摘要
【目的】通过量化不同时段烟叶的表观特征值,掌握烟叶含水率的变化规律,针对烤烟含水率时序数据的非线性特征及相邻时段烤烟含水率间的较强关联关系,实现对不同时段烟叶含水率预测。【方法】利用全自动色差计提取烟叶颜色特征[明度值(L^(*))、红度值(a^(*))、黄度值(b^(*))、饱和度(C^(*))和色相角(H^(*))]和图像信息技术提取纹理特征[纹理均值(m)、标准差(σ)、平滑度(R)、三阶矩(μ3)、一致性(U)以及纹理熵(e)]并加以分析。利用因子分析得出6个表观优度因子[明度值(L^(*))、红度值(a^(*))、黄度值(b^(*))、纹理均值(m)、标准差(σ)和纹理熵(e)]作为网络模型的样本集,利用Kolmogorov定理和试验法确定预测模型隐含层为13个节点,采用L-M(Levenberg-Marquardt)训练算法,进行有限次的训练拟合,建立拓扑结构为6-13-1的BP神经网络模型,对烟叶含水率进行预测。【结果】烟叶图像的颜色特征和纹理特征值所呈现的变化符合烟叶在烘烤变化的表观体现,表明烟叶的图像信息较好地符合神经网络模型输入层与输出层相关性的要求。所建立的烟叶含水率BP(Back Propagation)神经网络预测模型的决定系数R 2为0.9987,均方根误差RMSE为0.0118,能够较好地预测不同时段烟叶的含水率。【结论】利用BP神经网络模型的不同时段烟叶含水率预测模型,可为掌握烟叶含水率动态变化规律提出一种解决方案。
【Objective】By quantifying the apparent characteristic values of tobacco leaves at different time periods,the change law of tobacco water content was mastered,and the water content of tobacco leaves at different time periods was predicted according to the nonlinear characteristics of time series data of flue-cured tobacco water content and the strong correlation between flue-cured tobacco water content at adjacent periods.【Method】Image information technology was used to extract and analyze color of tobacco leaves[lightness value(L^(*)),red degrees(a^(*)),yellow(b^(*)),saturation degree value(C^(*))and hue Angle(H^(*))]and texture characteristic[texture feature mean(m),standard deviation(σ),smoothness(R),three moments(μ3),consistency(U)and texture entropy(e)].Six apparent factors[lightness values(L^(*))red degree value(a^(*))yellow degree(b^(*))texture means(m)standard(σ)and entropy(e)texture]from principal component analysis were used as sample sets of network model.Kolmogorov theorem and test method were used to determine the prediction model for 13 hidden layer nodes.L-M(Levenberg-Marquardt)training algorithm used training fitting of limited time,to establish the topological structure of BP neural network model for the 6-13-1 and predict water content of tobacco leaf.【Result】The changes of tobacco leaf image color features and texture features of the value were in accord with the tobacco leaf apparent reflect changes in baking process.This indicated that tobacco leaf image information conformed to the neural network input layer and output layer correlation model.The tobacco moisture content of the BP neural network prediction model(the determination coefficient R 2 was 0.9987 and RMSE root mean square error was 0.0118)could well predict the moisture content of tobacco leaf in the different periods.【Conclusion】Therefore,the different periods of model of BP neural network model for predicting water in the tobacco leaf could be used to master dynamic change law of the moisture content of tobacco leaf.
作者
陈飞程
杨懿德
李常军
杨洋
冉茂
鄢敏
江厚龙
汪代斌
CHEN Fei-cheng;YANG Yi-de;LI Chang-jun;YANG Yang;RAN Mao;YAN Min;JIANG Hou-long;WANG Dai-bin(College of Tobacco,Henan Agricultural University,Henan Zhengzhou 450002,China;Sichuan Province Tobacco Company Yibin City Company,Sichuan Yibin 644000,China;China Tobacco Corporation Chongqing Company,Chongqing 400023,China)
出处
《西南农业学报》
CSCD
北大核心
2021年第11期2378-2384,共7页
Southwest China Journal of Agricultural Sciences
基金
中国烟草总公司重庆市公司资助项目(A2020NY01-1303-1)
四川省烟草公司宜宾市公司资助项目(201951150020107)。
关键词
烟叶
图像处理
BP神经网络
含水率检测
Tobacco leaf
Image processing
BP neural network
Moisture content detection
