电子学报 ›› 2021, Vol. 49 ›› Issue (5): 953-963.DOI: 10.12263/DZXB.20200635

• 学术论文 • 上一篇    下一篇

基于深度学习的GPR B-SCAN图像双曲线检测方法

王辉1,2, 欧阳缮1, 廖可非1, 晋良念1   

  1. 1. 桂林电子科技大学信息与通信学院, 广西桂林 541004;
    2. 贺州学院人工智能学院, 广西贺州 542899
  • 收稿日期:2020-06-29 修回日期:2020-08-07 出版日期:2021-05-25 发布日期:2021-05-25
  • 通讯作者: 欧阳缮(通信作者) 男,1960年9月生,江西安福人,桂林电子科技大学教授,博士生导师.主要研究方向:探地雷达信号处理、智能信息处理. E-mail:hmoysh@guet.edu.cn
  • 作者简介:王辉 男,1982年10月生,湖北仙桃人,贺州学院副教授,现为桂林电子科技大学博士研究生.主要研究方向:探地雷达信号处理、深度学习. E-mail:syswangxueleng@163.com
  • 基金资助:
    国家自然科学基金(No.61871425);广西自然科学基金(No.2017GXNSFBA198032,No.2017GXNSFAA198050,No.2018JJA170185)

GPR B-SCAN Image Hyperbola Detection Method Based on Deep Learning

WANG Hui1,2, OUYANG Shan1, LIAO Ke-fei1, JIN Liang-nian1   

  1. 1. School of Information and Communication, Guilin University of Electronic Technology, Guilin, Guangxi 541004, China;
    2. School of Artificial Intelligence Hezhou University, Hezhou, Guangxi 542899, China
  • Received:2020-06-29 Revised:2020-08-07 Online:2021-05-25 Published:2021-05-25

摘要: 利用深度学习方法来处理探地雷达(Ground Penetrating Radar,GPR)数据以提高GPR B-SCAN双曲线检测准确率.为了解决数据集样本不够的问题采用循环生成对抗网络(Cycle Generative Adversarial Networks,CycleGAN)算法对GPR B-SCAN图像数据进行增强.采用Faster R-CNN算子来定位双曲线图像区域,充分利用双曲线结构对称性及其方向差异性特征,设计与之对应的卷积核模板,通过卷积运算实现对B-SCAN图像中双曲线目标的有效分割.对双曲线目标采用最小二乘法进行二次曲线拟合得到精确的双曲线图像.与基于迁移学习的方法、HOG算法以及基于Hough变换的B-SCAN检测算法等相比,本文方法得到的结果在综合指标度量F上是最优的.

关键词: 深度学习, 探地雷达, 循环对抗生成网络, 卷积运算

Abstract: In order to improve the accuracy of GPR(ground penetrating radar) B-SCAN hyperbola detection,the deep learning method is applied to the processing of GPR data.In order to solve the problem of insufficient samples in the data set,the GPR B-SCAN image data is augmented by using cycle generative adversarial networks algorithm (CycleGAN).The Faster R-CNN operator is used to locate the hyperbolic image area,making full use of the symmetry of the hyperbolic structure and its directional difference characteristics,designing the corresponding convolution kernel template,and realize effective segmentation of hyperbolic targets in B-SCAN images through convolution operation.The least square method is used to perform quadratic curve fitting on the hyperbolic target to obtain an accurate hyperbolic image.Compared with B-SCAN image detection algorithms such as transfer learning-based methods,HOG(histogram of oriented gradients) algorithm and Hough transform algorithm,the results obtained by the method in this paper are optimal on the comprehensive measurement index F.

Key words: deep learning, ground penetrating radar, cycle generative adversarial networks, convolution operation

中图分类号: