Improving Automatic Target Recognition through Dataset Augmentation using Generative Adversarial Networks in Synthetic Aperture Sonar
Abstract
"The proposed project aims to improve the capabilities of automated target recognition (ATR) forSynthetic Aperture Sonar (SAS). Deep learning techniques for SAS ATR typically suffer fromclass imbalance in image datasets, and the black-box nature of deep-learning models leads to poorinterpretability for this technology. The technical approach develops a method for generating noveland realistic SAS images of the seafloor to improve dataset augmentation for ATR. The projectcombines the benefit of renderers that can generate objects and environments with semantic controlwith generative adversarial networks (GANs) to help improve the outputs fidelity to real worldSAS imagery. In addition, the physics of sonar image formation will be encoded directly into GANstructure to enforce statistical and physical properties. This includes looking at how GANs performwith respect to the raw time series and/or the beamformed single-look complex image, and howthey can exploit physics-based knowledge of the properties in these data representations (i.e.real/complex are orthogonal, the power spectral density, etc). This will enable the networks tospeed up the generation of new SAS images, but still retain explicit control over key physicalparameters such as geometry, rotation, translation, scale, and scattering/illumination effects, whilebridging the gap towards physically-realistic machine learning for SAS image generation.Anticipated outcomes of this research include enhancing weak classes in SAS datasets andimproving target classification networks for SAS ATR. This research will also shed light on howdeep learning architectures characterize the physical aspects of SONAR images (e.g. scattering,frequency, propagation). This will have impact on DoD capabilities including technologyimprovement for SAS for the Navy."
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012330
Entities
People
- Suren Jayasuriya
Organizations
- Arizona State University
- Office of Naval Research
- United States Navy