Characterization of deep neural network feature space for inverse synthetic aperture radar automatic target recognition

Abstract

The Airborne Radar Systems and Techniques group at MIT Lincoln Laboratory trained neural networks to classify different targets at sea based on inverse synthetic aperture radar (ISAR) data. Simulated data was used to train these neural network based automatic target recognition (ATR) systems. The technical challenge of this project was to find a way to evaluate the quality and adequacy of a limited set of training data. Using simulated ISAR images to train neural networks, the project determined the minimum amount of variation in terms of parameters such as aspect angle to adequately train a neural network. Establishing a correspondence between training data variation and the resulting feature space of the data informed the minimum spanning-set of training data required for future data collects.