雪茄烟烟支外观质量检测方法  

An inspection method for cigar appearance quality

作　　者：张鹏飞[1] 陈姣文 郭洪滨 杨荣超[1] 贺琛[1] 张勍[1] 李青常[1] 李素晓[3] 慎龙舞[1] 苗芊[1] 薛芳 ZHANG Pengfei;CHEN Jiaowen;GUO Hongbin;YANG Rongchao;HE Chen;ZHANG Qing;LI Qingchang;LI Suxiao;SHEN Longwu;MIAO Qian;XUE Fang(Zhengzhou Tobacco Research Institute of CNTC,Zhengzhou 450001,China;Technology Center,China Tobacco Sichuan Industrial Co.,Ltd.,Chengdu 610066,China;School of Physics and Engineering,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]中国烟草总公司郑州烟草研究院,郑州450001 [2]四川中烟工业有限责任公司技术中心,成都市610066 [3]郑州大学物理学院,郑州市450001

出　　处：《烟草科技》2024年第1期91-97,共7页Tobacco Science & Technology

基　　金：国家烟草专卖局标准预研项目“卷烟盒条包装外观质量在线视觉检测设备性能指标检测及溯源方法研究”(国烟科[2023]35号);四川中烟工业有限责任公司科技项目“雪茄烟支外观质量视觉检测设备计量测试技术研究与应用”(2022510000340243)。

摘　　要：为解决雪茄烟烟支外观质量检测自动化程度低、人为影响因素大等问题,研制了一种雪茄烟烟支外观质量检测装置,并构建了基于机器视觉技术和深度学习的烟支外观质量缺陷检测模型。通过检测装置采集雪茄烟柱面缺陷图像,并利用随机平移、翻转等方法对图像进行均衡和增强,建立数据集;使用K-means聚类算法生成适用于柱面缺陷数据集的先验框;对YOLOv5模型进行训练,并验证检测效果;采用大津法求得最佳阈值对雪茄烟端面空头图像进行分割,通过统计烟丝缺失比例实现空头缺陷检测。结果表明:①采用YOLOv5模型对柱面缺陷进行检测,多类别平均准确率MAP(MeanAverage Precision)为87.7%,单张图像检测时间为13.1 ms/张,模型抗干扰能力强,检测精度和时间均优于对比模型。②YOLOv5模型能够准确识别和定位多尺寸、多种类以及密集缺陷,具有较强泛化能力和鲁棒性。③在不同亮度光照环境下,均能够实现空头缺陷检测。该方法可为提高雪茄烟产品质量提供支持。For the automation of cigar appearance inspection,a cigar appearance inspection device was developed,and a cigar appearance inspection model based on machine vision technology and deep learning was established.The images of defects of the cylindrical surface of cigars were captured by a test device.The captured images were equilibrated and augmented by random panning and flipping to create a data set.The K-means clustering algorithm was used to generate appropriate anchors for the cigar defect dataset.The YOLOv5 model was trained and verified.The optimal threshold was obtained using the OTSU method to segment the images of loose end cigars.The loose end defect was identified by statistically computing the proportion of tobacco loss from the cigar end.The results showed that:1)Using the YOLOv5 model to identify the defects of cylindrical surface,the mean average precision(MAP)reached 87.7%and the time for going through a single image was 13.1 ms.The YOLOv5 model featured strong anti-interference and advantages over other models in terms of inspection accuracy and efficiency.2)The YOLOv5 model was able to accurately identify and locate multi-size,multi-category and densely distributed defects with strong generalization ability and robustness.3)The loose end defect could be identified under different luminance lighting environments.This method supports the promotion of cigar quality.

关 键 词：雪茄烟 外观质量 机器视觉 深度学习 缺陷检测 

分 类 号：TS453[农业科学—烟草工业] TS46