Terahertz Polarimetric ISAR Target Recognition Based on a Dual-Branch Space-Scatter Interaction Transformer

Terahertz Polarimetric Inverse Synthetic Aperture Radar (Pol-ISAR) contains rich polarimetric scattering mechanisms and fine-grained spatial structures. However, how to effectively combine the scattering prior information from multi-polarimetric channels with deep spatial semantics remains a bottleneck issue for improving the recognition performance. To effectively combine the scattering characteristics and spatial characteristics of Pol-ISAR data, this article proposes a target recognition framework based on a Dual-Branch Spatial-Scattering Interaction Transformer (DB-SSIT). This framework first extracts the correlation between scattering mechanisms across different polarimetric channels through the polarimetric scattering branch, while capturing spatial information in the spatial feature branch using multi-scale convolutions and spatial attention. Subsequently, a local-global feature fusion module is introduced, which combines efficient multi-head self-attention with convolutional attention to achieve adaptive fusion of local details and global semantics. After training and testing on our terahertz polarization ISAR dataset, the results show that DB-SSIT performs stably in identifying various targets, with an overall accuracy rate of 99.38%, while the number of model parameters is only 1.91M. It can convert the tiny scattering differences in terahertz high-resolution imaging into discriminative features, thereby enhancing the recognition ability.