基于深度学习的烟叶等级分类及特征可视化  

Tobacco leaf grading and feature visualization based on deep learning

作　　者：鲁梦瑶 陈栋 周强 王志勇[3] 陈天恩 姜舒文 LU Mengyao;CHEN Dong;ZHOU Qiang;WANG Zhiyong;CHEN Tianen;JIANG Shuwen(Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;NONGXIN(Nanjing)Smart Agriculture Research Institute,Nanjing 211800,China;Anhui Wannan Tobacco Co.,Ltd.,Xuancheng 242000,Anhui,China;National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China)

机构地区：[1]北京市农林科学院信息技术研究中心,北京市100097 [2]农芯(南京)智慧农业研究院,南京市211800 [3]安徽皖南烟叶有限责任公司,安徽省宣城市242000 [4]国家农业信息化工程技术研究中心,北京市100097

出　　处：《烟草科技》2023年第6期92-100,共9页Tobacco Science & Technology

基　　金：安徽皖南烟叶有限责任公司科技重点项目“基于深度学习的人工智能技术在皖南烟区烟叶分级中的应用研究”(20180563006)、“烟叶自动分级关键技术研究与原型样机开发”(20190563001);云南省烟草公司科技计划项目“互联网+”烟叶生产管理及服务体系研究与应用(2020530000241027)。

摘　　要：为探索深度学习技术在烟叶图像上的特征提取效果,提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)模型的烟叶等级分类方法,并对模型关注的烟叶特征进行了可视化分析。通过图像预处理得到高分辨率的局部烟叶图像,以弥补全局烟叶图像缩放后导致烟叶细节信息丢失;利用改进的CNN模型VGG-16和ResNet-50分别提取烟叶全局和局部图像特征;构建分类器对烟叶全局和局部图像的特征向量进行分类和结果融合;采用类别激活图(Class Activation Map,CAM)技术绘制模型关注烟叶特征的热力图。结果表明:提出的方法对6个等级的烟叶分级准确率达到84.71%,单张烟叶图像测试时间为17.87 ms;特征热力图显示ResNet-50模型对烟叶病斑、皱褶、主脉和纹理走势等局部特征较为敏感。该方法可为实现烟叶快速、准确分级提供支持。To investigate the feature extraction effect of deep learning technology on tobacco leaf images,a tobacco leaf grading method based on a Convolutional Neural Network(CNN)model was proposed,and the tobacco leaf features involved in the model were visualized and analyzed.The local tobacco leaf images with high-resolution were obtained by image preprocessing to compensate for the loss of tobacco leaf details due to zooming the global tobacco leaf image.The modified CNN models VGG-16 and ResNet-50 were used to extract the features of the global and local tobacco leaf images respectively.A classifier was configured to classify and fuse the feature vectors of the global and local tobacco images.The Class Activation Map(CAM)method was used to draw the thermodynamic charts of the tobacco leaf features involved in the models.The results showed that the accuracy of the proposed method for grading tobacco leaves of six grades reached 84.71%,and the test time for a single tobacco leaf image was 17.87 ms.The thermodynamic charts of the features indicated that the ResNet-50 model was more sensitive to the local features of tobacco leaves,such as disease spot,wrinkle,main vein and texture trend.The proposed method provides support for the fast and accurate grading of tobacco leaves.

关 键 词：烟叶分级 烟叶图像 深度学习 CNN模型 特征可视化 类别激活图 

分 类 号：TS442[农业科学—烟草工业] TS46