基于卷积神经网络与生成对抗网络的地震噪声压制
Seismic Noise Suppression Based on Convolutional Neural Networks and Generative Adversarial Networks

目前，地震数据的深度学习去噪方法大多基于卷积神经网络，但此类方法受到卷积核局部操作的限制，缺乏对地震数据全局特征的有效捕捉，导致去噪效果不尽如人意。针对这一问题，本文提出了一种结合Swin-Transformer和生成对抗网络的去噪方法。该方法采用Swin-Transformer作为生成网络对地震数据进行去噪，Swin-Transformer的自注意力机制可进行全局操作，有效提取地震数据的全局特征，且与生成对抗网络的局部操作互为补充，提升了特征提取的能力，并有效避免了过度平滑引发的同相轴假象。通过将该方法应用于地震数据去噪，并与现有方法进行对比，实验结果表明，该方法在特征提取能力上具有明显优势，能够在有效抑制随机噪声的同时，恢复和保留更多的细节信息，从而提高了地震信号的信噪比。
Currently, most deep learning-based seismic data denoising methods rely on convolutional neural networks (CNNs). However, such methods are limited by the local operations of convolution kernels and lack effective capture of global features in seismic data, leading to unsatisfactory denoising results. To address this issue, this paper proposes a denoising method combining Swin-Transformer and Generative Adversarial Networks. The Swin-Transformer is used as the generator network to denoise seismic data. Its self-attention mechanism enables global operations, effectively extracting global features of seismic data, and complements the local operations of the GAN, enhancing feature extraction capabilities. This approach effectively avoids the aliasing of in-phase axes caused by excessive smoothing. By applying this method to seismic data denoising and comparing it with existing methods, experimental results show that this method has significant advantages in feature extraction. It can effectively suppress random noise while recovering and preserving more detailed information, thereby improving the signal-to-noise ratio (SNR) of seismic signals.