CottonBud-YOLOv5s轻量型棉花顶芽检测算法  

CottonBud-YOLOv5s lightweight cotton bud detection algorithm

作　　者：赵露强 彭强吉 兰玉彬 康建明 张敬文 代建龙[4] 陈玉龙 ZHAO Luqiang;PENG Qiangji;LAN Yubin;KANG Jianming;ZHANG Jingwen;DAI Jianlong;CHEN Yulong(College of Agricultural Engineering and Food Science,Shandong University of Technology,Zibo 255000,China;Shandong Academy of Agricultural Machinery Sciences,Jinan 250100,China;Shandong Key Laboratory of Intelligent Agricultural Equipment in Hilly and Mountainous Areas,Jinan 250100,China;Shandong Academy of Agricultural Sciences,Jinan 250100,China)

机构地区：[1]山东理工大学农业工程与食品科学学院,淄博255000 [2]山东省农业机械科学研究院,济南250100 [3]山东省丘陵山区智慧农业装备重点实验室,济南250100 [4]山东省农业科学院,济南250100

出　　处：《农业工程学报》2025年第4期175-184,共10页Transactions of the Chinese Society of Agricultural Engineering

基　　金：山东省农机研发制造推广应用一体化试点项目(NJYTHSD-202317)。

摘　　要：针对棉花机械打顶作业过程中,边缘移动设备算力受限实时性差,运动模糊、小目标遮挡导致难以检测的问题,该研究基于YOLOv5s模型提出CottonBud-YOLOv5s轻量型棉花顶芽检测模型,该模型采用ShuffleNetv2主干网络和DySample动态上采样模块替换原始模块降低计算量,提高模型检测速度;头部(head)和颈部(neck)分别引入ASFFHead检测头和GC(global context)全局注意力模块增强模型尺度不变性和上下文特征提取能力,提高小目标遮挡和运动模糊图像的检测性能。通过消融试验和模型对比试验,验证CottonBud-YOLOv5s棉花顶芽检测模型的可行性。试验结果表明:引入ASFFHead检测头和GC全局注意力机制后,小目标平均精度AP_(0.5:0.95)和平均召回率AR_(0.5:0.95)值比引入前分别提升3.6、2.1个百分点,中目标平均精度AP_(0.5:0.95)和平均召回率AR_(0.5:0.95)值分别提升4.1、3.5个百分点,大目标平均精度AP_(0.5:0.95)和平均召回率AR_(0.5:0.95)值分别提升6.5、5.9个百分点;与Faster-RCNN、TOOD、RTDETR、YOLOv3s、YOLOv5s、YOLOv9s和YOLOv10s检测模型相比检测速度分别提升26.4、26.7、24.2、24.8、11.5、18.6、15.6帧/s,平均精度均值分别提升14.0、13.3、5.5、0.9、0.8、0.2、1.5个百分点,召回率分别提升16.8、16.0、3.2、2.0、0.8、0.5、1.2个百分点,CottonBud-YOLOv5s模型平均精度均值达到97.9%,召回率达到97.2%,CPU检测速度达到27.9帧/s。由模型可视化分析可知CottonBud-YOLOv5s模型在单株、多株、运动模糊、小目标遮挡的整体检测性能优于其他检测模型。该模型具有较高的检测精度、鲁棒性和检测速度,适用于密植环境下棉花顶芽的精准检测,可为棉花机械化打顶提供视觉检测基础。Cotton mechanical topping is one of the most important cultural practices to improve crop yield during production.The shoots of cotton topping can be cut at about 10–20 cm from the top of plants.However,the performance of mechanical topping has been limited to computing power and real-time transport in several edge-moving devices at present.The detection can also be confined to the motion blur and small target occlusion.In this study,a lightweight detection model of a cotton bud(named CottonBud-YOLOv5s)was proposed using the well-known YOLOv5s architecture.Both performance and efficiency were optimized to detect the cotton buds in complex field environments.The ShuffleNetv2 backbone network was utilized to enhance the overall performance of the CottonBud-YOLOv5s model.The computational complexity was reduced to maintain the high accuracy of detection.In addition,the DySample dynamic upsampling module was integrated to replace the original ones.The computational costs were further reduced to improve the speed of detection.As such,the improved model was run more efficiently on edge devices with limited computing power.Real-time performance was also achieved during cotton mechanical topping.Moreover,the ASFFHead detection head and GC(global context)attention mechanism were also introduced into the head and neck components,in order to handle the varying object scales and complex contextual information.The scale invariance was significantly improved to extract the context-based features,which was crucial to detect the small targets that occluded or blurred due to the various motions in fields.Ultimately,the robustness of the model was improved to perform the best in real-world conditions.A series of ablation and comparison tests were conducted to validate the efficacy of the CottonBud-YOLOv5s model.The experimental results demonstrated that the introduction of the ASFFHead detection head and the GC global attention mechanism led to notable improvements in detection accuracy.Specifically,the average precision(AP)at 0

关 键 词：目标检测 遮挡 运动模糊 小目标 棉花顶芽 卷积神经网络 YOLOv5s 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程] S24[农业科学—农业电气化与自动化]