This is the authors' abstract. We don't add key points for this paper.
In recent years, HRRP-based radar automatic target recognition (RATR) using deep neural networks (DNNs) has garnered increasing attention. In real-world scenarios, test target categories frequently include both known and unknown classes, which is referred to as the open set recognition (OSR). Existing DNN-based classifiers tend to classify all captured HRRPs as known classes, presenting significant challenges for the practical application of HRRP-based RATR technology. Besides, extracting HRRP features at the optimal scale within DNNs is also essential for improving the HRRP-based OSR performance. To address these issues, we propose a systematic OSR solution named open set full-scale model (OSFSM). Firstly, OSFSM utilizes a set of convolutional kernels with prime-number lengths to efficiently extract HRRP features across all scales. Meanwhile, the learnable vectors are used to weight and filter the features at each scale. Secondly, we attribute the root cause of overconfidence in DNNs to the unconstrained magnitude of the logits. Therefore, OSFSM designs a parameter weight constraint (PWC) loss to implicitly suppress excessive growth in logits magnitude. Additionally, the prototype-based orientational (PBO) loss is proposed to reduce class cluster overlap. We validate the performance of OSFSM using two measured HRRP datasets and one simulated HRRP dataset. The OSFSM method achieved the optimal classification performance among the implemented methods.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Li et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: