基于YOLOv7-Sim和无人机遥感影像的烟株数量检测  

Tobacco Plant Number Detection Based on UAV Remote Sensing Image and YOLOv7-Sim

作　　者：耿利川 王忠丰[2] 秦永志 马莉 GENG Lichuan;WANG Zhongfeng;QIN Yongzhi;MA Li(College of Urban and Environment Sciences,Xuchang University,Xuchang,Henan 461000,China;92493 Troops,Huludao,Liaoning 125000,China;61206 Troops,Beijing 100042,China)

机构地区：[1]许昌学院城市与环境学院,河南许昌461000 [2]92493部队,辽宁葫芦岛125000 [3]61206部队,北京100042

出　　处：《中国烟草科学》2023年第4期94-102,共9页Chinese Tobacco Science

基　　金：智慧中原地理信息技术河南省协同创新中心,时空信息感知与融合技术自然资源部重点实验室联合基金项目(212109)。

摘　　要：植株数是用于监测作物生长状况和估测产量的重要田间表型性状。为实现烟草植株数高效自动清点,针对无人机遥感影像烟株检测中存在小尺寸聚集目标容易漏检的问题,提出了一种YOLOv7目标检测优化模型YOLOv7-Sim。首先引入SimAM注意力机制增强图像特征之间的聚合能力;然后加入小目标检测层强化算法对小目标的检测能力;再对定位损失函数进行优化,引入了EIOU定位损失函数;最后利用分块策略解决大图像检测中小目标容易采样丢失的问题。在VisDrone2019数据集和本文构造的UAVTob无人机遥感影像烟草数据集上的检测结果显示,检测均值平均精确率mAP@0.5提升了0.3%和6.3%,mAP@0.5:0.95提升了0.6%和18.3%,YOLOv7-Sim算法对无人机遥感影像中的烟株检测更具优越性。The plant number is an important field phenotypic trait in monitoring crop growth and estimating output.In order to establish an efficient tobacco plant number automatic counting technology,an optimized tobacco plant detection model YOLOv7-Sim based on YOLOv7 is proposed to solve the miss detection problem of small targets in UAV remote sensing images.First,the SimAM attention mechanism is introduced to enhance the aggregation ability between image features,and a small target detection layer is added to strengthen the detection ability of small targets,then EIOU is used to optimize the positioning loss function,and finally,a slicing strategy is used to solve the problem of small target sampling loss in large image detection.The experimental results on the Vis-Drone2019 dataset and the UAVTob dataset constructed in this study showed that the mean average accuracy rate mAP@0.5 of the detection results was increased by 0.3%and 6.3%,and the mean average accuracy rate mAP@0.5:0.95 was increased by 0.6%and 18.3%,which reflected the superiority of YOLOv7-Sim algorithm for tobacco detection in UAV remote sensing images.

关 键 词：YOLOv7 无人机遥感影像 烟草检测 深度学习 

分 类 号：S572[农业科学—烟草工业]