A method for strong noise suppression based on DC-UNet
-
摘要: 在成熟工业区采集地震数据的过程中,由于生产设备的持续运转,使得采集到的地震数据含有大量振幅很强的局部强噪声,难以用常规的去噪方法压制。将U-Net网络与空洞卷积结合,建立了适用于局部强噪声压制的空洞卷积DC-UNet网络。DC-UNet网络前端的循环空洞卷积块使用循环扩张的空洞卷积核提取不同尺度的强噪声特征信息,并且扩大了感受野;网络后端使用编码器提取强噪声特征,编码器还原强噪声细节特征。DC-UNet网络实现从含噪数据到噪声的非线性映射,通过从含噪数据减去学习到的强噪声,达到压制强噪声的目的。在GPU环境使用Pytorch框架进行训练,合成数据和实际数据实验结果表明,相较于DnCNN、U-Net、PCA-UNet网络,DC-UNet网络能更好地压制局部强噪声并且提高了信噪比。Abstract: Seismic data acquired from mature industrial areas frequently contain a large amount of local strong noise with high amplitude due to the continuous operation of production equipment.However,such local strong noise can be hardly suppressed using conventional denoising methods.This study integrated dilated convolution(DC) and U-Net into a DC-UNet network for suppressing local strong noise.For the circular DC blocks at the front end of the DC-Unet network,a circularly expanded DC kernel was used to extract the features of strong noise at different scales,with the receptive field being expanded.Meanwhile,an encoder was used at the back end of the network to extract the features of strong noise and restore the details of strong noise.Subsequently,the DC-UNet network was employed to perform a nonlinear mapping from noisy data to noise.On this basis,strong noise was suppressed by subtracting the learned strong noise from the noisy data.As indicated by the experimental results of synthetic and real data obtained from the training using the PyTorch framework in the GPU environment,the DC-UNet network can effectively suppress the local strong noise and improve the signal-to-noise ratio compared with DnCNN,U-Net,and PCA-UNet networks.
-
Key words:
- local strong noise /
- dilated convolution /
- convolutional neural network
-
-
计量
- 文章访问数: 73
- PDF下载数: 7
- 施引文献: 0
下载: