论文详情
在地震数据处理中,随机噪声压制是提高地震数据信噪比的关键。针对目前卷积神经网络大多关注局部特征以及在特征提取方面的局限性,提出了一种结合全局上下文和注意力机制的深度卷积神经网络(global context and attention-based deep convolutional neural network,GC-ADNet),并用残差学习压制地震数据随机噪声的方法。其中,全局上下文模块(global context block,GCBlock)既关注局部信息,又能提取全局上下文信息;注意力模块(Attention Block)不仅强调关键特征,还能高效提取隐藏在复杂背景中的噪声信息。加入残差学习和批量规范化方法加快了网络的训练和收敛速度,使用扩张卷积扩大上下文信息并降低计算成本。将GC-ADNet应用于合成和实际地震数据处理,并与现有的去噪方法进行了比较。实验结果表明,GC-ADNet能够更有效压制随机噪声,并保留更多局部细节信息。
The attenuation of random noise is a critical step for improving the signal-to-noise ratio of seismic data during seismic data processing.Existing convolutional neural networks mostly focus on local features and have limited feature extraction capability.In this work,a deep convolutional neural network is proposed that combines global context and attention-based deep convolutional neural network (GC-ADNet).Residual learning is used to suppress the random noise in seismic data.The global context block boosts the denoising performance by not only making the network focus on local information but also extracting global context information.Simultaneously,the attention module is used to highlight key information and efficiently extract the noise from in the complex background.In addition,the residual learning and batch normalization methods are included to speed up the training and convergence,and dilated convolution is used to enlarge the contextual information and reduce the computational cost.The proposed method was tested on both synthetic and actual seismic data.Compared with state-of-the-art denoising methods,the proposed method,and particularly the GC-ADNet,demonstrated superior random noise suppression and local detail retention capabilities.
河北省自然科学基金项目(F2019202364)和河北省教育厅资助重点项目(ZD2020304)共同资助。