融合四种注意力机制的多尺度残差地震数据去噪网络  

Multiscale residual seismic data denoising network with fusion of quadruple attentions

作　　者：高磊[1,2,3] 樊星灿 乔昊炜 闵帆 杨梅[1,2,3] Gao Lei;Fan Xingcan;Qiao Haowei;Min Fan;Yang Mei(School of Computer Science and Software Engineering,Southwest Petroleum University,Chengdu 610500,China;Institute for Artificial Intelligence,Southwest Petroleum University,Chengdu 610500,China;Lab of Machine Learning,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机与软件学院,成都610500 [2]西南石油大学人工智能研究院,成都610500 [3]西南石油大学机器学习研究中心,成都610500

出　　处：《南京大学学报(自然科学版)》2024年第5期763-775,共13页Journal of Nanjing University(Natural Science)

基　　金：南充市-西南石油大学市校科技战略合作专项资金(23XNSYSX0084)

摘　　要：去除随机噪声是地震数据处理的一个重要步骤.基于卷积神经网络的很多方法只考虑单尺度特征,不能自适应地线性聚合地震数据特征,因而难以去除复杂的噪声并保护弱信号.提出融合四种注意力机制的多尺度卷积残差地震去噪网络(MARN),它主要包括三个部分:单尺度特征提取层、多尺度特征取层、特征恢复层.单尺度特征提取层使用单个相同卷积核提取全局特征.多尺度特征提取层包含多个残差多尺度注意力特征提取块(RMSAB),每块由多个多轴注意力多尺度特征融合块(MAFB)组成. MAFB包含三个结构:特征提取结构通过四种注意力机制提取局部细特征,特征融合结构融合四种注意力机制提取的特征,特征传输结构传递特征至特征恢复层.特征恢复层融合提取的单尺度和多尺度特征,获得去噪地震数据.实验结果表明,MARN不仅能更具针对性地去除随机噪声,还能更好地保留弱信号.Random noise denoising is an important step in seismic data processing.Many methods based on convolutional neural networks only consider single-scale features and cannot adaptively linearly aggregate seismic data features,resulting in difficulties in removing complex noise and protecting weak signals.In this paper,a multi-scale convolutional residual seismic denoising network fusing quadruple attention mechanisms(MARN)is proposed.It consists of three main parts:a single-scale feature extraction layer,a multi-scale feature fetching layer,and a feature recovery layer.The single-scale feature extraction layer uses a single identical convolutional kernel to extract global coarse features.The multiscale feature extraction layer contains multiple residual multi-scale attention feature extraction blocks(RMSAB),each consisting of multiple multiaxial attention multi-scale feature fusion blocks(MAFB).The MAFBs contain three structures:the feature extraction structure extracts the local fine features through the four attentional mechanisms,the feature fusion structure fuses the features extracted by the four attentional mechanisms,and the feature transfer structure delivers the features to the feature recovery layer.The feature recovery layer fuses the extracted single-scale and multi-scale features to obtain denoised seismic data.The experimental results show that MARN can not only remove random noise in a more targeted way,but also retain the weak signals better.

关 键 词：去除随机噪声 卷积神经网络 多注意力机制 多尺度特征 残差网络 

分 类 号：P631[天文地球—地质矿产勘探]