盒装机制雪茄烟缺陷分类视觉检测方法  

A visual inspection method for integrity of machine-made cigars in packets

作　　者：杨心安 朱文魁[1] 张明建[1] 周博 王兵[1] 冯杨 高森 王珵珵 陈睿 YANG Xinan;ZHU Wenkui;ZHANG Mingjian;ZHOU Bo;WANG Bing;FENG Yang;GAO Sen;WANG Chengcheng;CHEN Rui(Zhengzhou Tobacco Research Institute of CNTC,Zhengzhou 450001,China;Great Wall Cigar Factory,China Tobacco Sichuan Industrial Co.,Ltd.,Shifang 618400,Sichuan,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长城雪茄烟厂,四川省什邡市618400

出　　处：《烟草科技》2023年第10期95-101,共7页Tobacco Science & Technology

基　　金：国家烟草专卖局科技项目“雪茄烟特色工艺技术研究”(212022AA0190);四川中烟科技项目“提升长城雪茄过程质量及工艺控制水平研究”(KJSB202110280002)。

摘　　要：为提高机制雪茄烟装盒过程中对缺支、断残、破损烟支的识别效率和准确率,建立了一种盒装机制雪茄缺陷的分类视觉检测方法。在雪茄包装机台设置图像采集工位并采集含有不同缺陷的雪茄小盒(未合盒)图像,建立缺支、断残检测数据集和破损检测数据集。对特征明显的缺支、断残数据集图像进行霍夫变换后,使用边缘检测和最小外接矩形法进行识别,判断盒内烟支数量和每支烟支长度是否符合要求;针对不同程度的破损数据集图像,在YOLOv5s模型的backbone部分引入金字塔压缩注意力机制(PSA)和EIoU损失函数(EIoU Loss)后对原始图像进行检测,判断烟支是否存在破损。结果表明:①缺支、断残检测方法的检测速率为21帧/s,对缺支、断残、无损3种类型样本的识别准确率均为100%;②改进后YOLOv5s模型的检测速率为30帧/s,将置信度阈值设置为0.6时,对无损样本和3种程度破损样本的识别准确率均为100%,与原始YOLOv5s模型相比,检测精度和召回率分别提高了1.80和0.93百分点;③分类视觉检测方法对单张图像的测试时间为0.080 s,可以实现对盒装机制雪茄缺陷的实时检测。该技术可为提高雪茄烟包装质量提供支持。In order to improve the efficiency and accuracy of integrity inspection of machine-made cigars in the packaging process,a visual inspection method for identifying various defects of machine-made cigars in packets was established.An image collection system was configured on a cigar packaging machine to capture the images of the uncovered packets containing cigars of different types of defects and create the dataset of missing and broken cigars and the dataset of damaged cigars separately.The images with typical features in the missing and broken cigar dataset were processed by Hough transform and then identified by edge detection and minimum enclosing rectangle method to judge whether the number of cigars in the packet and the length of each cigar met the specifications.After introducing the Pyramid Squeeze Attention Mechanism(PSA)and the EIoU Loss function into the backbone of YOLOv5s model,the original images were checked against the dataset of cigars with different degrees of damage to judge whether there were any damaged cigars.The results showed that:1)For the method for missing or broken cigars,it could go through 21 frames of image per second and the identification accuracy reached 100%.2)For the modified YOLOv5s model,it could go through 30 frames of image per second.When the confidence threshold was set at 0.6,the identification accuracy was 100%,and the inspection accuracy and recall rate increased by 1.80 and 0.93 percentage points,respectively,compared to the original YOLOv5s model.3)The time for going through a single frame of image was 0.080 s,which met the requirement of real-time inspection of machine-made cigars in packets.This technology supports the improvement of cigar packaging quality.

关 键 词：机制雪茄 机器视觉 缺陷检测 注意力机制 深度学习 

分 类 号：TS453[农业科学—烟草工业] TS46