基于改进YOLOv5的轻量级大豆田间杂草识别方法研究  

Research on Lightweight Soybean Field Weed Identification Method Based on Improved YOLOv5

作　　者：陈雪 谭峰[1] CHEN Xue;TAN Feng(Heilongjiang Bayi Agricultural University,Daqing 163319,China)

机构地区：[1]黑龙江八一农垦大学,黑龙江大庆163319

出　　处：《农机使用与维修》2025年第2期1-7,共7页Agricultural Machinery Using & Maintenance

基　　金：国家重点研发计划课题(2023YFD2301605);黑龙江省重点研发计划指导类项目(GZ20220020);黑龙江省自然科学基金项目(LH2023F043);黑龙江省高等教育教学改革研究项目(SJGY20210622);黑龙江八一农垦大学自然科学人才支持计划(ZRCPY202015)。

摘　　要：为对自然环境下大豆田间的大豆幼苗与杂草实现实时检测,提出一种基于改进YOLOv5的轻量级大豆田间杂草识别方法。该方法在YOLOv5s网络的基础上引入适用于边缘设备的轻量级卷积神经网络MobileNetV3,建立了YOLOv5s-MobileNetV3的大豆杂草识别模型。对含有大豆幼苗和杂草的图像进行数据增强预处理,以增强模型泛化能力,利用Bneck深度卷积神经网络架构提取特征,并通过随机梯度下降法作为优化训练模型。在对比试验中,分别将YOLOv5s的主干网络替换为EfficientNetV2和ShuffleNetV2,并选取传统的SSD、Faster R-CNN、YOLOv4等标准目标检测网络作为参照,进行性能评估。试验结果表明:YOLOv5s-MobileNetV3识别模型大小为7.8 MB,在CPU上单张图像的推理时间为6.9 ms,识别精确率为0.846,查全率为0.84,均值平均精度为0.875,F1分数为0.843。模型可快速识别大豆伴生杂草,为挂载式喷药、精准农业等领域,提供农作物与杂草快速准确的检测分类方法,以满足作物生长管理、变量施肥及靶向喷药的技术需求。In order to realize real-time detection of soybean seedlings and weeds in soybean fields under field environments,a lightweight soybean field weed recognition method based on improved YOLOv5 is proposed.The method introduces MobileNetV3,a lightweight convolutional neural network suitable for mobile and edge devices,on the basis of YOLOv5s network,and establishes a YOLOv5s-MobileNetV3 soybean weed recognition model.The captured images containing soybean seedlings and weeds were subjected to data enhancement preprocessing techniques to enhance the model generalization ability,features were extracted using the Bneck deep convolutional neural network architecture,and the model was trained by stochastic gradient descent method as an optimization.In the comparison experiments,the backbone network of YOLOv5s was replaced with EfficientNetV2 and ShuffleNetV2,respectively,and standard target detection networks such as traditional SSD,Faster R-CNN,and YOLOv4 were selected as references to perform a comprehensive performance evaluation.The experimental results show that:YOLOv5s-MobileNetV3 recognition model size is 7.8 MB,the inference time of a single image on CPU is 6.9 ms,the recognition accuracy is 0.846,the check-all rate is 0.84,the mean average accuracy is 0.875,and the F1 score is 0.843.The model can quickly recognize the accompanying weeds in soybean,which can be used for the purpose of mounted spraying,precision agriculture The model can quickly identify soybean companion weeds and provide a fast and accurate detection and classification method for crops and weeds in order to meet the technical needs of crop growth management,variable fertilization and targeted spraying.

关 键 词：大豆 苗草识别 目标检测算法 MobileNetV3 

分 类 号：S126[农业科学—农业基础科学]