SFUnet-DCNN: An Effective Approach for LPI Radar Waveform Recognition Under Low SNR Conditions

Low Probability of Intercept (LPI) radar waveform recognition is crucial in modern electronic defense, providing decision-makers with essential insights. However, LPI radar signals are highly susceptible to noise interference, especially in low Signal-to-Noise Ratio (SNR) environments, significantly reducing recognition accuracy. To address this challenge, we propose a novel LPI radar waveform recognition approach called SFUnet-DCNN, which can achieve robust recognition performance under low SNR scenarios. First, the LPI radar signals are transformed into Time-Frequency Images (TFIs) using the Choi-Williams Distribution (CWD). Subsequently, the TFIs at various SNR levels are fed into a Deep Convolutional Neural Network (DCNN) with Spatial Frequency Blocks (SFBs), which extracts effective features for the SNR estimation and the residual attention classification network. The SFUnet denoising network is activated specifically to reconstruct low SNR TFIs, which are then reprocessed by the DCNN. Finally, the residual attention classification network generates modulation-type predictions. Extensive simulation experiments demonstrate the impressive recognition performance of our approach compared with State-Of-The-Art (SOTA) baselines. Remarkably, even at an SNR of -10 dB, it achieves an overall recognition accuracy of up to 95%. Furthermore, we have verified the effectiveness of the proposed model using actual measurement data, emphasizing its practical applicability, which confirms that the proposed approach is a powerful tool for enhancing radar signal recognition in challenging environments. The code implementation of SFUnet-DCNN and the datasets will be available soon at https://github.com/stu-cjlu-sp/rsrc-for-pub/tree/main/SFUnet-DCNN.