Detection of objects with diverse geometric shapes in GPR images using deep-learning methods

Orhan Apaydln*, Turgay Işseven

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

Buried objects with regular geometric shapes appear as hyperbolic structures in ground-penetrating radar (GPR) images. The shapes of these hyperbolic structures differ depending on the geometric shapes of the objects. In this study, current deep learning-based object detection algorithms such as Faster R-CNN, YOLOv5, and single-shot detector are used to detect hyperbolic structures in GPR images and classify the buried object according to its geometric shape. A mixed data set is produced for training the models. A GPR measurement device is designed with a vector network analyzer and Vivaldi antenna pair to be used in laboratory measurements. Objects with rectangular and cylindrical geometric shapes are placed under the table and measurements are performed. The measurement scenarios created in the laboratory are modeled in the gprMax program and synthetic GPR data are produced. Data augmentation techniques such as flipping and resizing are applied to expand the data set. As a result of the training, three models successfully detect the objects and classify them according to their geometric shapes. The Faster R-CNN model gives the most accurate detection and classification with the metrics classification loss = 5.4 × 10-3, localization loss = 9 × 10-3, regularization loss = 5.1 × 10-5, [email protected] = 1, and [email protected]:0.95 = 1.

Original languageEnglish
Article number20220685
JournalOpen Geosciences
Volume16
Issue number1
DOIs
Publication statusPublished - 1 Jan 2024

Bibliographical note

Publisher Copyright:
© 2024 the author(s), published by De Gruyter.

Keywords

  • deep learning
  • ground-penetrating radar
  • object classification
  • object detection
  • rectangular and cylindrical objects

Fingerprint

Dive into the research topics of 'Detection of objects with diverse geometric shapes in GPR images using deep-learning methods'. Together they form a unique fingerprint.

Cite this