Underwater Signal Denoising Using Deep Learning Approach

In this work, we adapt statistical reasoning applied previously on image noise removal shown in Noise2Noise for the task of underwater acoustic signal denoising. Hydrophone channels across an array experience similar signal observation yet different noise realizations. In this context, we propose WaveN2N that is able to learn noise removal and clean signal reconstruction from multi-channels array data in a self-supervised learning setting. It can subsequently be used for practical signal enhancement for single hydrophone or shorter array configuration. Experiment using a vertical line array dataset showed that the trained model learnt to remove the water channel response and ambient noise based on the reverberant and noisy data. During testing, the model was able to achieve reverberation and noise reduction intelligently for single hydrophone signals.