摘要
Underwater acoustic target recognition based on ship-radiated noise remains challenging in practical maritime applications. Variations in recording conditions and vessel operating states often produce complex and non-uniform time-frequency structures, while environmental and propagation effects exacerbate intra-class variability and reduce recognition reliability. To improve the representation of such variable time-frequency characteristics, this study proposes MSFENet, a frequency-aware multi-scale network with expert adaptive fusion. Its multi-scale dilated residual backbone (MSDRNet) captures acoustic patterns with different temporal and spectral extents through progressive feature extraction and selective cross-scale fusion. The resulting high-level representations are refined by time-frequency coordinate attention (TF-CA), which enhances discriminative structures and suppresses irrelevant responses through axis-specific feature recalibration. At the decision stage, frequency expert adaptive fusion (FEAF) integrates global and band-specific predictions on a per-sample basis, adaptively balancing global acoustic representations against localized frequency-band evidence. Experiments on DeepShip and ShipsEar using STFT, Mel, and CQT representations show that MSFENet achieves its strongest and most consistent gains with STFT inputs. On DeepShip, it attains 80.99% accuracy and an 80.66% macro F1-score, outperforming the strongest comparison method by 3.24 and 3.80 percentage points, respectively. Additional experiments further demonstrate the effectiveness and favorable accuracy–complexity trade-off of the proposed framework.
| 源语言 | 英语 |
|---|---|
| 期刊论文编号 | 127542 |
| 期刊 | Ocean Engineering |
| 卷 | 366 |
| 期 | P1 |
| DOI | |
| 出版状态 | 已出版 - 15 10月 2026 |
学术指纹
探究 'A frequency-aware multi-scale network with expert adaptive fusion for underwater acoustic target recognition' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver