基于深度学习的探地雷达二维剖面图像结构特征检测方法

该文针对探地雷达(GPR) 2维剖面图像中目标特征提取困难及其识别精度较低等问题，采用深度学习方法来提取2维剖面图像中目标的特征双曲线。根据GPR工作的物理机制，设计了一种级联结构的卷积神经网络(CNN)，先检测并去除回波数据中的直达波干扰信号，再利用CNN得到B扫描(B-SCAN)图像的特征图，并对特征信号进行分类识别以提取目标的特征双曲线。同时，为处理各种干扰信号影响目标特征双曲线结构完整性的问题，提出了一种基于方向引导的特征数据补全方法，提高了目标特征双曲线识别的准确率。与方向梯度直方图(HOG)算法、单级式目标检测(YOLOV3)算法和更快速的区域卷积神经网络(Faster RCNN)算法相比，在综合评价指标F上该文方法的检测结果是最优的。

Structure Feature Detection Method for Ground Penetrating Radar Two-Dimensional Profile Image Based on Deep Learning

To solve the problem of the difficulty of target feature extraction and low recognition accuracy in Ground Penetrating Radar (GPR) two-dimensional profile, a deep learning method is used to extract the characteristic hyperbola of targets in B-SCAN image. Physics-based of GPR, a cascade Convolutional Neural Network (CNN) is designed to detect and remove the direct wave interference signal in the echo data. Then, the B-SCAN image is obtained by CNN, and the characteristic signals are classified and recognized to extract the characteristic hyperbola of the target. Meanwhile, in order to deal with the problem that the interference signals affect the structural integrity of the feature hyperbola, a feature data completion method based on directional guidance is proposed to improve the accuracy of the feature hyperbola recognition results. Compared with Histogram of Oriented Gradients(HOG) algorithm, You Only Look Once V3(YOLOV3) algorithm and Faster Region-based Convolutional Neural Network(Faster RCNN) algorithm, the detection result of the proposed method is the best in the comprehensive evaluation index F.