DCMNet: A Supervised Learning Framework for Radar Signal Modulation Recognition

Traditional radar signal modulation recognition (RSMR) methods struggle to achieve the required accuracy under low signal-to-noise ratio (SNR) conditions. To address this issue, a hybrid network architecture integrating deformable convolution and mamba (DCMNet) is proposed. Specifically, DCMNet employs a multi-view feature extraction structure that combines inverted deformable convolution (IDC) with a state space model (SSM), enabling dynamic adjustment of convolution kernel positions and capturing global information and dependencies in long sequence data. The cross-gated feature fusion (CGFF) mechanism effectively modulates and dynamically aggregates features from different perspectives. The lightweight design provides significant advantages in terms of network scale and deployment. Experimental results demonstrate that the proposed method achieves excellent performance on a dataset with ten different waveforms. Notably, at an SNR of −8 dB, the recognition accuracy exceeds 90%, significantly outperforming existing methods.