Source-linked AI summary

Buried object detection from B-scan ground penetrating radar data using Faster-RCNN

Minh-Tan Pham, Sébastien Lefèvre

arXiv:1803.08414v1cs.CV

TL;DR

Buried-object detection in GPR B-scans is difficult because available real training data are limited and classical methods produce false alarms or computational burdens. The paper adapts Faster-RCNN, pre-trains a CNN on grayscale Cifar-10, and trains on real plus gprMax-simulated radargrams. Preliminary results show better detection than classical HOG- and Haar-like-feature detectors and promising performance with few training samples.

  • Problem

    GPR buried-object detection has limited real training data, while classical methods can be computationally complex or produce false alarms and missed detections.

  • Method

    The paper pre-trains a CNN on grayscale Cifar-10 and uses its weights to train and fine-tune Faster-RCNN on real and simulated GPR data.

  • Results

    Faster-RCNN considerably outperforms HOG- and Haar-like-feature detectors on tested real data, while adding simulated data improves detection accuracy and confidence.

  • Takeaways & Limitations

    The proposed technique is promising for GPR data detection when only limited training samples are available.

  • Takeaways & Limitations

    The effectiveness was observed qualitatively, and quantitative evaluation plus finer localization of hyperbola apexes and prongs remained future work.

Abstract

from arXiv · show

In this paper, we adapt the Faster-RCNN framework for the detection of underground buried objects (i.e. hyperbola reflections) in B-scan ground penetrating radar (GPR) images. Due to the lack of real data for training, we propose to incorporate more simulated radargrams generated from different configurations using the gprMax toolbox. Our designed CNN is first pre-trained on the grayscale Cifar-10 database. Then, the Faster-RCNN framework based on the pre-trained CNN is trained and fine-tuned on both real and simulated GPR data. Preliminary detection results show that the proposed technique can provide significant improvements compared to classical computer vision methods and hence becomes quite promising to deal with this kind of specific GPR data even with few training samples.

1. INTRODUCTION

Buried-object detection in GPR B-scans is framed as hyperbola detection, but existing approaches face parameter complexity, false alarms, and limited end-to-end detection. The paper therefore applies Faster-RCNN after CNN pre-training and uses simulated data to address limited real training data.

  • GPR B-scan buried objects appear as reflected hyperbolic signatures that can be targeted with image-recognition methods.
  • Hough-based methods can incur high computational complexity because they handle and discretize many parameters.
  • Template, dictionary, HOG, and Haar-like approaches require substantial setup or still produce false alarms and missed detections.
  • Prior deep-learning work focused on classification or sliding-window patch detection rather than end-to-end hyperbola detection.
  • The paper adapts Faster-RCNN, pre-trains a CNN on grayscale Cifar-10, and incorporates simulated radargrams because real training data are limited.

2. DATA SETS

The dataset combines noisy real GPR radargrams with simulated scans generated by gprMax to expand training coverage. The simulated data vary object configurations while matching the real acquisition frequency and time range.

  • 2.1. Real collected data: About 100 real B-scans were collected at several French sites using a 300MHz antenna with a 100ns time range and penetration up to 7 meters.
  • 2.1. Real collected data: Real hyperbolas vary in shape and material response, with many weakly contrasted, asymmetric signatures disturbed by soil noise and clutter.
  • 2.2. Simulated data: Fifty simulated radargrams were generated with gprMax using varied object sizes, materials, positions, and depths.
  • 2.2. Simulated data: The simulations preserved the 300MHz antenna frequency and 100ns time range, then added noise estimated from real data.

3. APPLICATION OF FASTER-RCNN

The proposed application uses a two-stage pipeline: CNN pre-training followed by Faster-RCNN training and fine-tuning on real and simulated GPR images. Faster-RCNN combines region proposals with object classification and box regression.

  • The framework first pre-trains a designed CNN on grayscale Cifar-10, then trains and fine-tunes Faster-RCNN using real and simulated GPR data.
  • The CNN contains three convolutional layers, ReLU and max-pooling operations, and a 64-neuron fully connected layer.
  • Faster-RCNN uses a region proposal network to generate candidate regions and Fast-RCNN to classify regions and regress bounding boxes.
  • The region proposal and Fast-RCNN components share weights from the pre-trained CNN.

4. PRELIMINARY RESULTS

The evaluation tests simulated-only training, real-only training, and combined training with real-data testing. Preliminary results show confident simulated-data detections and better real-image performance when simulated data are added, relative to the classical detector.

  • The experiments used three scenarios: simulated-to-simulated, real-to-real, and combined real-plus-simulated training tested on real data.
  • 4.1. Performance on simulated data (scenario 1): On simulated data, the framework detected hyperbolas with high confidence scores and generally handled the scenario well.
  • 4.2. Performance on real data (scenarios 2 and 3): Adding simulated data produced higher-accuracy and higher-confidence real-image detections, with more good detections and fewer false alarms than real-only training.
  • 4.2. Performance on real data (scenarios 2 and 3): Faster-RCNN outperformed the HOG- and Haar-like-feature cascade detector on the tested real GPR image, whose boxes were unstable and had more false alarms.

5. CONCLUSION AND FURTHER WORK

The adapted Faster-RCNN detects buried objects in GPR B-scan data, with combined simulated and real training data outperforming classical-feature detectors on tested real data. The authors identify quantitative validation and finer hyperbola localization as further work.

  • Combining simulated and real radargrams enabled Faster-RCNN to perform well on tested real data and considerably outperform HOG- and Haar-like-feature detectors.
  • Faster-RCNN detection on real GPR data was compared with the classical COD detector.
  • The approach is promising for GPR detection when training samples are limited.
  • Quantitative evaluation remains necessary because the reported effectiveness was observed qualitatively, and hyperbola apex-and-prong detection is still needed for finer localization.
Loading 1803.08414v1…