Seismic facies identification method and device based on CBAM attention mechanism
By introducing the CBAM-U-Net model of the CBAM attention mechanism in seismic phase recognition, the problem of insufficient subjectivity and accuracy of manual interpretation in traditional seismic phase recognition technology is solved, and more efficient and accurate seismic phase recognition is achieved. in complex geological environment.
Patent Information
- Application Number
- CN202410960546.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-12-03
AI Technical Summary
The existing seismic phase recognition technology has subjectivity and uncertainty in manual interpretation, and traditional methods are difficult to meet the needs of high precision and high efficiency. Especially in unconventional oil and gas reservoir exploration, the applicability and reliability of traditional methods are insufficient.
The seismic phase recognition method based on the CBAM attention mechanism is adopted. By constructing the CBAM-U-Net model, combining the encoder-decoder structure and CBAM module of U-Net, the feature representation ability is enhanced, and Dice Loss is used as the loss function to embed attention. The force module focuses on seismic phase edge information and improves the accuracy of multi-task details segmentation.
It significantly improves the accuracy of seismic phase recognition and the meticulousness of edge information, reduces application difficulty, provides stronger intelligent detection support, and is suitable for complex geological environments.
Smart Images

Figure CN119068230A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of seismic phase identification, and in particular to a seismic phase identification method based on a CBAM attention mechanism, and a seismic phase identification device based on a CBAM attention mechanism. Background Art
[0002] Seismic phase identification technology is a powerful tool in oil exploration and exploration of certain salt mines and coal fields. Seismic phase is the manifestation of underground sedimentary phase in seismic data, which can indirectly reflect the characteristics of underground sedimentation. Its accurate identification is of great significance for sedimentary phase research and oil and gas exploration and development. Traditional seismic phase identification and division refers to the manual identification and division of sedimentary phase on seismic profile by professionals based on the characteristics and experience of sedimentary phase on seismic profile, commonly known as face-to-face method. This method has a large workload, high professional threshold requirements, and there is a certain degree of multi-solution in manual interpretation, which makes the interpretation results subjective. With the advancement of seismic exploration technology and the improvement of exploration and development requirements, the amount of seismic data has increased sharply and the demand for interpretation accuracy has become higher and higher, which is difficult to meet with manual methods. In recent years, thanks to the rapid development of computing resources, artificial intelligence technology represented by deep learning has also developed rapidly. This method has been widely used in many fields such as autonomous driving, language translation and medical image processing. The "face-to-face method" in traditional seismic phase analysis relies on the human visual system to work, while computer vision uses computers to simulate the human visual system and replace humans to complete the corresponding work. Due to the influence of human subjective factors, the results of seismic phase division have strong uncertainty. With the continuous improvement of computer performance, attribute analysis and waveform clustering have become the main methods for seismic phase identification. Compared with manual interpretation, attribute analysis has greatly improved the efficiency of seismic phase division. However, the relationship between seismic attributes and target geological bodies is very complex, and it is difficult to evaluate the reliability of classification results. The method based on seismic waveform clustering can eliminate the limitations of seismic attribute analysis. However, this method has high requirements for seismic data and is not suitable for areas with large variations in stratum thickness. As the exploration targets of major oil fields turn to unconventional oil and gas reservoirs, the applicability and reliability of traditional methods can no longer meet the requirements of existing seismic phase analysis. Therefore, it is particularly important to study a fast and effective seismic phase identification method.
[0003] To achieve this goal, convolutional neural networks (CNNs) have become the leading technology in the field of computer vision by virtue of their ability to automatically extract semantic information contained in images. They can not only analyze the intensity of a single pixel value, but also analyze patterns such as texture, edge, and shape in the image. In order to obtain more accurate seismic phase analysis results, researchers need to accurately locate the boundaries of the seismic phases, and the results of using the convolution classification algorithm on small blocks of seismic phases often show jagged edges at the edges of different seismic phases. The semantic segmentation technology can improve the smoothness of the edges of the seismic phase analysis results and improve the seismic phase analysis effect. Dramsch et al. used the pre-trained VGG16 model to predict seismic phases through the transfer learning method; these scholars divided the seismic profile into independent small windows, used the convolution layer for feature extraction, and then connected the fully connected layer to output the predicted seismic phase category as the prediction result of the small window. After cyclic prediction, the prediction result of the entire seismic profile was obtained. This method cannot effectively utilize the spatial distribution law of seismic phases, and the resolution and speed of prediction are mutually restricted. Alaudah et al. used an encoder-decoder for seismic phase identification, which improved the prediction speed of the network, but the prediction accuracy was still low; Fangyu Li et al. used a deep dilated convolutional neural network based on an attention mechanism for seismic phase identification. Through spectral decomposition, seismic records with different frequency components were input into the network, which improved the classification accuracy, but there was a problem of low MIoU (Mean Intersection over Union).
[0004] At present, seismic phase recognition technology is usually divided into methods based on convolution classification networks and methods based on semantic segmentation. In practical applications, methods based on convolution are more common than methods based on semantic segmentation. The reason why the representation method based on convolution classification is more popular is that the learning task using convolutional neural networks can better establish the mapping relationship between image features and targets, so that the analysis of seismic profiles can easily improve the recognition accuracy. However, the larger the data scale of the convolution classification network, the greater the corresponding training time cost; using the seismic phase recognition method based on semantic segmentation, image semantic segmentation can separate different targets in the image, so the semantic segmentation method can be used to perform pixel-level division of seismic phase images. However, when the model structure design is unreasonable, the boundary of the geological body identified by the seismic phase is not clear enough.
[0005] The attention mechanism in deep learning is a method that imitates the human visual and cognitive system, which allows the neural network to focus on relevant parts when processing input data. By introducing the attention mechanism, the neural network can automatically learn and selectively focus on important information in the input, improving the performance and generalization ability of the model. Therefore, this paper proposes a semantic segmentation network based on U-Net and CBAM (convolutional block attention module) attention mechanism. Summary of the invention
[0006] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a seismic phase identification method based on the CBAM attention mechanism, which can accurately identify the six seismic phases, maintain the edge information of the seismic phases so that the segmented picture is detailed and complete, providing strong support for the intelligent detection of seismic phases.
[0007] The technical solution of the present invention is: the seismic phase identification method based on the CBAM attention mechanism comprises the following steps:
[0008] (1) Data processing: selecting two-dimensional slices from the original data set as image data and corresponding label data, expanding the image data and label data into three-dimensional tensors, adjusting the image data and label data to the specified size using bilinear interpolation, and normalizing the image data;
[0009] (2) Construct a CBAM-U-Net model, extract and reconstruct features through the encoder-decoder structure of U-Net, and enhance the feature representation capability through the CBAM module. The network is based on U-Net and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer.
[0010] (3) Construct a CBAM attention module, which includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling respectively. Maximum pooling extracts texture detail features and reduces the influence of useless information, while average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using an activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance;
[0011] (4) Using Dice Loss as the loss function
[0012]
[0013] When training the image segmentation model, Dice Loss is used as the optimization target to minimize Dice Loss.
[0014] The present invention embeds an attention module in the U-Net network of the contraction-expansion path, so that the network focuses on the edge information of the seismic phase when processing data, improves the accuracy of multi-task detail segmentation, and improves the phenomenon of misclassification and omission; the CBAM lightweight attention module dynamically adjusts the weight of each feature channel, so that the model pays more attention to the edge information and overall trend of the seismic phase in the input image, and enhances the feature representation; CBAM-U-Net is more regular in boundary division, the recognition of thin layers is improved, the segmentation accuracy is significantly improved, CBAM-U-Net is the most accurate in identifying seismic phases, and the edge information of seismic phases is maintained to make the segmented picture detailed and complete; the results confirm the effectiveness and necessity of the attention module in multi-target semantic segmentation, provide strong support for the intelligent detection of seismic phases, can greatly reduce the application difficulty of seismic phase identification, and have a wide range of potential applications.
[0015] A seismic phase identification device based on the CBAM attention mechanism is also provided, the device comprising:
[0016] A data processing module, which selects two-dimensional slices from the original data set as image data and corresponding label data, expands the image data and label data into a three-dimensional tensor, adjusts the image data and label data to a specified size using bilinear interpolation, and normalizes the image data;
[0017] The CBAM-U-Net model uses the encoder-decoder structure of U-Net to extract and reconstruct features, and enhances the feature representation capability through the CBAM module. The network uses U-Net as the main framework and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer.
[0018] The CBAM attention module includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling, the maximum pooling extracts texture detail features and reduces the influence of useless information, and the average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using the activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance;
[0019] Loss function module, which uses Dice Loss as the loss function
[0020]
[0021] When training the image segmentation model, Dice Loss is used as the optimization target to minimize Dice
[0022] Loss. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 Four transformation methods for data augmentation are shown: (a) flipping, (b) stretching, (c) rotation, and (d) skewing.
[0024] Figure 2 The overall architecture of the CBAM-U-Net network is shown.
[0025] Figure 3 This is a schematic diagram of the CBAM module.
[0026] Figure 4 The CA model framework is shown.
[0027] Figure 5 The PA model framework is shown.
[0028] Figure 6 Comparison of seismic facies segmentation results is shown.
[0029] Figure 7 A flowchart of a seismic phase identification method based on the CBAM attention mechanism according to the present invention is shown. DETAILED DESCRIPTION
[0030] like Figure 7 As shown in the figure, this seismic phase identification method based on the CBAM attention mechanism includes the following steps:
[0031] (1) Data processing: selecting two-dimensional slices from the original data set as image data and corresponding label data, expanding the image data and label data into three-dimensional tensors, adjusting the image data and label data to the specified size using bilinear interpolation, and normalizing the image data;
[0032] (2) Construct a CBAM-U-Net model, extract and reconstruct features through the encoder-decoder structure of U-Net, and enhance the feature representation capability through the CBAM module. The network is based on U-Net and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer.
[0033] (3) Construct a CBAM attention module, which includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling respectively. Maximum pooling extracts texture detail features and reduces the influence of useless information, while average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using an activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance;
[0034] (4) Using Dice Loss as the loss function
[0035]
[0036] When training the image segmentation model, Dice Loss is used as the optimization target to minimize Dice Loss.
[0037] The present invention embeds an attention module in the U-Net network of the contraction-expansion path, so that the network focuses on the edge information of the seismic phase when processing data, improves the accuracy of multi-task detail segmentation, and improves the phenomenon of misclassification and omission; the CBAM lightweight attention module dynamically adjusts the weight of each feature channel, so that the model pays more attention to the edge information and overall trend of the seismic phase in the input image, and enhances the feature representation; CBAM-U-Net is more regular in boundary division, the recognition of thin layers is improved, the segmentation accuracy is significantly improved, CBAM-U-Net is the most accurate in identifying seismic phases, and the edge information of seismic phases is maintained to make the segmented picture detailed and complete; the results confirm the effectiveness and necessity of the attention module in multi-target semantic segmentation, provide strong support for the intelligent detection of seismic phases, can greatly reduce the application difficulty of seismic phase identification, and have a wide range of potential applications.
[0038] Preferably, in the step (1), the pixel values are scaled to the range of [0, 1], the normalized image data is converted to an integer type, and the pixel values are limited to the range of [0, 255]; the image data is normalized again, and the image data is copied to three channels to match a typical color image format.
[0039] Preferably, in step (1), the data set is cut into 1372 images with each sampling point of the training part as the center, and the number of samples of each class is equal. Four enhancement methods are used to generate different data cubes: flipping, stretching, rotation and tilting, and the images are randomly divided into training set, verification set and test set according to 7:2:1.
[0040] Preferably, in the step (2), a seismic profile is input, and the blue arrow indicates the convolution layer, which is first convolved, wherein the blue arrow represents a 3×3 convolution for feature extraction, and each encoding layer contains two convolution layers. The DoubleConv module is a basic convolution block, including two consecutive convolution layers, each convolution layer is followed by batch normalization and ReLU activation function, and rich features are extracted through two convolution operations, and batch normalization is used to stabilize the training, and the ReLU activation function introduces nonlinearity to improve the expression ability of the model;
[0041] Perform the maximum pooling operation on the feature map, downsample the image, repeat the downsampling 5 times to obtain 5 feature maps in total, and the feature map changes from large and flat to small and three-dimensional; reduce the spatial resolution of the feature map through the pooling layer and increase the receptive field of the feature; skip connection stores the output of each convolution block for feature fusion in the subsequent decoder path; first enhance the features of the 5-layer feature map obtained in the encoder stage through CBAM, focus on the edge information of the seismic phase and reduce the attention to non-edge information. Each decoder layer contains two convolution layers with a convolution kernel of 3 and an upsampling layer. After each upsampling, the upsampled feature map is spliced with the corresponding skip connection feature map. The spliced feature map is further processed by the convolution block DoubleConv and the CBAM module to ensure that the feature map size is consistent during splicing. If it is inconsistent, the size is adjusted using the image processing function;
[0042] The upsampling module in the U-Net model uses two convolutional layers and one bilinear upsampling layer to implement the upsampling process of the feature map. By connecting the feature map with the upsampled lower resolution feature map, more spatial information is integrated. At the same time, in each decoding layer, the feature map with the same size in the encoding layer is spliced on the channel through the jump connection and the output of the upper decoding layer; after passing the last convolutional layer, the output is the original size.
[0043] Map the feature map to the number of output channels to get the prediction result.
[0044] Preferably, in the channel attention module CAM of step (3), firstly, the input feature F is subjected to max-pooling and average-pooling operations, and the expressions of average-pooling and max-pooling are:
[0045]
[0046] Through the global average pooling and global maximum pooling layers, we obtain the channel descriptor F1 containing texture details and F2 containing more semantic information. Then, we send the two channel descriptors to the fully connected layer for fusion to obtain the channel attention map M∈R 1 ×1×c , after the Sigmod activation function is used to normalize the weights and multiply them channel by channel with the original features, the formula is as follows:
[0047]
[0048] Among them, MLP represents the fully connected layer, W0 and W1 represent the fully connected layer parameters, reducing the parameters to reduce the dimension of the channel and then increase the dimension, represents the Sigmod activation function, and the weights are normalized through the activation function and multiplied with the original features channel by channel.
[0049] Preferably, in the spatial attention module SAM of step (3), for the input feature layer, the maximum value and the average value are taken on the channel of each feature point, and then the two results are stacked, and the number of channels is adjusted by a convolution with a channel number of 1, and then a sigmod is taken to obtain the weight of each feature point in the input feature layer, which is between 0 and 1. After obtaining the weight, the weight is multiplied by the original input feature layer, and the formula is expressed as follows:
[0050]
[0051] Among them, f 7×7 Represents a convolution kernel of size 7*7, [] represents Concat connection, and F represents Sigmod activation function.
[0052] Preferably, the method further comprises step (5), wherein when the model performs image segmentation, each pixel is identified and assigned a unique label, and 6 representative images are extracted to verify the segmentation effect of each model on different types of cracks.
[0053] A seismic phase identification device based on the CBAM attention mechanism is also provided, the device comprising:
[0054] A data processing module, which selects two-dimensional slices from the original data set as image data and corresponding label data, expands the image data and label data into a three-dimensional tensor, adjusts the image data and label data to a specified size using bilinear interpolation, and normalizes the image data;
[0055] The CBAM-U-Net model uses the encoder-decoder structure of U-Net to extract and reconstruct features, and enhances the feature representation capability through the CBAM module. The network uses U-Net as the main framework and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer.
[0056] The CBAM attention module includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling, the maximum pooling extracts texture detail features and reduces the influence of useless information, and the average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using the activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance;
[0057] Loss function module, which uses Dice Loss as the loss function
[0058]
[0059] When training the image segmentation model, Dice Loss is used as the optimization objective to minimize Dice Loss.
[0060] Preferably, in the data processing module, the pixel values are scaled to the range of [0, 1], the normalized image data is converted to an integer type, and the pixel values are limited to the range of [0, 255]; the image data is normalized again, and the image data is copied to three channels to match a typical color image format;
[0061] The dataset is centered on each sampling point in the training part and cut into 1,372 images with an equal number of samples for each class. Four enhancement methods are used to generate different data cubes: flipping, stretching, rotation, and tilting. The images are randomly divided into training set, validation set, and test set in a ratio of 7:2:1.
[0062] Preferably, in the CBAM-U-Net model, a seismic profile is input, and the blue arrow indicates the convolution layer, which is first convolved, wherein the blue arrow represents a 3×3 convolution for feature extraction, and each encoding layer contains two convolution layers. The DoubleConv module is a basic convolution block, including two consecutive convolution layers, each convolution layer is followed by batch normalization and ReLU activation function, and rich features are extracted through two convolution operations, and batch normalization is used to stabilize the training, and the ReLU activation function introduces nonlinearity to improve the expression ability of the model;
[0063] Perform the maximum pooling operation on the feature map, downsample the image, repeat the downsampling 5 times to obtain 5 feature maps in total, and the feature map changes from large and flat to small and three-dimensional; reduce the spatial resolution of the feature map through the pooling layer and increase the receptive field of the feature; skip connection stores the output of each convolution block for feature fusion in the subsequent decoder path; first enhance the features of the 5-layer feature map obtained in the encoder stage through CBAM, focus on the edge information of the seismic phase and reduce the attention to non-edge information. Each decoder layer contains two convolution layers with a convolution kernel of 3 and an upsampling layer. After each upsampling, the upsampled feature map is spliced with the corresponding skip connection feature map. The spliced feature map is further processed by the convolution block DoubleConv and the CBAM module to ensure that the feature map size is consistent during splicing. If it is inconsistent, the size is adjusted using the image processing function;
[0064] The upsampling module in the U-Net model uses two convolutional layers and one bilinear upsampling layer to implement the upsampling process of the feature map. By connecting the feature map with the lower resolution feature map after upsampling, more spatial information is integrated. At the same time, in each decoding layer, the feature map with the same size in the encoding layer is spliced on the channel through jump connections and the output of the upper decoding layer; after passing through the last convolutional layer, the output is the original size, and the feature map is mapped to the number of output channels to obtain the prediction result.
[0065] The present invention is described in more detail below.
[0066] The present invention comprises the following steps:
[0067] 1. Data processing
[0068] The data comes from the open-source 3D seismic data of the Parihaka area of the New Zealand Public Seismological Survey. The data was also interpreted by geologists, with the interpretation range of the main survey line 4200-5000, a total of 801 profiles and 550 channels. Six seismic phases were divided: ① Baseline phase (low signal-to-noise ratio, weak reflection, and may contain volcanic rocks in some places); ② Slope mudstone phase A (mainly mudstone from the slope to the bottom of the basin, with strong amplitudes at the upper and lower boundaries; weak internal reflections); ③ Large-scale sedimentary phase (mixed with chaotic reflections and weak amplitude parallel reflections); ④ Slope mudstone phase B (mudstone and sandstone from the slope to the bottom of the basin, with strong amplitude parallel reflections); ⑤ Slope valley (strong amplitude, relatively low fluctuation); ⑥ Submarine canyon system (the eroded base is U-shaped, with high local fluctuations, and the interior is filled with low-amplitude parallel faults and chaotic reflections).
[0069] The above description of data processing is as follows:
[0070] 1. Select 2D slices from the original dataset as image data and corresponding label data. Expand the image data and label data into 3D tensors and use bilinear interpolation to resize the image data and label data to the specified size.
[0071] 2. Normalize the image data and scale the pixel values to the range of [0,1]. Convert the normalized image data to integer type and limit the pixel values to the range of [0,255].
[0072] 3. The image data is normalized again and copied to three channels to match the typical color image format.
[0073] 4. Limited training data severely limits the performance of deep neural networks, because models trained with too little data are prone to overfitting. Due to the complex underground structure and large geological differences between different regions, manual interpretation of field seismic data is time-consuming and laborious. Although Alaudah et al. have proposed some effective methods to generate training samples, there is also a simple and effective way to increase the data for constructing a data set. The data set is centered on each sampling point in the training part and cut into 1,372 images with an equal number of samples in each class. Four enhancement methods are used to generate different data cubes: flipping, stretching, rotation, and tilting. In order to ensure the independence of the data and avoid a picture appearing in both the training set and the test set after expansion, thereby generating false accuracy, the images are randomly divided into training set, validation set, and test set according to the ratio of 7:2:1. (See Figure 1 )
[0074] 2. Model construction
[0075] The CBAM-U-Net model combines the U-Net architecture and the CBAM (Convolutional Block Attention Module) module to improve the performance in image segmentation tasks. The model extracts and reconstructs features through the encoder-decoder structure of U-Net, and enhances the feature representation capability through the CBAM module, so that the model can more effectively focus on important feature areas in the image. The network is based on U-Net as the main framework and is divided into two parts: encoder and decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer (see Figure 2 ).
[0076] Input a seismic profile. The blue arrow indicates the convolution layer. First, convolution is performed on it. The blue arrow represents a 3×3 convolution for feature extraction. Each encoding layer contains two convolution layers. The DoubleConv module is a basic convolution block, which includes two consecutive convolution layers. Each convolution layer is followed by batch normalization and ReLU activation function. Rich features are extracted through two convolution operations, and batch normalization is used to stabilize training. The ReLU activation function introduces nonlinearity to improve the expression ability of the model.
[0077] The input image is extracted through a series of convolution blocks (DoubleConv) and the feature representation is enhanced through the CBAM module. In order to reduce the GPU memory and model parameter volume and increase the receptive field, it is crucial for image segmentation to enable 3×3 convolution to extract features over a larger receptive field. Next, the feature map is subjected to the maximum pooling operation to downsample the image. Repeating the downsampling five times, a total of five feature maps are obtained, and the feature maps change from large and flat to small and three-dimensional. The spatial resolution of the feature map is reduced through the pooling layer (MaxPool) to increase the receptive field of the feature. The skip connection stores the output of each convolution block for feature fusion in the subsequent decoder path. The 5-layer feature map obtained in the encoder stage is first enhanced through the CBAM (convolutional block attention module), focusing on the edge information of the seismic phase and reducing the attention to the non-edge information. Each decoder layer contains two convolution layers with a convolution kernel of 3 and an upsampling layer. After each upsampling, the upsampled feature map is spliced with the corresponding skip connection feature map. The concatenated feature maps are further processed by the convolution block (DoubleConv) and the CBAM module. Make sure the feature maps have the same size when concatenated. If not, use the image processing function (TF.resize) to resize them.
[0078] The upsampling module in the U-Net model. This module uses two convolutional layers and a bilinear upsampling layer to implement the upsampling process of the feature map. By connecting the feature map with the upsampled lower resolution feature map, more spatial information can be integrated. At the same time, at each decoding layer, the feature map of the same size in the encoding layer is spliced on the channel through a jump connection and the output of the upper decoding layer, thereby improving the problem of insufficient information during upsampling.
[0079] After the last convolutional layer, the output is of the original size, and the feature map is mapped to the number of output channels to obtain the final prediction result.
[0080] By combining U-Net and CBAM modules, the model not only has the multi-level feature fusion capability of U-Net, but also enhances the feature representation capability in channel and spatial dimensions through the CBAM module. The model can be more balanced in processing details and global information, improving the performance in image segmentation tasks.
[0081] 3. CBAM Attention Module
[0082] CBAM consists of two parts: a channel attention module CAM (channel attention module) that enhances the correlation between feature channels, and a spatial attention module SAM (spatial attention module) that performs adaptive weighting on each position of the feature map (see Figure 3 ).
[0083] 1) Channel Attention Module Channel Attention Module
[0084] Channel attention module such as Figure 4 As shown in the figure, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling respectively. Maximum pooling can extract texture detail features and reduce the influence of useless information. Average pooling can retain background semantic information, so two pooling operations are used to operate the input features to retain more complete feature information. Secondly, the two obtained feature information are sent to the fully connected layer (weight sharing), and the results are fused to obtain features with richer semantic details. Then, the activation function is used to normalize the weights (Sigmod is used here) and multiply the original features to obtain the final result.
[0085] Specifically, firstly, max-pooling and average-pooling operations are performed on the input feature F. The expressions of average-pooling and max-pooling are:
[0086]
[0087] Through the global average pooling and global maximum pooling layers, we obtain the channel descriptor F1 containing texture details and F2 containing more semantic information. Then, we send the two channel descriptors to the fully connected layer for fusion to obtain the channel attention map M∈R 1 ×1×c , after the Sigmod activation function is used to normalize the weights, the weights are multiplied channel by channel with the original features. The formula is as follows:
[0088]
[0089] Among them, MLP represents the fully connected layer, W0 and W1 represent the fully connected layer parameters, and the parameters are reduced to reduce the dimension of the channel and then increase the dimension. It represents the Sigmod activation function, which normalizes the weights and multiplies them with the original features channel by channel.
[0090] 2) Spatial Attention Module
[0091] The spatial attention module is Figure 5 As shown in the figure, its main function is to assign larger weights to important pixels for image segmentation to enhance the network segmentation performance.
[0092] Specifically, the maximum value and average value of each feature point in the input feature layer are taken. The two results are then stacked, and the number of channels is adjusted using a convolution with a channel number of 1. Then a sigmoid is taken, and the weight of each feature point in the input feature layer is obtained (between 0 and 1). After obtaining this weight, it is multiplied by the original input feature layer. The formula is as follows:
[0093]
[0094] Among them, f 7×7 The convolution kernel [] represents a 7*7 size, and represents a Concat connection. It represents a Sigmod activation function. Through CBAM attention, the channel features and spatial features that are important in the segmentation task can be given greater weight, and irrelevant features can be suppressed.
[0095] 4. Loss Function
[0096] Dice Loss is a loss function used to measure the performance of image segmentation models. It is based on the Dice coefficient, also known as the Dice similarity coefficient or The coefficient is usually used to evaluate the similarity between two sets, especially in medical image segmentation.
[0097] The calculation formula of Dice coefficient is as follows:
[0098] Among them, X is the binary result predicted by the model, Y is the true binary label, |X∩Y| is the number of pixels of the intersection of the predicted result and the true label, |X| and |Y| are the number of pixels of the predicted result and the true label respectively.
[0099] Dice Loss is a loss function derived from the Dice coefficient, and its general form is as follows:
[0100]
[0101] The smaller the Dice Loss, the higher the similarity between the model's prediction results and the true label, that is, the better the model's performance. Therefore, when training an image segmentation model, Dice Loss is usually used as an optimization target to improve the model's performance by minimizing Dice Loss.
[0102] 5. Model comparison
[0103] When the model performs image segmentation, it identifies each pixel and assigns it a unique label. Six representative images are extracted to verify the segmentation effect of each model on different types of cracks, such as Figure 6 shown.
[0104] Figure 6 (a) and (f) are seismic sections. Figure 6 (b) and (g) are labeled images of seismic profiles. From the test results, the U-Net network can correctly predict the categories of most pixels when segmenting the seismic phases in the image, but there are many omissions and misclassifications, and it is easy to confuse the first and third seismic phases. There are also misclassifications and some seismic phases cannot be identified, and the edge segmentation is incomplete. Specifically, Figure 6 In (h), the connection of the third seismic phase cannot be identified, the fourth seismic phase cannot be identified, and the edge identification of the sixth seismic phase is fuzzy. The recognition result of DeepLabv3+ network is more accurate but fails to maintain the boundary information well. Compared with U-Net network, it has fewer omissions and misclassifications, but the boundary is not clearly depicted. Figure 6 The thin layer in (d) cannot be identified. Figure 6 (i) The fault boundary cannot be accurately identified; CBAM-U-Net is more regular in boundary division, the recognition of thin layers is improved, and the segmentation accuracy is significantly improved. Compared with other models, CBAM-U-Net is the most accurate in identifying the six seismic phases, and the edge information of the seismic phase is maintained to make the segmentation picture detailed and complete. The results confirm the effectiveness and necessity of the attention module in multi-target semantic segmentation, providing strong support for the intelligent detection of seismic phases.
[0105] In order to verify the effectiveness of the proposed scheme, the open source 3D seismic data of the Parihaka area from the New Zealand Public Seismological Survey was used, and the data was divided into training set, validation set and test set to verify the effectiveness of the seismic phase identification method.
[0106] First, the 3D seismic data volume is segmented to generate regular 2D seismic sections. The 2D slices can be selected as image data and corresponding label data; then the image data and label data are expanded to a 3D tensor, and bilinear interpolation is used to adjust the image data and label data to the specified size. Finally, the data is normalized and enhanced to avoid overfitting of the model trained with too little data.
[0107] The processed data is input into the semantic segmentation model, and after the model training is completed, the semantic segmentation model can be tested for its ability to identify earthquake phases. Figure 6 It can be seen intuitively that CBAM-U-Net has the most accurate recognition of the six seismic phases, and maintains the edge information of the seismic phases to make the segmentation picture detailed and complete. The results confirm the effectiveness and necessity of the attention module in multi-target semantic segmentation, providing strong support for the intelligent detection of seismic phases.
[0108] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.
Claims
1. A seismic phase identification method based on CBAM attention mechanism, characterized by: The method comprises the following steps: (1) Data processing: selecting two-dimensional slices from the original data set as image data and corresponding label data, expanding the image data and label data into three-dimensional tensors, adjusting the image data and label data to the specified size using bilinear interpolation, and normalizing the image data; (2) Construct a CBAM-U-Net model, extract and reconstruct features through the encoder-decoder structure of U-Net, and enhance the feature representation capability through the CBAM module. The network is based on U-Net and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer. (3) Construct a CBAM attention module, which includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling respectively. Maximum pooling extracts texture detail features and reduces the influence of useless information, while average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using an activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance; (4) Using Dice Loss as the loss function When training the image segmentation model, Dice Loss is used as the optimization objective to minimize Dice Loss.
2. The seismic phase identification method based on the CBAM attention mechanism according to claim 1, characterized in that: In the step (1), the pixel values are scaled to the range of [0, 1], the normalized image data is converted to an integer type, and the pixel values are limited to the range of [0, 255]; the image data is normalized again, and the image data is copied to three channels to match a typical color image format.
3. The seismic phase identification method based on the CBAM attention mechanism according to claim 2, characterized in that: In the step (1), the data set is cut into 1372 images with each sampling point of the training part as the center, and the number of samples of each class is equal. Four enhancement methods are used to generate different data cubes: flipping, stretching, rotation and tilting. The images are randomly divided into training set, verification set and test set according to 7:2:
1.
4. The seismic phase identification method based on the CBAM attention mechanism according to claim 3, characterized in that: In the step (2), a seismic profile is input, and the blue arrow indicates the convolution layer. First, convolution is performed on it, where the blue arrow represents a 3×3 convolution for feature extraction. Each encoding layer contains two convolution layers. The DoubleConv module is a basic convolution block, including two consecutive convolution layers, each of which is followed by batch normalization and ReLU activation function. Rich features are extracted through two convolution operations, and batch normalization is used to stabilize training. The ReLU activation function introduces nonlinearity to improve the expression ability of the model; the feature map is subjected to maximum pooling operation, and the image is downsampled. The downsampling is repeated 5 times to obtain a total of 5 feature maps. The feature map changes from large and flat to small and Stereoscopic; reduce the spatial resolution of the feature map through the pooling layer and increase the receptive field of the feature; skip connection stores the output of each convolution block for feature fusion in the subsequent decoder path; the 5-layer feature map obtained in the encoder stage is first enhanced by CBAM, focusing on the edge information of the seismic phase and reducing the attention to non-edge information. Each decoder contains two convolution layers with a convolution kernel of 3 and an upsampling layer. After each upsampling, the upsampled feature map is spliced with the corresponding skip connection feature map. The spliced feature map is further processed by the convolution block DoubleConv and the CBAM module to ensure that the feature map size is consistent when splicing. If not, the image processing function is used to adjust the size; The upsampling module in the U-Net model uses two convolutional layers and one bilinear upsampling layer to implement the upsampling process of the feature map. By connecting the feature map with the upsampled lower resolution feature map, more spatial information is integrated. At the same time, at each decoding layer, the feature map of the same size in the encoding layer is spliced on the channel through skip connections and the output of the upper decoding layer; After the last convolutional layer, the output is of the original size, and the feature map is mapped to the number of output channels to obtain the prediction result.
5. The seismic phase identification method based on the CBAM attention mechanism according to claim 4, characterized in that: In the channel attention module CAM of step (3), firstly, the input feature F is subjected to max-pooling and average-pooling operations. The expressions for average-pooling and max-pooling are: The channel descriptor F1 containing texture details is obtained through the global average pooling and global maximum pooling layers, and F2 containing more semantic information is then sent to the fully connected layer for fusion to obtain the channel attention map M∈R 1×1×c , after the Sigmod activation function is used to normalize the weights and multiply them channel by channel with the original features, the formula is as follows: Among them, MLP represents the fully connected layer, W0 and W1 represent the fully connected layer parameters, reducing the parameters to reduce the dimension of the channel and then increase the dimension, represents the Sigmod activation function, and the weights are normalized through the activation function and multiplied with the original features channel by channel.
6. The seismic phase identification method based on the CBAM attention mechanism according to claim 5, characterized in that: In the spatial attention module SAM of step (3), for the input feature layer, the maximum value and the average value are taken on the channel of each feature point, and then the two results are stacked, and the number of channels is adjusted by a convolution with a channel number of 1, and then a sigmod is taken to obtain the weight of each feature point in the input feature layer, which is between 0 and 1. After obtaining this weight, this weight is multiplied by the original input feature layer. The formula is expressed as follows: Among them, f 7×7 Represents a convolution kernel of size 7*7, [] represents Concat connection, and F represents Sigmod activation function.
7. The seismic phase identification method based on the CBAM attention mechanism according to claim 6, characterized in that: The method also includes step (5), in which when the model performs image segmentation, each pixel is identified and assigned a unique label, and six representative images are extracted to verify the segmentation effect of each model on different types of cracks.
8. A seismic phase identification device based on the CBAM attention mechanism, characterized in that: The device includes: A data processing module, which selects two-dimensional slices from the original data set as image data and corresponding label data, expands the image data and label data into a three-dimensional tensor, adjusts the image data and label data to a specified size using bilinear interpolation, and normalizes the image data; The CBAM-U-Net model uses the encoder-decoder structure of U-Net to extract and reconstruct features, and enhances the feature representation capability through the CBAM module. The network uses U-Net as the main framework and is divided into an encoder and a decoder. Each encoder layer consists of two convolutional layers with a convolution kernel of 3, an attention module, and a downsampling layer. The CBAM attention module includes: a channel attention module CAM that enhances the correlation between feature channels, and a spatial attention module SAM that performs adaptive weighting on each position of the feature map; in the channel attention module, given a feature map F, the feature map is subjected to maximum average pooling and global average pooling, the maximum pooling extracts texture detail features and reduces the influence of useless information, and the average pooling retains background semantic information. The two types of feature information obtained are sent to the fully connected layer, and the results are fused to obtain features with richer semantic details. Subsequently, the weight normalization is multiplied by the original feature using the activation function to obtain the final result; in the spatial attention module, weights are assigned to important pixels for image segmentation to enhance the network segmentation performance; Loss function module, which uses Dice Loss as the loss function When training the image segmentation model, Dice Loss is used as the optimization target to minimize Dice Loss.
9. The seismic phase identification device based on the CBAM attention mechanism according to claim 8, characterized in that: In the data processing module, the pixel values are scaled to the range of [0, 1], the normalized image data is converted to an integer type, and the pixel values are limited to the range of [0, 255]; the image data is normalized again, and the image data is copied to three channels to match a typical color image format; The dataset is centered on each sampling point in the training part and cut into 1,372 images with an equal number of samples for each class. Four enhancement methods are used to generate different data cubes: flipping, stretching, rotation, and tilting. The images are randomly divided into training set, validation set, and test set in a ratio of 7:2:
1.
10. The seismic phase identification device based on the CBAM attention mechanism according to claim 9, characterized in that: In the CBAM-U-Net model, a seismic profile is input. The blue arrow indicates the convolution layer. First, convolution is performed on it. The blue arrow represents a 3×3 convolution for feature extraction. Each encoding layer contains two convolution layers. The DoubleConv module is a basic convolution block, including two consecutive convolution layers. Each convolution layer is followed by batch normalization and ReLU activation function. Rich features are extracted through two convolution operations, and batch normalization is used to stabilize training. The ReLU activation function introduces nonlinearity to improve the expression ability of the model. Perform the maximum pooling operation on the feature map, downsample the image, repeat the downsampling 5 times to obtain 5 feature maps in total, and the feature map changes from large and flat to small and three-dimensional; reduce the spatial resolution of the feature map through the pooling layer and increase the receptive field of the feature; skip connection stores the output of each convolution block for feature fusion in the subsequent decoder path; first enhance the features of the 5-layer feature map obtained in the encoder stage through CBAM, focus on the edge information of the seismic phase and reduce the attention to non-edge information. Each decoder layer contains two convolution layers with a convolution kernel of 3 and an upsampling layer. After each upsampling, the upsampled feature map is spliced with the corresponding skip connection feature map. The spliced feature map is further processed by the convolution block DoubleConv and the CBAM module to ensure that the feature map size is consistent during splicing. If it is inconsistent, the size is adjusted using the image processing function; The upsampling module in the U-Net model uses two convolutional layers and one bilinear upsampling layer to implement the upsampling process of the feature map. By connecting the feature map with the upsampled lower resolution feature map, more spatial information is integrated. At the same time, at each decoding layer, the feature map of the same size in the encoding layer is spliced on the channel through skip connections and the output of the upper decoding layer; After the last convolutional layer, the output is of the original size, and the feature map is mapped to the number of output channels to obtain the prediction result.
Citation Information
Cited By
Semantic segmentation method and system for laser radar point cloud of high-voltage transmission line
CN120107602A
Semantic segmentation method and system for lidar point cloud of high-voltage transmission line
CN120107602B
Seismic data interpolation method and system based on CBAM-Res2Unet network
CN120276031A
Human pelvis pose recognition method and device, electronic equipment and medium
CN120355761A
Human body pelvis posture recognition method and device, electronic equipment and medium
CN120355761B
