CN113484913A - Seismic data denoising method with multi-granularity feature fusion convolution neural network - Google Patents

Seismic data denoising method with multi-granularity feature fusion convolution neural network Download PDF

Info

Publication number
CN113484913A
CN113484913A CN202110939618.4A CN202110939618A CN113484913A CN 113484913 A CN113484913 A CN 113484913A CN 202110939618 A CN202110939618 A CN 202110939618A CN 113484913 A CN113484913 A CN 113484913A
Authority
CN
China
Prior art keywords
seismic data
denoising
neural network
channel
layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110939618.4A
Other languages
Chinese (zh)
Other versions
CN113484913B (en
Inventor
冯俊
李晓琴
乐芮含
刘斌
刘序志
雷竞雄
刘夕
周春花
李鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chengdu Univeristy of Technology
Original Assignee
Chengdu Univeristy of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chengdu Univeristy of Technology filed Critical Chengdu Univeristy of Technology
Priority to CN202110939618.4A priority Critical patent/CN113484913B/en
Publication of CN113484913A publication Critical patent/CN113484913A/en
Application granted granted Critical
Publication of CN113484913B publication Critical patent/CN113484913B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/36—Effecting static or dynamic corrections on records, e.g. correcting spread; Correlating seismic signals; Eliminating effects of unwanted energy
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00—Details of seismic processing or analysis
    • G01V2210/30—Noise handling
    • G01V2210/32—Noise reduction
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Acoustics & Sound (AREA)
  • Environmental & Geological Engineering (AREA)
  • Geology (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Geophysics (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

The invention discloses a seismic data denoising method of a multi-granularity characteristic fusion convolution neural network, which comprises the steps of adding Gaussian white noise with the same standard deviation into clean seismic data to obtain a data set to be trained, and inputting the data set into the neural network; extracting approximate characteristics of seismic data of each channel from different granularities through a multi-granularity characteristic fusion block; converting multi-channel seismic data containing approximate features into a predictable single-channel vector; obtaining denoised seismic data and network parameters by calculating the minimum value of the loss function, and obtaining a denoising model of Gaussian white noise of the current standard deviation; adding Gaussian white noise with different standard deviations into clean seismic data, repeating the steps to obtain denoising models of the Gaussian white noise with different standard deviations, and completing model training; and selecting a denoising model for denoising the seismic data to be denoised according to the standard deviation of the seismic data. The invention can improve the denoising effect, well reserve the texture information and reduce the parameter adjusting time.

Description

Seismic data denoising method with multi-granularity feature fusion convolution neural network
Technical Field
The invention relates to the field of image processing, in particular to a seismic data denoising method of a multi-granularity feature fusion convolution neural network.
Background
Seismic data play an important role in the exploration of subsurface structures, such as oil and gas exploration. However, the process of acquiring seismic data is subject to various disturbances, resulting in the acquired seismic data being contaminated by random noise. The noise contained in the seismic data degrades the quality of the seismic data, which makes subsequent interpretation and analysis of the seismic data difficult. Meanwhile, seismic data denoising is a hot problem in seismic data processing, and has been widely studied in recent years. So far, common seismic denoising methods can be classified into the following four categories: prediction-based methods, transform-based methods, low-rank approximation-based methods, neural network-based methods. Prediction-based methods assume that the useful signal is predictable and that random noise is unpredictable, and therefore separation of the signal and noise can be achieved by constructing separation filters, such as f-x prediction filtering, t-x prediction filtering, non-stationary prediction filtering, and the like. Transform-based methods use the distinguishable differences between signal and noise in the transform domain for denoising, such as wavelet transforms, curvelet transforms, shear wave transforms, and the like. The low rank approximation based approach assumes that the noise-free seismic data is low rank. Since adding random noise to clean seismic data will increase the rank of the matrix, removing random noise from the seismic can be considered a low rank matrix approximation problem, such as: high Order Singular Value Decomposition (HOSVD), weighted kernel norm minimization method (WNNM), etc. Although these conventional seismic data denoising methods can improve the quality of seismic data to some extent, they often require that the seismic data contain specific characteristics. Moreover, the denoising performance of the traditional method depends on parameter adjustment, and when the data set is large, a lot of time is spent on parameter adjustment.
Disclosure of Invention
Aiming at the defects in the prior art, the seismic data denoising method with the multi-granularity feature fusion convolutional neural network provided by the invention solves the problems of long parameter adjusting time and poor denoising effect in the prior art.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that:
the seismic data denoising method for the multi-granularity feature fusion convolution neural network is provided, and comprises the following steps:
s1, adding Gaussian white noise with the same standard deviation into clean seismic data to obtain a data set to be trained, namely single-channel noise-containing seismic data;
s2, inputting the single-channel noisy seismic data into a neural network, and converting the single-channel noisy seismic data into multi-channel seismic data through the head of the neural network;
s3, extracting approximate characteristics of the seismic data of each channel from different granularities through a multi-granularity characteristic fusion block, namely the middle part of a neural network;
s4, converting the multi-channel seismic data containing the approximate characteristics into predictable single-channel seismic data at the tail part of the neural network;
s5, constructing a loss function through clean seismic data and predictable single-channel seismic data;
s6, calculating the minimum value of the loss function through a gradient descent method to obtain denoised seismic data and network parameters, and obtaining a denoising model of Gaussian white noise of the current standard deviation;
s7, sequentially adding the Gaussian white noises with different standard deviations into the clean seismic data, repeating the steps S1 to S6 to obtain denoising models of the Gaussian white noises with different standard deviations, and finishing model training;
s8, estimating the standard deviation of the seismic data to be denoised, and selecting the trained denoising model for denoising according to the standard deviation.
Further, the specific method of step S1 is:
according to the formula:
Y＝X+N
N～Ν(μ,σ2)
obtaining single-channel noise-containing seismic data Y; where X is clean seismic data, N is random noise, N (μ, σ)2) Representing a standard normal distribution obeying a mean value of μ and a standard deviation of σ.
Further, the specific method of step S2 includes:
according to the formula:
Figure BDA0003214286050000031
obtaining multi-channel seismic data
Figure BDA0003214286050000032
Wherein the multi-channel seismic data has 16 channels, Relu (a) is an activation function, Cov (a) is a convolution layer, w is the width of the seismic data, h is the height of the seismic data, Y is1×w×hSeismic data including the number, width and height of channels are obtained by processing single-channel noise-containing seismic data Y.
Further, the head of the neural network in step S2 includes the convolutional layer and the Relu activation function.
Further, the specific method of step S3 includes:
according to the formula:
Figure BDA0003214286050000033
obtaining multi-channel seismic data H containing approximate characteristics16×w×h(ii) a Where body (-) is the processing of input data for a network containing three multi-granular feature fusion blocks.
Further, the middle part of the neural network in step S3 includes three multi-granularity fusion blocks connected in sequence; the multi-granularity fusion block comprises three parallel multi-granularity feature extraction parts and a feature fusion part which are connected in sequence; the multi-granularity feature extraction part comprises a first convolution layer and a first batch normalization layer, a second convolution layer and a second batch normalization layer, and a third convolution layer and a third batch normalization layer and a first activation function layer which are sequentially connected; the feature fusion part comprises a fusion layer, a fourth convolution layer, a fourth batch processing normalization layer, a second activation function layer, a fifth convolution layer, a fifth batch processing normalization layer and a third activation function layer which are sequentially connected.
Further, the specific method of step S4 includes:
according to the formula:
Figure BDA0003214286050000041
obtaining predictable single-channel seismic data
Figure BDA0003214286050000042
Tail () is the data processing process at the tail of the neural network; wherein the tail of the neural network contains the Cov (-) convolutional layer and the Relu (-) activation function.
Further, the specific method for constructing the loss function in step S5 is as follows:
according to the formula:
Figure BDA0003214286050000043
obtaining a loss function loss; where θ is a network parameter, XijIs clean seismic data having a width i and a height j,
Figure BDA0003214286050000044
the width is a predictable single-channel seismic data with i and the height is j, w is the width of the seismic data, and h is the height of the seismic data.
Further, the specific method for calculating the minimum value of the loss function through the gradient descent method to obtain the denoised seismic data and the network parameters in the step S6 is as follows:
s6-1, randomly initializing a network parameter theta;
s6-2, according to the formula:
Figure BDA0003214286050000045
obtaining a derivative of the loss function with respect to the network parameter;
s6-3, according to the formula:
Figure BDA0003214286050000046
updating the network parameters; wherein n is a positive integer and α is a learning rate;
s6-4, judging whether the two changes before and after the loss function are smaller than a preset threshold value, if so, judging that the loss function is converged to obtain a minimum value, and obtaining a corresponding network parameter theta and predictable single-channel seismic data
Figure BDA0003214286050000051
Namely de-noised seismic data; otherwise, the step S6-3 is returned to continue the iterative updating.
The invention has the beneficial effects that:
1. the invention provides a multi-granularity feature fusion convolutional neural network denoising method based on convolutional neural network denoising, which is used for denoising seismic data; the method utilizes convolution kernels with different sizes to extract the characteristics of the seismic data from different granularities, then fuses the extracted characteristics, and makes full use of the local self-similarity of the seismic data to perform denoising, thereby improving the denoising effect.
2. Compared with the traditional denoising method, the method disclosed by the invention has the advantages that the parameter adjustment is more intelligent in the experiment, the parameter adjustment is not required to be carried out on a certain seismic data, and a large amount of time is not required to be spent on parameter adjustment when the data volume is large.
3. The invention improves the existing denoising method based on the convolutional neural network, the existing method based on the convolutional neural network usually uses a convolutional kernel with fixed size in a specific layer, which limits that the convolutional kernels with different sizes can not extract features from different granularities in the same region, and the method can utilize the convolutional kernels with different sizes to extract the features with different granularities in parallel.
4. The method can effectively remove Gaussian noise in the seismic image and well keep texture detail information; the method can not only denoise seismic data, but also be applicable to general natural images.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a network architecture diagram of a multi-granularity feature fusion block of the present invention;
FIG. 3 is a graph of synthetic seismic data simulated by the present invention;
FIG. 4 is a graph of simulated noisy synthetic seismic data according to the present invention;
FIG. 5 is a simulated seismic data section de-noising map of the BM3D method;
FIG. 6 is a simulated seismic data profile de-noising map of the NLH method;
FIG. 7 is a simulated seismic data profile de-noising map of the WNNM method;
FIG. 8 is a cross-sectional de-noising diagram of a simulated seismic data by the DNCNN method;
FIG. 9 is a cross-sectional de-noising diagram of a simulated seismic data for the ADNET method;
FIG. 10 is a simulated seismic data profile de-noising map of the method of the present invention;
FIG. 11 is an original single-channel signal;
FIG. 12 is a noisy single-channel signal;
FIG. 13 is a single channel signal after BM3D simulation seismic data denoising;
FIG. 14 is a single-channel signal after NLH simulation seismic data denoising;
FIG. 15 is a single channel signal after WNNM simulated seismic data denoising;
FIG. 16 is a single-channel signal after denoising of DNCNN simulated seismic data;
FIG. 17 is a single channel signal after ADNET simulation seismic data denoising;
FIG. 18 is a single channel signal after denoising of seismic data according to the present invention;
FIG. 19 is actual seismic data used with the present invention;
FIG. 20 shows the denoising result of BM3D actual seismic data;
FIG. 21 is the residual of BM 3D;
FIG. 22 shows the denoising result of the NLH actual seismic data;
FIG. 23 is the residual of NLH;
FIG. 24 shows the denoising result of WNNM actual seismic data;
FIG. 25 is the residual of WNNM;
FIG. 26 shows the de-noising result of DNCNN actual seismic data;
FIG. 27 shows the residual DNCNN;
FIG. 28 shows the de-noising result of ADNET actual seismic data;
fig. 29 is the residue of ADNET;
FIG. 30 is a graph of the de-noising result of the actual seismic data of the present invention;
FIG. 31 shows the residual error of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
As shown in fig. 1, the seismic data denoising method of the multi-granularity feature fusion convolutional neural network includes the following steps:
s1, adding Gaussian white noise with the same standard deviation into clean seismic data to obtain a data set to be trained, namely single-channel noise-containing seismic data;
s2, inputting the single-channel noisy seismic data into a neural network, and converting the single-channel noisy seismic data into multi-channel seismic data through the head of the neural network;
s3, extracting approximate characteristics of the seismic data of each channel from different granularities through a multi-granularity characteristic fusion block, namely the middle part of a neural network;
s4, converting the multi-channel seismic data containing the approximate characteristics into predictable single-channel seismic data at the tail part of the neural network;
s5, constructing a loss function through clean seismic data and predictable single-channel seismic data;
s6, calculating the minimum value of the loss function through a gradient descent method to obtain denoised seismic data and network parameters, and obtaining a denoising model of Gaussian white noise of the current standard deviation;
s7, sequentially adding the Gaussian white noises with different standard deviations into the clean seismic data, repeating the steps S1 to S6 to obtain denoising models of the Gaussian white noises with different standard deviations, and finishing model training;
s8, estimating the standard deviation of the seismic data to be denoised, and selecting the trained denoising model for denoising according to the standard deviation.
The specific method of step S1 is:
according to the formula:
Y＝X+N
N～Ν(μ,σ2)
obtaining single-channel noise-containing seismic data Y; where X is clean seismic data, N is random noise, N (μ, σ)2) Representing a standard normal distribution obeying a mean value of μ and a standard deviation of σ.
The invention adopts the Rake wavelets to synthesize the clean seismic data, and the formula of the signal x (t) of the synthesized clean seismic data is as follows:
Figure BDA0003214286050000081
wherein A is amplitude, fmIs the dominant frequency, t0Is the initial time, t is the time, pi is 180 degrees, and e is the natural logarithm.
The invention adopts the main frequency range of 10Hz-65Hz, the amplitude range of 0.1-1, the speed range of 300m/s-8000m/s and the interval of 0.02s, and randomly selects parameters to generate the synthetic seismic data. And finally obtaining 85 pieces of synthesized clean seismic data, wherein each seismic data signal consists of 256 channels and 256 sampling points. The pixel values of the synthesized data are normalized to 0-255, as in the conventional denoising method. Of all the generated synthetic data, 2 of them were randomly selected as test data, and the other 83 were used as training data. In order to train the model, white gaussian noise with the same standard deviation is added to the clean training data, and a training set of the noise level is obtained.
The specific method of step S2 includes:
according to the formula:
Figure BDA0003214286050000082
obtaining multi-channel seismic data
Figure BDA0003214286050000083
Wherein the multi-channel seismic data has 16 channels, Relu (a) is an activation function, Cov (a) is a convolution layer, w is the width of the seismic data, h is the height of the seismic data, Y is1×w×hSeismic data including the number, width and height of channels are obtained by processing single-channel noise-containing seismic data Y.
The head of the neural network in step S2 includes a convolutional layer and a Relu activation function; the input channel of the header is 1, the size of the convolution layer kernel is 3, and the number of kernels is 16.
The specific method of step S3 includes:
according to the formula:
Figure BDA0003214286050000091
obtaining multi-channel seismic data H containing approximate characteristics16×w×h(ii) a Where body (-) is the processing of input data for a network containing three multi-granular feature fusion blocks.
As shown in fig. 2, the Multi-Feature is a Multi-granularity Feature, Fusion is Fusion, Cov is a convolutional layer, BN is a batch normalization layer, Relu is an activation layer, and the middle of the neural network in step S3 includes three Multi-granularity Fusion blocks connected in sequence; the multi-granularity fusion block comprises three parallel multi-granularity feature extraction parts and a feature fusion part which are connected in sequence; the multi-granularity feature extraction part comprises a first convolution layer and a first batch normalization layer, a second convolution layer and a second batch normalization layer, and a third convolution layer and a third batch normalization layer and a first activation function layer which are sequentially connected; the feature fusion part comprises a fusion layer, a fourth convolution layer, a fourth batch processing normalization layer, a second activation function layer, a fifth convolution layer, a fifth batch processing normalization layer and a third activation function layer which are sequentially connected; the sizes of all the convolutional layer kernels in the three parallel multi-granularity feature extraction parts are respectively 3, 5 and 7, and the number of the kernels is 16.
The specific method of step S4 includes:
according to the formula:
Figure BDA0003214286050000092
obtaining predictable single-channel seismic data
Figure BDA0003214286050000093
Tail () is the data processing process at the tail of the neural network; wherein the tail of the neural network contains the Cov (-) convolutional layer and the Relu (-) activation function.
The tail of the neural network in step S4 includes the convolutional layer and the Relu activation function.
The specific method for constructing the loss function in step S5 is as follows:
according to the formula:
Figure BDA0003214286050000101
obtaining a loss function loss; where theta is a parameter of the network,Xijis clean seismic data having a width i and a height j,
Figure BDA0003214286050000102
is a predictable single-channel seismic data with width i and height j, which is abbreviated as
Figure BDA0003214286050000103
w is the width of the seismic data and h is the height of the seismic data.
The specific method for calculating the minimum value of the loss function through the gradient descent method to obtain the denoised seismic data and the network parameters in the step S6 is as follows:
s6-1, randomly initializing a network parameter theta;
s6-2, according to the formula:
Figure BDA0003214286050000104
obtaining a derivative of the loss function with respect to the network parameter;
s6-3, according to the formula:
Figure BDA0003214286050000105
updating the network parameters; wherein n is a positive integer and α is a learning rate;
s6-4, judging whether the two changes before and after the loss function are smaller than a preset threshold value, if so, judging that the loss function is converged to obtain a minimum value, and obtaining a corresponding network parameter theta and predictable single-channel noise-containing seismic data
Figure BDA0003214286050000106
I.e. denoised seismic data, abbreviated as
Figure BDA0003214286050000107
Otherwise, the step S6-3 is returned to continue the iterative updating.
In order to verify the denoising effect of the invention, the seismic data denoising Method (MFFCNN) of the multi-granularity feature fusion convolutional neural network of the invention is compared with advanced denoising methods such as block matching and three-dimensional filtering (BM3D), non-local blind denoising (NLH), weighted kernel norm minimization (WNNM), deep convolutional neural network (DnCNN) and attention-directed convolutional neural network (ADNET). In all methods, DnCNN, ADNET, MFFCNN are neural network based models that require training, while others do not. In the experiments, all comparison methods used default parameters of the original code.
To evaluate the performance of the proposed MFFCNN and comparison methods, the present invention uses different metrics on the synthetic and actual seismic data sets. For synthetic seismic data, the method calculates the Structural Similarity (SSIM), the peak signal-to-noise ratio (PSNR) and the signal-to-noise ratio (SNR), extracts a single-channel signal for display, and draws an amplitude spectrum to further evaluate the denoising effect of the method. The SSIM, PSNR and SNR calculation formulas are as follows:
Figure BDA0003214286050000111
c1＝(Lk1)2
c2＝(Lk2)2
wherein X is clean seismic data,
Figure BDA0003214286050000112
for de-noised seismic data, uXAnd
Figure BDA0003214286050000113
are respectively X and
Figure BDA0003214286050000114
is determined by the average value of (a) of (b),
Figure BDA0003214286050000115
is X and
Figure BDA0003214286050000116
covariance of (a)XAnd
Figure BDA0003214286050000117
are respectively X and
Figure BDA0003214286050000118
variance of (1), L is 255, k1＝0.01，k2＝0.03。
Figure BDA0003214286050000119
Wherein XijIs clean seismic data having a width i and a height j,
Figure BDA00032142860500001110
for the denoised single-channel seismic data with the width i and the height j, max (·) is the maximum gray value, generally 255, and since the data used in the present invention includes 256 channels and 256 sampling points, M is 256, and lg is a base-10 logarithm.
Figure BDA00032142860500001111
Wherein xiThe signal of the ith clean seismic data,
Figure BDA0003214286050000121
is xiAnd f, denoising the seismic signal, wherein l is the length of the signal. For actual seismic data, the denoising effect is evaluated by utilizing the residual error between the actual seismic data containing noise and the denoised actual seismic data.
To evaluate the effectiveness of MFFCNN and its comparison method, 8 data sets were constructed with 50-190 different noise levels of noise added to the training data and test data at 20 intervals. For models based on a deep neural network, including DnCNN, ADNET and MFFCNN, 8 denoising models with different noise levels are trained for each model, and test data corresponding to the noise levels are denoised by using the trained models.
As shown in table 1, table 1 shows PSNR values and SSIM values after denoising seismic data with different noise levels by the proposed MFFCNN and contrast method, and the maximum value of each evaluation index is highlighted in bold in the table. As can be seen from table 1, the MFFCNN method proposed by the present invention has higher PSNR and SSIM values at different noise levels compared to other advanced comparison methods. Therefore, the MFFCNN provided by the invention has certain advantages in PSNR and SSIM. With the increase of noise, the PSNR value of the method is always above 32.401dB, and the SSIM value is always greater than 0.9631.
TABLE 1
Figure BDA0003214286050000122
Figure BDA0003214286050000131
As shown in fig. 3 to 10, the results of denoising a synthetic seismic section with a noise standard deviation of 110 by the MFCNN and the contrast method are shown. Fig. 3 and 4 show clean seismic data and seismic data with gaussian noise, respectively. Artifacts are present in fig. 5; in fig. 6, there is still noise that is not removed, and there are many point-like information missing; there is a phenomenon of missing edge information in fig. 7; FIGS. 8 and 9 fail to recover the flat background of the seismic section; fig. 10 clearly better preserves the useful information of the seismic data and background, demonstrating the effectiveness of the proposed method.
As shown in fig. 11 to 18, the abscissa Time is Time, ms is Time unit millisecond, and the ordinate Amplitude is Amplitude; to further verify the MFFCNN method of the present invention, a 118 th single-channel signal randomly extracted from the denoised seismic data in fig. 3 to 10 is shown. The method has the advantages that the fitting error is relatively least obvious, the original single trace can be better fitted, and the noise is effectively inhibited.
As shown in fig. 19, actual seismic data used by the present invention is shown. In order to further verify the MFFCNN denoising effect provided by the invention, the MFFCNN and the comparison method are applied to actual seismic data denoising and are compared. Similar to other methods, the invention standardizes the actual seismic data to 0-255 for experiment.
As shown in fig. 20 to 31, the results and residuals for the actual seismic data by MFFCNN and six comparison methods are shown. The residual is the difference between the noisy seismic data and the denoised seismic data. In fig. 20 and fig. 21, some noise in the seismic data after BM3D denoising is not suppressed, and the residual error obtained by BM3D has obvious lateral texture, which means that BM3D removes some useful information while denoising. In fig. 22, 23, 24 and 25, the denoised results of the NLH and WNNM methods are too smooth, and the residuals denoised by the NLH and WNNM methods have no obvious texture from the residual results, which means that the two methods remove the noise in the seismic section and do not remove the effective information of the seismic section. In fig. 26, fig. 27, fig. 28 and fig. 29, the field seismic sections after denoising of DNCNN and ADNET still have a lot of noise, and the residual errors after denoising of DNCNN and ADNET have no obvious texture. Compared with other methods, in fig. 30 and fig. 31, the denoising result of the MFFCNN has better suppression on noise, better visual effect, clear texture structure information, and no edge blur, and meanwhile, the residual error obtained by the MFFCNN denoising method provided by the present invention has no obvious texture, which means that the method can effectively remove noise without removing effective information.
The invention provides a multi-granularity feature fusion convolutional neural network denoising method based on convolutional neural network denoising, which is used for denoising seismic data; the method utilizes convolution kernels with different sizes to extract the characteristics of the seismic data from different granularities, then fuses the extracted characteristics, and makes full use of the local self-similarity of the seismic data to perform denoising, thereby improving the denoising effect. Compared with the traditional denoising method, the method disclosed by the invention has the advantages that the parameter adjustment is more intelligent in the experiment, the parameter adjustment is not required to be carried out on a certain seismic data, and a large amount of time is not required to be spent on parameter adjustment when the data volume is large. The invention improves the existing denoising method based on the convolutional neural network, the existing method based on the convolutional neural network usually uses a convolutional kernel with fixed size in a specific layer, which limits that the convolutional kernels with different sizes can not extract features from different granularities in the same region, and the method can utilize the convolutional kernels with different sizes to extract the features with different granularities in parallel. The method can effectively remove Gaussian noise in the seismic image and well keep texture detail information; the method can not only denoise seismic data, but also be applicable to general natural images.

Claims (9)

1. A seismic data denoising method of a multi-granularity feature fusion convolution neural network is characterized by comprising the following steps:
s1, adding Gaussian white noise with the same standard deviation into clean seismic data to obtain a data set to be trained, namely single-channel noise-containing seismic data;
s2, inputting the single-channel noisy seismic data into a neural network, and converting the single-channel noisy seismic data into multi-channel seismic data through the head of the neural network;
s3, extracting approximate characteristics of the seismic data of each channel from different granularities through a multi-granularity characteristic fusion block, namely the middle part of a neural network;
s4, converting the multi-channel seismic data containing the approximate characteristics into predictable single-channel seismic data at the tail part of the neural network;
s5, constructing a loss function through clean seismic data and predictable single-channel seismic data;
s6, calculating the minimum value of the loss function through a gradient descent method to obtain denoised seismic data and network parameters, and obtaining a denoising model of Gaussian white noise of the current standard deviation;
s7, sequentially adding the Gaussian white noises with different standard deviations into the clean seismic data, repeating the steps S1 to S6 to obtain denoising models of the Gaussian white noises with different standard deviations, and finishing model training;
s8, estimating the standard deviation of the seismic data to be denoised, and selecting the trained denoising model for denoising according to the standard deviation.
2. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 1, wherein the specific method of step S1 is:
according to the formula:
Y＝X+N
N～Ν(μ,σ2)
obtaining single-channel noise-containing seismic data Y; where X is clean seismic data, N is random noise, N (μ, σ)2) Representing a standard normal distribution obeying a mean value of μ and a standard deviation of σ.
3. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 1, wherein the specific method of step S2 comprises:
according to the formula:
Figure FDA0003214286040000021
obtaining multi-channel seismic data
Figure FDA0003214286040000022
Wherein the multi-channel seismic data has 16 channels, Relu (a) is an activation function, Cov (a) is a convolution layer, w is the width of the seismic data, h is the height of the seismic data, Y is1×w×hSeismic data including the number, width and height of channels are obtained by processing single-channel noise-containing seismic data Y.
4. The method for denoising seismic data of a multi-granularity feature fusion convolution neural network according to claim 1, wherein the head of the neural network in the step S2 comprises a convolution layer and a Relu activation function.
5. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 3, wherein the specific method of step S3 comprises:
according to the formula:
Figure FDA0003214286040000023
obtaining multi-channel seismic data H containing approximate characteristics16×w×h(ii) a Where body (-) is the processing of input data for a network containing three multi-granular feature fusion blocks.
6. The method for denoising seismic data of a multi-granularity feature fusion convolution neural network as claimed in claim 1, wherein the middle part of the neural network in step S3 includes three multi-granularity fusion blocks connected in sequence; the multi-granularity fusion block comprises three parallel multi-granularity feature extraction parts and a feature fusion part which are connected in sequence; the multi-granularity feature extraction part comprises a first convolution layer and a first batch normalization layer, a second convolution layer and a second batch normalization layer, and a third convolution layer and a third batch normalization layer and a first activation function layer which are sequentially connected; the feature fusion part comprises a fusion layer, a fourth convolution layer, a fourth batch processing normalization layer, a second activation function layer, a fifth convolution layer, a fifth batch processing normalization layer and a third activation function layer which are sequentially connected.
7. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 5, wherein the specific method of step S4 comprises:
according to the formula:
Figure FDA0003214286040000031
obtaining predictable single-channel seismic data
Figure FDA0003214286040000032
Whereintail (-) is the data processing process of the tail of the neural network; wherein the tail of the neural network contains the Cov (-) convolutional layer and the Relu (-) activation function.
8. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 1, wherein the specific method for constructing the loss function in step S5 is as follows:
according to the formula:
Figure FDA0003214286040000033
obtaining a loss function loss; where θ is a network parameter, XijIs clean seismic data having a width i and a height j,
Figure FDA0003214286040000034
the width is a predictable single-channel seismic data with i and the height is j, w is the width of the seismic data, and h is the height of the seismic data.
9. The method for denoising seismic data of a multi-granularity feature fusion convolutional neural network as claimed in claim 8, wherein the specific method for calculating the minimum value of the loss function through the gradient descent method to obtain the denoised seismic data and the network parameters in step S6 is as follows:
s6-1, randomly initializing a network parameter theta;
s6-2, according to the formula:
Figure FDA0003214286040000035
obtaining a derivative of the loss function with respect to the network parameter;
s6-3, according to the formula:
Figure FDA0003214286040000041
updating the network parameters; wherein n is a positive integer and α is a learning rate;
s6-4, judging whether the two changes before and after the loss function are smaller than a preset threshold value, if so, judging that the loss function is converged to obtain a minimum value, and obtaining a corresponding network parameter theta and predictable single-channel seismic data
Figure FDA0003214286040000042
Namely de-noised seismic data; otherwise, the step S6-3 is returned to continue the iterative updating.
CN202110939618.4A 2021-08-16 2021-08-16 Seismic data denoising method for multi-granularity feature fusion convolutional neural network Active CN113484913B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110939618.4A CN113484913B (en) 2021-08-16 2021-08-16 Seismic data denoising method for multi-granularity feature fusion convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110939618.4A CN113484913B (en) 2021-08-16 2021-08-16 Seismic data denoising method for multi-granularity feature fusion convolutional neural network

Publications (2)

Publication Number Publication Date
CN113484913A true CN113484913A (en) 2021-10-08
CN113484913B CN113484913B (en) 2023-06-16

Family

ID=77946590

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110939618.4A Active CN113484913B (en) 2021-08-16 2021-08-16 Seismic data denoising method for multi-granularity feature fusion convolutional neural network

Country Status (1)

Country Link
CN (1) CN113484913B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118071210A (en) * 2024-04-17 2024-05-24 成都理工大学 Ecological environment vulnerability assessment method combining CNN and PPM

Citations (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10032256B1 (en) * 2016-11-18 2018-07-24 The Florida State University Research Foundation, Inc. System and method for image processing using automatically estimated tuning parameters
CN108845352A (en) * 2018-06-27 2018-11-20 吉林大学 Desert Denoising of Seismic Data method based on VMD approximate entropy and multi-layer perception (MLP)
CN110045419A (en) * 2019-05-21 2019-07-23 西南石油大学 A kind of perceptron residual error autoencoder network seismic data denoising method
CN110058305A (en) * 2019-05-24 2019-07-26 吉林大学 A kind of DAS seismic data noise-reduction method based on convolutional neural networks
CN110680278A (en) * 2019-09-10 2020-01-14 广州视源电子科技股份有限公司 Electrocardiosignal recognition device based on convolutional neural network
CN111368710A (en) * 2020-02-27 2020-07-03 东北石油大学 Seismic data random noise suppression method combined with deep learning
CN111580162A (en) * 2020-05-21 2020-08-25 长江大学 Seismic data random noise suppression method based on residual convolutional neural network
CN111967506A (en) * 2020-07-31 2020-11-20 西安工程大学 Electroencephalogram signal classification method for optimizing BP neural network by artificial bee colony
CN112596104A (en) * 2020-12-09 2021-04-02 成都理工大学 Seismic data denoising method combining tensor decomposition and total variation
CN112965113A (en) * 2021-02-01 2021-06-15 冉曾令 Seismic data signal-to-noise ratio improving method
CN113093272A (en) * 2021-03-29 2021-07-09 吉林大学 Time domain full waveform inversion method based on convolutional coding
CN113156513A (en) * 2021-04-14 2021-07-23 吉林大学 Convolutional neural network seismic signal denoising method based on attention guidance
CN113204051A (en) * 2021-06-10 2021-08-03 成都理工大学 Low-rank tensor seismic data denoising method based on variational modal decomposition
CN113208614A (en) * 2021-04-30 2021-08-06 南方科技大学 Electroencephalogram noise reduction method and device and readable storage medium
US20210302603A1 (en) * 2020-03-30 2021-09-30 Qingdao university of technology Performance-level seismic motion hazard analysis method based on three-layer dataset neural network

Patent Citations (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10032256B1 (en) * 2016-11-18 2018-07-24 The Florida State University Research Foundation, Inc. System and method for image processing using automatically estimated tuning parameters
CN108845352A (en) * 2018-06-27 2018-11-20 吉林大学 Desert Denoising of Seismic Data method based on VMD approximate entropy and multi-layer perception (MLP)
CN110045419A (en) * 2019-05-21 2019-07-23 西南石油大学 A kind of perceptron residual error autoencoder network seismic data denoising method
CN110058305A (en) * 2019-05-24 2019-07-26 吉林大学 A kind of DAS seismic data noise-reduction method based on convolutional neural networks
CN110680278A (en) * 2019-09-10 2020-01-14 广州视源电子科技股份有限公司 Electrocardiosignal recognition device based on convolutional neural network
CN111368710A (en) * 2020-02-27 2020-07-03 东北石油大学 Seismic data random noise suppression method combined with deep learning
US20210302603A1 (en) * 2020-03-30 2021-09-30 Qingdao university of technology Performance-level seismic motion hazard analysis method based on three-layer dataset neural network
CN111580162A (en) * 2020-05-21 2020-08-25 长江大学 Seismic data random noise suppression method based on residual convolutional neural network
CN111967506A (en) * 2020-07-31 2020-11-20 西安工程大学 Electroencephalogram signal classification method for optimizing BP neural network by artificial bee colony
CN112596104A (en) * 2020-12-09 2021-04-02 成都理工大学 Seismic data denoising method combining tensor decomposition and total variation
CN112965113A (en) * 2021-02-01 2021-06-15 冉曾令 Seismic data signal-to-noise ratio improving method
CN113093272A (en) * 2021-03-29 2021-07-09 吉林大学 Time domain full waveform inversion method based on convolutional coding
CN113156513A (en) * 2021-04-14 2021-07-23 吉林大学 Convolutional neural network seismic signal denoising method based on attention guidance
CN113208614A (en) * 2021-04-30 2021-08-06 南方科技大学 Electroencephalogram noise reduction method and device and readable storage medium
CN113204051A (en) * 2021-06-10 2021-08-03 成都理工大学 Low-rank tensor seismic data denoising method based on variational modal decomposition

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
FENG QIAN: "Multidimensional Seismic Data Denoising Using Framelet-Based Order-p Tensor Deep Learning", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》, no. 60 *
VINEELA CHANDRA DODDA: "Simultaneous Seismic Data Denoising and Reconstruction With Attention-Based Wavelet-Convolutional Neural Network", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》, no. 61 *
张世立: "基于深度学习的AutoEncoder地震信号去噪、重建及压缩研究", 《中国优秀硕士学位论文全文数据库基础科学辑》, no. 6 *
李晓琴: "基于中值滤波和张量分解的地震资料去噪方法", 《2020年中国地球科学联合学术年会论文集（二十）》, no. 1 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118071210A (en) * 2024-04-17 2024-05-24 成都理工大学 Ecological environment vulnerability assessment method combining CNN and PPM

Also Published As

Publication number Publication date
CN113484913B (en) 2023-06-16

Similar Documents

Publication Publication Date Title
CN105913393B (en) A kind of adaptive wavelet threshold image de-noising method and device
Zhang et al. Joint image denoising using adaptive principal component analysis and self-similarity
CN103020916B (en) Image denoising method combining two-dimensional Hilbert transform and BEMD
CN111028163A (en) Convolution neural network-based combined image denoising and weak light enhancement method
CN109242799B (en) Variable-threshold wavelet denoising method
CN110706181A (en) Image denoising method and system based on multi-scale expansion convolution residual error network
CN103116873B (en) Image denoising method
CN111915486B (en) Confrontation sample defense method based on image super-resolution reconstruction
CN110490816B (en) Underwater heterogeneous information data noise reduction method
CN105913382B (en) The high-fidelity anisotropic filtering method of threshold value optimizing
CN112596104B (en) Seismic data denoising method combining tensor decomposition and total variation
CN114677300A (en) Hyperspectral image depth noise reduction method and system based on two-stage learning framework
CN102930508A (en) Image residual signal based non-local mean value image de-noising method
CN115082336A (en) SAR image speckle suppression method based on machine learning
CN113484913B (en) Seismic data denoising method for multi-granularity feature fusion convolutional neural network
CN111652810A (en) Image denoising method based on wavelet domain singular value differential model
CN116385312A (en) Low-illumination image denoising method based on phase correlation
CN113109873B (en) Desert seismic signal noise suppression method based on rank residual error constraint
CN102509268B (en) Immune-clonal-selection-based nonsubsampled contourlet domain image denoising method
CN115689958A (en) Synthetic radar image denoising method based on deep learning
CN110008633B (en) Highway noise suppression method and system based on artificial intelligence deep neural network
CN114076986B (en) Multi-scale dictionary learning sparse denoising method
CN113379641A (en) Single image rain removing method and system based on self-coding convolutional neural network
Qiaoman et al. Application of adaptive median filter and wavelet transform to dongba manuscript images denoising
CN109271971B (en) Noise reduction method for time sequence financial data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant