主动太赫兹成像中的多目标分割与检测识别方法  

Multi-object Segmentation,Detection and Recognition in Active Terahertz Imaging

作　　者：薛飞 梁栋[1] 喻洋 潘家兴 吴天鹏 XUE Fei;LIANG Dong;YU Yang;PAN Jia-xing;WU Tian-peng(College of Computer Science and Technology, NanjingUniversity of Aeronautics and Astronautics, Nanjing 211106, China;Institute of Electronic Engineering, China Academy of Engineering Physics, Mianyang 621900, China;Microsystem and Terahertz Research Center, China Academy of Engineering Physics, Chengdu 610200, China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106 [2]中国工程物理研究院电子工程研究所,四川绵阳621900 [3]中国工程物理研究院微系统与太赫兹研究中心,四川成都610200

出　　处：《红外》2020年第2期13-25,共13页Infrared

摘　　要：针对主动太赫兹成像中存在的图像品质差以及藏匿物品类别多样、训练样本稀缺且类别不平衡等问题,提出了基于用条件生成对抗网络构建的Mask-CGANs模型的目标分割网络和基于RetinaNet的目标检测识别网络,实现了太赫兹图像中藏匿物品的多目标分割和检测识别。针对分割任务提出的约束损失函数和网络结构,使模型在召回率和虚警率之间达到平衡且降低了对训练样本规模的要求。针对检测任务采用的损失函数提高了训练样本不平衡条件下的检测精度。Aiming at the problems in the active terahertz(THz)imaging such as the poor image quality,the variety of hidden objects and the scarcity and imbalance of training samples,the objects segmentation networks based on the conditional generative adversarial networks′model Mask-CGANs and the objects detection and recognition networks based on the RetinaNet are built,which realizes the multi-object segmentation,detection and recognition of hidden objects in the THz imaging.The constraint loss functions and the networks structures proposed for the segmentation task make the model keep the balance between the recall rate and the false alarm rate,and the requirement of training sample size is reduced.The loss functions used for the detection task improve the detection accuracy under the condition of unbalanced training samples.

关 键 词：太赫兹成像 条件生成对抗网络 目标分割 目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]