Lightweight underwater acoustic time-frequency separation network for efficient marine target recognition
Chen, M.; Lu, Y.; Cheng, L.; Zhu, R.; Tao, K.; Li, Y.; Abdel Wahab, M.A. (2026). Lightweight underwater acoustic time-frequency separation network for efficient marine target recognition. Ocean Eng. 343: 123234. https://dx.doi.org/10.1016/j.oceaneng.2025.123234
In: Ocean Engineering. Pergamon: Elmsford. ISSN 0029-8018; e-ISSN 1873-5258, more
|  |
| Author keywords |
Underwater acoustic target recognition;Time-frequency separation convolutions;Marine environment monitoring;Lightweight neural networks;Ocean engineering;attention mechanism |
| Authors | | Top |
- Chen, M.
- Lu, Y.
- Cheng, L.
- Zhu, R.
|
- Tao, K.
- Li, Y.
- Abdel Wahab, M.A., more
|
|
| Abstract |
Underwater acoustic target recognition is a core technology in ocean engineering, essential for marine security, underwater navigation, and resource development, with growing demand driven by expanding marine activities and autonomous platforms. Traditional underwater acoustic target recognition methods process time-frequency spectrograms using standard two-dimensional convolutions, failing to exploit signal physical properties due to coupled time-frequency processing. This paper presents our proposed lightweight Underwater Acoustic TimeFrequency Separation Network UATFSN implementing decoupled modeling through our designed TimeFrequency Separate Convolutions TFSC modules. TFSC splits feature maps into frequency path and time path, employing anisotropic convolution kernels respectively to extract harmonic features in frequency dimension and transient features in time dimension, incorporating adaptive residual normalization for enhanced stability. Experimental results demonstrate UATFSN achieves 98.23 % accuracy on DeepShip dataset, surpassing the strongest baseline, ResNet34, by 2.04 percentage points while requiring only 0.05M parameters, representing 99.77 % reduction compared to ResNet34. On ShipsEar dataset, UATFSN attains 98.75 % accuracy with 99.26 % computational reduction. Robustness evaluation confirms 84.20 % accuracy under 0 dB SNR and 78.17 % accuracy using 5 % training data. Our proposed UATFSN represents the first breakthrough achieving simultaneous accuracy and computational efficiency in underwater acoustic recognition, providing critical technological support for intelligent deployment on resource-constrained marine platforms. Our proposed UATFSN achieves a remarkable balance between accuracy and computational efficiency in underwater acoustic recognition, providing an effective technological solution for intelligent deployment on resource-constrained marine platforms. |
|
