Detection and classification of near-surface buried objects in GPR images with deep learning-based Faster R-CNN and YOLOv5 methods

Abstract

In this study, the detection of buried objects in GPR images using the deep learning-based Faster R-CNN and YOLOv5 methods and their classification according to their geometric shapes are investigated. Buried objects in the near surface may have different geometric shapes. Such objects can be imaged using Ground Penetrating Radar (GPR). As research materials, a rectangular prism-shaped aluminum-coated box and a cylindrical rod are used for laboratory measurements. A simulated underground model has been created in a laboratory environment, and GPR measurements have been performed. A radar device is designed for measurements using a Vector Network Analyzer (VNA) and a Vivaldi antenna pair. The scenarios for the measurements in the laboratory environment are modeled in the gprMax program, and synthetic GPR images are generated. The dataset consists of both actual measurements and synthetic data. Deep learning-based Faster R-CNN and YOLOv5 methods are popular techniques used for object detection in images. The GPR images used for training in these methods are augmented by using flipping and resizing techniques, and the dataset is expanded. Subsequently, hyperbolic structures of objects in GPR images are labeled as "rectangular" and "cylindrical" based on their geometric shapes. The training process is then carried out using these methods, resulting in the detection of buried objects in GPR images with high accuracy and classification based on their geometric shapes as "rectangular" and "cylindrical". The performances of the two different methods are compared, revealing that Faster R-CNN achieved higher accuracy, while the YOLOv5 method exhibited faster detection.