Considerations for Modeling Backgrounds in Synthetic Aperture Sonar Imagery

Synthetic aperture sonar (SAS) imagery may be used to detect and classify objects on the seafloor. Recent advances in high-frequency sonar models allow realistic-looking simulated data that could potentially be used for training sonar image classifiers. For simulated data to be used effectively in training, it must bear sufficient similarity to experimental data. Experimental SAS images contain scattering from the object, as well as the shadow that it casts, that provide the important feature for detection and classification. But the images also contain a background, often produced by scattering from the seafloor. This work describes a set of considerations for modeling backgrounds in SAS imagery for the purposes of training convolutional neural network-based image classifiers. Both texture and the overall intensity of the background play roles when simulating imagery for training, but the sensitivity to each depends on the architecture of the particular network.