基于一维残差卷积的烟叶分级方法研究  

Research on tobacco grade classification based on one⁃dimensional residual convolution

作　　者：孙祥洪 罗智勇 SUN Xianghong;LUO Zhiyong(Information Center,China Tobacco Jiangxi Industrial Co.,Ltd.,Nanchang 330096,China;College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]江西中烟工业有限责任公司技术中心,江西南昌330096 [2]青岛科技大学信息科学技术学院,山东青岛266061

出　　处：《现代电子技术》2024年第2期165-170,共6页Modern Electronics Technique

摘　　要：在烟叶分级过程中,由于人为主观性、分级标准不一致等因素导致分级结果不一致。针对以上问题,提出一种一维残差卷积的烟叶等级分类模型。首先,改进VGG16网络,将方形矩阵卷积核和池化窗口改为适应于一维光谱数据的向量卷积核和池化窗口。然后,利用BasicBlock残差模块替换多层卷积叠加的结构,对光谱数据进行更深层的提取,防止梯度消失问题。最后,在卷积层后面接入BN层模块,通过归一化的方式,防止卷积计算后由于数据分布分散而导致的网络效率降低问题。选取B2V、B1F、C4F、C1L和X2L等5种不同等级的烟叶样本的近红外光谱数据进行实验。结果表明,所提方法对5种等级烟叶训练集和测试集的平均分类准确率分别为98.0%和97.3%,明显高于其他方法。该方法在一定程度上解决了烟叶人工分级带来的误差,减少了人力输出,提高了效率。In the tobacco grading process,inconsistent grading results are often observed due to factors such as human subjectivity and inconsistent grading standards.To address these issues,a tobacco grade classification model based on one‐dimensional residual convolution is proposed.The VGG16 network is improved by replacing the square matrix convolutional kernels and pooling windows with vector convolution kernel and pooling window suitable for one‐dimensional spectral data.The BasicBlock residual module is employed to replace the structure of multi‐layer convolutional stacking for deeper extraction of spectral data and prevention of gradient vanishing issues.A BN layer module is added behind the convolutional layer to prevent the network efficiency reduction caused by scattered data distribution after convolutional computation by means of the normalization way.The near‐infrared spectral data of five different grades of tobacco leaf samples,including B2V,B1F,C4F,C1L,and X2L are selected for experiments.The results show that the average classification accuracy of the training and testing sets for five levels of tobacco leaves in the proposed method is 98.0%and 97.3%,respectively,which is significantly higher than those of other methods.This method to some extent can solve the errors caused by manual grading of tobacco leaves,reduce manpower output,and improve efficiency.

关 键 词：烟叶分级 残差卷积神经网络 残差模块 近红外光谱 数据特征提取 数据采集 

分 类 号：TN219‐34[电子电信—物理电子学] TP391[自动化与计算机技术—计算机应用技术]