Training SAR-ATR Models for Reliable Operation in Open-World Environments

Training deep learning-based synthetic aperture radar automatic target recognition (SAR-ATR) systems for use in an “open-world” operating environment has, thus far proven difficult. Most SAR-ATR are designed to achieve maximum accuracy a limited set of classes, yet ignore the implications encountering novel classes during deployment. Even worse, standard learning training objectives fundamentally inherit closed-world assumption, and provide no guidance how handle out-of-distribution (OOD) data. In this work, we develop procedure called adversarial outlier exposure (AdvOE) codesign ATR system OOD detection. Our method introduces large, diverse, unlabeled auxiliary dataset containing samples from set. The AdvOE objective encourages neural network learn robust features in-distribution data, while also promoting entropy predictions adversarially perturbed versions We experiment with recent SAMPLE dataset, find our nearly doubles detection performance over baseline key settings, excels when using only As compared several other advanced techniques, affords significant improvements both classification statistics. Finally, conduct extensive experiments that measure effect granularity on rates; discuss different algorithms; analysis technique validate findings interpret problem new perspective.