基于多光谱与深度学习技术的皱褶初烤烟叶等级识别  

Grade identification of folded flue-cured tobacco based on multi-spectrum with deep learning

作　　者：叶大鹏[1,2] 黄俊昆 秦华 翁海勇 卢敏瑞[3] 王芳 李庆 YE Dapeng;HUANG Junkun;QIN Hua;WENG Haiyong;LU Minrui;WANG Fang;LI Qing(College of Mechanical and Electrical Engineering,Fujian Agriculture and Forestry University,Fuzhou,Fujian 350002,China;Fujian Colleges and Universities Engineering Research Center of Modern Agricultural Equipment,Fuzhou,Fujian 350002,China;Fujian Wuyi Tobacco Leaf Co.LTD,Nanping,Fujian 350005,China)

机构地区：[1]福建农林大学机电工程学院,福建福州350002 [2]现代农业装备福建省高校工程研究中心,福建福州350002 [3]福建武夷烟叶有限公司,福建南平350005

出　　处：《福建农林大学学报（自然科学版）》2022年第2期282-288,共7页Journal of Fujian Agriculture and Forestry University:Natural Science Edition

基　　金：福建省农业工程高原学科建设项目(712018014)。

摘　　要：为解决皱褶初烤烟叶分级过程中存在的人工成本高、分级效率低等问题,本研究提出一种基于多光谱成像技术结合深度学习的皱褶初烤烟叶等级识别方法.通过使用连续投影算法选取皱褶初烤烟叶高光谱反射图像的12个特征波段,并以此为基础搭建多光谱成像设备,采集了B2F、C2F、C3F和C4F四个等级皱褶初烤烟叶的多光谱反射图像,形成模型训练和验证的数据集.通过改造VGG11网络模型,用上述训练集构建烟叶等级识别网络(tobacco gradeidentification network, TGIN)模型,用验证集测试该模型烟叶等级识别的正确率为99.8%.针对不同年份的烟叶样本,引入迁移学习方法后,其识别的平均正确率也可达到99.4%.结果显示,TGIN模型能够实现皱褶初烤烟叶等级的快速识别,可为皱褶初烤烟叶自动化等级识别提供理论基础和技术支持.In order to solve the problems of high labor cost and low classification efficiency in the grading process of folded flue-cured tobacco, a grade identification method was proposed based on multispectral imaging technology with deep learning. The continuous projection algorithm was used to screen 12 characteristic bands of hyperspectral reflectance images of folded flue-cured tobacco, and a multispectral imaging device was built on this basis. The multispectral reflectance images of 4 grades of folded flue-cured tobacco(B2 F, C2 F, C3 F and C4 F) were collected to form a dataset for model training and verification. By transforming the VGG11 network model, a tobacco grade identification network(TGIN) model was constructed with previous training set, with the average accuracy of the verification set reaching 99.8%. For tobacco samples collected in different years, the recognition accuracy also reached at 99.4 % after introduced with transfer learning method. To summarize, the TGIN model can quickly identify the grade of folded flue-cured tobacco, to provide a theoretical basis and technical support for the automatic grade identification of folded flue-cured tobacco.

关 键 词：皱褶初烤烟叶 深度学习 多光谱成像技术 烟叶等级识别网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]