可逆网络的地震数据分辨率增强与去噪  

Resolution enhancement and denoising of seismic data based on invertible network

作　　者：闵帆[1] 王林蓉 MIN Fan;WANG Linrong(Lab of Machine Learning,School of Computer Science and Software Engineering,Chengdu,Sichuan 610500,China)

机构地区：[1]西南石油大学计算机与软件学院机器学习实验室,四川成都610500

出　　处：《闽南师范大学学报（自然科学版）》2024年第2期20-33,共14页Journal of Minnan Normal University：Natural Science

基　　金：南充市校办科技合作项目(23XNSYSX0062,23XNSYSX0013)。

摘　　要：基于卷积神经网络的地震超分辨率方法表现良好,但在下采样过程中会丢失高频信息,无法解决一对多不适定性问题.为此,提出了一种基于可逆网络的地震超分辨率和去噪算法(SRInvNet)缓解该问题.在前向过程中,SRInvNet被训练为将有噪声的低分辨率地震图像转换为干净的降尺度图像和噪声-高频信息混合的潜在表示;丢弃潜在的表示来完全去除噪声,并从高斯分布中随机采样一个新的变量来恢复高频信息.在后向过程中,该变量和降尺度图像作为输入来恢复干净的高分辨率图像.结果表明,SRInvNet的性能和参数量均优于最新的超分辨率算法CAUC和SeisGAN.The seismic super-resolution method based on convolutional neural networks performs well but high-frequency information will be lost during downsampling,which can not address the problem of one-to-many ill posedness.A seismic super-resolution and denoising algorithm based on invertible network(SRInvNet)is proposed to alleviate this problem.In the forward process,SRInvNet is trained to transform noisy low-resolution seismic images into clean downscaled images and a latent representation of noise high-frequency information mixture.The latent representation is discarded to completely remove the noise,and a new variable is randomly sampled from the Gaussian distribution to recover high-frequency information.In the backward pass,the variable and the downscaled image are used as inputs to restore a clean high-resolution image.The results indicate that SRInvNet outperforms the latest super-resolution algorithms CAUC and SeisGAN in terms of quality and number of parameters.

关 键 词：地震数据 可逆网络 地震分辨率 地震去噪 

分 类 号：P631[天文地球—地质矿产勘探]