An Open-Set Recognition Approach for SAR Targets Using Only Classification Scores

The focus of this paper is the open-set recognition problem of Synthetic Aperture Radar (SAR) targets, and a simple and robust open-set recognition approach is proposed that uses only a simple convolutional neural classification network. The proposed approach constructs the D-SCORE feature and uses the statistical method to model the D-SCORE obtained in the training phase. This allows for the identification of the threshold for distinguishing between known and unknown classes, and ultimately realizes the open-set recognition of SAR targets. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) data set demonstrate the efficacy of the proposed approach in achieving enhanced open-set recognition performance.