Introduction: The Challenge of GPR Interpretation
Ground Penetrating Radar (GPR) is a staple of non-destructive testing (NDT), used to locate buried utilities or structural rebars without invasive digging. However, raw GPR data, known as B-scans, does not provide a direct visual map of the subsurface. Instead, it consists of abstract, overlapping hyperbolas that require expert interpretation.
While mathematical "migration" techniques exist to collapse these hyperbolas back into point-like objects, they are computationally expensive and often fail to reveal the actual physical dimensions, such as the diameter, of the buried objects. This paper introduces MigrationNet, a deep-learning-based framework designed to automate this process and reconstruct object geometry with high efficiency.
Methods: The MigrationNet Architecture
MigrationNet replaces traditional back-projection math with a specialized neural network pipeline consisting of three main stages:
- Noise Removal via Segmentation: A UNet-based model first processes the raw B-scan to create a mask, isolating hyperbolic features and filtering out background clutter.
- Sparse Back Projection Aggregation: The authors utilize a "Multi-Resolution Spatial Encoder" (MSRE), which stacks back-projected data at different spatial resolutions (256, 128, and 64 channels). This allows the network to simultaneously process fine details and broader geometric context.
- Hybrid Loss Function: To ensure the output accurately represents physical objects, the model uses a two-level loss: a pixel-level cross-entropy loss for location and a structural similarity loss to enforce the expected circular cross-sections of pipes or rebars.
Experimental Setup and Datasets
Obtaining "ground truth" for buried objects in the real world is difficult and often destructive. Consequently, the authors used gprMax, an open-source electromagnetic wave simulator, to generate a robust training dataset.
The synthetic dataset included 12 different concrete slab models containing perfectly conducting (PEC) circular bars of varying sizes (diameters of 1cm, 2cm, and 3cm) and depths. The final dataset consisted of 628 B-scan images, providing a comprehensive benchmark for training and testing the migration results.
Key Results and Performance
MigrationNet demonstrates significant advantages over traditional migration methods in terms of speed, accuracy, and robustness:
- Speed: MigrationNet processed a scan in 0.0347 seconds, compared to 3.47 seconds for conventional migration—a 100x speedup.
- Accuracy: The model achieved a 95.70% pixel accuracy and an 89.97% Mean IOU (Intersection over Union), significantly outperforming baseline methods.
- Noise Robustness: In tests involving Gaussian, Salt & Pepper, and Speckle noise, MigrationNet maintained low error rates (RMSE) while traditional methods saw rapid performance degradation.
Conclusion and Future Work
MigrationNet successfully mimics the physical migration process through a multi-resolution neural network, offering a fast and robust solution for subsurface imaging. While the results on synthetic data are impressive, future work will focus on real-world data integration. Real-world materials like soil and concrete introduce complexities such as varying moisture levels and air pockets that can distort signals. Moving toward real-time, on-site implementation for field engineers remains the next major hurdle.
Summary updated