Transductive Multi-Prototype Network for HRRP Recognition with Missing Aspects

The High-Resolution Range Profile (HRRP) plays a pivotal role in radar automatic target recognition. However, due to its aspect sensitivity, the absence of aspects within the training data can lead to a decline in recognition accuracy. To mitigate this issue, we introduce a novel Transductive Multi-Prototype Network (TMP-Net), leveraging the intrinsic similarity within test data to propagate labels along a manifold, thereby achieving classification. TMP-Net addresses the aspect sensitivity by utilizing multiple prototypes per class. Additionally, a loss function term is proposed to ensure the existence of the manifold. Experiments on the MSATR dataset demonstrate that TMP-Net outperforms existing methods by 10.22% in recognition accuracy even with only 10% of aspects represented in the training set. Furthermore, analysis reveals the advantages of employing multiple prototypes.