Understanding model fidelity for training synthetic aperture sonar image classifiers

Abstract

Convolutional neural networks (CNNs) are increasingly employed for classification tasks in automated target recognition (ATR) algorithms for synthetic aperture sonar (SAS) images. Training ATR algorithms, however, requires many unique observations of the targets of interest. Data available for training is often limited, and acquiring additional training data through experimentation is prohibitively expensive. Increasingly, training with simulated data is being considered as an alternative, but the fidelity required of the models that generate this data is not yet known. SAS imagery typically contains significant complexity from countless physical mechanisms. Simulating complex scenes requires multiple models to account for these different effects, but some effects may not require accurate modeling for the purposes of training an ATR algorithm. This presentation will describe a study to investigate the fidelity of models required so that simulated data may be used interchangeably with experimental data for training CNNs. Using in-air experimentation, a high-fidelity data set was developed with multiple degrees of complexity. A high-frequency sonar signal model was then used to generate complementary simulated data. This approach allows for specific physical features in the data to be individually isolated, enabling detailed exploration of the relationships between model fidelity and CNN architecture.