On the Development of an Acoustic Image Dataset for Unexploded Ordnance Classification Using Front-Looking Sonar and
Piotr Ściegienka1,2, Marcin Blachnik3
1Joint Doctoral School, Silesian University of Technology, 44-100 Gliwice, Poland.
Sensors (Basel, Switzerland)
|September 28, 2024
Summary
Researchers created a sonar acoustic image dataset for classifying unexploded ordnances (UXOs) and non-UXOs. Pre-training entire networks achieved 98% accuracy, with VGG models performing best.
Area of Science:
- Marine robotics and autonomous systems
- Artificial intelligence and machine learning
- Sonar imaging and signal processing
Background:
- Underwater unexploded ordnance (UXO) detection is critical for maritime safety and infrastructure development.
- Current detection methods often rely on visual or magnetic surveys, which have limitations in certain environments.
- Acoustic imaging offers a promising alternative for UXO identification, but requires robust datasets and advanced classification algorithms.
Purpose of the Study:
- To develop a comprehensive dataset of forward-looking sonar acoustic images for classifying unexploded ordnances (UXOs) and non-UXOs.
- To establish a benchmark for evaluating machine learning models in UXO detection using acoustic data.
- To investigate the effectiveness of various deep learning architectures and transfer learning techniques for this classification task.
Main Methods:
- A dataset of 69,444 acoustic images (512x399 resolution) was generated using digital twin simulations in Gazebo, featuring UXO and non-UXO classes.
- State-of-the-art image classification models including VGG16, ResNet34/50, ViT, RegNet, and Swin Transformer were evaluated.
- Experiments involved two neural network pre-training strategies: final layers only, and the entire network.
Main Results:
- All evaluated models achieved comparable performance, reaching 98% balanced accuracy when the entire network was pre-trained.
- Pre-training the entire network was found to be essential for achieving high classification accuracy.
- The VGG model unexpectedly yielded the highest accuracy among the tested architectures.
Conclusions:
- A high-performance dataset for acoustic-based UXO classification has been successfully developed.
- Deep learning models, particularly when fully pre-trained, demonstrate significant potential for accurate UXO detection using sonar imagery.
- The VGG model's superior performance suggests specific architectural advantages for this acoustic classification task, warranting further investigation.
