Line Spectrum Enhancement of Underwater Acoustic Targets Based on a Time-Frequency Attention Network

Deep learning-based line spectrum enhancement methods have received increasing attention for improving the detection performance of underwater low-noise targets using passive sonar. Among them, Long Short-Term Memory (LSTM)-based line spectrum enhancement networks have high flexibility due to their nonlinear processing capabilities in time and frequency domains. However, their performance requires further improvement. Therefore, a Time-Frequency Attention Network (TFA-Net) is proposed herein. The line spectrum enhancement effect of the LOw-Frequency Analysis Record (LOFAR) spectrum can be improved by incorporating the time and frequency-domain attention mechanisms into LSTM networks, In TFA-Net, the time-domain attention mechanism utilizes the correlation between the hidden states of LSTM to increase the model’s attention in the time domain, while the frequency-domain attention mechanism increases the model’s attention in the frequency domain by designing the full link layer of the shrinkage sub-network in deep residual shrinkage networks as a one-dimensional convolutional layer. Compared to LSTM, TFA-Net has a higher system signal-to-noise ratio gain: when the input signal-to-noise ratio is –3 dB and –11 dB, the system signal-to-noise ratio gain is increased from 2.17 to 12.56 dB and from 0.71 to 10.6 dB, respectively. Experimental results based on simulated and real data show that TFA-Net could effectively improve the line spectrum enhancement effect of the LOFAR spectrum and address the problem of detecting underwater low-noise targets. © 2024 Journal of Pattern Recognition and Artificial Intelligence. All rights reserved.