Synthetic Data Pipeline for Enhancing Machine Learning on Coherent Underwater Acoustic Images
Abstract
Machine learning (ML) can aid human analysts in interpreting large quantities of high-resolution synthetic aperture sonar (SAS) imagery of the seafloor collected by uncrewed underwater vehicles (UUVs). Moreover, it is vital to in-situ perception algorithms for UUV autonomy. Contemporary ML, such as deep learning, requires abundant and diverse training data for success. However, it is infeasible to create a sufficiently large training database of real SAS imagery encompassing all possible seafloor environments and conditions. This is a barrier to adoption of ML and human trust. We posit that this issue can be mitigated by augmentation with synthetic data produced by realistic SAS simulation.A recent break-through has facilitated the rapid but realistic simulation of raw SAS data. The new approach captures the important coherent acoustic wave physics and produces similar raw time-series data as existing point orfacet-based diffraction models (such as the US Navy s MASTODON), except an order-of-magnitude faster. This enables the generation of a sufficiently large, diverse, and realistic synthetic database for augmenting ML training. In this work, we will investigate how SAS data simulation is best leveraged to bridge the gap between lack of available real training data with the necessity of abundant data for good ML performance. To this end, we will address several open challenges and research questions: How does simulation realism affect ML performance (i.e., how good is good-enough?) Which aspects of the simulation physics are a necessity for good ML performance? What does the availability of abundant synthetic data mean for our ML algorithms?
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312315
Entities
People
- Alan J. Hunter
Organizations
- Office of Naval Research
- United States Navy
- University of Bath