Deep Open World SAR Target Recognition with Regular Polytope Networks

Traditional synthetic aperture radar (SAR) automatic target recognition (ATR) algorithms operate under the assumption that all relevant target labels are available during training, and that the number of targets of interest does not change over time. This means that the classifiers are guaranteed to make errors when presented with unknown target classes and cannot be efficiently updated to accommodate new classes. Open world recognition (OWR) addresses these limitations by rejecting unknown targets at test time and later adding them to the set of known classes via class incremental learning (CIL). Several OWR models have been proposed for generic visual tasks, but SAR ATR applications have been largely unexplored. Moreover, relatively few OWR approaches leverage deep learning in general. Regular polytope networks (RPNs) were recently proposed for CIL which pre-allocate and freeze the weights in the classifier layer of a deep network such that they correspond to the vertices of high-dimensional regular polytopes. This reduces catastrophic forgetting, the tendency for deep networks to become biased toward new data during CIL. In this paper, we employ an RPN as a deep feature extractor and study its effectiveness for open world SAR ATR when paired with a state-of-the-art OWR classifier. We also propose to use center loss to obtain more compact, stationary class clusters. We present initial results using the MSTAR targets dataset that serve to motivate future work on deep open world SAR ATR.