12:00 PM - 12:15 PM
[SEM14-12] Automatic Detection of Cylindrical Objects in Radargram Using Faster-Regional-Convolutional Neural Network Algorithm
Keywords:Ground-Penetrating Radar, object detection, radargram, Convolutional Neural Network, gprMax
Ground-Penetrating Radar (GPR) is a non-destructive geophysical technique that uses high-frequency electromagnetic waves to image substructures and other man-artificial objects like cylindrical targets. Such cylindrical objects such as pipes and cables show a hyperbolic signal pattern in the GPR section called radargram. The typical shape of the hyperbolic reflections is dependent on the depth and properties contrast of the buried objects and their host material. Identifying such a target is time-consuming and limits the procedure for the later interpretation phase. This paper proposed a new method to automatically detect hyperbola-formed features in the radargram by combining object detection methods and digital image processing techniques. We took three steps of work. The first step was pre-processing and then converting the data to a raster format. In the second step, the so-called Faster-Regional Convolutional Neural Network (Faster R-CNN) model extracted the hyperbola segments as a set of rectangular boundary boxes. The Faster R-CNN was used to train synthetic data simulated by the gprMax software. The third step is to estimate the coordinates of the hyperbola apex using a search window algorithm on a digital image. In this simulation, we used Berlage-wavelet, which offers flexibility to model a quasi-real GPR wavelet. We examined the proposed method with the real dataset from The IFSTTAR Geophysical Test Site. Our study showed the applicability of the proposed technique to the automatic detection of the location of cylindrical objects with a minimal amount of run-time.