基于残差神经网络的烟丝类型识别方法的建立  

A method for identifying types of tobacco strands based on residual neural network

作　　者：钟宇 周明珠[2] 徐燕 刘德祥 王宏强 董浩[2,3,4] 禹舰 李晓辉[2] 杨进[2] 邢军[2] ZHONG Yu;ZHOU Mingzhu;XU Yan;LIU Dexiang;WANG Hongqiang;DONG Hao;YU Jian;LI Xiaohui;YANG Jin;XING Jun(Xinjiang Tobacco Quality Supervision and Test Station,Urumqi 830026,China;China National Tobacco Quality Supervision and Test Center,Zhengzhou 450001,China;Opto-Electronics Technology Research Center,Anhui Institute of Optics and Fine Mechanics,Hefei Institutes of Physical Science,Chinese Academy of Sciences,Hefei 230031,China;University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]新疆维吾尔自治区烟草质量监督检测站,乌鲁木齐830026 [2]国家烟草质量监督检验中心,郑州450001 [3]中国科学院合肥物质科学研究院安徽光学精密机械研究所光电子技术研究中心,合肥市230031 [4]中国科学技术大学,合肥市230026

出　　处：《烟草科技》2021年第5期82-89,共8页Tobacco Science & Technology

基　　金：国家烟草专卖局科技重大专项项目“卷烟产品鉴别大数据构建及应用研究”[110201901026(SJ-05)]。

摘　　要：为快速、准确地识别叶丝、梗丝、膨胀叶丝、再造烟叶丝等烟丝类型,利用各类烟丝图像特征差异,以残差神经网络为基础构建了识别模型,并对模型的预训练权值、优化算法、学习率等超参数进行了研究,结果表明:①基于残差神经网络的识别方法可以有效识别4种类型烟丝,相比基于卷积神经网络的识别方法,模型具有更高的识别率、泛化能力与鲁棒性。②较优超参数对模型的训练速度及表现影响显著,通过训练得到的模型在测试集上的准确率及召回率均高于96%,且与训练集表现差异较小。该方法可为提高烟丝类型识别效率和准确性提供支持。For rapidly and accurately identifying expanded tobacco and the strands of tobacco strips,stems,expanded tobacco and reconstituted tobacco,an identification model on the basis of residual neural network was developed by their image characteristics.The pre-training weights,optimized algorithm,learning rate and other super parameters of the model were studied as well.The results indicated that:(1)The developed method could effectively identify the above mentioned four types of tobacco strands,and the developed model presented higher identification rate,generalization capability and robustness comparing with the model based on convolutional neural network.(2)The super parameters influenced the training speed and performance of the model obviously,and the identification rate and recall rate of the model trained with the optimized super parameters were higher than 96%for the test set,which were close to that of the training set.This method provides a support for promoting the efficiency and accuracy of tobacco strands identification.

关 键 词：烟丝类型 识别 残差神经网络 卷积神经网络 学习率 

分 类 号：TS439[农业科学—烟草工业]