An underwater acoustic target recognition system with band splitting and sub-band weighting

Current widely used data-driven machine learning underwater acoustic target recognition (UATR) methods encounter issues such as blurring of energy distribution and loss of position information of the original frequency bands. To mitigate these issues, a band-split weighting network (BSWNet) is proposed in this paper. BSWNet is an UATR system that employs band splitting to enhance the performance of UATR tasks by accurately modeling the relationships between different sub-band features. The system incorporates two innovative modules: band position encoding (BPE) and band weight attention (BWA). BPE preserves the positional information of the original sub-bands, while BWA extracts the importance weights of these sub-bands. Experiments on the ShipsEar and DeepShip datasets demonstrate that BSWNet outperforms baseline systems with over 10% increase in accuracy under -10 dB SNR, which exhibit enhanced noise resilience. Ablation studies confirm the positive impact of both BPE and BWA modules on the network's performance.