Lightweight and Robust Radar Waveform Recognition Based on RepNRS-LPI-Net

Abstract

To address the degradation of low-probability-of-intercept (LPI) radar waveform recognition caused by noise dispersion and the masking of modulation-dependent structures in low-SNR Choi–Williams distribution (CWD) images, this paper proposes RepNRS-LPI-Net, an integrated framework for robust recognition and lightweight deployment. CWD converts each received waveform into a two-dimensional time–frequency image that characterizes temporal evolution, frequency variation, and localized energy distribution. The proposed RepDW block integrates 3 × 3, 1 × 3, and 3 × 1 depthwise branches with an identity branch during training to capture joint time–frequency, temporal-direction, and frequency-direction responses while preserving informative features. These branches are then algebraically fused for efficient deployment. In addition, NRS-ECA combines channel recalibration with channel-dependent soft shrinkage to attenuate weakly supported noise-like activations without assuming that all weak responses are noise. Focal modulation and label smoothing are conservatively adopted as auxiliary training strategies to address difficulty imbalance and confidence regularization. Experimental results show that RepNRS-LPI-Net achieves 79.553350% overall accuracy and 49.137529% low-SNR accuracy, while the deployment form reduces the parameter count to 37,142 and the learned-layer computation to 15,722,496 MACs. These results indicate that RepNRS-LPI-Net improves measured recognition performance while substantially reducing deployment complexity under the modeled multipath, Rayleigh-fading, Doppler, and AWGN conditions.