CN110221346A - A kind of data noise drawing method based on the full convolutional neural networks of residual block - Google Patents

Abstract

The invention discloses a method for suppressing data noise based on a residual block full convolutional neural network. Both the training set and the test set for suppressing seismic noise by applying deep learning methods come from the same data set, which limits the generalization of the model. To solve the generalization problem, the design idea of the network structure is to fuse double residual blocks on the basis of the Unet network to enhance the network's ability to capture random noise. The present invention is based on an end-to-end encoding-decoding network structure, takes noisy seismic data as input, and extracts the essential features of random noise from multiple convolutional layers and residual blocks to form a code; The multilayer and residual blocks constitute the decoding, and the output of the network is the noise-suppressed seismic data. Compared with the current seismic data denoising method, due to the fusion of the double residual block, the extracted random noise features are digested and learned twice, and the essential features of the noise are learned more fully, so it has obvious generalization. Advantages, not only can effectively suppress random noise, but also protect effective signals.

Description

A Data Noise Suppression Method Based on Residual Block Fully Convolutional Neural Network

technical field

The invention relates to the technical field of data noise suppression, in particular to seismic random noise suppression.

Background technique

Traditional seismic data denoising methods include f-k domain filtering, f-x domain denoising, wavelet transform, curvelet transform and discrete cosine transform, etc. The above methods have been widely used in the field of seismic data denoising, but there are still problems such as insufficient denoising ability and destruction of effective signals.

In recent years, with the development of deep learning technology, researchers have proposed a method for denoising seismic data using deep learning technology. Different from traditional denoising methods, deep learning belongs to the category of statistical learning. Statistical learning can learn the essential characteristics of effective signals and noises based on noise samples, and fit a model that can classify effective signals and noises. Due to this advantage of statistical learning, the paper "Wang Yuqing, Lu Wenkai, Liu Jinlin, etc. Seismic Random Noise Suppression Based on Data Augmentation and CNN [J]. Acta Geophysics, 2019, 62(1): 421-433." The method based on the Unet convolutional neural network suppresses the random noise of seismic data and has achieved certain results, but the problem is that in the post-stack seismic data test experiment, the training set and the test set come from the same data set, and the generalization of the model is limited. limit. The literature "Ma J.Deep Learning for Attenuating Random and Coherence Noise Simultaneously[C]//80th EAGEConference and Exhibition 2018.2018." proposes that the method of denoising seismic data based on DnCNN can suppress Gaussian noise, but the data sets used for training and testing are both For synthetic data, good noise suppression effects may not necessarily be achieved in real seismic data sets. The document "Liu J, Lu W, Zhang P.Random Noise Attenuation Using Convolutional Neural Networks[C]//80th EAGE Conference and Exhibition2018.2018." proposes a seismic random noise suppression method based on the Unet convolutional neural network, but the data sets are all from Synthetic data, generalization is limited in real seismic data.

Contents of the invention

The invention discloses a data noise suppression method based on a residual block full convolutional neural network, which is characterized in that it comprises the following steps:

Step 1: making a training set and a test set of data, characterized in that noise-containing data and noise-free data of different noise levels are used as a training set, and another piece of data with a different distribution from the training set data is selected as a test set;

Step 2: Design an end-to-end network structure for encoding and decoding, which is characterized in that the encoding process consists of 5 sets of double residual blocks of different scales, and each set of residual blocks consists of 5 convolutional layers and 1 pooling Layer composition; the decoding process is symmetrical to the encoding process, consisting of 4 groups of double residual blocks of different scales, each group of residual blocks is composed of 1 deconvolution layer and 5 convolution layers, and fused with the extracted data from the corresponding encoding stage noise characteristics;

Step 3: Train the network and save the network model;

Step 4: Adjust parameters and select the final ideal model;

Step 5: Use the ideal denoising model obtained by training to suppress the noise of the data, and the output is the noise-suppressed data.

In the step 2, the scale sizes of the 5 groups of double residual blocks in the coding part are 256×256, 128×128, 64×64, 32×32, 16×16, and the 4 groups of double residual blocks in the decoding part The block sizes are 32×32, 64×64, 128×128, 256×256, respectively.

Compared with the current seismic data denoising method, the present invention has the following advantages:

(1) RUnet (ie: residual block fully convolutional neural network) can not only effectively suppress random noise, but also protect effective signals;

(2) The innovation of the present invention is that compared with the Unet convolutional neural network, due to the fusion of the residual block, the extracted random noise features are digested and learned twice, and the essential features of the noise are learned more fully, so It has obvious advantages in generalization.

Description of drawings

Fig. 1 is the flow chart that the present invention suppresses random noise of seismic data, will contain the noisy seismic data of different noise levels as the training set of training model, build RUnet network, training model, select according to peak signal-to-noise ratio and signal-to-noise ratio index Ideal model, the test set is input into the ideal model, and the output is the data after seismic random noise suppression;

Fig. 2 is the structural diagram of the dual residual block fused in the network structure of the present invention, x _i-2 represents the input of the residual block; f( _xi-2 ) represents the noise feature extracted by the two-layer convolutional layer; x _i represents The output feature value of the residual block, that is, x _i =f( _xi-2 )+ _xi-2 ; CONV denotes a convolutional layer.

Fig. 3 is the structural diagram of the original Unet convolutional neural network;

Fig. 4 is a structure diagram of the RUnet convolutional neural network of the present invention. On the basic structure of the original Unet convolutional neural network, the double residual block is fused, and the noise feature is subjected to secondary feature learning. The dotted line in the figure is the double residual Bad block;

Fig. 5 is the denoising example of the specific implementation of the present invention: a is the original data of the training set; b is the noise-added data of the training set; c is the denoising result of RUnet in the same data set; d is the noisy data of the test set; e is RUnet Test results in different data sets; f is the noise removed by RUnet.

Detailed ways

In order to effectively remove random noise in seismic data, this paper proposes a RUnet convolutional neural network denoising model, including the following steps.

Step 1: The noise-containing seismic data with different noise levels and the preprocessed 3D post-stack seismic data are used as the training set. The specific steps are as follows:

(1) Select 256 traces from the Parihaka post-stack 3D seismic data volume, and the sampling points are 256 seismic data slices;

(2) Add 20%, 25%, and 30% Gaussian random noise to the seismic data respectively, and use the corresponding preprocessed seismic data as a training set, where the noise-added seismic data is used as input, and the preprocessed seismic data is used as label, the sample size is 900;

Step 2: The network as a whole includes an encoding process and a decoding process. The encoding process consists of 5 groups of residual blocks, each group of residual blocks consists of 5 convolutional layers and 1 pooling layer, and encodes [256×256]-dimensional input data into [16×16]-dimensional feature information, The convolution kernel size is set to 3×3, and the stride is set to 1. After each residual block operation, the size of the feature map is compressed to 1/2 of the previous operation, and accordingly the number of channels of the feature map is twice that of the previous residual block operation, ensuring that the feature information is not lost. The feature decoding process consists of 4 groups of residual blocks, each residual block consists of 1 deconvolution layer and 5 convolution layers, and the [16×16] dimension feature information generated by the encoding process is upsampled to [256 ×256]-dimensional output data. Symmetrical to the encoding process, after each residual operation, the size of the feature map is upsampled to twice that of the previous residual operation, and the number of channels of the feature map becomes 1/2 of the previous residual operation. The feature map of the corresponding position of the encoding part is added to the feature map of the decoding part in order to fuse feature information of different scales. The final output is completed by a convolutional layer with a convolution kernel size of 1×1 and a step size of 1 and an activation function tanh, which is similar to a fully connected layer;

Step 3: Input the training set obtained in step 1 into the network model built in step 2 through the queue, use error backpropagation, and use the mean square error loss function to measure the distance between the real value of the network output and the label value, using The stochastic gradient descent algorithm is used to adjust the weights between neurons to minimize the loss function, and the network denoising effect is judged by quantitative peak signal-to-noise ratio, signal-to-noise ratio and qualitative visual experience. After obtaining the optimal denoising effect, save the network The parameters of the model;

The formula for mean square difference is:

where y is the true value of the network output, is the corresponding label value, the smaller the mean square error, the closer the real value is to the predicted value, and the better the learning effect of the network on the training set;

Step 4: Select several seismic data samples in the same 3D data volume as the test set, input them into the network model obtained in step 3, and judge the denoising effect of the network by quantitative peak signal-to-noise ratio, signal-to-noise ratio and qualitative visual experience , if the requirements are not met, return to step 3 to continue training or adjust the parameters to retrain the network until the requirements are met, and save the final ideal network model;

Step 5: Use the saved network model to remove the noise of the Kerry seismic data volume, and the output is the denoised seismic data.

Claims (2)

1. The present invention discloses a method for suppressing data noise based on a residual block full convolutional neural network, characterized in that it comprises the following steps: Step 1: making a training set and a test set of data, characterized in that noise-containing data and noise-free data of different noise levels are used as a training set, and another piece of data with a different distribution from the training set data is selected as a test set; Step 2: Design an end-to-end network structure for encoding and decoding, which is characterized in that the encoding process consists of 5 sets of double residual blocks of different scales, and each set of residual blocks consists of 5 convolutional layers and 1 pooling Layer composition; the decoding process is symmetrical to the encoding process, consisting of 4 groups of double residual blocks of different scales, each group of residual blocks is composed of 1 deconvolution layer and 5 convolution layers, and fused with the extracted data from the corresponding encoding stage noise characteristics; Step 3: Train the network and save the network model; Step 4: Adjust parameters and select the final ideal model; Step 5: Use the ideal denoising model obtained by training to suppress the noise of the data, and the output is the noise-suppressed data. 2. In claim 1, the scale sizes of the 5 groups of double residual blocks of the encoding part in the step 2 are respectively 256 × 256, 128 × 128, 64 × 64, 32 × 32, 16 × 16, the decoding part The scale sizes of the four groups of double residual blocks are 32×32, 64×64, 128×128, and 256×256, respectively.