Understanding Model Fidelity for ATR Data Augmentation

Convolutional neural networks (CNNs) are increasingly employed in remote sensing applications such as synthetic aperture sonar (SAS), where the available sensor imagery may be limited. These networks can be applied for recognition and classification tasks in automa,ted target recognition (ATR) algorithms. Successful training of ATR algorithms, however, generally requires many unique observations, of the targets of interest. The increasing fidelity of target and environmental modeling suggests that it may be possible to augmen,t ATR training data sets with additional data that is generated through modeling and simulation. This project will investigate the,fidelity that is required of simulated data to be used interchangeably with experimental data for training ATR algorithms.Using prec,isely controlled in-air acoustic instrumentation, experimental SAS data will be collected on scenes with increasing complexity. Sev,eral environmental and target models will be used to generate complementarysimulated data of the same types of scenes. CNNs of vary,ing complexity will then be trained on subsets and combinations of these two data sets at multiple levels of resolution and fidelity,. By quantifying the performance of the networks and comparing the mutual information between the experimental and synthetic data,,this research will seek to identify the relationships between network architecture, scene complexity, model fidelity, and algorithm,performance.DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited.