基于YOLOv5的毫米波图像目标检测方法研究  

Research on Object Detection Method of Millimeter Wave Image based on YOLOv5

作　　者：张格菲 李春宇[1] 刘金坤 屈音璇 ZHANG Ge-fei;LI Chun-yu;LIU Jin-kun;QU Yin-xuan(School of Investigation,People􀆳s Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学侦查学院,北京100038

出　　处：《宇航计测技术》2021年第5期41-45,共5页Journal of Astronautic Metrology and Measurement

基　　金：国家重点研发计划(2019YFF0303405);公安部技术研究计划(2019JSYJC21);中央高校基本科研业务费项目(2020JKF502)资助。

摘　　要：毫米波成像技术是安全检查领域中的新兴方向,研究符合此应用场景的目标检测算法并提升相应的检测速度与检测准确率,具有很高的应用意义和价值。本文提出了一种基于YOLOv5深度学习模型的实时目标检测算法,用来检测安检人员身上隐匿的违禁品,该方法采用GIOU_Loss损失函数,提升了衡量检测框相交的能力。此外,对网络结构进行了一定的优化和改进,同时增加了毫米波数据增强预处理功能,用以加速函数收敛,从而形成了针对毫米波图像进行物品检测的深度神经网络[1],以提高目标检测效果。实验结果表明,该方法能够有效地检测毫米波图像中的危险品,具备自动识别和实时检测的优势。Millimeter wave imaging technology is an emerging direction in the field of security inspection.Researching object detection algorithms that meet this application scenario and improving the corresponding detection speed and detection accuracy have high application significance and value.This paper proposes a real-time object detection algorithm based on the YOLOv5 deep learning model to detect contraband hidden by security personnel.The method uses the GIOU_Loss loss function to improve the ability to measure the intersection of the detection frames.In addition,the network structure is optimized and the improvement and the addition of the millimeter wave data enhancement preprocessing function to accelerate the convergence,thus forming a deep neural network for millimeter wave images for item detection to improve the object detection effect.Experimental results show that this method can effectively detect dangerous goods in millimeter-wave images,and has the advantages of automatic identification and real-time detection.

关 键 词：深度学习 毫米波图像 目标检测 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]