CN109782339A - A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network - Google Patents
A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network Download PDFInfo
- Publication number
- CN109782339A CN109782339A CN201910032839.6A CN201910032839A CN109782339A CN 109782339 A CN109782339 A CN 109782339A CN 201910032839 A CN201910032839 A CN 201910032839A CN 109782339 A CN109782339 A CN 109782339A
- Authority
- CN
- China
- Prior art keywords
- training
- network
- dncnn
- seismic data
- noise
- 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.)
- Pending
Links
Landscapes
- Geophysics And Detection Of Objects (AREA)
Abstract
The present invention discloses a kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network: 01: construction 3D-DnCNN network, the tentatively selected higher region of one piece of signal-to-noise ratio is as training region, remaining area is as test zone, and it is suppressed using random noise of the state-of-the-art stochastic noise suppression method of current industry to training region, using the pressed data of noise as label data, training sample pair is constructed；02: the training sample pair obtained using the further screening step 01 of gradient-structure tensor, to obtain quality more preferably training sample；03: the training sample that step 02 obtains being trained to 3D-DnCNN network is sent into, after the completion of training, uses the random noise of 3D-DnCNN network compacting test zone.The present invention solves the interference problem of random noise in poststack three dimensional seismic data, has suppressed random noise and three-dimensional arcuation imaging noise.In addition, the method for the present invention can carry out parallel processing, and there is good adaptivity, meets industrial large-scale calculations demand.
Description
Technical field
The present invention relates to field of signal processing, in particular to a kind of random noise about in field of seismic exploration is suppressed
Method.
Background technique
With the continuous improvement of oil and gas prospect degree and the lasting extension of Exploration Domain, oil-gas exploration it is main
Object also from it is different before, the main object of China's oil-gas exploration has turned to slit formation, hidden-type and deep-seated oil gas reservoir.This
Class oil-gas reservoir belongs to the maximum Oil Reservoir Types of exploration and development difficulty, is faced with such as surface conditions and subsurface structure complexity, preserves
A series of common exploration challenges such as thickness degree is thin and heterogeneity is strong.Above-mentioned challenge forces us most from seismic exploration data
More useful informations are possibly excavated, seismic exploration technique, seismic data processing technique and Explanation Accuracy are proposed higher
Requirement, while high s/n ratio, high-resolution and high fidelity also have become pursuing a goal for seismic prospecting.Therefore, it suppresses
Random noise in seismic data is the important step during seismic data processing and explanation.
The concept of deep learning is derived from the research of artificial neural network, and Hinton in 2006 et al. will contain multiple hidden layers
Neural network be described as a kind of deep learning structure and promoted, open deep learning grinding in academia and industry
Study carefully tide.Deep learning the fields such as natural language processing, image procossing, man-machine game and health medical treatment obtain it is huge at
Function, but due to stratigraphic structure is complicated, lithology multiplicity, target bury depth and signal-to-noise ratio is weak etc., at other signals
Reason, more stringent requirements are proposed to signal processing method for seismic signal.Up to the present, although deep learning is at seismic signal
Some applications are had been achieved in reason, and the effect obtained in some aspects has been over traditional seismic data processing side
Method, but in seismic data processing and reservoir prediction field, it is generally still at an early stage.
The prior art:
SVD.This method is a kind of method for carrying out carrying out signal reconstruction based on the corresponding feature vector of characteristic value, using having
Imitate the correlation of signal, the corresponding characteristic value of identification noise；Then seismic signal is rebuild using the corresponding feature vector of characteristic value,
Remove random noise.
The shortcomings that prior art:
(1) this method denoising effect depends on the correlation of seismic data axis in the same direction, for curved axis denoising effect in the same direction
Fruit is poor；
(2) this method eigenvalue is improper is easy to cause denoising to be not thorough or cause to damage to useful signal.
Summary of the invention
The purpose of the present invention is to provide one kind to be based on 3D-DnCNN (three-dimensional denoising
Convolutional neural network, three-dimensional denoising convolutional neural networks) network poststack three dimensional seismic data it is random
Noise drawing method, to solve the above technical problems.The present invention constructs 3D-DnCNN network model, is explored using seismic signal
The characteristics of process, has selected the higher region of signal-to-noise ratio as training region, remaining region is as test zone, and with current work
The state-of-the-art conventional method of industry suppresses the random noise in the training region, using gradient-structure tensor to the sample in training region
It is screened, constructs ideal training sample pair, be sent into 3D-DnCNN network and be trained, after training up, utilize 3D-
The random noise of DnCNN compacting test zone.
To achieve the goals above, the present invention adopts the following technical scheme:
A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network, comprising the following steps:
Step 01: constructing 3D-DnCNN network, one piece of 3 D stereo region conduct in arbitrarily selected three dimensional seismic data
Training region Ytraining, remaining area is as test zone Ytest, and using stochastic noise suppression method to training region
YtrainingRandom noise suppressed, using the pressed data of noise as label data, construct training sample pair；
Step 02: the training sample pair obtained using the further screening step 01 of gradient-structure tensor, to obtain quality more
Excellent training sample；
Step 03: the training sample that step 02 obtains is trained to 3D-DnCNN network is sent into, after the completion of training,
Using the random noise of 3D-DnCNN network compacting test zone, the random noise compacting of three dimensional seismic data is completed.
Further, step 01, comprising:
The input data of 3D-DnCNN network is original three dimensional seismic data y；And y=x+v, wherein x indicates effectively letter
Number, v indicates noise, is mapped using one residual error of residual error learning trainingIt is that this residual error maps the result is that net
The noise that network learns, in this way, just having obtained the seismic data useful signal of network output
The convolution kernel of 3D-DnCNN network uses three-dimensional structure, and convolution filter is dimensioned to 3 × 3 × 3 and moves
Except all pond layers；Zero padding is used, if stride is set as 1, zero padding setting are as follows:
In formula, K indicates the size of filter；
The case where when not being 1 for stride setting, output ruler of any given convolutional layer after convolution operation in network
It is very little to calculate according to the following formula:
In formula, O indicates the size of output；W indicates original size；The size of K expression filter；P indicates the size of filling；
The size of S expression stride；
3D-DnCNN network uses residual error and learns and criticize normalization technology；
Residual error learns solve gradient the phenomenon that back-propagation process disappears in the structural level of network, learns to residual error
Basic unit calculates partial derivative and obtains:
Criticizing normalization reduces internal covariant transfer, ensure that the training with model, between layers in approximate phase
Continue to learn under same input distribution, accelerates network training；The forward conduction formula for criticizing normalization network layer is as follows:
In formula,β(k)=E [x(k)]；
The chain type derivation for criticizing normalization layer is shown below:
The higher region of signal-to-noise ratio is selected from actual seismic data as training region Ytraining, remaining region conduct
Test zone Ytest；Then it is suppressed using random noise of the random noise compacting conventional method to training region, by noise
Pressed data are as label data X 'training, construct training sample pair.
Further, step 02, comprising:
The calculating step of gradient-structure tensor are as follows:
01), to training area data X 'trainingHilbert transform is carried out to be shown below:
02) X ', is calculatedtrainingInstantaneous phase ψ and instantaneous frequency A:
03), centered on each point in ψ, square set is constructedWherein, w is indicated just
The length and width of cube, b indicate that the height of square, N indicate the sum of square；Then respectively along time, crossline and
The gradient-structure tensor of each square in inline direction calculating square set Z, is shown below:
In formula, gx,gy,gzRespectively indicate ziGradient along the direction time, crossline and inline,Indicate ziMiddle institute
Average value a little；
0.4), to giEigenvalues Decomposition is carried out, λ is obtained1 i,λ2 i,λ3 i, tomography confidence level C is constructed, as follows:
0.5) training sample X ' further, is screened to tomography confidence level C setting hard -threshold λtraining, hard threshold function formula
It is as follows:
In formula:For hard threshold function, λ is hard -threshold.
Further, step 03 includes: to be trained the training sample that step 03 obtains to 3D-DnCNN network is sent into,
Optimize loss function using stochastic gradient descent, loss function is shown below:
After the completion of training, using the random noise of 3D-DnCNN network compacting test zone, three dimensional seismic data is completed
Random noise compacting.
Further, λ=0.65.
Further, training region is not less than respectively in the length of time orientation, main profile direction and cross-track direction
Entire three dimensional seismic data is 1/5th of direction total length.
Compared with the existing technology, the invention has the following advantages: using method of the invention by being instructed to small range
The study for practicing regional earthquake data, obtains the compacting ability to entire 3D seismic data random noise.Method of the invention
The random noise of Gaussian distributed and non-gaussian distribution in seismic data can be suppressed, has very strong adaptivity, especially
It is to have compacting ability more stronger than conventional method for the three-dimensional arcuation imaging noise in actual seismic data.Meanwhile the present invention
Method due to used GPU carry out parallel computation, have faster noise pressing speed.
Detailed description of the invention
Fig. 1 is 3D-DnCNN network structure；
Fig. 2 is the flow chart of 3D-DnCNN network denoising；
Fig. 3 is practical noisy seismic data 2600ms slice；
The original seismic profile of Fig. 4 A；
Fig. 4 B is tomography confidence calculations result；
Fig. 4 C-4J is respectively 0.80,0.75,0.70,0.65,0.60,0.55,0.50,0.45 screening knot using threshold value
Fruit；
Fig. 5 A is the denoising result using 3D-CNN network on 2600ms isochronous surface；
Fig. 5 B is the noise removed on 2600ms isochronous surface using 3D-CNN network；
Fig. 6 is the probability density distribution of the noise removed using 3D-CNN network；
Fig. 7 is flow diagram of the invention.
Specific embodiment
The present invention will be further described in detail with reference to the accompanying drawings and detailed description.
The present invention provides a kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network, first
3D-DnCNN network is constructed, the tentatively selected higher region of one piece of signal-to-noise ratio is as training region, and remaining area is as test section
Domain, and suppressed using random noise of the state-of-the-art stochastic noise suppression method of current industry to training region, it will make an uproar
The pressed data of sound construct training sample as label data, then further screen training sample using gradient-structure tensor
This, obtains quality more preferably training sample pair, is finally trained obtained training sample to 3D-DnCNN network is sent into, to
After the completion of training, the random noise of 3D-DnCNN network compacting test zone is used.
As shown in Figure 1, the convolution kernel of 3D-DnCNN network uses three-dimensional structure, convolution filter is dimensioned to 3 × 3
× 3 and remove all pond layers.In order to ensure the size for the seismic data that the seismic data and network of network inputs export
Unanimously, the present invention is distributed using zero padding.The input data of 3D-DnCNN network is original three dimensional seismic data y.And y=x+
V, wherein x indicates that useful signal, v indicate noise.It is mapped using one residual error of residual error learning trainingThis residual error is reflected
It is penetrating the result is that the noise that e-learning arrives.In this way, the present invention has just obtained the seismic data useful signal of network output
As shown in Fig. 2, in conjunction with the surface conditions in seismic prospecting work progress, by observing three dimensional seismic data, on ground
One piece of 3 D stereo region in shake data in arbitrarily selected three dimensional seismic data is as training region Ytraining, remaining region is made
For test zone Ytest.Then using the state-of-the-art random noise compacting conventional method of current industry to the random of training region
Noise is suppressed, using the pressed data of noise as label data X 'training, construct training sample pair.Then it uses
The further Screening Samples of gradient-structure tensor are sent into 3D-DnCNN network and are trained, optimized using stochastic gradient descent and lost
Function.Training region is not less than entire 3-D seismics in the length of time orientation, main profile direction and cross-track direction respectively
Data in direction total length 1/5th.
Loss function is shown below:
In formula, N indicates the total number of training sample pair.
Stochastic gradient descent optimizes the parameter Θ in network by minimizing loss function, is shown below:
In formula, x1…NIt is training dataset.In order to improve the parallelization degree of program, this optimization process substep is executed, often
One step takes m sample x from training data concentration1…mIt is trained, here it is small lot stochastic gradient descents.Using small lot with
The decline of machine gradient, parameter Θ are updated according to following formula:
After the completion of training, using the random noise of 3D-DnCNN network compacting test zone, obtain finally suppressing random
The denoising result of noise.
As shown in figure 3, training sample has very important status in deep learning, the denoising of network is directly affected
Energy.And completely clean actual seismic data XtrainingIt can not obtain, therefore, the training sample of actual seismic data constructs
It is relative complex.During actual seismic exploration, earth's surface situation differs greatly, geological structure complexity, acquisition time difference etc.
Factor causes the seismic data signal-to-noise ratio of different zones different.In addition to this, height is used in the exploration process of partial region
The exploration mode of density acquisition, can make the signal-to-noise ratio in this region be apparently higher than other regions.In summary advantage, knot
The surface conditions in seismic prospecting work progress are closed, it is arbitrarily selected from actual seismic data by observing three dimensional seismic data
One piece of 3 D stereo region lineups continuity in three dimensional seismic data is good, and the few region of random noise is as training region
Ytraining, as shown in the rectangle frame in the lower left corner in figure.Training region is in time orientation, main profile direction and cross-track direction
Length is not less than entire three dimensional seismic data 1/5th of direction total length respectively.
As shown in Fig. 4 A to Fig. 4 J, the invention proposes use gradient-structure tensor screening training sample available strategy,
Different threshold values are set and obtain the training sample of different quality.The calculating of gradient-structure tensor mainly comprises the steps that
101 couples of trained area data X 'trainingHilbert transform is carried out to be shown below:
102. calculating X 'trainingInstantaneous phase ψ and instantaneous frequency A:
103. constructing square set centered on each point in ψWherein, w is indicated just
The length and width of cube, b indicate that the height of square, N indicate the sum of square.Then respectively along time, crossline and
The gradient-structure tensor of each square in inline direction calculating square set Z, is shown below:
In formula, gx,gy,gzRespectively indicate ziGradient along the direction time, crossline and inline,Indicate ziMiddle institute
Average value a little.
104 couples of giEigenvalues Decomposition is carried out, λ is obtained1 i,λ2 i,λ3 i, tomography confidence level C is constructed, as follows:
105 couples of tomography confidence level C setting hard -threshold λ further screen training sample X 'training, hard threshold function formula is such as
Under:
In formula:For hard threshold function, λ is hard -threshold.
In the present embodiment, the number of iterations that hard -threshold λ is chosen for 0.65,3D-DnCNn network training is selected as 30 times, instruction
Practicing sample size is 52062 pairs, and the number of every batch of training is 6, and learning rate is set to 0.001.
The invention has the following beneficial effects:
1) present invention has hi-fi to useful signal, and useful signal can be effectively protected；
2) the method for the present invention can carry out random noise compacting for different seismic signals, have good adaptive
Property；
3) the method for the present invention is to be met industrial-scale based on GPU parallel processing and calculated demand.
It is applied in the random noise compacting of actual seismic data below with analysis and processing method provided by the invention, hair
Existing method of the invention can not only effectively Attenuating Random Noise, and useful signal has compared with hi-fi, is subsequent money
The analysis of material lays the foundation.
As shown in figures 5a and 5b, present invention application 3D-DnCNN network is in three-dimensional poststack actual seismic data by small
The study of range training region seismic data, has successfully suppressed random noise, has obtained to entire 3D seismic data
The compacting ability of random noise.
To the noise removed in Fig. 5 B, as shown in fig. 6, it is for statistical analysis, it can be found that the probability of network removal noise
Density Distribution is equal to the probability density distribution of conventional method removal noise, shows again network and obtains to entire three-dimensional folded
The compacting ability of seismic data random noise afterwards.And the kurtosis value of network and conventional method removal noise is all greater than 0, shows
The probability density distribution of removal noise all obeys super-Gaussian distribution.
In above real data example, random noise pressure is carried out to three-dimensional poststack seismic data using method of the invention
System, can not only effective Attenuating Random Noise, and useful signal has compared with hi-fi, establishes for the analysis of subsequent data
Basis.
Finally, it should be noted that the above real data example is to the purpose of the present invention, technical solution and beneficial effect
Further verifying is provided, this only belongs to specific implementation example of the invention, it is not intended to limit the scope of protection of the present invention,
Within the spirit and principles in the present invention, any modification made, improvement or equivalent replacement etc., should all be in protection model of the invention
In enclosing.
Claims (6)
1. a kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network, which is characterized in that including
Following steps:
Step 01: constructing 3D-DnCNN network, one piece of 3 D stereo region in arbitrarily selected three dimensional seismic data is as training
Region Ytraining, remaining area is as test zone Ytest, and using stochastic noise suppression method to training region Ytraining's
Random noise is suppressed, and using the pressed data of noise as label data, constructs training sample pair；
Step 02: the training sample pair obtained using the further screening step 01 of gradient-structure tensor, to obtain quality more preferably
Training sample；
Step 03: the training sample that step 02 obtains being trained to 3D-DnCNN network is sent into, after the completion of training, is used
3D-DnCNN network suppresses the random noise of test zone, completes the random noise compacting of three dimensional seismic data.
2. a kind of poststack three dimensional seismic data random noise compacting side based on 3D-DnCNN network as described in claim 1
Method, which is characterized in that step 01, comprising:
The input data of 3D-DnCNN network is original three dimensional seismic data y；And y=x+v, wherein x indicates useful signal, v
It indicates noise, is mapped using one residual error of residual error learning trainingIt is that this residual error maps the result is that network science
The noise practised, in this way, just having obtained the seismic data useful signal of network output
The convolution kernel of 3D-DnCNN network uses three-dimensional structure, and convolution filter is dimensioned to 3 × 3 × 3 and removes institute
Some pond layers；Zero padding is used, if stride is set as 1, zero padding setting are as follows:
In formula, K indicates the size of filter；
The case where when not being 1 for stride setting, Output Size root of any given convolutional layer after convolution operation in network
It is calculated according to following formula:
In formula, O indicates the size of output；W indicates original size；The size of K expression filter；P indicates the size of filling；S table
Show the size of stride；
3D-DnCNN network uses residual error and learns and criticize normalization technology；
Residual error learns solve gradient the phenomenon that back-propagation process disappears in the structural level of network, learns to residual error basic
Unit calculates partial derivative and obtains:
Criticizing normalization reduces internal covariant transfer, ensure that the training with model, between layers approximately uniform
Continue to learn under input distribution, accelerates network training；The forward conduction formula for criticizing normalization network layer is as follows:
In formula,β(k)=E [x(k)]；
The chain type derivation for criticizing normalization layer is shown below:
The higher region of signal-to-noise ratio is selected from actual seismic data as training region Ytraining, remaining region is as test
Region Ytest；Then it is suppressed using random noise of the random noise compacting conventional method to training region, noise is suppressed
Data afterwards are as label data X 'training, construct training sample pair.
3. a kind of poststack three dimensional seismic data random noise compacting side based on 3D-DnCNN network as claimed in claim 2
Method, which is characterized in that step 02, comprising:
The calculating step of gradient-structure tensor are as follows:
01), to training area data X 'trainingHilbert transform is carried out to be shown below:
02) X ', is calculatedtrainingInstantaneous phase ψ and instantaneous frequency A:
03), centered on each point in ψ, square set is constructedWherein, w indicates square
Length and width, b indicate square height, N indicate square sum；Then respectively along time, crossline and the side inline
To the gradient-structure tensor for calculating each square in square set Z, it is shown below:
In formula, gx,gy,gzRespectively indicate ziGradient along the direction time, crossline and inline,Indicate ziMiddle all the points
Average value；
0.4), to giEigenvalues Decomposition is carried out, λ is obtained1 i,λ2 i,λ3 i, tomography confidence level C is constructed, as follows:
0.5) training sample X ' further, is screened to tomography confidence level C setting hard -threshold λtraining, hard threshold function formula is such as
Under:
In formula:For hard threshold function, λ is hard -threshold.
4. a kind of poststack three dimensional seismic data random noise compacting side based on 3D-DnCNN network as claimed in claim 3
Method, which is characterized in that step 03 includes: to be trained the training sample that step 03 obtains to 3D-DnCNN network is sent into, and is made
Optimize loss function with stochastic gradient descent, loss function is shown below:
After the completion of training, using 3D-DnCNN network compacting test zone random noise, complete three dimensional seismic data with
The compacting of machine noise.
5. a kind of poststack three dimensional seismic data random noise compacting side based on 3D-DnCNN network as claimed in claim 3
Method, which is characterized in that λ=0.65.
6. a kind of poststack three dimensional seismic data random noise compacting side based on 3D-DnCNN network as described in claim 1
Method, which is characterized in that training region is not less than entirely respectively in the length of time orientation, main profile direction and cross-track direction
Three dimensional seismic data in direction total length 1/5th.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910032839.6A CN109782339A (en) | 2019-01-14 | 2019-01-14 | A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910032839.6A CN109782339A (en) | 2019-01-14 | 2019-01-14 | A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109782339A true CN109782339A (en) | 2019-05-21 |
Family
ID=66500434
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910032839.6A Pending CN109782339A (en) | 2019-01-14 | 2019-01-14 | A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109782339A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN111563846A (en) * | 2019-10-23 | 2020-08-21 | 哈尔滨理工大学 | Image denoising method based on compressed neural network |
| CN111708087A (en) * | 2020-07-09 | 2020-09-25 | 中国科学技术大学 | Method for suppressing seismic data noise based on DnCNN neural network |
| CN112965113A (en) * | 2021-02-01 | 2021-06-15 | 冉曾令 | Seismic data signal-to-noise ratio improving method |
| CN113805235A (en) * | 2020-06-11 | 2021-12-17 | 中国石油天然气股份有限公司 | Three-dimensional seismic facies identification method and device based on convolutional neural network |
| CN113900146A (en) * | 2020-07-06 | 2022-01-07 | 中国石油天然气股份有限公司 | Surface wave pressing method and system |
| CN114236610A (en) * | 2021-12-24 | 2022-03-25 | 西安交通大学 | Iterative seismic data unmixing method and system based on depth convolution network prior |
| CN114325821A (en) * | 2021-12-23 | 2022-04-12 | 西安交通大学 | Method and system for suppressing strong scattering noise in pre-stack seismic data based on 3D-SNACNN network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103489159A (en) * | 2013-09-02 | 2014-01-01 | 电子科技大学 | Three-dimensional seismic data image denoising method based on trilateral structure guide smoothing |
| CN104122584A (en) * | 2014-08-08 | 2014-10-29 | 中国石油集团川庆钻探工程有限公司地球物理勘探公司 | Method and device for determining directionality according to seismic data |
| CN107248144A (en) * | 2017-04-27 | 2017-10-13 | 东南大学 | A kind of image de-noising method based on compression-type convolutional neural networks |
| CN108549102A (en) * | 2018-03-29 | 2018-09-18 | 西安交通大学 | The earth formation Curvature Estimation method of the more window analyses of joint gradient-structure tensor sum |
| WO2019005180A1 (en) * | 2017-06-26 | 2019-01-03 | Elekta, Inc. | Method for improving cone-beam ct image quality using a deep convolutional neural network |
-
2019
- 2019-01-14 CN CN201910032839.6A patent/CN109782339A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103489159A (en) * | 2013-09-02 | 2014-01-01 | 电子科技大学 | Three-dimensional seismic data image denoising method based on trilateral structure guide smoothing |
| CN104122584A (en) * | 2014-08-08 | 2014-10-29 | 中国石油集团川庆钻探工程有限公司地球物理勘探公司 | Method and device for determining directionality according to seismic data |
| CN107248144A (en) * | 2017-04-27 | 2017-10-13 | 东南大学 | A kind of image de-noising method based on compression-type convolutional neural networks |
| WO2019005180A1 (en) * | 2017-06-26 | 2019-01-03 | Elekta, Inc. | Method for improving cone-beam ct image quality using a deep convolutional neural network |
| CN108549102A (en) * | 2018-03-29 | 2018-09-18 | 西安交通大学 | The earth formation Curvature Estimation method of the more window analyses of joint gradient-structure tensor sum |
Non-Patent Citations (3)
| Title |
|---|
| DAWEI LIU ET AL.: "Random noise suppression in seismic data: what can deep learning do?", 《SEG INTERNATIONAL EXPOSITION AND 88TH ANNUAL MEETING》 * |
| PETER BAKKER: "Image structure analysis for seismic interpretation", 《DELFT UNIVERSITY OF TECHNOLOGY》 * |
| 陈双全等: "地震数据结构张量相干计算方法", 《石油物探》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN111563846A (en) * | 2019-10-23 | 2020-08-21 | 哈尔滨理工大学 | Image denoising method based on compressed neural network |
| CN111563846B (en) * | 2019-10-23 | 2021-10-26 | 哈尔滨理工大学 | Image denoising method based on compressed neural network |
| CN113805235A (en) * | 2020-06-11 | 2021-12-17 | 中国石油天然气股份有限公司 | Three-dimensional seismic facies identification method and device based on convolutional neural network |
| CN113805235B (en) * | 2020-06-11 | 2023-12-26 | 中国石油天然气股份有限公司 | Three-dimensional seismic phase identification method and device based on convolutional neural network |
| CN113900146A (en) * | 2020-07-06 | 2022-01-07 | 中国石油天然气股份有限公司 | Surface wave pressing method and system |
| CN111708087A (en) * | 2020-07-09 | 2020-09-25 | 中国科学技术大学 | Method for suppressing seismic data noise based on DnCNN neural network |
| CN112965113A (en) * | 2021-02-01 | 2021-06-15 | 冉曾令 | Seismic data signal-to-noise ratio improving method |
| CN114325821A (en) * | 2021-12-23 | 2022-04-12 | 西安交通大学 | Method and system for suppressing strong scattering noise in pre-stack seismic data based on 3D-SNACNN network |
| CN114236610A (en) * | 2021-12-24 | 2022-03-25 | 西安交通大学 | Iterative seismic data unmixing method and system based on depth convolution network prior |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109782339A (en) | A kind of poststack three dimensional seismic data stochastic noise suppression method based on 3D-DnCNN network | |
| Liu et al. | Random noise suppression in seismic data: What can deep learning do? | |
| CN107589448B (en) | A kind of multitrace seismogram reflection coefficient sequence Simultaneous Inversion method | |
| Yang et al. | Deep learning seismic random noise attenuation via improved residual convolutional neural network | |
| CN109143356A (en) | A kind of ADAPTIVE MIXED norm dictionary learning seismic impedance inversion | |
| CN105974468B (en) | A kind of method that can be carried out at the same time five dimension Reconstruction of seismic data and noise compacting | |
| Jiang et al. | A convolutional autoencoder method for simultaneous seismic data reconstruction and denoising | |
| CN109709603A (en) | Seismic horizon identification and method for tracing, system | |
| Liu et al. | An unsupervised deep learning method for denoising prestack random noise | |
| CN112946749B (en) | Method for suppressing seismic multiples based on data augmentation training deep neural network | |
| CN110308484A (en) | A kind of chromatography conversion method and system based on deep learning intelligent screening first arrival | |
| CN107607992A (en) | More ripple matching process based on convolutional neural networks | |
| CN104932010B (en) | A kind of diffracted wave separation method based on the sparse Radon transformation of shortcut fringing | |
| CN107870355B (en) | A kind of kirchhoff type beam deviation method under MODEL OVER COMPLEX TOPOGRAPHY | |
| CN105607122B (en) | A kind of earthquake texture blending and Enhancement Method based on full variation geological data decomposition model | |
| CN108693555A (en) | Intelligent time-varying blind deconvolution wideband processing method and processing device | |
| CN110471111A (en) | A kind of normal-moveout spectrum automatic pick method based on convolutional neural networks | |
| CN103954998B (en) | Residual amplitude compensating method based on AVO | |
| Liu et al. | Microseismic event detection and classification based on convolutional neural network | |
| CN111045077A (en) | Full waveform inversion method of land seismic data | |
| Zhou et al. | Absorption attenuation compensation using an end-to-end deep neural network | |
| CN108037533B (en) | A kind of seismic prospecting noise drawing method recombinating expectation log-likelihood based on block | |
| Kazei et al. | Mapping full seismic waveforms to vertical velocity profiles by deep learning | |
| CN116068644A (en) | Method for improving resolution and noise reduction of seismic data by using generation countermeasure network | |
| CN106950597B (en) | Mixing source data separation method based on the filtering of three sides |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190521 |
|
| RJ01 | Rejection of invention patent application after publication |