Open-Set Recognition of Radar One-Dimensional Range Profiles Based on Attention Mechanism and Multi-Task Joint Training

Abstract

Radar automatic target recognition (RATR) based on high-resolution range profiles (HRRP) has recently attracted increasing attention. The closed-set assumption prevalent in HRRP research breaks down in the presence of unseen classes, and conventional open-set methods are limited to unknown rejection rather than identification. We propose an open-set recognition strategy based on multihead self-attention mechanisms and multitask collaborative sparse training (MHA–CoST) to address this challenge. MHA–CoST combines CNN-multihead attention feature extraction with sparsely linked classification and reconstruction heads, whilst employing distance-based postprocessing to eliminate samples beyond recognised classes and subsequently categorise the unknowns. This strategy enables the HRRP recognition system to not only identify known classes but also effectively exclude unknown categories whilst achieving finer discrimination among different unknown classes. The proposed method exhibits superior rejection and classification capabilities for unknown categories compared to conventional open-set recognition techniques, significantly improving recognition efficiency and accuracy.