基于深度学习的卷烟牌号识别方法  

Recognition of cigarette brand based on deep learning

作　　者：陈智斌[1] 农英雄[1] 梁冬[1] 孙忱[1] 韦屹 钟征燕[1] CHEN Zhibin;NONG Yingxiong;LIANG Dong;SUN Chen;WEI Yi;ZHONG Zhengyan(Information Center,China Tobacco Guangxi Industrial Co.,Ltd.,Nanning 530001,China)

机构地区：[1]广西中烟工业有限责任公司信息中心,南宁市西乡塘区530001

出　　处：《烟草科技》2020年第12期75-82,共8页Tobacco Science & Technology

基　　金：广西中烟工业有限责任公司科技计划项目“基于图像识别的卷烟零售数据采集与分析技术研究”(GXZYCX2019E002)。

摘　　要：为满足自动化采集卷烟零售数据需求,提出了一种基于深度学习的卷烟牌号识别方法。首先设计一种卷烟检测模型,检测卷烟在图像中的位置和姿态,并通过姿态矫正生成卷烟正向图像;其次构建卷烟特征提取模型,提取卷烟正向图像的特征;最后生成欧式距离阈值,通过阈值在卷烟特征库检索出与待识别特征最相符的卷烟特征,由卷烟特征与卷烟牌号的对应关系得出待识别特征的卷烟牌号。以广西南宁市若干卷烟零售店为例进行测试,结果表明:对卷烟牌号的识别准确率达到98.0%。说明所建卷烟牌号识别方法可满足卷烟零售数据自动化采集的需求,具备良好的泛化性能。In order to meet the demands of automatic collection of cigarette retail data,a method for cigarette brand recognition was proposed on the basis of deep learning.Firstly,design a detection model to detect the position and posture of cigarette packet image,and adjust the packet to right its image.Secondly,develop a feature extraction model to extract the feature of the image.Finally,generate a Euclidean distance threshold to search the cigarette feature with the highest similarity via the threshold from the cigarette feature database,and the cigarette brand associated with that feature was taken as the cigarette brand to be identified.The proposed method was tested in some cigarette retail stores in Nanning City of Guangxi Province,the results showed that the accuracy of the method for cigarette brand recognition was 98.0%,which indicated that the proposed method could meet the demands of automatic collection of cigarette retail data and had good generalization performance.

关 键 词：卷烟 牌号识别 深度学习 卷烟检测 姿态矫正 

分 类 号：TP315[自动化与计算机技术—计算机软件与理论]