基于PMMS-Net和叶绿素荧光成像的绿豆叶斑病抗性鉴定方法  

Identification method of resistance to mung bean leaf spot disease based on PMMS-Net and chlorophyll fluorescence imaging

作　　者：李洁 高尚兵 余骥远 陈新[2] 李士丛 袁星星[2] Li Jie;Gao Shangbing;Yu Jiyuan;Chen Xin;Li Shicong;Yuan Xingxing(Faculty of Computer and Software Engineering,Huaiyin Institute of Technology,Huai′an,223003,China;Institute of Economic Crops,Jiangsu Academy of Agricultural Sciences,Nanjing,210014,China)

机构地区：[1]淮阴工学院计算机与软件工程学院,江苏淮安223003 [2]江苏省农业科学院经济作物研究所,南京市210014

出　　处：《中国农机化学报》2024年第8期210-216,共7页Journal of Chinese Agricultural Mechanization

基　　金：科技部重点研发政府间国际合作项目(2019YFE0109100);江苏省一带一路国际合作项目(BZ2022005);国家食用豆产业技术体系生物防治与综合防控岗位科学家(CARS—08—G15);江苏省种业揭榜挂帅项目(JBGS[2021]004);江苏省林业科技创新与推广项目(LYKJ[2021]22);江苏省研究生科研与实践创新计划项目(SJCX_231859)。

摘　　要：考虑到相近发病指数的绿豆叶片病斑区域特征区分不明显,在检测类似大小的病斑时,使用固定尺度的卷积核检测效果不是很好,故设计一种并行多分支多尺度卷积神经网络(PMMS-Net)模型。该模型先使用并行多分支多尺度特征融合模块获取丰富的病斑特征;然后采用坐标注意力机制,使模型能更好地定位到病变区域,实现对感兴趣区域的选择性强调;最后使用特征充分提取模块,将深度可分离卷积与普通卷积结合,实现对特征的充分提取,进一步优化特征提取效果。试验数据集由绿豆叶斑病的叶绿素荧光图像构成,其中包含五种抗性类型的绿豆叶斑病图像。结果表明,本文提出的方法在试验数据集上训练迭代1000次,所耗费时间仅比AlexNet多0.8倍,验证准确率却比AlexNet高出18.9%,本模型在该数据集上的验证准确率为87.8%,平均特异度为96.92%,参数内存仅为0.54 MB。本文提出的方法有利于将该模型部署在移动终端等资源受限的嵌入式设备上,为绿豆叶斑病的抗性鉴定提供一种新方法。Considering that the regional characteristics of mung bean leaf disease spots with similar disease index are not clearly differentiated,the effect of using fixed scale convolution kernel to detect disease spots with similar size is not very good,so a parallel multi branch multi⁃scale convolution neural network(PMMS-Net)model is designed.Firstly,the model uses parallel multi branch and multi⁃scale feature fusion module to obtain rich plaque features.Then the coordinate attention mechanism is used to make the model better locate the lesion area and realize selective emphasis on the region of interest.Finally,full feature extraction module is used to combine the depth separable convolution with the ordinary convolution to achieve the full extraction of features and further optimize the effect of feature extraction.The experimental data set consists of chlorophyll fluorescence images of mung bean leaf spot disease,including five resistance types of mung bean leaf spot disease images.The results show that the method proposed in this paper takes only 0.8 times more time to train 1000 iterations on the dataset than AlexNet,but the verification accuracy is 18.9%higher than AlexNet.The verification accuracy of this model on the dataset is 87.8%,the average specificity is 96.92%,and the parameter memory is only 0.54 MB.The method proposed in this paper is conducive to deploying the model on embedded devices with limited resources,such as mobile terminals,and provides a new method for identification of resistance to mung bean leaf spot.

关 键 词：绿豆叶斑病 抗性鉴定 叶绿素荧光图像 坐标注意力机制 深度可分离卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S435.22[自动化与计算机技术—计算机科学与技术]