Suppressing short time marine ambient noise based on deep complex unet to enhance the vessel radiation signal in LOFAR spectrogram

Abstract

UNet-type networks have demonstrated good performance in the field of denoising. In this paper, we applied a DCUNet network specifically for denoising underwater acoustic signals, which are characterized by their nonlinear, non-smooth and non-Gaussian features. The process involves transforming noisy data into LOFAR spectrograms for input into DCUnet, redesigning the network structure based on the features of underwater acoustic signals. Subsequently, a Noise2Noise training method was employed to reconstruct the underwater background noise through the end-to-end architecture. The effectiveness of the algorithm was validated on publicly available datasets after augmentation. Extensive experimental results show that our method achieves an SNR improvement of over 10 dB and is capable of restoring signals with an initial SNR of −20 dB, demonstrating better performance compared to traditional denoising algorithms. In addition, the method is verified using the public datasets and long-distance single-frequency experiments. In conclusion, the DCUNet model exhibit effectiveness in underwater acoustic noise suppression and robustness in different data.