摘要
为了解决绿豆叶斑病不同病害等级之间容易混淆的问题,本研究以感染不同程度叶斑病的绿豆叶片叶绿素荧光图像为研究对象,提出了多模块串联卷积神经网络(Multi-Module Sequential Convolutional Neural Network,MMS-Net)模型。该模型主要由本研究搭建的Sub模块和Wave模块串联堆叠组成,并且在每个Sub模块中和每个Wave模块结尾处加入混合注意力机制CBAM,在减少非叶斑病特征干扰的同时,对相似病斑进行更为细致的特征提取,提高病害识别的准确率。在相同条件下,与经典的卷积神经网络模型(VGG16、GoogLeNet、ResNet50)以及流行的轻量级卷积神经网络模型(MobileNetV2、MobileNeXt、MobileNetV3、Shuffle-NetV2)进行比较,本研究提出的MMS-Net模型参数量仅为11.43 M,测试准确率达到91.25%,均高于其他模型,分类效果最优。通过分析精度、召回率、F1分数等评价指标可以看出MMS-Net模型具有较好的鲁棒性和泛化能力。本研究结果可为绿豆等作物的抗病种质资源识别和筛选提供新思路。
In order to solve the problem of confusion among different disease levels of mung bean leaf spot,a Multi-Module Sequential Convolutional Neural Network(MMS-Net)model was proposed based on chlorophyll fluorescence imaging of mung bean leaves infected by the disease.The model was mainly composed of the Sub modules and Wave modules proposed in this article,and the Convolutional Block Attention Module(CBAM)was added into each Sub module and at the end of each Wave module,which could detect similar disease spot features in more detail and reduce the mixing of non-leaf spot features at the same time,thereby improved the accuracy rate of disease recognition.Under the same conditions,compared with several classic convolutional neural network models(VGG16,GoogLeNet,ResNet50)and popular lightweight convolutional neural network models(MobileNetV2,MobileNeXt,MobileNetv3,ShuffleNetV2),the parameter size of the MMS-Net model was only 11.43 M and the test accuracy was 91.25%,which were higher than those in the other models,so it showed the best classification effect.By analyzing evaluation indicators such as precision,recall rate and F1-score,it was concluded that the MMS-Net model exhibited better robustness and generaliza-tion ability,which could provide new ideas for screening disease-resistant germplasm resources of mung bean and other crops.
作者
张浩淼
高尚兵
蒋东山
李洁
袁星星
陈新
刘金洋
Zhang Haomiao;Gao Shangbing;Jiang Dongshan;Li Jie;Yuan Xingxing;Chen Xin;Liu Jinyang(Faculty of Computer and Software Engineering,Huaiyin Institute of Technology,Huai’an 223001,China;Institute of Economic Crops,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China)
出处
《山东农业科学》
北大核心
2024年第9期133-141,共9页
Shandong Agricultural Sciences
基金
国家自然科学基金面上项目(62076107)
科技部重点研发政府间国际合作项目“抗黄花叶病毒病绿豆新品种选育与示范推广”(2019YFE0109100)
江苏省一带一路国际合作项目“抗黄花叶病毒病绿豆新品种及绿色增产增效技术合作研发及海外应用示范”(BZ2022005)
江苏省种业揭榜挂帅项目“双抗杂交绿豆新种质创制及关键基因挖掘利用”(JBGS〔2021〕004)
江苏省研究生科研与实践创新计划项目(SJCX24_2145)。
关键词
绿豆叶斑病
病害等级
卷积神经网络
叶绿素荧光成像
注意力机制
Mung bean leaf spot
Disease degree
Convolutional neural network
Chlorophyll fluores-cence imaging
Attention mechanism
作者简介
张浩淼(2000—),男,硕士研究生,研究方向为深度学习和计算机视觉。E-mail:zhang760920@163.com;通信作者:高尚兵(1981—),男,博士,教授,研究方向为深度学习、计算机视觉、模式识别和数据挖掘。E-mail:11060036@hyit.edu.cn。
