% over the baseline approach. The study's findings can help in crop object detection. The chilli pepper flower dataset: https://drive.google.com/file/d/1cKNie_iAzx-K4iPLQRVdyiOKV1d9zHrF/view?usp=drive_link The source code is available in https://drive.google.com/drive/folders/1ubNnKu7PWYAdUXvbs4Z2OBAVcSAQ3WLd?usp=drive_link .
在封闭环境（如温室）中种植时，人工授粉可以显著提高授粉成功率，增加辣椒坐果率和果实品质。然而，人工授粉会增加生产成本，同时还需要特定的操作技能。精确高效地识别辣椒花是开发机器人授粉器或授粉无人机的关键步骤。在本文中，我们提出了一种基于YOLOv8的辣椒花检测方法，该方法融合了多尺度、注意力和条件信息。首先，将包含边缘信息的CBAM结构集成到主干网络中，以扩大特征提取感受野，促进长距离依赖的学习。然后使用BERT模型对条件信息进行编码，并通过ELAN层将其集成到主干网络中，以辅助训练和推理过程。最后，应用改进的MPDIoU来提高检测精度，同时增加灵活性。实验结果表明，该改进增强了网络深度，将参数数量从4M减少到2.85M，同时比基线方法提高了3.1%的平均精度均值（mAP）。该研究结果有助于作物目标检测。辣椒花数据集：https://drive.google.com/file/d/1cKNie_iAzx-K4iPLQRVdyiOKV1d9zHrF/view?usp=drive_link 源代码可在https://drive.google.com/drive/folders/1ubNnKu7PWYAdUXvbs4Z2OBAVcSAQ3WLd?usp=drive_link 中获取。
