Multiscale Attention-Based Whistle Segmentation for Biomimetic Communication

This paper proposes a deep learning-based segmentation model for biomimetic underwater acoustic communication systems, capable of extracting dolphin whistle signals from underwater acoustic data containing diverse noise sources while reducing contour discontinuity. The proposed model is based on a U-Net architecture and incorporates the Convolutional Block Attention Module (CBAM) and the Multi-Scale Combining Spatial Attention Module (MSC-SAM) to capture whistle features at multiple resolutions. CBAM is applied to skip connections to emphasize meaningful channels and spatial information related to whistle contours. In addition, MSC-SAM is designed to integrate spatial attention maps generated at each resolution, thereby reducing whistle contour discontinuity that occurs during the segmentation process. The proposed model is evaluated using acoustic data collected by the National Oceanic and Atmospheric Administration (NOAA) and compared with existing models, including U-Net, ResU-Net, U-Net++, FPN, PSPNet, DeepLabv3+, and MA-Net. Experimental results show that the proposed model achieves improved segmentation performance in terms of pixel accuracy, IoU, and Dice score. For the proposed whistle-level evaluation criteria, namely Discontinuous Whistles and Missed Whistles, the proposed model reduces discontinuous whistles and missed detections by 65.6% and 61.1%, respectively, relative to U-Net.