摘要
提出了一种基于卷积神经网络(convolutional neural networks, CNN)与梯度类激活热力图(gradient class activation map, Grad-CAM)的探地雷达公路地下目标检测方法。首先使用标记好的探地雷达图像数据集训练一个用于图像分类的CNN,然后基于训练完成的CNN对图像计算Grad-CAM激活图,将获得的激活图进行二值化,定位目标位置。构建了一个包括了5 000张探地雷达图像的数据集,并使用该数据集进行实验,其中4 000张图像用来训练模型,1 000张用来测试。在1 000张测试数据中,各个类别的召回率分别为:管线目标99.2%,地下空洞98.5%,无目标图像98.8%。目标定位结果与实际位置非常吻合。这些结果表明该方法能够有效的检测探地雷达图像中的目标。
A ground penetrating radar target detection method based on CNN and Grad-CAM is proposed.Firstly a CNN should be trained using the labeled GPR image dataset for image classification,then based on the trained CNN,the Grad-CAM feature map can be calculated on the image,and the obtained feature map can be binarized by thresholding to obtain the bounding box of the area of interest and locate the target position.We constructed a data set including 5000GPR images and used the data set for experiments.Of these,4000images were used to train the model and 1000were used for testing.In the 1000test data,the recall rate of each category is:99.2%of pipeline targets,98.5%of underground holes,and 98.8%of non-target images.The target positioning result is in good agreement with the actual position.These results show that the method can effectively detect targets in GPR images.
作者
赵迪
叶盛波
周斌
Zhao Di;Ye Shengbo;Zhou Bin(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100190,China;Key laboratory of electromagnetic radiation and sensing technology,Chinese Academy of Sciences,Beijing 100190,China;School of electrical and electronic engineering,University of Chinese Academy of Sciences,Beijing 100048,China)
出处
《电子测量技术》
2020年第10期113-118,共6页
Electronic Measurement Technology
基金
北京市科技计划课题实施方案(Z181100000118004)
载人航天预研项目(030201)
作者简介
赵迪,硕士研究生,主要研究方向为探地雷达图像处理,机器学习技术。E-mail:zhaodi171@mails.ucas.ac.cn;叶盛波,副研究员,主要研究方向为超宽带探地雷达系统设计。E-mail:sbye@mails.ie.ac.cn;周斌,研究员,主要研究方向为超宽带雷达系统设计及信号处理。E-mail:zhb@mail.ie.ac.cn
