YOLO-STM:基于Swin-Transformer与MSDA的X光安检图像危险品识别网络  

YOLO-STM:A network model for identifying prohibited items in X-ray security inspection images based on Swin-Transformer and MSDA

作　　者：刘俊杰 冯鹏[1,2] 廖望[1,2] 王羲 LIU Junjie;FENG Peng;LIAO Wang;WANG Xi(The Key Lab of Optoelectronic Technology and Systems,Ministry of Education,Chongqing University,Chongqing,400044,China;ICT NDTEngineering Research Center,Ministry of Education,Chongqing University,Chongqing400044,China;Chongqing Rail Transit(Group)Co.Ltd,Chongqing,401120;School of Instrumentation Science and Engineering,Harbin Institute of Technology,Harbin150001,China)

机构地区：[1]重庆大学光电技术及系统教育部重点实验室,重庆400044 [2]重庆大学工业CT无损检测教育部工程中心,重庆400044 [3]重庆市轨道交通(集团)有限公司,重庆401120 [4]哈尔滨工业大学仪器科学与工程学院,哈尔滨150001

出　　处：《中国体视学与图像分析》2024年第3期230-241,共12页Chinese Journal of Stereology and Image Analysis

基　　金：重庆市科委技术创新与应用发展专项(No.cstc2021jscx-gksbX0056);北京市自然科学基金-海淀原始创新联合基金(L242114)

摘　　要：当前,为保证轨道交通的正常运行,安检已然成为必不可少的重要保障措施,但大多数安检图像中对危险品的判别主要依靠安检员的人工判图,难以保证安检的准确性和实时性。为此,本文开展了基于深度学习的X光安检图像危险品识别算法研究,以实现对违禁物品高精度和智能化的检测。首先,在重庆市轨道交通站内实地采集了含9类违禁物品的安检图像,并使用ACE自动色彩增强算法实现数据增广,构建了X光危险品安检图像数据集CRTXray;然后,针对小目标危险品识别效果欠佳、不同安检设备成像色彩和清晰度不同的问题,提出了基于Swin Transformer和多尺度空洞注意力机制(MSDA)的改进危险品检测模型YOLO-STM。利用Swin Transformer的滑动窗口机制,增强高维特征图的感受野,提高模型对重叠危险品的检测能力;通过MSDA不同的头部,使用不同的空洞率执行滑动窗口注意力,改善不同尺度下的物品形状和纹理等特征的捕获性能。实验结果表明,YOLO-STM模型对于数据集中所有危险品的平均精确度为89.2%,平均识别精度为91.8%,对枪支类的识别精度达到99.5%。相比YOLO v8模型的平均精确度和平均识别精度分别提高了3.0%和4.0%。To ensure the smooth operation of rail transit,security screening has become an essential safe-guarding measure.However,the identification of hazardous goods in most security screening images mainly relies on the subjective judgment of security inspectors,which is difficult to ensure its accuracy and timeliness.Therefore,this work studied on the deep learning-based algorithm for the recognition of dangerous goods in X-ray security inspection images,aiming to realize high-precision and intelligent detection of prohibited items.Firstly,the security inspection images containing nine kinds of prohibited items were collected in Chongqing rail transit stations.The dataset was augmented using an automatic color enhancement algorithm to enhance its diversity.The obtained dataset is referred as CRTXray.Then,addressing the issues of the poor identification performance of small-target dangerous goods and variation in imaging color and clarity across different security devices,an improved dangerous goods detection model—YOLO-STM based on Swin Transformer and Multi-scale Dilated Attention(MSDA)is proposed.Experimental results show that the YOLO-STM model achieves an average accuracy of 89.2%,and an average recognition accuracy of 91.8%at 50%IoU for all dangerous goods in the data set.Specifically,the model exhibits a recognition accuracy of 99.5%at 50%IoU for guns.Compared with the YOLO v8 model,the YOLO-STM model demonstrates improvements in average accuracy by 3.0%and average recognition accuracy by 4.0%.

关 键 词：深度学习 X光图像 危险品检测 YOLO v8 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]