Open Set Recognition in Synthetic Aperture Radar Using the Openmax Classifier

Abstract

Automatic Target Recognition (ATR) using Synthetic Aperture Radar (SAR) images has received a lot of attention in the past two decades. The prevailing assumption in most of the classification studies is the “closed set” modeling. However, the system might having to operate in an open set environment, in which unknown targets may be given to the system for classification. To tackle this problem, the Openmax classifier has been recently introduced in optical domain to enable convolutional neural networks (CNNs) to distinguish between open set and closed set classes. To the best knowledge of the authors, Openmax has not been yet examined in the context of SAR or ISAR images. In this work, we address the open set recognition problem in the radar domain. We evaluate the performance of the Openmax classifier using real images from the SAMPLE dataset, which is a subset of the well-known MSTAR dataset. A special emphasis has been given to the tail-fitting procedure that plays a major role in the Openmax scores calculation. Moreover, the conventional performance indexes under different global thresholds are also analyzed.