Shearlet域深度残差CNN用于沙漠地震信号去噪

由于沙漠地震信号中含有较强的随机噪声，从而给沙漠地震数据的处理和解释带来了很大的困难。针对上述问题，提出了一种基于Shearlet 变换的深度残差卷积神经网络( ST-CNN: Deep Ｒesidual Convolutional NeuralNetwork for Shearlet Transform) 模型，实现沙漠地震信号的随机噪声压制。在训练阶段，将沙漠地震信号经Shearlet 分解后的系数作为输入，将随机噪声经Shearlet 分解后的系数作为标签，通过卷积神经网络( CNN: Convolutional Neural Network) 学习输入和标签之间的映射关系; 在测试阶段，利用此映射关系即可从沙漠地震信号系数中预测出噪声系数，并间接地获得有效信号系数，最后通过Shearlet 反变换获得有效信号。通过与传统的Shearlet 硬阈值去噪算法对比，发现该算法可把沙漠地震信号的信噪比从－ 4． 48 dB 提高到14． 15 dB，具有更好的去噪效果。

Shearlet Domain Deep Ｒesidual CNN for Removing Noise from Desert Seismic Signals

Desert seismic signals contain strong random noise，which brings great trouble to the processing and interpretation of desert seismic signals． In order to solve this technical problem，Deep Ｒesidual Convolutional Neural Network for Shearlet Transform model is proposed for the implementation of the desert seismic signal random noise suppression． In training phase，the Shearlet coefficients of desert seismic data are taken as inputs，and the Shearlet coefficients of random noise are taken as labels． Through network training，the mapping relationship between them could be learned by a deep CNN ( Convolutional Neural Network) ． In test phase，the coefficients of random noise can be predicted from the coefficients of desert seismic data by the mapping relationship，and thereafter the effective signals coefficients is obtained indirectly． Finally the effective signals can be reconstructed by inverse Shearlet transform． By comparing with the traditional Shearlet hard threshold denoising algorithm，the proposed algorithm has improved the SNＲ of the desert seismic signals from － 4． 48 dB to 14． 15 dB and has better denoising performance．