FocuS-MN: Focusing on Underwater Signal Denoising via Sequential Memory Networks with Learnable Resampling

The coupling of non-stationary marine noise and complex ship-radiated signals makes high-fidelity signal recovery exceptionally difficult. Existing deep learning methods often prioritize objective metrics, such as the Scale-Invariant Signal-to-Noise Ratio (SI-SNR), but fail to maintain the integrity of narrow-band line spectral data. We propose FocuS-MN, an end-to-end framework that combines learnable resampling with Feedforward Sequential Memory Network (FSMN)-based temporal modeling for precise waveform reconstruction. The model is optimized using a two-stage training strategy to ensure stable magnitude estimation and waveform consistency. On the ShipsEar dataset, FocuS-MN shows strong generalization to unseen vessel types. At a −5 dB Signal-to-Noise Ratio (SNR), it achieves a Signal-to-Distortion Ratio (SDR) of 3.77 dB and a Segmental Signal-to-Noise Ratio (SSNR) of 3.83 dB. Power Spectral Density (PSD) analysis further confirms that FocuS-MN recovers fine-grained line spectral structures, proving its effectiveness in both noise suppression and signal fidelity.