DenoiseNet: Deep Generator and Discriminator Learning Network With Self-Attention Applied to Ocean Data

Well-organized seismic signals play a significant role in the subsequent seismic data processing. The multiscale learning of characteristic signals of complex structures by deep convolutional neural networks has obvious benefits in reducing random noise in seismic data. However, deep convolutional neural networks also have shortcomings. It cannot discover effective features in seismic data structures or recover high-quality seismic signals just using convolution. Therefore, the article presents a generative adversarial network (GAN) architecture in conjunction with the U-Net network. To produce the mapping connection between clean seismic signals and noisy seismic data, the U-Net network is employed as the G network of GAN. Incorporating a self-attention mechanism to strengthen the correlation between seismic data, with the goal of improving the network's reconstruction capacity on the continuity of seismic signals. The intelligent denoising of seismic data enabled by denoising network with self-attention GAN (DsGAN) enhances labor efficiency when compared to traditional approaches. When compared to the optimal state of current models such as denoising convolutional neural network (DnCNN), denoising network GAN (DnGAN), the peak signal-to-noise ratio (PSNR) is enhanced by 1.52 dB of the DsGAN model, according to experimental data from simulated and actual seismic data. Experiments show that the network has the ability to learn complex unknown noise, and has strong generalization and robustness.

Automated summary

This article presents a generative adversarial network architecture combined with the U-Net network that incorporates a self-attention mechanism to strengthen the correlation between seismic data, aiming to improve the network's reconstruction capacity on the continuity of seismic signals. The intelligent denoising of seismic data enabled by this network enhances labor efficiency compared to traditional approaches and shows strong generalization and robustness.