A novel noise-robust and lightweight underwater acoustic target recognition method based on BSCQT and DCAM

To address the challenges of low recognition accuracy and high computational overhead in noisy underwater environments, this paper proposes a novel noise-robust and lightweight underwater acoustic target recognition method based on Band-Specific Constant Q Transform (BSCQT) and Dynamic Context-Aware Masking (DCAM). First, BSCQT achieves effective noise suppression and feature extraction through multi-band adaptive weighting and feature concatenation. Then, by combining frequency-adaptive pooling granularity with traditional lightweight context-aware masking, a dynamic context-aware masking (DCAM) mechanism is constructed to implement adaptive attention on BSCQT features, improving recognition accuracy while maintaining low computational complexity. Furthermore, a Dynamic Context-Aware Masking Network (DCAMNet) is developed based on DCAM for hierarchical feature learning, integrating cascaded DCAM dense TDNN blocks for efficient information transmission. Finally, within the DCAMNet architecture, target recognition is accomplished through global pooling and fully connected classification layers. Extensive experimental results demonstrate that the proposed method achieves 99.23% recognition accuracy with only 0.55G Floating Point Operations (FLOPs) computational complexity, showing significant improvement in recognition efficiency compared to existing state-of-the-art methods and verifying the effectiveness of our approach.