摘要
卷积神经网络(CNN)是近几年来最常用的用于有监督学习的深度学习算法之一.DnCNN是一种针对图像去噪问题提出的卷积神经网络.U-Net是一种U型结构的卷积神经网络,特点是融合了不同层次的特征,并在不同问题取得了不错的效果.近几年来多种CNN结构被广泛运用于地震数据的随机噪声压制中,但在地震去噪问题上选择CNN结构时有一定的盲目性.针对这一问题,本文对DnCNN和U-Net这两种常用于地震随机噪声压制的网络结构进行了多方面的对比.网络训练时添加不同方差的高斯噪声构成训练集,并分别对地震数据进行了噪声学习和信号学习两种学习方式.同时将CNN的去噪方法与传统的曲波变换(Curvelet)方法在合成数据和实际数据上进行随机噪声的压制效果对比.实验结果表明,在选择的数据上,CNN相对于曲波变换方法能够更好的去除地震数据中的随机噪声,并保护有效信号,同时U-Net比DnCNN能够在地震随机噪声压制中取得更好的效果.
Convolutional Neural Network(CNN) is one of the commonly-used deep learning algorithms for supervised learning in recent years. DnCNN is one type of CNN proposed for image denoising. U-Net is one type of CNN with a U-shaped structure, which is characterized by fusion of features at different levels and has achieved good results in different problems. In recent years, a variety of CNN structures have been widely used in the suppression of random noise in seismic data, but there is a certain degree of blindness when choosing CNN structures for seismic denoising. To solve this problem, this paper compares the performance of DnCNN and U-Net, which are two network structures commonly used for seismic random noise suppression. In the network training, Gaussian noise with different variances is added to form the training set, and two learning methods of noise learning and signal learning are carried out on seismic data. At the same time, the CNN denoising method is compared with the traditional Curvelet transform method to suppress the random noise on the synthetic data and the actual data. The experimental results show that the CNN structure can better remove the random noise in the seismic data and protect the effective signal than the Curvelet transform method on our data. Furthermore, U-Net can achieve better results in the suppression of seismic random noise than DnCNN.
作者
高好天
孙宁娜
孙可奕
李钟晓
秦宁
李振春
GAO HaoTian;SUN NingNa;SUN KeYi;LI ZhongXiao;QIN Ning;LI ZhenChun(School of Electronic Information,Qingdao University,Qingdao 266071,China;Shengli Oilfield Geophysical Research Institute of Sinopec,Dongying 257022,China;School of Geosciences,China University of Petroleum(East China),Qingdao 266580,China)
出处
《地球物理学进展》
CSCD
北大核心
2021年第6期2441-2453,共13页
Progress in Geophysics
基金
中国石油重大科技项目“塔里木盆地深层复杂高陡构造与碳酸盐岩储层地震速度建模及成像关键技术研究”(ZD2019-183-003)
国家自然科学基金项目“基于卷积神经网络的多次波自适应相减方法”(41804110)联合资助。
关键词
卷积神经网络
深度学习
地震随机噪声压制
Convolutional neural network
Deep learning
Seismic random noise suppression
作者简介
第一作者:高好天,男,1997年生,硕士研究生,主要从事深度学习理论及其在地震数据处理中的应用研究.E-mail:17806242857@163.com;通讯作者:李钟晓,男,1987年生,博士,青岛大学电子信息学院讲师,主要从事地震信号处理研究.E-mail:thulzx@163.com。
