NHPL: Neural Collapse Guided Hierarchical Prototypical Learning for Radar HRRP Open Set Recognition

High-resolution range profile (HRRP) has emerged as a promising approach in real-time radar ground target recognition. To address the real-world operational requirements, it is demanding to develop open set recognition (OSR) that identifies both known and unknown classes of targets. In this paper, we propose a neural collapse-guided hierarchical prototypical learning (NHPL) method for HRRP OSR. The key of prototypical learning is constructing class-specific prototypes in the feature space as the representatives. However, this is challenging in HRRP recognition due to the intrinsic aspect-sensitivity, which results in variation of feature distribution and deviation of estimated prototypes. To tackle the above issues, we employ a hierarchical classification strategy which decomposes the OSR task into a multiple level recognition framework according to the hierarchical semantic taxonomy of classes. Within each level, discriminate features can be more effectively learned, thereby constraining the variation. Furthermore, inspired by neural collapse theory, we estimate class prototypes using the weight vectors of the trained network's final layer. This approach aligns prototypes closely with features of known classes while staying distant from features of unknown classes, thereby reducing prototype deviation. Extensive experiments on measured HRRPs demonstrate that the proposed method outperforms existing methods in terms of accuracy and robustness for recognizing unknown classes.