基于改进YOLOv3-Tiny的毫米波图像目标检测  

Millimeter wave image object detection based on improved YOLOv3-Tiny

作　　者：陈国平[1] 彭之玲 黄超意 管春[1] Chen Guoping;Peng Zhiling;Huang Chaoyi;Guan Chun(School of Optoelectronic Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学光电工程学院,重庆400065

出　　处：《电子测量技术》2021年第21期163-167,共5页Electronic Measurement Technology

基　　金：国家自然科学基金(61671091)项目资助。

摘　　要：毫米波是一种不具有电离辐射的电磁波,其能够穿透绝缘衣物布料和对人体无害的特性使得毫米波在人体安检领域有着巨大的发展前景。将深度学习方法运用至毫米波图像目标检测领域,提出一种基于改进YOLOv3-Tiny的毫米波图像目标检测方法。首先,在特征提取网络中增加卷积层提升网络深度,并增加至3个不同尺度的预测层加强对毫米波图像目标的检测能力;然后,在特征金字塔(FPN)中引入注意力机制(CBAM),使网络更关注毫米波图像中待测目标的特征,忽略背景噪声冗余的特征。结果表明,改进后的网络平均准确率可达93.4%,单帧检测速度为15 ms,模型参数仅为38.7M,为毫米波安检系统高精度、小型化的研究提供了参考价值。The millimeter-wave is an electromagnetic wave without ionizing radiation.It can penetrate the insulating cloth and is harmless to the human body.These characteristics make the millimeter wave have a wide range of application prospects in the field of public safety.Apply deep learning to the field of millimeter-wave image object detection,a millimeter-wave image object detection method based on improved YOLOv3-Tiny is proposed.Firstly,add convolutional layers to the feature extraction network to increase the depth of the network and increases to 3 different scale prediction layers to enhance the detection ability of millimeter-wave image object.Then,the convolutional block attention module is introduced in the feature pyramid network to make the network pay more attention to the features of targets and ignore the characteristics of redundant background noise.The results show that the improved network has mean average accuracy up to 93.4%,single frame detection speed is 15 ms,model parameters are only 38.7 M,which provides a reference value for the research of high precision and miniaturization of millimeter wave security system.

关 键 词：毫米波图像 目标检测 注意力机制 YOLOv3-Tiny 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]