基于卷积神经网络的毫米波图像人体隐匿物检测  

Convolutional-neural-network-based Human Concealed Object Detection for Millimeter Wave Images

作　　者：骆尚 吴晓峰[1,2] 杨明辉[3] 王斌[1,2] 孙晓玮[3] LUO Shang;WU Xiaofeng;YANG Minghui;WANG Bin;SUN Xiaowei(Key Laboratory for Information Science of Electromagnetic Waves,Fudan University,Shanghai 200433,China;Research Center of Smart Networks and Systems,School of Information Science and Technology,Fudan University,Shanghai 200433,China;Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences,Shanghai 200050,China)

机构地区：[1]复旦大学电磁波信息科学教育部重点实验室,上海200433 [2]复旦大学信息学院智慧网络与系统研究中心,上海200433 [3]中国科学院上海微系统与信息技术研究所,上海200050

出　　处：《复旦学报（自然科学版）》2018年第4期442-452,共11页Journal of Fudan University：Natural Science

基　　金：国家自然科学基金(61572133)

摘　　要：本文提出一种基于卷积神经网络的毫米波图像中人体隐匿物检测算法.首先,通过滑动窗选取局部毫米波图像得到正负训练样本,相较于传统的IoU,利用文中提出的IoG指标,不仅可以解决目前毫米波图像数据量不足的问题,而且尽可能扩大正负样本之间的特征差别使得模型泛化能力得到提高.其次,由人工扩展数据集训练的卷积神经网络模型可自动学习到不同种类的隐匿物信息.最后,叠加所有的局部可疑区域得到可疑目标热度图,从而进一步得到隐匿物的位置.使用本文方法对中科院上海微系统提供的728张毫米波图像进行了隐匿物检测实验,训练阶段在局部图像块的验证集上达到了95.6%的查准率和97.6%的召回率,表明了本文提出的IoG指标的有效性和重要性;测试阶段在全图的测试集上达到了88.5%的检测率和13.5%的虚警率,相较于现有的方法,所提出的算法检测率更高、虚警率更低,表明本文方法的有效性和稳健性.This paper proposes a concealed object detection method based on Convolutional Neural Network（CNN）for millimeter wave images.Firstly,the sliding window method is used to extract positive and negative samples for CNN training,compared to common intersection over union（IoU）,the proposed intersection over ground-truth（IoG）index not only generates enough samples for training,but also can enlarge the feature gap between positive and negative samples for better generalization.Secondly,the CNN model trained by the augmented datasets can learn information of different kinds of concealed objects,and then all suspicious patches superimpose and form a suspicious object heat map for the location of concealed objects.The proposed method is evaluated on millimeter wave images provided by Shanghai Institute of Microsystem and Information Technology.Precision of 95.6% and recall of 97.6% on validation patches in training phase demonstrated the importance and effectiveness of the proposed IoG index.True positive rate（TPR）of 88.5% and false positive rate（FPR）of 13.5% demonstrated that our work achieved higher TPR and lower TPR compared with the existing methods in test phase,which validated the effectiveness and robustness of our work.

关 键 词：毫米波图像 隐匿物检测 卷积神经网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]