一种基于改进YOLOv5s的麦穗检测计数方法  

A wheat ear detection and counting method based on improved YOLOv5s

作　　者：仝召茂 陈学海[1,2] 汪本福 马志艳 杨光友[1,2] TONG Zhaomao;CHEN Xuehai;WANG Benfu;MA Zhiyan;YANG Guangyou(Institute of Agricultural Machinery,Hubei University of Technology,Wuhan 430068,China;Hubei Engineering Research Center for Intellectualization of Agricultural Equipment,Wuhan 430068,China;Hubei Provincial Key Laboratory of Grain Crop Germplasm Innovation and Genetic Improvement/Key Laboratory of Crop Molecular Breeding of the Ministry of Agriculture and Rural Affairs,Wuhan 430068,China)

机构地区：[1]湖北工业大学农机工程研究设计院,湖北武汉430068 [2]湖北省农机装备智能化工程技术研究中心,湖北武汉430068 [3]粮食作物种质创新与遗传改良湖北省重点实验室/农业农村部作物分子育种重点实验室,湖北武汉430068

出　　处：《南京农业大学学报》2024年第6期1202-1211,共10页Journal of Nanjing Agricultural University

基　　金：国家重点研发计划项目(2022YFD2301005-03);湖北省科技创新人才计划项目(2023DJC088);湖北省农机装备补短板核心技术应用攻关项目(HBSNYT202208)。

摘　　要：[目的]为实现对田间麦穗的实时准确计数,本文提出一种基于改进YOLOv5s的麦穗检测计数方法。[方法]通过C2f模块获得更加丰富的梯度流,增强模型细粒度特征提取能力,并在网络关键部位引入CoordConv坐标卷积,加大对坐标信息关注程度,提升模型对麦穗位置的感知能力,同时考虑到麦穗检测任务中中小尺寸麦穗居多,采用Inner CIoU损失函数加快模型收敛速度。[结果]在公开数据集Global Wheat Head Detection(GWHD)上对上述方法进行试验,结果表明,本文改进模型的精确率、召回率、平均精度均值mAP0.5分别为93.5%、91.6%和95.9%,参数量、计算量、每秒帧数分别为12.4 MB、27.5 GFLOPs和34。[结论]本文改进模型在精确率、召回率、平均精度均值mAP0.5等指标上较原始YOLOv5s模型分别增加1.0、1.2和1.3百分点,并且优于YOLOv7-tiny、YOLOv8s等模型,可满足检测的实时性要求。同时改进后模型在处理遮挡、重叠等复杂情况时都比原始模型表现更优,具有良好的鲁棒性。[Objectives]To achieve real-time and accurate counting of wheat ears in the field,this paper proposed a wheat ear detection and counting method based on improved YOLOv5s.[Methods]By using the C2f module to obtain richer gradient flows,the model's ability to extract fine-grained features was enhanced,and CoordConv coordinate convolution was introduced in key parts of the network to increase attention to coordinate information and improve the model's perception of wheat ear position.At the same time,considering that wheat ear detection tasks mainly involved small and medium-sized wheat ears,the Inner CIoU loss function was adopted to accelerate the convergence speed of the model.[Results]The above methods were tested on the publicly available dataset Global Wheat Head Detection(GWHD),and the results showed that the accuracy,recall,and average accuracy mean mAP0.5 of the proposed model were 93.5%,91.6%,and 95.9%,respectively.The parameter count,computational complexity,and frames per second were 12.4 MB,27.5 GFLOPs,and 34,respectively.[Conclusions]The model proposed in this article improved accuracy,recall,and average precision mean mAP0.5 by 1.0,1.2,and 1.3 percentage points respectively compared to the original YOLOv5s model,and was superior to YOLOv7 tiny,YOLOv8s,and other models,while meeting the real-time detection requirements.In addition,the improved model performed better than the original model in dealing with complex situations such as occlusion and overlap,and had good robustness.

关 键 词：麦穗计数 估产 作物表型 YOLO 目标检测 

分 类 号：S126[农业科学—农业基础科学] TP391.4[自动化与计算机技术—计算机应用技术]