A bi-directional cascaded transformer network for underwater narrowband signal enhancement

The narrowband components of ship-radiated noise are critical for the passive detection and identification of ship targets. However, the intricate underwater environment poses challenges for conventional acoustic signal processing methods, particularly at low signal-to-noise ratios. Previous studies have suggested the use of deep learning for denoising, but there is a significant lack of research on underwater narrowband signals. In response, this paper introduces a bi-directional cascaded transformer network (BCT-Net) with two branches simultaneously extracting features from target signals and ambient noise. Leveraging cascaded attention mechanisms, BCT-Net is able to detect narrowband features at signal-to-noise ratios as low as -20 dB. Through a frequency-guided attention module and bi-directional cross-attention-based interaction module, the BCT-Net excels in the extraction of the target features. This methodology enhances noise suppression by leveraging the interaction of features extracted from both the target and noise branches. Operating at a fine-grained level, the model reconstructs subtle frequency variations while ensuring the separation of the target signals and ambient noise. Ablation experiments underscore the unique contributions of each module, which together significantly enhance denoising performance. Our proposed BCT-Net surpasses the existing methods across various evaluation metrics, emphasizing its superiority when it comes to narrowband signal enhancement.