Realistic Augmentation for Mid-Frequency Sonar (RAMS)

Signal Systems Corporation proposes the innovative Realistic Augmentation of Mid-frequency Sonar (RAMS) to generate large quantities of physically plausible sonar ping images for the SQQ-89(A) system. Our approach incorporates two multiscale deep predictive GANs trained on time series data and constructed B-scans. RAMS trains on data from both high-fidelity simulations and real-world recordings to generate novel matched filter responses. Augmented data is processed to produce realistic target and clutter echo images and clusters with automatic ground truth across a comprehensive sampling of bathymetric profiles, environmental scattering, and dynamic weather conditions. RAMS uses innovative single and multiple ping training to make the most use of limited real and high-fidelity data. RAMS will achieve a cost efficiency an order of magnitude better than physics-based simulations (or real-world data collections) to supply robust training data for active sonar single and multiple ping classifiers used by the US Navy.ÝIn Phase I we will use a Sonar Simulation Toolkit mid-frequency simulation to drive our feasibility studies combined with in-house unclassified active sonar data sets. Our image quality control pipeline incorporates Multi-Scale Structural Similarity Index, FrÈchet measures of feature distribution similarity, and area under ROC curve for trained single and multiple ping classifiers.