A study on quality, architecture, and performance relationships in sonar classifiers trained with simulated data

Convolutional neural networks (CNNs) are commonly used to classify objects observed in synthetic aperture sonar imagery. Controlled laboratory data and complementary simulated imagery produced with physics-based acoustic scattering models were used to investigate the relationships between classification accuracy, simulation quality, and network architecture. Three experiments investigated simulation errors that arise from imperfect models, inaccurate knowledge of target shape, and inaccurate knowledge of target appearance due to partial burial. In a matched environment, results show CNN-based classifiers can be designed to be insensitive to simulation errors of a certain scale, allowing coarse, imperfect simulations for training.