基于深度学习的复杂场景中卷烟烟盒检测与识别方法  

A detection and recognition method of cigarette cases in complex scene based on deep learning

作　　者：单宇翔[1] 龙涛 楼卫东[1] 陆海良 胡红春[3] 李轩 SHAN Yuxiang;LONG Tao;LOU Weidong;LU Hailiang;HU Hongchun;LI Xuan(Information Center of China Tobacco Zhejiang Industrial Co.,Ltd.,Hangzhou 310008,China;Shenzhen AiMall Tech.,Nanshan District,Shenzhen 518000,China;Staff Development Institute of China Tobacco Corporation,Henan 450008,China)

机构地区：[1]浙江中烟工业有限责任公司信息中心,杭州市310008 [2]深圳爱莫科技有限公司,深圳市518000 [3]中国烟草总公司职工进修学院,郑州450008

出　　处：《中国烟草学报》2021年第5期71-80,共10页Acta Tabacaria Sinica

基　　金：浙江中烟工业有限责任公司基于智能图像识别的营销机器人流程自动化(RPA)关键技术及应用研究(ZJZY2021E001)。

摘　　要：近年来,深度学习技术在图像识别领域取得了巨大的成功,在零售、金融、安全、自动驾驶等行业得到了广泛的应用。本论文提出了一套在复杂场景条件下可行有效的卷烟烟盒图像识别框架,利用深度学习技术实现营销活动或陈列活动中的卷烟烟盒图片的自动化识别。本论文提出的框架包含基于可见区域实例分割的烟盒位置检测、烟盒区域正则化和烟盒规格识别几个关键性步骤,构建了鲁棒性强、精确度高的卷烟烟盒图像识别方法。在大规模测试数据集上,其中包括陈列条件、拍摄条件等外部环境影响造成烟盒图片模糊、角度大、遮挡、玻璃反光、部分遮挡、随意摆放,以及部分不同规格的烟盒图像外观差异微小等场景,本论文的方法都表现出了稳定的识别能力,平均识别精度超过95%,达到了在实际应用场景大规模推广使用的条件。In recent years,deep learning has achieved great success in image recognition and has been widely applied in fields such as retail,finance,security,self-driving,etc.In this paper,we present a novel deep learning-based recognition framework of cigarette cases in complex scene.The framework consists of three key steps:1)cigarette case localization based visible region segmentation;2)cigarette case region regularization based on image transformation;3)cigarette case recognition based on deep Convolutional Neural Network.We carried out the validation experiment on a large dataset collected from real application scenario,with cigarette case images exhibiting large view point changes,occlusion,light reflection,incomplete texture,random rotation,and visual similarity in appearance,etc.The result showed that the proposed method is capable of handling various challenges and achieved a robust recognition accuracy over 95%,which meets the requirements of real large-scale application scenarios.

关 键 词：深度学习 图像识别 烟盒识别 陈列分析 大数据 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] TS47[农业科学—烟草工业]