应用MCCW-YOLOv7-tiny研究轻量级玉米田间杂草识别算法  

Research on lightweight cornfield weed identification algorithm based on MCCW-YOLOv7-tiny

作　　者：王希如 贾仁山 曹玉莹 刘银川 高新悦 吴佳鑫 贾银江[1] WANG Xiru;JIA Renshan;CAO Yuying;LIU Yinchuan;GAO Xinyue;WU Jiaxin;JIA Yinjiang(College of Electrical and Information Engineering,Northeast Agricultural University,Harbin 150030,China;College of General Education,Heilongjiang Polytechnic,Harbin 150086,China)

机构地区：[1]东北农业大学电气与信息学院,哈尔滨150030 [2]黑龙江职业学院通识教育学院,哈尔滨150086

出　　处：《东北农业大学学报》2025年第1期124-138,共15页Journal of Northeast Agricultural University

基　　金：黑龙江省揭榜挂帅科技攻关项目(20212XJ05A0201)。

摘　　要：针对玉米田间杂草传统目标检测模型存在体积大、实时性差、精准度低、移动端部署难等问题,提出了改进的轻量级目标检测算法MCCW-YOLOv7-tiny。通过将YOLOv7-tiny主干网络CSPDarknet替换为MobileNet V3模块,降低模型冗余和参数量,满足实时性要求。添加并行网络至主干网络,并在检测头部引入CBAM注意力机制,增强对小目标的关注,提高整体识别精度。损失函数改进为WIoUv3,以动态平衡样本质量,提升模型泛化能力。结果表明:MCCW-YOLOv7-tiny较YOLOv7-tiny,mAP从93.7%提升至95.3%,计算复杂度从13.3 GFLOPs降至6.2 GFLOPs,模型参数量为3.71 M,为复杂环境下的玉米田间杂草检测提供有效技术支持。In response to the issues of large model size,poor real-time performance,low accuracy,and challenges in deploying on mobile devices in traditional weed detection models for corn fields,an improved lightweight object detection algorithm,MCCW-YOLOv7-tiny was propose.This algorithm reduced model redundancy and parameter count by replacing the YOLOv7-tiny backbone network,CSPDarknet,with the MobileNet V3 module,thereby meeting real-time requirements.A parallel network was added to the backbone,and the CBAM attention mechanism was introduced into the detection head to enhance the model's focued on small objects,thereby improving overall recognition accuracy.The loss function was modified to WIoUv3,dynamically balancing sample quality and enhancing the model's generalization ability.Experimental results showed that MCCW-YOLOv7-tiny outperformed YOLOv7-tiny,with the mAP increasing from 93.7%to 95.3%,computational complexity decreasing from 13.3 GFLOPs to 6.2 GFLOPs,and the model's parameter size reduced to 3.71 M.This provided effective technical support for weed detection in corn fields under complex conditions.

关 键 词：YOLOv7-tiny MobileNet V3 CBAM注意力机制 WIoUv3 玉米 杂草识别 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]