Ground Penetrating Radar Image Analysis for Underground Barrier Detection by Combining YOLOv12 with Channel-wise Attention and Denoising Auto-Encoder

1. Overview

This repository contains the official implementation of the paper "Ground Penetrating Radar Image Analysis for Underground Barrier Detection by Combining YOLOv12 with Channel-wise Attention and Denoising Auto-Encoder".

The method integrates:

• YOLOv12: a SOTA real-time one-stage object detector

• Denoising Autoencoder: to suppress clutter and preserve hyperbolic GPR signatures

• CBAM: to highlight informative channel-spatial features

By combining denoising and attention mechanisms, the framework enhances detection performance in challenging subsurface environments where GPR B-scan images are affected by noise, clutter, and overlapping reflections.

Key Features

• AE-enhanced preprocessing removes heterogeneous noise (soil irregularities, sensor interference), while maintaining the structural integrity of hyperbolic reflections.

• CBAM: Channel-wise + spatial attention refines multi-scale features and improves detection robustness under cluttered backgrounds.

• End-to-End Detection Pipeline: AE output consists of: original image, reconstructed image, and residual map. They are combined into 3-channel input for YOLOv12

Dataset

The dataset is privately provided by the company, which contains GPR B-scan images. It contains 4 types of pipeline: water, sewage, rain, and gas. In this study, we focus on the gas pipeline only, which offers distinct radar signatures.

2. Reproducibility

3. Reference

📃 Paper

4. Cite

Contributors

---

---