LAUSR

In this paper, an underwater ambient noise (UWAN) removing framework based on generative adversarial network (GAN) to recover underlying ship-engine data is proposed. Unlike previously reported GAN-based denoising techniques, the proposed underwater ambient-noise removing GAN (UWAR-GAN) is modelled using both, log magnitudes as well as phase information with significant time-frequency resolution. The as-developed network integrates U-net generator with articulate of skip connections, a patch discriminator and an additional adversarial correlation loss function. The proposed technique has been rigorously evaluated using publicly available real-time and simulated noisy data. Moreover, the UWAR-GAN shows excellent performance in SNR (signal-to-noise ratio), SSIM (structural similarity), MOS (mean opinion score) and RMSE (root mean square error) with 45.1%, 40.7%, 64.3% and 61.1% improvement, respectively, when compared with statistical method such as Wiener. Similarly, in comparison to GAN-based method, the proposed framework still outperforms with an improvement of 26.5%, 39.1%, 7.6% and 35.3% for SNR, SSIM, MOS and RMSE, respectively. Finally, the UWAR-GAN is assessed on very low SNR such as −10 dB with four common UWAN models, which infers that the proposed framework can be a robust candidate for the real-time underwater applications.