Machine learning mask R-CNN for GPR B-scans

Abstract

Ground Penetrating Radar (GPR) is a commonly used technology to detect underground objects and environments for different purposes. The output generated from moving GPR across the ground surface is called the GPR scan. Realistic subsurface surroundings can be visualized, mapped, and monitored with these data. In B-scans, underground objects are represented in hyperbolic signatures. However, the process of manual recognition of hyperbolas presented in the B-scan is difficult and tedious due to the noisy and complex nature of subterranean environments. Mask R-CNN will be implemented to perform object detection and instance segmentation with supervised learning. The procedures include data preparation, training, testing and evaluation of prediction results.