融合部分卷积和注意力机制对抗网络模型的地震数据重建

以生成对抗网络(GAN)为代表的深度学习模型在地震数据重建中取得了较好效果,但普通GAN网络的重建结果常存在模糊、假频等缺点。主要原因是：普通卷积模型在对缺失较大的数据进行卷积时,其卷积结果主要受缺失区域的影响,而有效区域的影响较小;且普通卷积模型属于局部操作,卷积结果主要受卷积核内数据的影响,而相距较远的数据对其影响甚微。为此,文中提出了融合部分卷积和注意力模型的改进GAN网络。首先,在卷积过程中引入一个比例因子r实现部分卷积,从而强化有效区域对卷积结果的影响;然后,利用注意力机制选择余弦相似度高的有效(背景)数据,以突破卷积距离的限制,使更多背景数据参与缺失区域的重建。数据处理结果表明,所提方法显著改善了重建数据中的模糊、假频等现象。

Seismic data reconstruction based on partial convolution and attentional mechanism adversarial network model

The deep learning model represented by generative adversal network (GAN) has achieved good results in seismic data reconstruction,but the reconstruction results of ordinary GAN networks have some shortcomings such as ambiguity and false frequency. The main reasons are as follows： during convolution,part of the convolution kernel slides into the missing region,so that the convolution result is affected by the region without data; Secondly,due to the limitation of the size of the convolution kernel,the convolution result is mainly affected by the data in the convolution kernel,and the effective information of distant locations cannot be obtained. In order to solve these two problems,this paper uses the part of the convolution thought and attention to improve GAN network model,Firstly,a scale factor r is introduced into the convolution process to achieve partial convolution,so as to shrink the convolution result and strengthen the influence of the effective region on the convolution result. Secondly,the attention mechanism is used to select the background data with high cosine similarity,which can break through the convolution distance limit and make more effective background data promote the reconstruction of the missing foreground data region,The data processing results show that the proposed method can improve the problems of ambiguity and false frequency in reconstructed data.