基于注意力机制的X光安检图像危险物品检测  

Detection of dangerous objects in X-ray security inspectionimages based on attention mechanism

作　　者：郭豆豆 李国权[1,2] 黄正文 吴建[1] 庞宇[2] GUO Doudou;LI Guoquan;HUANG Zhengwen;WU Jian;PANG Yu(School of Communications and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Key Laboratory of Photoelectric Information Sensing and Transmission Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学光电信息感测与传输技术重点实验室,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2023年第6期1117-1126,共10页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家重点研发计划项目(2019YFC1511300);国家自然科学基金项目(61971079)。

摘　　要：针对X光行李安检系统中危险物品检测上存在的误检、漏检等问题,提出了一种有效利用上下文信息和增强特征表征能力的注意力融合网络(attention fusion network,AFN)。在YOLOv3-SPP架构上融合多压缩激活(multi-squeeze excitation,MSE)模块和多融合全局注意力(multi-fusion global attention,MFGA)模块,将特征提取网络提取的特征与MSE提取的通道信息融合,得到具备通道全局性的语义特征;将MFGA模块设置在各个检测网络分支,有效地融合具有深层特征的通道及空间信息,使多尺度特征具备3维全局性。在公开的SIXray数据集上进行测试表明,提出的方法有效提高了中等目标和大目标的检测精度和召回率,模型的精度均值为51.1%,比经典的YOLOv3-SPP算法提高了1.4百分点,证明了通道注意力和空间注意力可有效增强输入特征图中用于检测危险物品的细节信息,提高模型对中等目标和大目标的检测性能。Aiming at the problems of false detection and missing detection of dangerous objects in X-ray baggage security inspection system,we propose an attention fusion network(AFN),which effectively uses context information and enhances the ability of feature representation.The network integrates multi-squeeze excitation(MSE)module and multi-fusion global attention(MFGA)module on the YOLOv3-SPP architecture.Firstly,the feature extracted by the feature extraction network is fused with the channel information extracted by MSE to obtain the semantic feature with channel globality The MFGA module is set in each detection network branch to effectively integrate the channel and spatial information with deep features,so that the multi-scale features have three-dimensional globality.Testing on the public SIXray data set shows that the proposed method effectively improves the detection accuracy and recall rate of medium and large targets.The average accuracy of the model is 51.1%,which is 1.4%higher than the classical YOLOv3-SPP algorithm.It shows that channel attention and spatial attention can effectively enhance the detail information used to detect dangerous goods in the input feature map,and improve the detection performance of the model for medium and large targets.

关 键 词：X光行李安检系统 危险物品检测 通道注意力 空间注意力 通道全局性 

分 类 号：TN919[电子电信—通信与信息系统] TP183[电子电信—信息与通信工程]