Seismic image interpretation method based on attention mechanism

Abstract

The present invention relates to the field of seismic image interpretation and discloses a seismic image interpretation method based on an attention mechanism. By adjusting the attention on local seismic attributes and spatial information, pixels and attributes useful for segmentation are strengthened and useless pixels or attributes are suppressed, the training process can be accelerated and the interpretation performance can be improved.

Description

Technical Field

[0001] The present invention relates to the field of seismic image interpretation, and particularly to a seismic image interpretation method based on an attention mechanism. Background Art

[0002] Seismic exploration can understand the underground geological structure according to seismic reflection signals. Seismic attributes are geometric, kinematic, or statistical characteristics of seismic waves obtained by mathematical transformation of seismic data, and can be regarded as components reflecting different geological characteristics in seismic data; there are many types of seismic attributes, and fusing different seismic attributes can improve the reliability of seismic interpretation results. However, since different attributes may play different roles in different interpretation tasks, and some attributes may also be redundant, how to consider the roles of different attributes in interpretation to improve the interpretation effect is an important issue.

[0003] Seismic data can be processed as images. By calculating multiple seismic attributes and taking each attribute as a channel, a multi-channel image can be obtained. Since image processing techniques, especially deep learning-based image processing, can achieve good results, the application of these techniques to seismic image interpretation has also attracted the attention of many researchers. For example, in the literature "Research on the Recognition Method of Abnormal Bodies in 3D Seismic Images", recognition methods for abnormal bodies are proposed from two aspects: feature extraction and segmentation methods, including edge enhancement based on principal component analysis; fusing different seismic attributes, modeling them in different ways, and then realizing segmentation, etc. However, in this technology, seismic attributes based on edge features and seismic attributes based on regional features are fused through determined coefficients, and their respective coefficients need to be set by themselves before calculation. Therefore, the value of the coefficient will directly affect the segmentation effect.

[0004] The patent "A Method for Processing Seismic Fault Images" (CN111382799A) includes the following steps: obtaining a seismic fault image data set; preprocessing the seismic fault image data set; constructing a seismic fault image processing network; training the seismic fault image processing network; obtaining seismic fault image data to be processed; processing the seismic fault image. This patent mainly aims to solve the problem of insufficient labeled seismic data, generate natural-like geological data from a small amount of labeled seismic data, and use a three-dimensional convolutional integral network to judge unknown data, thereby improving the efficiency of seismic fault recognition. However, this patent uses the original seismic data as input, and the features extracted by the CNN network have insufficient interpretability. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a seismic image interpretation method based on an attention mechanism.

[0006] To solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A seismic image interpretation method based on an attention mechanism comprises the following steps:

[0008] Step 1: Collect the data set of seismic images and the corresponding interpretation results, and record the i-th seismic image as where i=1,2,…,U r , U r represents the number of seismic images in the dataset, The length, width and number of seismic attributes of the seismic image are respectively; the seismic image is uniformly divided into two parts with length and width of w u The segmented images are used to form training samples, where the number of seismic attributes of the segmented images is still The number of segmented images is denoted as U u ;

[0009] Step 2: Set the structure and objective function of the interpretation network F. The objective function is:

[0010]

[0011] Among them, n p is the number of pixels of the segmented image; C is the number of categories; Indicates whether the classification is correct. If the true category of the i-th pixel in the segmented image is c (c = 1, 2, ..., C), then otherwise Represents the probability of classifying the i-th pixel of the segmented image into category c;

[0012] Step 3: Input the segmented image in the training sample into the interpretation network F, and use the gradient descent method to train until the objective function converges, obtain the network parameters, and obtain the trained interpretation network G;

[0013] Step 4: Divide the seismic image X to be interpreted into segments with length and width w u After the segmented image is obtained, it is input into the trained interpretation network G to obtain the interpretation result. The interpretation results of all the segmented images can be combined to obtain the interpretation result of the seismic image X to be interpreted.

[0014] Specifically, the explanation network F in step 2 includes F 1 Module, F 2 Modules and F 3 Module; F 1 The module takes the segmented image as input and obtains the enhanced feature map h according to the local attribute attention mechanism. 1 ; F 2 Module to enhance the feature map h 1As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2 ; F 3 The module is a classifier based on the image segmentation network, F 3 Module to enhance the feature map h 2 As input, the final explanation result is obtained through feature learning

[0015] Specifically, F 1 The module takes the segmented image as input and obtains the enhanced feature map h according to the local attribute attention mechanism. 1 The specific steps include:

[0016] The input segmented image u X is further divided into two parts: length and width are h p and w p The image block is calculated to calculate the average descriptor of the a-th earthquake attribute of the image block. And the maximum descriptor where v a (i,j) represents the pixel value at (i,j) in the ath earthquake attribute of the image block, The average descriptor and the maximum descriptor of the earthquake attributes form vectors ave z and max z, then the joint descriptor con z= ave z+ max z, joint descriptor con The dimension of z is Joint Descriptor Graph Joint description subgraph Z h The number of rows and columns are

[0017] Generating Attention Map Among them, H r represents a 1×1 convolution with a dimensionality reduction ratio of r, H i To restore the dimension to 1×1 convolution, σ and δ represent the Sigmoid function and ReLU function respectively, F U Represents an upsampling operation;

[0018] Then the enhanced feature map is obtained Where ◇ represents element-wise multiplication.

[0019] Specifically, F 2 Module to enhance the feature map h 1 As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2The following steps are included:

[0020] Calculate the enhanced feature map h 1 Average descriptor at spatial point (i, j) ave z i,j and the maximum descriptor max z i,j :

[0021]

[0022]

[0023] Average Descriptor ave z i,j The corresponding average description subgraph is Maximum Descriptor max z i,j The corresponding maximum description subgraph

[0024] Then the enhanced feature map h 2 =h 1 +h 1 σ[H c Z con ],Z con Indicates that ave Z and max The joint descriptor graph obtained by concatenation of Z, H c Indicates that the joint description subgraph Z con The dimension of is converted to 1 by a 1×1 convolution.

[0025] Compared with the prior art, the beneficial technical effects of the present invention are:

[0026] The present invention proposes a seismic image interpretation method based on the attention mechanism, which can accelerate the training process and improve the interpretation performance by adjusting the attention of local seismic attributes and spatial information, strengthening pixels and attributes useful for segmentation, and suppressing useless pixels or attributes. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0028] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0029] A seismic image interpretation method based on an attention mechanism comprises the following steps:

[0030] S1: Collect the data set of seismic images and the corresponding interpretation results. The i-th seismic image is where i=1,2,…,Ur , U r represents the number of seismic images in the dataset, are the length, width and number of seismic attributes of the seismic image respectively; since seismic images usually reflect underground geological structure information over a large area, in order to make full use of their information, the seismic image is firstly divided into uniformly sized images with length and width of w u The image has the same number of channels. The total number of segmented images obtained after segmentation is U u ;

[0031] S2: Set the structure and objective function of the explanation network F. The objective function is:

[0032]

[0033] Among them, n p is the number of pixels of the segmented image; C is the number of categories; Indicates whether the classification is correct. If the true category of the i-th pixel in the segmented image is c (c = 1, 2, ..., C), then otherwise It represents the probability of classifying the i-th pixel of the segmented image into category c.

[0034] Specifically, the explanation network F in step 2 includes F 1 Module, F 2 Modules and F 3 Module; where F 1 The module has a length and width of w u The segmented image is taken as input, and the enhanced feature map h is obtained according to the local attribute attention mechanism. 1 ; F 2 Module to enhance the feature map h 1 As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2 ; F 3 The module is a classifier, which can use any image segmentation network. One available structure is based on the Unet neural network structure, including multi-layer feature extraction and multi-layer feature fusion. 3 Module to enhance the feature map h 2 As input, the final explanation result is obtained through feature learning

[0035] Among them, F 1 The module takes the segmented image as input and obtains the enhanced feature map h according to the local attribute attention mechanism. 1 The specific steps include:

[0036] The input segmented imageu X is further divided into two parts: length and width are h p and w p The image block is calculated to calculate the average descriptor of the a-th earthquake attribute of the image block. And the maximum descriptor max z z =max({v a (1,1),…,v a (1,w p ),…,v a (h p ,1),…,v a (h p ,w p )}); where v a (i,j) represents the pixel value at (i,j) in the ath earthquake attribute of the image block, The average descriptor and the maximum descriptor of the earthquake attributes form vectors ave z and max z, then the joint descriptor con z= ave z+ max z, joint descriptor con The dimension of z is Joint Descriptor Graph Joint description subgraph Z h The number of rows and columns are

[0037] Generating Attention Map Among them, H r represents a 1×1 convolution with a dimensionality reduction ratio of r, H i To restore the dimension to 1×1 convolution, σ and δ represent the Sigmoid function and ReLU function respectively, F U Represents an upsampling operation;

[0038] Then the enhanced feature map is obtained Where ◇ represents element-wise multiplication.

[0039] Among them, F 2 Module to enhance the feature map h 1 As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2 The following steps are included:

[0040] Calculate the enhanced feature map h 1 Average descriptor at spatial point (i, j) ave z i,j and the maximum descriptor max zi,j :

[0041]

[0042]

[0043] Average Descriptor ave z i,j The corresponding average description subgraph is Maximum Descriptor max z i,j The corresponding maximum description subgraph

[0044] Then the enhanced feature map h 2 =h 1 +h 1 σ[H c Z con ],Z con Indicates that ave Z and max The joint descriptor graph obtained by concatenation of Z, H c Indicates that the joint description subgraph Z con The dimension of is converted to 1 by a 1×1 convolution.

[0045] S3: Input the segmented image in the training sample into the interpretation network F, and use the gradient descent method to train until the objective function converges, obtain the network parameters, and obtain the trained interpretation network G.

[0046] S4: Divide the seismic image X to be interpreted into two parts with length and width w u After the segmented image is obtained, it is input into the trained interpretation network G to obtain the interpretation result. The interpretation results of all the segmented images can be combined to obtain the interpretation result of the seismic image X to be interpreted.

[0047] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0048] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A seismic image interpretation method based on the attention mechanism, comprising the following steps: Step 1: Collect the data set of seismic images and the corresponding interpretation results, and record the i-th seismic image as X i , where i=1,2,…,U r , U r represents the number of seismic images in the dataset, M r ,N r , are the length, width and number of seismic attributes of the seismic image, respectively; The seismic image is uniformly divided into two parts with length and width of w. u The segmented images are used to form training samples, where the number of seismic attributes of the segmented images is still The number of segmented images is denoted as U u ; Step 2: Set the structure of the interpretation network F and the objective function, and the objective function is: Among them, n p is the number of pixels of the segmented image; C is the number of categories; Indicates whether the classification is correct. If the true category of the i-th pixel in the segmented image is c, c = 1, 2, ..., C, then otherwise Represents the probability of classifying the i-th pixel of the segmented image into category c; Step 3: Input the segmented images in the training samples into the interpretation network F, and use the gradient descent method for training until the objective function converges to obtain the network parameters, and at the same time obtain the trained interpretation network G; Step 4: Divide the seismic image X to be interpreted into segments with length and width w u After the segmented image is obtained, it is input into the trained interpretation network G to obtain the interpretation result. The interpretation results of all the segmented images can be combined to obtain the interpretation result of the seismic image X to be interpreted; The explanation network F in step 2 includes F 1 Module, F 2 Modules and F 3 Module; F 1 The module takes the segmented image as input and obtains the enhanced feature map h according to the local attribute attention mechanism. 1 ; F 2 Module to enhance the feature map h 1 As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2 ; F 3 The module is a classifier based on the image segmentation network, F 3 Module to enhance the feature map h 2 As input, the final explanation result is obtained through feature learning 2. The seismic image interpretation method based on the attention mechanism according to claim 1, characterized in that: F 1 The module takes the segmented image as input and obtains the enhanced feature map h according to the local attribute attention mechanism. 1 The specific steps include: The input segmented image u X is further divided into two parts: length and width are h p and w p The image block is calculated to calculate the average descriptor of the a-th earthquake attribute of the image block. And the maximum descriptor max z a =max({v a (1,1),…,v a (1,w p ),…,v a (h p ,1),…,v a (h p ,w p )}); where v a (i,j) represents the pixel value at (i,j) in the ath earthquake attribute of the image block, The average descriptor and the maximum descriptor of the earthquake attributes form vectors ave z and max z, then the joint descriptor con z= ave z+ max z, joint descriptor con The dimension of z is Joint Descriptor Graph Joint description subgraph Z h The number of rows and columns are Generating Attention Map Among them, H r represents a 1×1 convolution with a dimensionality reduction ratio of r, H i To restore the dimension to 1×1 convolution, σ and δ represent the Sigmoid function and ReLU function respectively, F U Represents an upsampling operation; Then the enhanced feature map is obtained in Represents element-wise multiplication.

3. The seismic image interpretation method based on the attention mechanism according to claim 2, characterized in that: F 2 Module to enhance the feature map h 1 As input, the feature is further enhanced according to the spatial attention mechanism to obtain the enhanced feature map h 2 The following steps are included: Calculate the enhanced feature map h 1 Average descriptor at spatial point (i, j) and the maximum descriptor Average Descriptor The corresponding average description subgraph is Maximum Descriptor The corresponding maximum description subgraph Then the enhanced feature map h 2 =h 1 +h 1 σ[H c Z con ],Z con Indicates that ave Z and max The joint descriptor graph obtained by concatenation of Z, H c Indicates that the joint description subgraph Z con The dimension is converted to 1 by a 1×1 convolution.

Images