Transformer-based dual-modal fusion for radar target recognition

Abstract

Radar automatic target recognition (ATR) faces challenges such as heterogeneous amplitude–electromagnetic modalities, long-sequence modeling difficulty, and class imbalance. To address these issues, this paper proposes a Transformer-based dual-modal fusion framework. A structured electromagnetic sequence Transformer encoder is designed to capture long-range dependencies in radar electromagnetic signals. A bidirectional cross-attention module is introduced for fine-grained interaction between amplitude (AD) and electromagnetic (ED) features. Furthermore, prototype contrastive learning is integrated with classification loss to enhance intra-class compactness and inter-class separability. Experiments on the IMD ISAR dataset demonstrate that the proposed method achieves 98.27% accuracy, outperforming existing baseline models.