PulseFormer: A Transformer-Based Architecture for Panoptic Segmentation in Radar Word Extraction

Radar word extraction is a crucial task in electronic intelligence for multifunction radars (MFRs). Although neural networks have shown potential in improving radar word extraction, existing methods still struggle with complex electromagnetic environments and background interference. To conquer these issues, we propose a more practical radar word extraction framework integrating panoptic segmentation to eliminate background interruption. Furthermore, an efficient Transformer-based feature extraction module is designed for pulse streams (named PulseFormer), which enhances robustness to missing or spurious pulses. PulseFormer segments the pulse stream into fixed-length patches, leveraging temporal dependencies to capture the structural features of radar words. These latent features are then processed through a tailored projection header to distinguish radar words from background noise, complementing radar word segmentation and recognition. Simulation experiments on the Mercury radar, a representative MFR, demonstrate that our method outperforms baseline approaches in challenging electromagnetic environments, including scenarios with corrupted data and variable pulse train lengths. In addition, by exploiting the fixed structure of radar words, our method can extract unknown radar words by leveraging information embedded in known radar words.