基于知识蒸馏的多教师棉田杂草检测模型  

Multi-teacher cotton field weed detection model based on knowledge distillation

作　　者：朱养鑫 郝珊珊 郑伟健 金诚谦 印祥 周鹏 ZHU Yangxin;HAO Shanshan;ZHENG Weijian;JIN Chengqian;YIN Xiang;ZHOU Peng(School of Agricultural Engineering and Food Science,Shandong University of Technology,Zibo 255000,China;Shandong University of Technology Institute of Modern Agricultural Equipment,Zibo 255000,China;Shandong Provincial Key Laboratory of Smart Agricultural Technology and Intelligent Agricultural Machinery Equipment for Field Crops,Zibo 255000,China;Nanjing Institute of Agricultural Mechanization,Ministry of Agriculture and Rural Affairs,Nanjing 210014,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]山东理工大学现代农业装备研究院,淄博255000 [3]山东省大田作物智慧农业技术与智能农机装备重点实验室,淄博255000 [4]农业农村部南京农业机械化研究所,南京210014

出　　处：《农业工程学报》2025年第7期200-210,共11页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(62305197);山东省重点研发计划项目(2022SFGC0201);国家重点研发计划项目(2021YFD200050205)。

摘　　要：杂草对农业生产的负面影响日益严重,高效的杂草检测方法对推动农业可持续发展至关重要。然而,现有的杂草检测模型在检测精度和效率方面仍存在提升空间。该研究基于YOLOv5,引入知识蒸馏技术,并整合多个教师模型的知识,旨在提高模型在棉田杂草检测中的精度和实时检测能力。首先,提出了一种基于温度系数的Logits软投票机制,通过动态调整教师模型的蒸馏损失权重,有效融合各教师模型的优势。同时,还提出了基于注意力机制的多教师特征融合方法,动态加权突出关键特征,抑制冗余信息。结果表明,该研究提出的YOLOv5s-CWD在模型精度和实时检测性能上均表现优异,F1分数(F1-score, F_(1))为94.5%,平均精度均值(mean average precision, mAP)在验证集和测试集上分别为96.8%和93.6%,帧率(frames per second, FPS)为46.71帧/s,浮点运算数(giga floating point operations, GFLOPs)为4.1,模型大小为2.9 MB。与YOLOv5s相比,YOLOv5s-CWD在验证集上的m AP仅降低0.9个百分点,FPS提高约57.22%,计算复杂度降低约74.38%,模型大小减少约79.86%。与YOLOv7相比,YOLOv5s-CWD在验证集上的mAP降低1.3个百分点,FPS提高约1 076.57%,计算复杂度和模型大小大幅优化。与YOLOv10s相比,YOLOv5s-MGD在验证集上的mAP降低0.9个百分点,FPS提高约143.41%,计算复杂度降低约83.27%,模型大小降低约82.42%。综上所述,YOLOv5s-CWD在保证高精度的同时,显著提升了检测速度、计算效率和存储性能,适用于在性能受限设备上实时检测,为棉田杂草检测提供了有力的技术支持。As the negative impact of weeds on agricultural production becomes increasingly severe,efficient weed detection methods are crucial for promoting sustainable agricultural development.However,existing weed detection models still face challenges in terms of accuracy and efficiency.To address this issue,this paper,based on YOLOv5,introduced Knowledge Distillation techniques and integrated the"knowledge"from multiple-teacher models,aiming to enhance performance and realtime detection ability of the weed detection model.Firstly,a temperature coefficient-based soft voting mechanism for logits(TS_Logits)was proposed,which dynamically adjusted the distillation loss weights of the teacher models,effectively integrating the strengths of each teacher model.Meanwhile,a multi-teacher feature fusion method based on attention mechanisms(AT_Feature)was also proposed,which dynamically weighted key features and suppressed redundant information.In this paper,the open-source CottonWeedDet12 dataset was used,with data augmentation techniques applied for random expansion,ultimately obtaining 9210 weed images.These images were divided into training and validation sets in an 8:2 ratio,and a test set consisting of 554 actual weed images was also incorporated.Based on YOLOv5s,channel pruning techniques were applied to obtain a student model with a size of 2.9 MB,an F_(1)-score( F_(1))of 94.5%,and a mean average precision of validation set(mAP_(val))of 96.8%.Meanwhile,YOLOv5s,YOLOv7 and YOLOv10s were used as teacher models,and three Logits-based knowledge distillation methods(KD,Luminet,and DKD)and ten Feature-based knowledge distillation methods(FitNet,AT,NST,PKT,RKD,VID,SemCKD,CWD,MGD,and FGD)were applied.The experimental results showed that the student model achieved the best performance when YOLOv5s was used as the teacher model,further proving that using the same model architecture facilitated knowledge transfer between the teacher and student models,thereby enhancing the distillation effect.Additionally,this study evaluated the effec

关 键 词：杂草检测 图像处理 深度学习 YOLO 知识蒸馏 

分 类 号：S126[农业科学—农业基础科学] S233.3[自动化与计算机技术—控制理论与控制工程] TP181[自动化与计算机技术—控制科学与工程]