Automated Detection of Reflection Hyperbolas in Complex GPR Images With No A Priori Knowledge on the Medium
Automated Detection of Reflection Hyperbolas in Complex GPR Images With No A Priori Knowledge on the Medium
复制标题
DOI：
10.1109/tgrs.2015.2462727
复制
发表时间：
2016-01-01
影响因子：
8.2
通讯作者：
Lambot, Sebastien
中科院分区：
文献类型：
--
作者：
Mertens, Laurence;Persico, Raffaele;Lambot, Sebastien
在本文中，我们提出了一个自动检测算法，以及病态的地质雷达反射双曲线的复杂介质，校准与人类识别的原则。该算法通过将双曲线的解析函数拟合到用Canny滤波器检测到的轮廓边缘点来检测双曲线的顶点。双曲线的存在是使用一组精心选择的标准来确定的，这些标准经过校准，以适应人类大脑的模糊区。通过在理论双曲线周围定义一个缓冲区，进一步考虑了场双曲线的固有畸形。首先，在实验室中对树根和PVC管道以及树根系统的现场图像进行了测试。时域和频域雷达都在地面上使用。在MATLAB环境（单个1.96 GHz处理器）中对10 000个边缘点进行约1-3分钟的计算后，结果显示误报率和未检测率最大分别为20%和28%。与商业软件提供的半自动双曲线检测相比，这些比率较低。其次，我们进行了灵敏度分析，以估计忽略物体半径的双曲线方程拟合的有效性。拟合是接近的，但是从忽略半径的分析方程推导相对介电常数导致高误差。该算法计算时间短，性能好，适用于复杂环境。
In this paper, we propose an automated detection algorithm for well-and ill-shaped ground-penetrating radar reflection hyperbolas for complex media, calibrated with human recognition principles. The algorithm detects the apex of the hyperbolas by fitting an analytical function of a hyperbola to the profile edge dots detected with a Canny filter. The existence of a hyperbola is determined using a set of carefully chosen criteria calibrated in order to fit the ambiguity zone for the human brain. The inherent misshapedness of field hyperbolas is further considered by defining a buffer zone around the theoretical hyperbola. First, the method was tested in the laboratory over tree roots and PVC pipes and on field images over tree root systems. Both time- and frequency-domain radars were used on-ground. After around 1-3 min of computation time for 10 000 edge dots in a MATLAB environment (single 1.96-GHz processor), the results showed rates of false alarm and nondetection of maximum 20% and 28%, respectively. In comparison with the semiautomated hyperbola detection provided by a commercial software, these rates were lower. Second, we conducted a sensitivity analysis to estimate the validity of the fitting of a hyperbola equation neglecting the object radius. The fitting was close, but the derivation of the relative permittivity from the analytical equation neglecting the radius led to high errors. In conclusion, owing to the low computational time and its good performances, the proposed algorithm is suitable for complex environments.