一种面向交通安全的违禁品轻量级检测方法  

A Lightweight Detection Method of Contraband Aimed at Traffic Safety

作　　者：翁成康 黄贤明 黄海洋 WENG Chengkang;HUANG Xianming;HUANG Haiyang(College of Computer Science,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007

出　　处：《湖南工业大学学报》2025年第1期72-78,共7页Journal of Hunan University of Technology

基　　金：教育部中国高校产学研创新基金资助项目(2020ITA05043);湖南省教育厅科学研究基金资助项目(21C0409,2023DT002)。

摘　　要：针对X光安检图像中摆放杂乱、故意遮挡、小型不规则物品检测等问题,以及安检工作对实时性、快速性的要求,基于YOLO v5s网络模型提出了一种结合改进的轻量级实时违禁品监测方法LRCD,以辅助安检人员快速进行检测。通过在模型的主干中使用DenseOne模块替代YOLO v5s主干中的C3模块,进而丰富特征,提高网络的特征表达能力;为了提高推理速度,使用SimSPPF替换YOLO v5s主干中的SPPF;同时引入WIoU(Wise-IoU)损失函数,抑制了冗余特征对检测网络的影响,增强了网络获取违禁品中包含的多尺度特征的能力。在针对X光下行李物品图片的EDS数据集中进行测试,mAP达68.96%,FPS达136.9,对比近年来被广泛使用的其他经典目标检测模型,分别平均提升了6.35%与66.7%。In view of the flaws of disorderly placement,intentional obstruction,and detection of small irregular items in X-ray security inspection images,as well as the requirements for real-time and fast security inspection purpose,a lightweight real-time contraband detection method(LRCD)has thus been proposed based on YOLO v5s network model to assist security personnel in rapid detection.By replacing the C3 module in the YOLO v5s backbone with the DenseOne module in the model backbone,the features can be enriched and the network’s feature expression ability can be improved;with SPPF(spatial pyramid pooling-fast)in YOLO v5s backbone replaced with SimSPPF for an improvement of the inference speed.Meanwhile,the WIoU(Wise IoU)loss function is introduced for an suppression of the influence of redundant features on the detection network,thus enhancing the network’s ability to obtain multi-scale features contained in contraband goods.In the EDS(endogenous domain shift)dataset for X-ray images of luggage and items,the mAP(mean average precision)reaches 68.96%and FPS reaches 136.9.Compared with other widely used classic object detection models in recent years,the average improvements can be as high as 6.35%and 66.7%,respectively.

关 键 词：轻量化 卷积神经网络 X光图像 注意力机制 违禁品检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]