基于改进YOLOv7的线虫智能识别研究  

Intelligent Identification of Nematodes Based on Improved YOLOv7 Neural Network

作　　者：牛成文 侯华鑫 谢雯媛 王秀丽[1] 殷汝枭 曲建平 王己光 周波[4,5] NIU Cheng-wen;HOU Hua-xin;XIE Wen-yuan;WANG Xiu-li;YIN Ru-xiao;QV Jian-ping;WANG Ji-guang;ZHOU Bo(College of Information Science and Engineering/Shandong Agricultural University,Tai'an 271018,China;School of Finance/Shandong University of Finance and Economics,Ji'nan 250014,China;College of Computer Science and Technology/Harbin Engineering University,Ha'erbin 150001,China;College of Life Sciences/Shandong Agricultural University,Tai'an 271018,China;Shandong Future Biotechnology Co.Ltd.,Tai'an 271000,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018 [2]山东财经大学金融学院,山东济南82911086 [3]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001 [4]山东农业大学生命科学学院,山东泰安271018 [5]山东未来生物科技有限公司,山东泰安271000

出　　处：《山东农业大学学报(自然科学版)》2024年第1期100-107,共8页Journal of Shandong Agricultural University:Natural Science Edition

基　　金：一种线虫高通量智能识别系统;供港叶菜土壤连作障碍预警与综合修复技术研究(2021BBF02006);马铃薯疮痴病发生机理及防控关键技术研究(2023BCFO1015)。

摘　　要：线虫存活率是杀线试剂活性测试的重要指标,目前线虫计数多以显微镜下的人工识别方式为主,存在耗时长、准确率低、工作量大等问题,利用卷积神经网络实现线虫的智能识别与计数是解决上述问题的重要途径。本文基于YOLOv7网络架构进行了三方面改进:主干网络添加ECA注意力机制;用EIoU替换原模型损失函数;将原本的激活函数替换为Mish激活函数。对比试验测试发现,改进后YOLOv7模型的mAP达到了95.3%,与SSD、Faster-RCNN等经典目标检测算法相比分别提高12.3、6.2个百分点,在准确率、召回率和F1因子上分别提高了0.6、2.4和1.5个百分点,且减少了冗余信息的干扰,增强了多尺度目标的特征提取能力;提高了重叠黏连线虫目标的检测能力和回归精度。此外,本文基于Vue、SpringBoot等技术开发了一款线虫存活状态检测系统,将该系统与本文改进后的模型部署到服务器,为研究人员提供了方便、高效的线虫死/活状态在线智能识别与计数服务。Nematode survival rate is an important indicator for assessing the effectiveness of nematicidal reagents.Currently,nematode counting relies heavily on manual identification under a microscope,which is time-consuming,inaccurate and labor-intensive,etc.The use of a convolutional neural network to achieve intelligent identification and counting of nematodes is a crucial method to solve the above problems.We proposes an improved YOLOv7 neural network model with three improvements:adding ECA attention mechanism modules to the main network;optimizing the loss function of the original model by EIoU and replacing the original activation function with the Mish activation function.Comparative experimental tests reveal that the mAP of the improved YOLOv7 model reaches 95.3%,which is 12.3 and 6.2 percentage points higher than that of the classical target detection algorithms,such as SSD and Faster-RCNN,and 0.6,2.4 and 1.5 percentage points higher than that of SSD,Faster-RCNN and other classic target detection algorithms in terms of the accuracy rate,the recall and the F1 factor,respectively.Additionally,the model reduces redundancy,enhances multiscale feature extraction,and improves detection and regression precision for overlapping nematodes.In order to implement the research,the improved model was deployed to the server,and an application system for nematode survival status detection was developed using Vue,SpringBoot and other technologies,these provide the convenient and efficient nematode identification and counting service for researchers.

关 键 词：目标检测 注意力机制 损失函数 激活函数 线虫 

分 类 号：S154.386[农业科学—土壤学] TP391.41[农业科学—农业基础科学]