基于局域特征相似性度量的卷烟制品外观缺陷检测  

Appearance quality inspection of cigarette products based on local characteristic similarity metric

作　　者：冯东 李志刚 何爱民 杨旭东 王澍 董浩[4] 张龙 FENG Dong;LI Zhigang;HE Aimin;YANG Xudong;WANG Shu;DONG Hao;ZHANG Long(Institute of Physical Science and Information Technology,Anhui University,Hefei 230601,China;Hefei Institute of Physical Science,CAS,Hefei 230031,China;China Tobacco Hebei Industrial Co.,Ltd.,Shijiazhuang 050051,China;China National Tobacco Quality Supervision&Test Center,Zhengzhou 450001,China)

机构地区：[1]安徽大学物质科学与信息技术研究院,合肥市230601 [2]中科院合肥物质科学研究院,合肥市230031 [3]河北中烟工业有限责任公司,石家庄市050051 [4]国家烟草质量监督检验中心,郑州450001

出　　处：《烟草科技》2023年第4期82-90,共9页Tobacco Science & Technology

基　　金：国家烟草专卖局科技重大专项项目“烟草行业质量监控大数据构建及应用研究”[110202101080(SJ-04)];国家烟草专卖局标准项目“卷烟包灰性能测试方法”(YC/T2021169);烟草行业标准制修订项目“卷烟包灰性能测试方法”(2021HB002)。

摘　　要：为解决深度学习技术在卷烟制品外观缺陷检测中存在人工标注繁琐、目标形态随机性大等问题,提出了一种基于局域特征相似性度量(Local Characteristic Similarity Metric,LCSM)的图像算法。LCSM算法只需要对正常样本进行训练,使用特定的卷积神经网络提取正常样本的特征并构建图像每个区域的数据特征分布,再提取测试图像局域特征向量并采用Wasserstein距离度量其特征分布与对应正常样本特征分布之间的相似性,从而判断测试图像是否存在缺陷。采集并制作了卷烟小盒、烟用胶囊和烟支3个数据集用于验证LCSM算法性能。结果表明:LCSM算法在3个数据集的缺陷检测准确率分别达到98.75%、99.50%和98.50%,与近期报道的Skip-GANomaly、STPM以及IGD算法相比,分别提高14、2和2百分点。该方法可为提高卷烟制品外观缺陷检测准确率提供技术支持。In order to rationalize and facilitate the application of deep learning technology into the appearance quality inspection of cigarette products,an image algorithm based on Local Characteristic Similarity Metric(LCSM)was proposed,with the LCSM algorithm the only need is training normal samples.A specific convolutional neural network was used to extract the image features from the normal samples and construct the data feature distribution of each region in the image.The local characteristic vector of the test image is extracted,and Wasserstein distance is used to measure the similarity between its feature distribution and that of the corresponding normal sample so as to judge whether there is any defect in the test image.The data sets of cigarette packet,breakable capsule,and cigarette were established for verifying the performance of LCSM algorithm.The results showed that the inspection accuracies of LCSM algorithm for the three data sets reached 98.75%,99.50%,and 98.50%,which were higher than the recently reported accuracies of Skip-GANomaly,STPM,and IGD algorithms by 14,2,and 2 percentage points,respectively.This method provides technical support for improving the inspection accuracy of appearance defects of cigarette products.

关 键 词：卷烟制品 外观质量 缺陷检测 深度学习 局域特征 相似性度量 

分 类 号：TS473[农业科学—烟草工业]