Multi-label classification with deep learning techniques applied to the B-Scan images of GPR
The ground penetrating radars (GPR) are now widely used for the detection of buried objects in areas such as: geology, archaeology and civil engineering. It has the advantage of allowing detection by a non-destructive technique. The principle for time domain GPR consists in emitting electromagnetic...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Académie des sciences
2024-09-01
|
| Series: | Comptes Rendus. Physique |
| Subjects: | |
| Online Access: | https://comptes-rendus.academie-sciences.fr/physique/articles/10.5802/crphys.193/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The ground penetrating radars (GPR) are now widely used for the detection of buried objects in areas such as: geology, archaeology and civil engineering. It has the advantage of allowing detection by a non-destructive technique. The principle for time domain GPR consists in emitting electromagnetic pulses in the ground, these one are then diffracted by the targets to be detected. A single GPR signal trace captured at a position of the radar is a 1D signal called Ascan. A set of Ascan radar waveforms captured at a certain number of different consecutive positions along a particular direction will form a 2D image called B-scan. They show response shapes of hyperbolic type and their analysis give many characteristics. For example, in the case of buried pipes, a specific processing allows to find their diameter, their nature as well as the electrical characteristics of the ground. However, these approaches often require complex post-processing of the Bscan, which can be time-consuming and therefore makes it difficult to perform real-time characterization at the expense of such methods. With the emergence of deep neural networks and with a learning phase on a large number of Bscan, it becomes possible to extract almost instantaneously the characteristics of GPR radar data. In this study, a multi-label classification (MLC) model based on transfer learning and data augmentation has been developed to generate multiple information elements on the same image and to realize classification. Three deep learning models: VGG-16, ResNet-50 and adapted CNN were used as pre-trained models for transfer learning. The networks were trained on a synthetic dataset created in this study and evaluated on a set of performance metrics. |
|---|---|
| ISSN: | 1878-1535 |