Deep Learning Based Advanced SAR Automatic Target Recognition Algorithms

As artificial intelligence / machine learning (AI/ML) algorithms have been successfully applied to simple radar target detection and classification problems where targets are easily discernable or separable or a large database for training is available, the challenging problems are emerging.  These challenges include adversarial perturbed radar imagery classification, measured data limited target recognition, and out of library target detection and classification.  Hence, the focus of this talk will be highlighting three recent advancements on deep learning based synthetic Aperture Radar (SAR) imagery classification for automatic target recognition (ATR).  First, we will present how SAR signals/imagery can be modified by various noise sources and thus the loss of classification accuracy.  We then propose adversarial training (AT) to mitigate this issue. The second part of the talk highlights how limited amount of measured data or fully synthetic data can be used to train a deep neural network for target classification. We achieved 95% target recognition accuracy on measured data using fully synthetic data for training.  Finally, we will present out of library target classification using adversarial outlier exposure (AdvOE) algorithm.  In all of these three research, we will be using AFRL public released civilian datadome (CVDome) and SAMLE datasets for training and testing the deep neural networks model.