Dual-axis spectrum attention network: A robust model for underwater acoustic signal denoising

In the field of underwater acoustics, analyzing underwater targets through acoustic signals constitutes a critical task. However, the complex and diverse marine environments lead to sparse distribution of target-related discriminative patterns within acoustic signals, thereby constraining the construction of accurate acoustic systems. To effectively analyze target-radiated noise features in such environments, denoising methods based on attention mechanisms have become increasingly prominent. In this work, the Dual-Axis Spectrum Attention Network (DASANet) is proposed as a denoising model that applies an encoder-decoder structure. Based on the characteristics of underwater target-radiated noise in the temporal and frequency domains, DASANet integrates a Temporal and Frequency Axes Self-Attention (TFASA) module in the encoder to enhance narrowband line spectra and periodic modulation features that reflect mechanical operations and propeller structures. To further recover spectral details, the decoder incorporates Gated Cross-Attention (GCA) modules, dynamically capturing and refining target-related representations. DASANet was thoroughly evaluated on the Shipsear dataset, demonstrating superior performance with 11.43 dB improvement in signal-to-distortion ratio and 8.85 dB increase in scale-invariant signal-to-noise ratio.