Frequency line detection in spectrograms using a deep neural network with attention.

Area of Science:

• Signal Processing
• Underwater Acoustics
• Machine Learning

Background:

• Detecting weak and time-varying frequency lines in underwater acoustic signals is challenging, especially at low signal-to-noise ratios (SNRs).
• Existing methods often struggle with noise and accuracy in complex acoustic environments.

Purpose of the Study:

• To propose a novel deep learning network, FLDNet, for robust detection of multiple weak and time-varying frequency lines in underwater acoustic signals.
• To enhance the accuracy and performance of frequency line detection under low SNR conditions.

Main Methods:

• An encoder-decoder architecture forms the basis of FLDNet, extracting multilevel features and reconstructing frequency lines.
• Attention-based feature fusion modules integrate deep and shallow features to mitigate noise and improve reconstruction.
• A composite loss function leveraging frequency line continuity further refines detection performance.

Main Results:

• FLDNet successfully detects frequency lines in both simulated and measured underwater acoustic signals.
• The network demonstrates superior performance compared to state-of-the-art methods, even at SNRs as low as -28 dB.
• Experimental results validate the effectiveness of the attention mechanisms and composite loss function.

Conclusions:

• FLDNet offers a significant advancement in detecting weak and time-varying frequency lines in challenging underwater acoustic environments.
• The proposed network provides a robust and accurate solution for underwater acoustic signal analysis.
• FLDNet's performance at extremely low SNRs opens new possibilities for acoustic monitoring and detection.