Automatic hyperbola detection and fitting in GPR B-scan image

Abstract

Detecting buried objects from ground penetrating radar (GPR) profiles often requires manual interaction and plenty of time. This paper presents an automatic scheme for buried objects detection and localization. First, a trained deep learning framework — Faster R-CNN with data augmentation strategy is applied to identify hyperbolic signatures from a gray GPR B-scan image, which is capable of not only recognizing whether a B-scan profile contains traces of buried object, but also detecting candidate hyperbola region. Then, the detected rectangle region is extracted and transformed to a binary image, a novel double cluster seeking estimate (DCSE) algorithm is proposed to separate object point cluster from each other and enable the identification of hyperbolic signatures. Subsequently, a column-based transverse filter points (CTFP) method is utilized to extract hyperbola fitting points automatically from the validated point cluster. Downward opening hyperbola fitting is carried out and their respective peaks are obtained finally. The proposed scheme is able to extract information from GPR B-scan images automatically and efficiently; it is validated significant performance in the analysis of synthetic and on-site GPR data sets. • Faster R-CNN with data augmentation strategy is applied to pick hyperbola region from the gray GPR B-scan image. • A novel double cluster seeking estimate (DCSE) algorithm is proposed to separate object point cluster from each other. • A column-based transverse filter points (CTFP) method is utilized to extract hyperbola fitting points automatically. • The experiments show significant performance in the analysis of synthetic and on-site GPR data sets.