Multi-layer Multi-head Self-attention Model for Radar HRRP Target Recognition

Radar automatic target recognition (RATR) based on high-resolution range profile (HRRP) has attracted much attention in recent years. We present a Multi-head Attention model to characterizes the long-term temporary dependency across the range cells of HRRP, which utilizes stacked attention modules to aggregate features with weights obtained by aligning range cells in a single HRRP, resulting in more discriminative latent representations. We further develop two efficient pooling strategies to obtain the final HRRP representations for the downstream tasks. Experiments on measured data show that the proposed method brings better performance over the corresponding models.