一种用于深度学习断层识别的高分辨率地震图像合成方法  

A Seismic Image Synthesis Method for Deep Learning Fault Recognition

作　　者：江穗华 段友祥[1] 孙歧峰[1] JIANG Suihua;DUAN Youxiang;SUN Qifeng(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,青岛266580

出　　处：《计算机与数字工程》2023年第9期2152-2158,共7页Computer & Digital Engineering

摘　　要：断层识别是研究断层构造基础,也是地震资料解释和油藏地质分析的关键环节,针对断层智能化识别研究中的训练数据不足、数据质量差、标注困难等问题,论文提出了一种基于Pix2pix框架的高分辨率断层地震图像合成方法。首先以断层位置信息建立地震记录的标签映射,作为输入生成对抗模型的约束信息;其次通过全局生成器与局部生成器实现由粗糙到精细的图像生成,同时通过一个影响因子去控制多级判别器对不同抽象级别特征的重视程度。最后通过与其他方法的对比实验,论文方法具有更好的视觉感知效果,以及更优的FID,PSNR与SSIM定量化评价指标,可作为监督学习的训练数据来源。Fault identification is not only the basis of fault structure research,but also the key link of seismic data interpretation and reservoir geological analysis.In order to solve the problems of insufficient training data,poor data quality and difficult labeling in intelligent fault identification,a high-resolution fault seismic image synthesis method based on Pix2pix framework is proposed in this paper.Firstly,the label mapping of seismic records is established with fault location information as input to generate constraint information of adversarial model.Secondly,a global generator and a local generator are used to generate images from rough to fine,and an impact factor is used to control how much attention multi-level discriminators attach to features at different abstraction levels.Finally,through the comparative experiment with other methods,the proposed method has better visual perception effect and better quantitative evaluation indexes of FID,PSNR and SSIM,which can be used as the training data source of supervised learning.

关 键 词：地震图像 地震记录标签映射 条件生成对抗网络 Pix2pixHD模型 图像生成 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]