Self-supervised denoising model based on deep audio prior using single noisy marine mammal sound sample

Due to the limited signal acquisition conditions in the marine environment, it is difficult to repeatedly sample the same target and obtain samples with high quality. Therefore, the denoising algorithm based on self-supervised method is the most commonly used approach. In this paper, we propose a prior based denoising network (PriorDeNet), which is a two-stage model for denoising a single noisy sample without ground truth. In the first stage, the potential statistical prior is obtained by the proposed model structure, so that the signal and noise in the sample can have different fitting speeds, and the application of tricks such as dropout significantly improves the prior. Consequently, the single-sample problem can be transformed into a multi-sample problem. In the second stage, a denoising model that only uses noisy samples is constructed, which uses intermediate outputs generated in the first stage to obtain the denoised sample, and designs a loss function based on the coefficient of variation to match their features. Through the tests on real-world and synthetic samples with different signal-to-noise ratios, we can prove that this method outperforms existing marine bioacoustic signal denoising algorithms such as Wiener filter, spectral subtraction and wavelet-based denoising method in terms of MSE, PSNR and SSIM, and is not sensitive to the change of signal-to-noise ratio of noisy samples.