基于高光谱成像技术和深度学习的烤后烟叶品种分类判别  

Classification and Identification of Tobacco Varieties after Curing Based on Hyperspectral Imaging and Deep Learning

作　　者：孟令峰 邢富康 韦克苏 王爱华[1] 程昌新[4] 曹娜 林跃平 王松峰[1] 朱荣光[2] MENG Lingfeng;XING Fukang;WEI Kesu;WANG Aihua;CHENG Changxin;CAO Na;LIN Yueping;WANG Songfeng;ZHU Rongguang(Institute of Tobacco Research of CAAS/Key Laboratory of Tobacco Biology and Processing,Ministry of Agriculture and Rural Affairs,Qingdao 266101,China;College of Mechanical and Electrical Engineering,Shihezi University,Shihezi 832003,Xinjiang,China;Guizhou Academy of Tobacco Science,Guiyang 550081,China;Hongyun Honghe Tobacco Co.,Ltd,Kunming 650202,China;Baoshan Branch of Yunnan Province Tobacco Company,Banshan 678000,Yunnan,China)

机构地区：[1]中国农业科学院烟草研究所/农业农村部烟草生物学与加工重点实验室,青岛266101 [2]石河子大学机械电气工程学院,新疆石河子832003 [3]贵州省烟草科学研究院,贵阳550081 [4]红云红河烟草(集团)有限责任公司,昆明650202 [5]云南省烟草公司保山市公司,云南保山678000

出　　处：《中国烟草科学》2024年第4期83-92,共10页Chinese Tobacco Science

基　　金：中国农业科学院科技创新工程(ASTIP-TRIC03);中国烟草总公司重点项目(110202102007);红云红河集团项目(HYHH2019YL02)。

摘　　要：为了实现对不同品种烤后烟叶的快速判别,采集红花大金元、K326、云烟105三种烟叶的高光谱图像,利用阈值分割和边缘提取算法选择感兴趣区域,获得烟叶的光谱数据。分别使用卷积平滑结合一阶导数、标准正态变量变换、多元散射校正和中心化等4种常用方法对原始光谱数据进行预处理,在此基础上建立多层感知机网络(Multilayer Perceptron,MLP)、卷积神经网络(Convolutional Neural Network,CNN)和支持向量机(Support Vector Machine,SVM)模型,进行烤后烟叶品种分类的比较研究。结果表明,所建立的模型均可对烟叶的品种进行有效判别,其中一阶导数预处理在多个模型中表现均优于其他3种预处理方法,基于CNN建立的分类模型性能优于其他模型,两者结合的性能最优,其测试集正确率和准确率分别为97.9%和98%,表明高光谱结合CNN可以实现烟叶品种的定性判别。上述研究为后续基于高光谱技术开发针对烤后烟叶的快速检测系统提供了理论依据和技术支撑。In order to realize rapid identification of different varieties of cured tobacco leaves,hyperspectral images of Honghua Dajinyuan,K326 and Yunyan 105 tobacco leaves were collected,and the region of interest was selected by threshold segmentation and edge extraction algorithm to obtain spectral data of tobacco leaves.The raw spectral data were preprocessed by four common methods:convolution smoothing combined with first-order derivative,standard normal variable transformation,multivariate scattering correction and centralization.On this basis,Multilayer Perceptron(MLP)network was established.MLP,Convolutional Neural Network(CNN)and Support Vector Machine(SVM)models were used to compare the classification of varieties using cured tobacco leaves.The results showed that all the established models could effectively distinguish the varieties of tobacco leaves,among which the first-order derivative pretreatment performed better than the other three pretreatment methods in several models,and the classification model based on CNN performed better than the other models.The combined performance of the two models was the best,and the correctness and accuracy of the test set were 97.9%and 98%,respectively.Hyperspectrum combined with CNN could realize the qualitative identification of tobacco varieties.The results of this study provide theoretical basis and technical support for the subsequent development of rapid detection system for cured tobacco based on hyperspectral technology.

关 键 词：烤后烟叶 品种分类 高光谱图像 深度学习 机器学习 

分 类 号：S572[农业科学—烟草工业] S126[农业科学—作物学]