Detecting Vertices of Hyperbolas in GPR Data Using Fully Convolutional Neural Network

Hyperbola is a distinctive pattern in the B-scan image of ground penetrating radar (GPR). Accurate and automatic localization of the hyperbolas' vertices is of great importance for the interpretation of GPR B-scan data, however, which is still a challenging problem due to the irregular hyperbola shapes in complex detection scenarios. In this paper, a novel method for detecting the vertices of hyperbolas in GPR B-scan image is proposed using a modified fully convolutional neural network (FCN). First, a simulated training dataset approximating the real data is created, which ensures the reliability of the network for real measurements. Second, a lightweight network based on fully convolutional architecture is proposed to mask the vertex regions of the hyperbolas, without resizing the input B-scan image. Finally, the density-based spatial clustering of applications with noise DBSCAN) algorithm is employed to cluster the masked results, thereby obtaining the coordinates of the hyperbola vertices. The effectiveness of the proposed method is validated by simulations and experiments.