MULTI-SCALE ATTENTION AND CONTRASTIVE LEARNING FOR SIGNAL MODULATION CLASSIFICATION

Classification of radar signals under low SNR conditions is challenging, especially for modulation types such as LFM slope direction and Barker code length. A network with multiscale attention and contrastive learning (MSAC-Net), is proposed, which integrates multi-scale attention and contrastive learning to enhance feature discrimination. The multi-scale attention module focuses on discriminative time-frequency patterns, making the slope differences between LFM-down and LFM-up more pronounced, while the contrastive learning module enhances the separability of Barker codes with different lengths. Our method significantly improves classification accuracy, achieving a 3% to 5% improvement across different SNR levels and demonstrating remarkable performance in challenging environments.