基于改进YOLOv7的香烟品类检测技术研究  

Research on cigarette category recognition technology based on improved YOLOv7

作　　者：陈佳林 唐利 游强 罗强 代雨根 李嘉炜 赵歆波[3] CHEN Jialin;TANG Li;YOU Qiang;LUO Qiang;DAI Yugen;LI Jiawei;ZHAO Xinbo(Bazhong City Tobacco Monopoly Administration,Bazhong 636600,China;Luzhou City Tobacco Monopoly Administration,Luzhou 646000,China;School of Northwestern Polytechnical University,School of Computer Science,Xi'an 710129,China)

机构地区：[1]巴中市烟草专卖局(公司),巴中636600 [2]泸州市烟草专卖局(公司),泸州646000 [3]西北工业大学计算机学院,西安710129

出　　处：《中国体视学与图像分析》2024年第3期252-261,共10页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然科学基金资助项目(No.61871326);中国烟草四川省科技项目专项资助(No.SCYC20222961871326)

摘　　要：本文针对烟草行业在稽查、盘库等香烟品类检测工作场景中存在的工作量大、效率低和时间成本高的问题,对基于深度学习的香烟品类自动检测技术展开研究。首先,实场采集10 000余张香烟盘库图像,建立首个常见香烟品规图像检测数据集CCD-12K。其次,由于现有方法检测的准确度不高难以满足实际应用的需要,提出一种融合可变形卷积DCNv2和自注意力的香烟品类检测模型YOLOv7-Cr。检测模型将可变形卷积引入到特征提取网络,以增强模型的感受野和对物体形变的感知能力。最后,使用BRA注意力机制来聚合香烟图像中的关键性特征,计算前后特征信息,以增强网络对前景和背景的辨别能力。测试结果表明,在CCD-12K数据集上,YOLOv7-Cr对香烟品类图像检测的均值平均精度(mAP)达到92.3%,相较于YOLOv7提升了2.9个百分点,满足执法工作中检测均值精度高于90%的需求。To overcome the challenges encountered by the cigarette industry,characterized by high workload,low efficiency,and high time and cost in inspection and inventory management,this study explores the application of deep learning-based cigarette product classification technology.Initially,10000 images of cigarette inventory were collected from the field to establish the first common cigarette product image recognition dataset,termed CCD-12K.Subsequently,a novel model named YOLOv7-Cr was proposed,which integrates deformable convolution DCNv2 and self-attention mechanisms for cigarette product detection.The model incorporates deformable convolution into the feature extraction network to enhance the model's receptive field and its capacity to discern shape variations.Additionally,it employs BRA attention to aggregate key features from cigarette images,calculating the information of both foreground and background features to improve the network's discrimination ability between foreground and background.The test results demonstrate that on the CCD-12K dataset,YOLOv7-Cr achieved an average precision(mAP)of 92.3%for cigarette product recognition,representing a 2.9 percentage point improvement over YOLOv7,effectively fulfilling the requirement of mean accuracy exceeding 90%in law enforcement applications.

关 键 词：香烟品类检测 YOLOv7 可变形卷积 BRA注意力机制 

分 类 号：P315.69[天文地球—地震学]