Seismic data interpolation method and system based on CBAM-Res2Unet network

Abstract

The invention belongs to the technical field of oil-gas exploration and seismic data processing, and discloses a seismic data interpolation method and system based on a CBAM-Res2Unet network. According to the method, the problems that the receptive field is limited and low-level characteristics in jump connection are not selectively transmitted due to a fixed convolution kernel in a traditional ResUNet network are solved. The method comprises the following steps: generating a seismic record through wave equation forward modeling, and constructing a training set containing random missing data; a CBAM-Res2Unet network is designed, an encoder and a decoder of the CBAM-Res2Unet network adopt Res2Net modules containing compression excitation modules, and the multi-scale feature extraction capability is enhanced through grouping convolution and layered residual connection; introducing a CBAM module in jump connection, and adaptively weighting features by using a channel and a space attention mechanism; a spectrum normalization layer is added to the deep layer of the network to stabilize training. The system comprises a data acquisition module, a network generation module, a training module and an interpolation module. According to the invention, through multi-scale feature fusion and attention mechanism and stability optimization, high-precision interpolation of complex missing seismic data is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil and gas exploration and seismic data processing, and provides a seismic data interpolation method and system based on the CBAM-Res2Unet network. Background Art

[0002] Complex acquisition conditions usually lead to missing seismic traces, which will affect subsequent steps such as AVO analysis, time-lapse analysis, and fluid detection. Therefore, interpolation is necessary for a better understanding of the reservoir. To obtain reliable interpretation results, it is necessary to interpolate the missing seismic traces. Therefore, interpolation is of great significance.

[0003] The seismic data interpolation method based on the CBAM-Res2Unet network is one of the important methods for deep learning-based seismic data interpolation. This method uses the missing seismic data to make a training set to train the network, and uses the trained network to interpolate the missing seismic data. The CBAM-Res2Unet network not only helps to prevent overfitting, but also can perform interpolation more effectively. However, the fixed size of the convolutional kernel in the traditional ResUnet network results in a limited receptive field and the non-selective transmission of low-level features in the skip connection, making it difficult for the network to capture long-distance correlation information. This will affect the training effect of the network, and further affect the interpolation result, resulting in discontinuity in the interpolated seismic data.

[0004] Therefore, in the prior art, the seismic data interpolation method based on the ResUnet network has the following technical problems:

[0005] The fixed size of the convolutional kernel in the ResUnet network results in a limited receptive field and the non-selective transmission of low-level features in the skip connection, making it difficult for the network to capture long-distance correlation information. This leads to discontinuity in the interpolated seismic data, that is, the quality of the processing result is difficult to guarantee. Summary of the Invention

[0006] Aiming at the problems studied above, the purpose of the present invention is to provide a seismic data interpolation method and system based on the CBAM-Res2Unet network, and solve the problems that the fixed size of the convolutional kernel in the ResUnet network of the prior art results in a limited receptive field and the non-selective transmission of low-level features in the skip connection.

[0007] To achieve the above purpose, the present invention adopts the following technical solutions:

[0008] The present invention provides a seismic data interpolation method based on CBAM-Res2Unet, including the following steps:

[0009] Step 1: Obtain seismic data and form a training set and a validation set, where the seismic data is a seismic shot gather record;

[0010] Step 2: Obtain a preset CBAM-Res2Unet network, which includes an encoder module, a decoder module, and a skip connection layer, where:

[0011] The encoder module contains five layers, and each layer sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and a downsampling layer, where a spectral normalization layer is added outside the convolutional layer of the fifth layer;

[0012] The decoder module contains five layers, and each layer sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and an upsampling layer;

[0013] The skip connection layer sets a CBAM module between the corresponding levels of the encoder and the decoder. Each CBAM module enhances the feature expression ability by adaptively learning channel and spatial attention weights;

[0014] Step 3: Train and validate the CBAM-Res2Unet network based on the training set and the validation set:

[0015] Step 4: Interpolate the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

[0016] In the above technical solution, Step 1 includes the following steps:

[0017] Step 1.1: Perform forward numerical simulation based on the wave equation to obtain seismic records;

[0018] Step 1.2: Merge all the seismic records obtained in Step 1.1 to obtain a large seismic data set, and then generate two copies of it. Among them, Copy 1 is the seismic records with 30%, 50%, and 70% randomly missing, and Copy 2 is the corresponding complete data, that is, the target data;

[0019] Step 1.3: Randomly select multiple pairs of data from Copy 1 and Copy 2 with a preset grid size to form a training pair set A, and finally use the training pair set A as the final training set and validation set.

[0020] In the above technical solution, the preset CBAM-Res2Unet network in Step 2 includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence, and a skip connection layer for connecting the encoder and the decoder;

[0021] Among them, the encoder module: It contains a 5-layer connected encoder structure. Each layer of the encoder sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and a downsampling layer. Each layer includes a residual connection that directly adds the input feature map to the output of the convolutional layer of this layer. The residual connection includes a convolutional layer for capturing the feature image. A spectral normalization layer is added outside each of the two convolutional layers in the fifth layer (one convolutional layer before and one after the Res2Net module in the fifth layer).

[0022] The convolutional layer performs a convolution operation on the input feature map to retain detailed features and information. By means of feature extraction and residual connection, the size of the feature map is gradually reduced while important information is retained.

[0023] The Res2Net module performs a grouped convolution operation on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer.

[0024] The squeeze-and-excitation module is after the second convolutional layer and before the residual connection in the Res2Net module, and is used to weight the features of different channels to enhance the key features.

[0025] For the two convolutional layers in the fifth layer: A spectral normalization layer is added outside each of the two convolutional layers, which is used to enhance the stability of the network.

[0026] The skip connection layer includes: Skip connections including four CBAM modules, and each skip connection includes a CBAM module.

[0027] The CBAM module adaptively learns channel and spatial attention weights to improve the feature expression ability of the network and can capture the correlation between features in different dimensions.

[0028] The decoder module includes: A decoder connected by five convolutional layers and an upsampling layer. Each decoder includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and an upsampling layer.

[0029] In the above technical solution, the Res2Net module performs a grouped convolution operation on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer. The operations of the Res2Net module include:

[0030] The input feature map is divided into four groups. The convolutional output of the previous group is cascaded with the input feature map of the next group, and feature extraction is performed using a convolution operation. The receptive field is expanded through hierarchical residual connections.

[0031] Finally, all group outputs are fused through a 1×1 convolution, and its mathematical expression is:

[0032]

[0033] Among them, x represents the input feature map, F(x) represents the output feature map of the previous set of convolutions, and G i (x) represents the feature transformation operation of the i-th branch, and W i represents the weight of the i-th branch, and F(y) represents the final output.

[0034] Add the CBAM module to each skip connection layer. By adaptively learning the channel and spatial attention weights, the feature expression ability of the network can be improved, and the correlation between features can be captured in different dimensions. The operations of the CBAM module include:

[0035] Channel attention stage: Perform global max pooling and average pooling on the input feature map respectively, and generate channel weights through a shared MLP. The calculation formula is:

[0036]

[0037] Among them, M c (F) is the channel attention function, δ represents the activation function Sigmoid, F is the feature input, represents the average pooling operation of c channels for each channel, represents the max pooling operation of c channels, W0 and W1 are the network parameters in the multi-layer perceptron MLP respectively, AvgPool(F) represents global average pooling for each channel of the feature map F, and MaxPool(F) represents global max pooling for each channel of the feature map F;

[0038] Spatial attention stage: Perform max and average pooling on the channel-weighted feature map, and generate spatial weights through a 7×7 convolution. The calculation formula is:

[0039]

[0040] The final output is the double-weighted feature map of channel and spatial attention, and F″ is:

[0041]

[0042] Among them, M s (F) is the spatial attention function, and f 7×7 represents the 7x7 convolution kernel, where F′ is the feature after passing through the spatial attention mechanism, and F″ is the refined feature map, represents the global average pooling result of the input feature map F, represents the global max pooling result of the input feature map F.

[0043] The compression excitation module is used to weight the features of different channels. The operations of the compression excitation module include the following steps:

[0044] Step A: Perform a convolution operation on the input feature map X to generate a feature map U, expressed as:

[0045] F tr : X → U, X ∈ R W″*H″*C″ , U ∈ R W′*H′*C′

[0046] where F tr is a convolution transformation, where R represents the set of real numbers, W′ and W″ represent the widths, H′ and H″ represent the heights, C′ and C″ represent the number of channels. The convolution formula for the c-th channel of the feature map U is as follows:

[0047]

[0048] where v c represents the c-th convolution kernel, x s represents the s-th input covered by the current convolution kernel, u c represents the s-th output, and K represents the number of convolution kernels;

[0049] Step B: Perform a squeezing operation on the intermediate feature map U to obtain global information, and generate a channel description vector z ∈ R C , specifically calculated as:

[0050] where z c represents the average value of the c channels of this layer, and u c (i, j) represents the feature value of the c-th channel at the position (i, j);

[0051] Step C: Perform an excitation operation on the channel description vector z c , and generate a channel weight vector s c ∈ R C , and the calculation process is:

[0052] s c = sigmoid(W2 · ReLU(W1 · z c )) where W1 and W2 represent linear layers, and z represents the average value of the channels of each layer;

[0053] Step D: Perform a per-channel multiplication of the channel weight vector s and the intermediate feature map U, and output the refined feature map Calculated as:

[0054] where s c represents the weight value of the c-th channel, and u cIt is the corresponding channel feature map.

[0055] A spectral normalization layer is added outside each convolutional layer in the fifth layer to limit the spectral norm of the weight matrix in the fifth layer. The specific steps of the spectral normalization process are as follows:

[0056] Step A: Impose a Lipschitz constraint on the weight matrix of the fifth-layer convolutional layer

[0057] The mapping relationship of the convolutional layer is expressed as g: h in → h out , where h in represents the input data, and h out represents the output data. The Lipschitz constant is defined as:

[0058]

[0059] where h is the input feature map, ||·|| represents the L2 norm, represents the calculation for all non-zero input vectors h. The network mapping relationship is expressed in terms of the weight W as:

[0060] g(h) = Wh

[0061] According to the above formula, the Lipschitz constant is determined by the largest singular value σ max (W):

[0062] |g|Lip = σ max (W)

[0063] Step B: Use the power iteration method to calculate the spectral norm of the weight matrix (the largest singular value of the matrix)

[0064] Initialize the random vectors and Update the approximate maximum singular value through iteration:

[0065]

[0066] where, represents the left singular vector, represents the right singular vector, and T represents the transpose;

[0067] Calculate the approximate value of the spectral norm after iteration:

[0068]

[0069] Step C: Normalize the weight matrix

[0070] Based on the approximate spectral norm in Step B, perform a normalization operation on the weight matrix W:

[0071]

[0072] Replace the original weight matrix with the normalized weight matrix to ensure that the Lipschitz constant of the network layer is 1.

[0073] The present invention provides a seismic data interpolation system based on a CBAM-Res2Unet network, comprising:

[0074] An acquisition module: acquiring seismic data and forming a training set and a validation set, wherein the seismic data is a seismic shot gather record;

[0075] A network generation module: obtaining a preset CBAM-Res2Unet network;

[0076] A training module: training and validating the CBAM-Res2Unet network based on the training set and the validation set;

[0077] An interpolation module: interpolating the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

[0078] In the above system, the specific implementation steps of the acquisition module are as follows:

[0079] Step 1.1: Performing forward numerical simulation based on the wave equation to obtain seismic records;

[0080] Step 1.2: Merging all the seismic records obtained in Step 1.1 to obtain a large seismic data set, and then generating two copies thereof, wherein Copy 1 is the seismic record with 30%, 50% and 70% randomly missing, and Copy 2 is the corresponding complete data, that is, the target data;

[0081] Step 1.3: Randomly selecting multiple pairs of data from Copy 1 and Copy 2 with a preset size grid to form a training pair set A, and finally using the training pair set A as the final training set and validation set.

[0082] In the above system, the CBAM-Res2Unet network includes an input layer, an encoder module, a decoder module and an output layer connected in sequence, and a skip connection layer for connecting the encoder and the decoder;

[0083] Among them, the encoder module: includes 5 connected encoder structures, each layer of the encoder includes a convolutional layer, a Res2Net module containing a squeeze-and-excitation module and a downsampling layer in sequence, each layer includes a residual connection for directly adding the input feature map to the output of the convolutional layer of this layer, the residual connection includes a convolutional layer for capturing the feature image, and a spectral normalization layer is added outside each of the two convolutional layers of the fifth layer,

[0084] The convolutional layer performs convolutional operations on the input feature map to retain detailed features and information. Through feature extraction and residual connection, the size of the feature map is gradually reduced while important information is retained;

[0085] The Res2Net module performs grouped convolutional operations on the input feature map, improving the multi-scale representation ability at a finer granularity level and increasing the receptive field of each network layer;

[0086] The squeeze-and-excitation module is located after the second convolutional layer and before the residual connection in the Res2Net module, and is used to weight the features of different channels to enhance key features;

[0087] In the two convolutional layers of the fifth layer: a spectral normalization layer is added outside each of the two convolutional layers to enhance the stability of the network;

[0088] The skip connection layer includes: skip connections including four CBAM modules, and each skip connection includes a CBAM module,

[0089] The CBAM module adaptively learns channel and spatial attention weights to improve the feature expression ability of the network and can capture the correlation between features in different dimensions;

[0090] The decoder module includes: a decoder connected by five convolutional layers and an upsampling layer, and each decoder includes a convolutional layer, a Res2Net module containing a squeeze-and-excitation module, and an upsampling layer.

[0091] In the above system, the Res2Net module performs grouped convolutional operations on the input feature map, improving the multi-scale representation ability at a finer granularity level and increasing the receptive field of each network layer. The operations of the Res2Net module include:

[0092] The input feature map is divided into four groups. The output of the previous group of convolutions is concatenated with the input feature map of the next group, and feature extraction is performed using convolutional operations. The receptive field is expanded through hierarchical residual connections;

[0093] Finally, all group outputs are fused through 1×1 convolutions, and its mathematical expression is:

[0094]

[0095] Among them, x represents the input feature map, F(x) represents the output feature map of the previous group of convolutions, G i (x) represents the feature transformation operation of the i-th branch, W i represents the weight of the i-th branch, and F(y) represents the final output.

[0096] Add the CBAM module to each skip connection layer. By adaptively learning the channel and spatial attention weights, the feature expression ability of the network is improved, and the correlation between features can be captured in different dimensions. The operations of the CBAM module include:

[0097] Channel attention stage: Perform global max pooling and average pooling on the input feature map respectively, and generate channel weights through a shared MLP. The calculation formula is:

[0098]

[0099] where M c (F) is the channel attention function, δ represents the activation function Sigmoid, F is the feature input, represents the average pooling operation of c channels for each channel, represents the max pooling operation of c channels, W0 and W1 are the network parameters in the multi-layer perceptron MLP respectively, AvgPool(F) represents global average pooling for each channel of the feature map F, and MaxPool(F) represents global max pooling for each channel of the feature map F;

[0100] Spatial attention stage: Perform max and average pooling on the channel-weighted feature map, and generate spatial weights through a 7×7 convolution. The calculation formula is:

[0101]

[0102] The final output is the double-weighted feature map of channel and spatial attention, and F″ is:

[0103]

[0104] where, M s (F) is the spatial attention function, f 7×7 represents the 7x7 convolution kernel, where F′ is the feature after the spatial attention mechanism, and F″ is the refined feature map, represents the global average pooling result of the input feature map F, represents the global max pooling result of the input feature map F.

[0105] The squeeze-and-excitation module is used to weight the features of different channels. The operations of the squeeze-and-excitation module include the following steps:

[0106] Step A: Perform a convolution operation on the input feature map X to generate a feature map U, which is expressed as:

[0107] F tr : X→U, X∈R W″*H″*C″ , U∈R W′*H′*C′

[0108] Among them, F tr is a convolution transform, where R represents the set of real numbers, W′ and W″ represent widths, H′ and H″ represent heights, C′ and C″ represent the number of channels, and the convolution formula for the c-th channel of the feature map U is as follows:

[0109]

[0110] Among them, v c represents the c-th convolution kernel, x s represents the s-th input covered by the current convolution kernel, u c represents the s-th output, and K represents the number of convolution kernels;

[0111] Step B: Perform a squeezing operation on the intermediate feature map U to obtain global information, and generate a channel description vector z ∈ R C , and the specific calculation is:

[0112] Among them, z c represents the average value of the c channels of this layer, and u c (i, j) represents the eigenvalue of the c-th channel at position (i, j);

[0113] Step C: Perform an excitation operation on the channel description vector z c to generate a channel weight vector s c ∈ R C , and the calculation process is:

[0114] s c = sigmoid(W2·ReLU(W1·z c )) where W1 and W2 represent linear layers, and z represents the average value of the channels of each layer;

[0115] Step D: Perform a per-channel multiplication of the channel weight vector s and the intermediate feature map U, and output the refined feature map The calculation is:

[0116]

[0117] Among them, s c represents the weight value of the c-th channel, and u c is the corresponding channel feature map.

[0118] A spectral normalization layer is added outside each convolution layer of the fifth layer to limit the spectral norm of the weight matrix of the fifth layer. The specific steps of the spectral normalization process are:

[0119] Step A: Impose a Lipschitz constraint on the weight matrix of the fifth-layer convolution layer

[0120] The mapping relationship of the convolutional layer is represented as g: h in →h out , where h in represents the input data, and h out represents the output data. The Lipschitz constant is defined as:

[0121]

[0122] where h is the input feature map, ||·|| represents the L2 norm, represents the calculation for all non-zero input vectors h. The network mapping relationship is represented by the weight W as:

[0123] g(h) = Wh

[0124] According to the above formula, the Lipschitz constant is determined by the largest singular value σ max (W):

[0125] |g|Lip = σ max (W)

[0126] Step B: Use the power iteration method to calculate the spectral norm (the largest singular value of the matrix) of the weight matrix

[0127] Initialize the random vectors and Update the approximate largest singular value through iteration:

[0128]

[0129] where represents the left singular vector, represents the right singular vector, and T represents the transpose;

[0130] Calculate the approximate value of the spectral norm after iteration:

[0131]

[0132] Step C: Normalize the weight matrix

[0133] Based on the approximate spectral norm in Step B, perform a normalization operation on the weight matrix W:

[0134]

[0135] Replace the original weight W with the normalized weight matrix to ensure that the Lipschitz constant of the network layer is 1.

[0136] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0137] Interpolate seismic data using the CBAM-Res2Unet network. Based on ResUnet, a Res2Net module incorporating a squeeze-and-excitation module, a CBAM module, and a spectral normalization layer are added. The CBAM module adaptively learns channel and spatial attention weights to enhance the network's feature representation ability, capturing the correlation between features in different dimensions and thus improving the interpolation ability. The Res2Net module enhances the multi-scale representation ability at a finer granularity level and increases the receptive field of each network layer to improve the interpolation ability. The squeeze-and-excitation module explicitly models the interdependence between feature channels by learning the importance of each channel. It weights each channel according to the importance of the features on that channel and then highlights the key features, achieving the effect of improving the signal-to-noise ratio of seismic data interpolation. The spectral normalization layer decomposes the eigenvalues of the weight matrix and then normalizes them. During the optimization of the neural network, the parameter changes are more stable, less prone to gradient explosion, and the interpolation performance of the network is improved. That is, the CBAM-Res2Unet network technology obtained after improvement maximally preserves the weight matrix information; the squeeze-and-excitation module retains the detailed information of the image while realizing the function of extracting the segmented image. Through the improvement of the network performance, the present invention finally achieves the effect of improving the signal-to-noise ratio of seismic data interpolation. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 is a flowchart of a method for interpolating seismic data based on the CBAM-Res2Unet network of the present invention;

[0139] Figure 2 is the CBAM-Res2Unet network architecture according to the present invention;

[0140] Figure 3 is a seismic profile of the target data of the test data of the present invention;

[0141] Figure 4 is a seismic profile of the missing seismic data of the test data of the present invention;

[0142] Figure 5 is the interpolation result of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0143] The following will give a detailed description of the embodiments of the present invention. Although the present invention will be described and explained in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments only. On the contrary, any modifications or equivalent replacements made to the present invention shall be covered within the scope of the claims of the present invention.

[0144] In addition, for a better illustration of the present invention, numerous specific details are given in the following specific implementation manners. Those skilled in the art will understand that the present invention can also be implemented without these specific details.

[0145] Technical problems solved by the present invention: Solve the problems that the receptive field is limited due to the fixed size of the convolutional kernel in the existing ResUnet network and the non-selective transmission of low-level features in the skip connection, and improve the signal-to-noise ratio of seismic data interpolation.

[0146] A seismic data interpolation method based on a CBAM-Res2Unet network, comprising the following steps:

[0147] Step 1: Obtain seismic data and form a training set and a validation set. Among them, the seismic data is a seismic shot gather record; it includes the following steps:

[0148] Step 1.1: Perform forward numerical simulation based on the wave equation to obtain seismic records;

[0149] Step 1.2: Merge all the seismic records obtained in Step 1.1 to obtain a large seismic data collection, and then generate two copies of it. Among them, Copy 1 is the seismic record with 30%, 50%, and 70% randomly missing, and Copy 2 is the corresponding complete data, that is, the target data;

[0150] Step 1.3: Randomly select multiple pairs of data from Copy 1 and Copy 2 with a preset grid size to form a training pair set A, and finally use the training pair set A as the final training set and validation set.

[0151] Step 2: Obtain a preset CBAM-Res2Unet network; the main structure of the preset CBAM-Res2Unet network is the ResUnet network, which includes an encoder, a decoder, and a skip connection layer. Residual connections and Res2Net modules containing squeeze-and-excitation modules are adopted in the encoder and decoder, and a spectral normalization layer is added outside each of the two convolutional layers in the fifth layer of the network. At the same time, a CBAM module is added to each layer of the skip connection.

[0152] The specific steps are as follows:

[0153] The main network structure is constructed using the ResUnet network, including: five encoders and five decoders, that is, four encoders connected by downsampling and four decoders connected by upsampling, an input layer connected to the first encoder and an output layer connected to the fifth decoder. The first encoder to the fifth encoder are respectively jump-connected to the first decoder to the fifth decoder. Each encoder includes a convolutional layer and a Res2Net module containing a squeeze-and-excitation module, and each decoder includes an upsampling layer and a Res2Net module containing a squeeze-and-excitation module. Their convolutional kernel sizes are all 3×3. A CBAM module is included in the skip connection layer. The number of channels of the five encoders are 32, 64, 128, 256, 256, 512 in sequence, and the number of channels of the five decoders are 512, 256, 128, 64, 32 in sequence. A Res2Net module containing a squeeze-and-excitation module is added after the convolutional layer in each layer of the encoder and decoder to perform grouped convolution operation on the input feature map. The specific steps of grouped convolution are as follows:

[0154] We divided the input features into 4 groups. A set of filters first extracts features from a set of input feature maps, and then sends the output feature maps of the previous set and another set of input features Figure 1 to the next set of filters. This process is repeated several times until all input feature maps are processed. Finally, the feature maps of all groups are concatenated and sent to another 1×1 filter to fully fuse the information. Along any path from the input feature map to the output feature map, when passing through a 3×3 filter, the equivalent receptive field will increase, and due to the combination effect, many equivalent feature scales are obtained, that is:

[0155]

[0156] Among them, x represents the input feature map, F(x) represents the output feature map of the previous convolution, G i (x) represents the feature transformation operation of the i-th branch, W i represents the weight of the i-th branch, and F(y) represents the final output.

[0157] The CBAM module can adaptively learn channel and spatial attention weights. The specific steps are as follows:

[0158] First, the input feature map passes through the channel attention module. For the input feature map, global max pooling and global average pooling operations are first performed on each channel to calculate the maximum eigenvalue and average eigenvalue on each channel. Two vectors containing the number of channels are generated, representing the global maximum feature and average feature of each channel respectively. The feature vectors after global max pooling and average pooling are input into a shared fully connected layer. The fully connected layer is used to learn the attention weights for each channel. The global maximum feature vector and the average feature vector are intersected to obtain the final attention weight vector. To ensure that the attention weights are between 0 and 1, the Sigmoid activation function is applied to generate the channel attention weights. These weights will be applied to each channel of the original feature map. Using the obtained attention weights, multiply them with each channel of the original feature map to obtain the attention-weighted channel feature map. M c (F) is:

[0159]

[0160] where M c (F) is the channel attention function, δ represents the Sigmoid activation function, F is the feature input, represents the average pooling operation for c channels of each channel, represents the max pooling operation for c channels, W0 and W1 are the network parameters in the multi-layer perceptron MLP respectively, AvgPool(F) represents global average pooling for each channel of the feature map F, and MaxPool(F) represents global max pooling for each channel of the feature map F;

[0161] Spatial attention stage: Perform max and average pooling on the channel-weighted feature map, and generate spatial weights through a 7×7 convolution. The calculation formula is:

[0162]

[0163] The final output is the double-weighted feature map of channel and spatial attention, and F″ is:

[0164]

[0165] where, M s (F) is the spatial attention function, f 7×7 represents the 7x7 convolution kernel, where F′ is the feature after the spatial attention mechanism, and F″ is the refined feature map, represents the global average pooling result of the input feature map F, represents the global max pooling result of the input feature map F.

[0166] The squeeze-and-excitation module processes the weight information of different channels. The specific steps of the squeeze-and-excitation process are:

[0167] Step A: Perform a convolution operation on the input feature map X to generate a feature map U, expressed as:

[0168] F tr : X → U, X ∈ R W″*H″*C″ , U ∈ R W′*H′*C′

[0169] where F tr is a convolution transform, where R represents the set of real numbers, W′ and W″ represent widths, H′ and H″ represent heights, C′ and C″ represent the number of channels, and the convolution formula for the c-th channel of the feature map U is as follows:

[0170]

[0171] where v c represents the c-th convolution kernel, x s represents the s-th input covered by the current convolution kernel, u c represents the s-th output, and K represents the number of convolution kernels;

[0172] Step B: Perform a squeezing operation on the intermediate feature map U to obtain global information, and generate a channel description vector z ∈ R C , specifically calculated as:

[0173] where z c represents the average value of the c channels of this layer, and u c (i, j) represents the feature value of the c-th channel at position (i, j);

[0174] Step C: Perform an excitation operation on the channel description vector z c to generate a channel weight vector s c ∈ R C , and the calculation process is:

[0175] s c = sigmoid(W2 · ReLU(W1 · z c )) where W1 and W2 represent linear layers, and z represents the average value of the channels of each layer;

[0176] Step D: Perform a per-channel multiplication of the channel weight vector s and the intermediate feature map U, and output the refined feature map Calculated as:

[0177]

[0178] where s c represents the weight value of the c-th channel, and u c is the corresponding channel feature map.

[0179] Spectrum normalization layer, which restricts the spectral norm of the weight matrix of the fifth convolutional layer. The specific steps of spectrum normalization processing are as follows:

[0180] Step A: Impose Lipschitz constraint on the weight matrix of the fifth convolutional layer

[0181] The mapping relationship of the convolutional layer is expressed as g: h in →h out where h in represents the input data, and h out represents the output data. The Lipschitz constant is defined as:

[0182]

[0183] where h is the input feature map, ||·|| represents the L2 norm, represents the calculation for all non-zero input vectors h. The network mapping relationship is expressed in terms of the weight W as:

[0184] g(h) = Wh

[0185] According to the above formula, the Lipschitz constant is determined by the largest singular value σ max (W):

[0186] |g|Lip = σ max (W)

[0187] Step B: Use the power iteration method to calculate the spectral norm (the largest singular value of the matrix)

[0188] Initialize the random vectors and Update the approximate largest singular value through iteration:

[0189]

[0190] where represents the left singular vector, represents the right singular vector, and T represents the transpose;

[0191] Calculate the approximate value of the spectral norm after iteration:

[0192]

[0193] Step C: Normalize the weight matrix

[0194] Based on the approximate spectral norm in Step B, perform a normalization operation on the weight matrix W:

[0195]

[0196] The normalized weight matrix Replace the original weight W to ensure that the Lipschitz constant of the network layer is 1.

[0197] Step 3: Train and validate the network based on the training set and the validation set;

[0198] Step 4: Interpolate the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

[0199] Step 4.1: Select some seismic records of the model and create missing data for testing;

[0200] Step 4.2: Input the seismic data to be interpolated into the optimal model selected in Step 3.2 for testing.

[0201] Example 1

[0202] The present invention provides a seismic data interpolation system based on a CBAM-Res2Unet network, including:

[0203] Acquisition module: Obtain seismic data and form a training set and a validation set, where the seismic data is seismic shot gather records;

[0204] Network generation module: Obtain a preset CBAM-Res2Unet network;

[0205] Training module: Train and validate the CBAM-Res2Unet network based on the training set and the validation set;

[0206] Interpolation module: Interpolate the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

[0207] In the above system, the specific implementation steps of the acquisition module are:

[0208] Step 1.1: Perform forward numerical simulation based on the wave equation to obtain seismic records;

[0209] Step 1.2: Merge all the seismic records obtained in Step 1.1 to obtain a large seismic data collection, and then generate two copies of it. Among them, Copy 1 is the seismic records with 30%, 50%, and 70% randomly missing, and Copy 2 is the corresponding complete data, that is, the target data;

[0210] Step 1.3: Randomly select multiple pairs of data from Copy 1 and Copy 2 with a preset size grid to form a training pair set A, and finally use the training pair set A as the final training set and validation set.

[0211] In the above system, the CBAM-Res2Unet network includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence, as well as a skip connection layer for connecting the encoder and the decoder;

[0212] Among them, the encoder module: consists of 5 connected encoder structures. Each layer of the encoder includes a convolutional layer, a Res2Net module containing a squeeze-and-excitation module, and a downsampling layer in sequence. Each layer includes a residual connection for directly adding the input feature map to the output of the convolutional layer of this layer. The residual connection includes a convolutional layer for capturing feature images. A spectral normalization layer is added outside each of the two convolutional layers in the fifth layer.

[0213] The convolutional layer performs convolutional operations on the input feature map to retain detailed features and information. By means of feature extraction and residual connection, the size of the feature map is gradually reduced while important information is retained;

[0214] The Res2Net module performs grouped convolutional operations on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer;

[0215] The squeeze-and-excitation module is after the second convolutional layer in the Res2Net module and before the residual connection in the module, and is used to weight features of different channels to enhance key features;

[0216] In the two convolutional layers of the fifth layer: a spectral normalization layer is added outside each of the two convolutional layers, which is used to enhance the stability of the network;

[0217] The skip connection layer includes: skip connections including four CBAM modules, and each skip connection includes a CBAM module.

[0218] The CBAM module adaptively learns channel and spatial attention weights to improve the feature expression ability of the network and can capture the correlation between features in different dimensions;

[0219] The decoder module includes: a decoder connected with five convolutional layers and an upsampling layer. Each decoder includes a convolutional layer, a Res2Net module containing a squeeze-and-excitation module, and an upsampling layer.

[0220] In the above system, the Res2Net module performs grouped convolutional operations on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer. The operations of the Res2Net module include:

[0221] Dividing the input feature map into four groups, cascading the convolutional output of the previous group with the input feature map of the next group, performing feature extraction using convolutional operations, and expanding the receptive field through hierarchical residual connections.

[0222] Finally, all group outputs are fused through 1×1 convolution, and its mathematical expression is:

[0223]

[0224] Among them, x represents the input feature map, F(x) represents the output feature map of the previous group of convolution, and G i (x) represents the feature transformation operation of the i-th branch, and W i represents the weight of the i-th branch, and F(y) represents the final output.

[0225] Add the CBAM module to each skip connection layer. By adaptively learning the channel and spatial attention weights, the feature expression ability of the network is improved, and the correlation between features can be captured in different dimensions. The operations of the CBAM module include:

[0226] Channel attention stage: Perform global max pooling and average pooling on the input feature map respectively, and generate channel weights through a shared MLP. The calculation formula is:

[0227]

[0228] Among them, M c (F) is the channel attention function, δ represents the activation function Sigmoid, F is the feature input, represents the average pooling operation of c channels for each channel, represents the max pooling operation of c channels, W0 and W1 are the network parameters in the multi-layer perceptron MLP respectively, AvgPool(F) represents global average pooling for each channel of the feature map F, and MaxPool(F) represents global max pooling for each channel of the feature map F;

[0229] Spatial attention stage: Perform max and average pooling on the channel-weighted feature map, and generate spatial weights through a 7×7 convolution. The calculation formula is:

[0230]

[0231] The final output is the double-weighted feature map of channel and spatial attention, and F″ is:

[0232]

[0233] Among them, M s (F) is the spatial attention function, and f 7×7 represents the 7x7 convolution kernel, where F′ is the feature after the spatial attention mechanism, and F″ is the refined feature map, represents the global average pooling result of the input feature map F, Represents the global maximum pooling result of the input feature map F.

[0234] The squeeze-and-excitation module is used to weight the features of different channels. The operations of the squeeze-and-excitation module include the following steps:

[0235] Step A: Perform a convolution operation on the input feature map X to generate a feature map U, expressed as:

[0236] F tr : X → U, X ∈ R W″*H″*C″ , U ∈ R W′*H′*C′

[0237] where F tr is a convolution transformation, where R represents the set of real numbers, W′, W″ represent the widths, H′, H″ represent the heights, C, C represent the number of channels, and the convolution formula for the c-th channel of the feature map U is as follows:

[0238]

[0239] where v c represents the c-th convolution kernel, x s represents the s-th input covered by the current convolution kernel, u c represents the s-th output, and K represents the number of convolution kernels;

[0240] Step B: Perform a squeezing operation on the intermediate feature map U to obtain global information, and generate a channel description vector z ∈ R C , and the specific calculation is:

[0241] where z c represents the average value of the c channels of this layer, and u c (i, j) represents the feature value of the c-th channel at the position (i, j);

[0242] Step C: Perform an excitation operation on the channel description vector z c to generate a channel weight vector s c ∈ R C , and the calculation process is:

[0243] s c = sigmoid(W2 · ReLU(W1 · z c )) where W1, W2 represent linear layers, and z represents the average value of the channels of each layer;

[0244] Step D: Perform a per-channel multiplication of the channel weight vector s and the intermediate feature map U, and output the refined feature map The calculation is:

[0245]

[0246] where s c represents the weight value of the c-th channel, and u c is the corresponding channel feature map.

[0247] A spectral normalization layer is added outside each convolutional layer in the fifth layer to limit the spectral norm of the weight matrix of the fifth layer. The specific steps of the spectral normalization process are as follows:

[0248] Step A: Impose a Lipschitz constraint on the weight matrix of the fifth-layer convolutional layer

[0249] The mapping relationship of the convolutional layer is expressed as g: h in → h out , where h in represents the input data, and h out represents the output data. The Lipschitz constant is defined as:

[0250]

[0251] where h is the input feature map, ||·|| represents the L2 norm, represents the calculation for all non-zero input vectors h. The network mapping relationship is expressed in terms of the weight W as:

[0252] g(h) = Wh

[0253] According to the above formula, the Lipschitz constant is determined by the largest singular value σ max (W) of the matrix W:

[0254] |g|Lip = σ max (W)

[0255] Step B: Use the power iteration method to calculate the spectral norm (the largest singular value of the matrix) of the weight matrix

[0256] Initialize the random vectors and Update the approximate maximum singular value through iteration:

[0257]

[0258] where, represents the left singular vector, represents the right singular vector, and T represents the transpose;

[0259] Calculate the approximate value of the spectral norm after iteration:

[0260]

[0261] Step C: Normalize the weight matrix

[0262] Based on the approximate spectral norm in step B, perform a normalization operation on the weight matrix W:

[0263]

[0264] Replace the original weight W with the normalized weight matrix to ensure that the Lipschitz constant of the network layer is 1.

[0265] Since the above embodiments are adopted, the following key technical effects are summarized:

[0266] 1. Improve interpolation performance: This technical solution improves the performance of seismic data interpolation by introducing the Res2Net module, CBAM module, spectral normalization, and squeeze-and-excitation module into the ResUnet network. The CBAM module adaptively learns channel and spatial attention weights to improve the feature expression ability of the network, can capture the correlation between features in different dimensions, and thus enhance the interpolation ability. The Res2Net module improves the multi-scale representation ability at a finer-grained level and increases the receptive field of each network layer to improve the interpolation ability. The spectral normalization layer enhances the stability of the network by constraining the spectral norm of the weight matrix, while the squeeze-and-excitation module enhances the key features by weighting the features of different channels, thereby improving the interpolation effect.

[0267] 2. Improve network training stability: The traditional ResUnet network may have problems with unstable training during training. By adding a squeeze-and-excitation module and a spectral normalization layer to the network, this technical solution enhances the training stability of the network, reduces the risk of gradient explosion, and enables the network to learn the features of seismic data more stably.

[0268] 3. Retain detailed information: When processing the feature map, the CBAM module, Res2Net module, and squeeze-and-excitation module not only enhance the key features but also retain the detailed information of the image by increasing the receptive field. This makes the interpolated seismic data not only complete but also restores more details of the original data, which is beneficial for subsequent seismic data analysis and interpretation.

[0269] 4. Effectively handle complex data missing: The present invention constructs a training set by using seismic records generated by forward numerical simulation based on the wave equation, combined with seismic records with randomly added missing parts and complete seismic data. This method can effectively handle complex missing situations that may be encountered in the actual acquisition process, and improves the adaptability and practicality of the interpolation method.

[0270] 5. Systematic processing flow: The present invention not only includes the design of the network structure, but also covers the entire process from seismic data acquisition, training set construction to network training and interpolation processing. This systematic processing flow ensures the coherence and consistency from data preparation to the final interpolation result.

[0271] Compared with the prior art, the advantage of this technical solution is that through the improved network structure and training process, more stable and efficient seismic data interpolation is achieved, while more data details are retained, which is of great significance for improving the accuracy and efficiency of seismic data processing.

[0272] As Figure 4-5 shown, the interpolated result is relatively close to the target data, proving the correctness of this method. The present invention uses the improved CBAM-Res2Unet network to perform interpolation processing on seismic data. Based on the CBAM-Res2Unet network, more information can be extracted from the data, avoiding the problem of unstable network training. The interpolation performance of the traditional ResUnet network is significantly improved, achieving the effect of improving the signal-to-noise ratio of seismic data interpolation.

[0273] The above are only representative embodiments among the numerous specific application scopes of the present invention, and do not constitute any limitation to the protection scope of the present invention. Any technical solution formed by transformation or equivalent replacement falls within the scope of the present invention's rights protection.

Claims

1. A seismic data interpolation method based on the CBAM-Res2Unet network, characterized in that, It includes the following steps: Step 1: Obtain seismic data and form a training set and a validation set. The seismic data is seismic shot gather records; Step 2: Obtain a preset CBAM-Res2Unet network, which includes an encoder module, a decoder module, and a skip connection layer, where: The encoder module contains five layers. Each layer sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and a downsampling layer. A spectral normalization layer is added outside the convolutional layer of the fifth layer; The decoder module contains five layers. Each layer sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and an upsampling layer; The skip connection layer sets a CBAM module between the corresponding levels of the encoder and the decoder. Each CBAM module enhances the feature expression ability by adaptively learning channel and spatial attention weights; Step 3: Train and validate the CBAM-Res2Unet network based on the training set and the validation set: Step 4: Interpolate the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

2. The seismic data interpolation method based on the CBAM-Res2Unet network according to claim 1, characterized in that: The said Step 1 includes the following steps: Step 1.1: Perform forward numerical simulation based on the wave equation to obtain seismic records; Step 1.2: Merge all the seismic records obtained in Step 1.1 to get a large seismic data set, and then generate two copies. Among them, Copy 1 is the seismic records with 30%, 50%, and 70% randomly missing, and Copy 2 is the corresponding complete data, that is, the target data; Step 1.3: Randomly select multiple pairs of data from Copy 1 and Copy 2 with a preset grid size to form a training pair set A. Finally, use the training pair set A as the final training set and validation set.

3. The seismic data interpolation method based on the CBAM-Res2Unet network according to claim 2, characterized in that: The preset CBAM-Res2Unet network in the said Step 2 includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence, and a skip connection layer for connecting the encoder and the decoder; Among them, the encoder module: contains 5 connected encoder structures. Each encoder layer sequentially includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and a downsampling layer. Each layer includes a residual connection for directly adding the input feature map to the output of the convolutional layer of this layer. The residual connection includes a convolutional layer for capturing the feature image. A spectral normalization layer is added outside each of the two convolutional layers of the fifth layer, The convolutional layer performs convolutional operations on the input feature map to retain detailed features and information. By means of feature extraction and residual connection, the size of the feature map is gradually reduced while important information is retained; The Res2Net module performs grouped convolutional operations on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer; The squeeze-and-excitation module is after the second convolutional layer in the Res2Net module and before the residual connection in the module, and is used to weight the features of different channels to enhance the key features; In the two convolutional layers of the fifth layer: A spectral normalization layer is added outside each of the two convolutional layers to enhance the stability of the network; The skip connection layer includes: a skip connection including four CBAM modules, and each skip connection includes a CBAM module. The CBAM module adaptively learns channel and spatial attention weights to improve the feature expression ability of the network, and can capture the correlation between features in different dimensions. The decoder module includes: a decoder including five convolutional layers connected to an upsampling layer, and each decoder includes a convolutional layer, a Res2Net module containing a squeeze-and-excitation module, and an upsampling layer.

4. A seismic data interpolation method based on the CBAM-Res2Unet network according to claim 3, characterized in that: The operations of the Res2Net module include: Dividing the input feature map into four groups, cascading the convolutional output of the previous group with the input feature map of the next group, performing feature extraction using convolutional operations, and expanding the receptive field through hierarchical residual connections. Finally, fusing the outputs of all groups through a 1×1 convolution, and its mathematical expression is: Among them, x represents the input feature map, F(x) represents the output feature map of the previous set of convolutions, and G i (x) represents the feature transformation operation of the i-th branch, and W i represents the weight of the i-th branch, and F(y) represents the final output.

5. A seismic data interpolation method based on the CBAM-Res2Unet network according to claim 3, characterized in that: The operations of the CBAM module include: Channel attention stage: performing global max pooling and average pooling on the input feature map respectively, generating channel weights through a shared MLP, and the calculation formula is: Among which M c (F) is the channel attention function, δ represents the activation function Sigmoid, and F is the feature input. represents the average pooling operation of c channels for each channel. represents the max pooling operation of c channels, W0 and W1 are the network parameters in the multi-layer perceptron MLP respectively, AvgPool(F) represents global average pooling for each channel of the feature map F, and MaxPool(F) represents global max pooling for each channel of the feature map F. Spatial attention stage: performing max and average pooling on the channel-weighted feature map, generating spatial weights through a 7×7 convolution, and the calculation formula is: The final output is a feature map double-weighted by channel and spatial attention, and F″ is: Among them, M s (F) is the spatial attention function, f 7×7 represents a 7x7 convolutional kernel, where F′ is the feature after passing through the spatial attention mechanism, and F″ is the refined feature map. represents the global average pooling result of the input feature map F. represents the global maximum pooling result of the input feature map F.

6. The seismic data interpolation method based on the CBAM-Res2Unet network according to claim 3, wherein: The operations of the squeeze-and-excitation module include the following steps: Step A: performing a convolutional operation on the input feature map X to generate a feature map U, which is expressed as: F tr : X → U, X ∈ R W″*H″*C″ , U ∈ R W′*H′*C′ Among them, F tr is a convolution transform, where R represents the set of real numbers, W′ and W″ represent widths, H′ and H″ represent heights, C′ and C″ represent the number of channels, and the convolution formula for the c-th channel of the feature map U is as follows: where v c represents the c-th convolutional kernel, x s represents the s-th input covered by the current convolutional kernel, u c represents the s-th output, and K represents the number of convolutional kernels; Step B: Perform a squeezing operation on the intermediate feature map U to obtain global information, and generate a channel description vector z ∈ R C , and the specific calculation is as follows: where z c represents the average value of the c channels of this layer, and u c (i, j) represents the eigenvalue of the c-th channel at position (i, j); Step C: Perform an excitation operation on the channel description vector z c to generate a channel weight vector s through two fully-connected layers c ∈R C . The calculation process is as follows: s c = sigmoid(W2 · ReLU(W1 · z c )) where W1 and W2 represent linear layers, and z represents the channel average of each layer; Step D: Perform a channel-wise multiplication of the channel weight vector s and the intermediate feature map U, and output the refined feature map Calculated as: where s c represents the weight value of the c-th channel, and u c is the corresponding channel feature map.

7. A seismic data interpolation method based on the CBAM-Res2Unet network according to claim 3, characterized in that: A spectral normalization layer is added outside each convolutional layer in the fifth layer to limit the spectral norm of the weight matrix of the fifth layer. The specific steps of spectral normalization processing are: Step 7-1: imposing a Lipschitz constraint on the weight matrix of the fifth-layer convolutional layer The mapping relationship of the convolutional layer is expressed as g: h in →h out , where h in represents the input data, and h out represents the output data. The Lipschitz constant is defined as: where h is the input feature map, and ||·|| represents the L2 norm, which represents the calculation for all non-zero input vectors h, and the network mapping relationship is represented by the weight W as: g(h) = Wh According to the above formula, the Lipschitz constant is determined by the largest singular value σ max (W) of the matrix W: |g|Lip=σ max (W) Step 7-2: using the power iteration method to calculate the spectral norm (the largest singular value of the matrix) of the weight matrix Initialize the random vector and Update the approximate maximum singular value by iteration: Among them, represents the left singular vector, represents the right singular vector, and T represents the transpose; Calculating the approximate value of the spectral norm after iteration: Step 7-3: normalizing the weight matrix Based on the approximate spectral norm in Step 7-2, performing a normalization operation on the weight matrix W: Replace the original weight matrix with the normalized weight matrix to ensure that the Lipschitz constant of the network layer is 1.

8. A seismic data interpolation system based on the CBAM-Res2Unet network, characterized in that, including: Acquisition module: acquiring seismic data and forming a training set and a validation set, where the seismic data is seismic shot gather records. Network generation module: obtaining a preset CBAM-Res2Unet network. Training module: training and validating the CBAM-Res2Unet network based on the training set and the validation set. Interpolation module: interpolating the seismic data to be interpolated based on the trained CBAM-Res2Unet network.

9. The seismic data interpolation system based on the CBAM-Res2Unet network according to claim 8, wherein The specific implementation steps of the acquisition module are: Step 1.1: performing forward numerical simulation based on the wave equation to obtain seismic records. Step 1.2: merging all the seismic records obtained in Step 1.1 to obtain a large seismic data collection, and then generating two copies of it. Among them, copy 1 is the seismic records with 30%, 50%, and 70% randomly missing, and copy 2 is the corresponding complete data, that is, the target data. Step 1.3: randomly selecting multiple pairs of data from copy 1 and copy 2 with a preset grid size to form a training pair set A, and finally using the training pair set A as the final training set and validation set.

10. A seismic data interpolation system based on the CBAM-Res2Unet network according to claim 9, characterized in that, The CBAM-Res2Unet network includes an input layer, an encoder module, a decoder module, and an output layer connected in sequence, as well as a skip connection layer for connecting the encoder and the decoder; Among them, the encoder module: contains 5 connected encoder structures. Each layer of the encoder includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and a downsampling layer in sequence. Each layer includes a residual connection for directly adding the input feature map to the output of the convolutional layer of this layer. The residual connection includes a convolutional layer for capturing the feature image. A spectral normalization layer is added outside each of the two convolutional layers in the fifth layer. The convolutional layer performs a convolution operation on the input feature map to retain detailed features and information. By means of feature extraction and residual connection, the size of the feature map is gradually reduced while important information is retained; The Res2Net module performs a grouped convolution operation on the input feature map, improves the multi-scale representation ability at a finer granularity level, and increases the receptive field of each network layer; The squeeze-and-excitation module is after the second convolutional layer and before the residual connection in the Res2Net module, and is used to weight the features of different channels to enhance the key features; In the two convolutional layers of the fifth layer: A spectral normalization layer is added outside each of the two convolutional layers to enhance the stability of the network; The skip connection layer includes: skip connections including four CBAM modules, and each skip connection includes a CBAM module. The CBAM module adaptively learns the channel and spatial attention weights to improve the feature expression ability of the network and can capture the correlation between features in different dimensions; The decoder module includes: a decoder connected with five convolutional layers and an upsampling layer. Each decoder includes a convolutional layer, a Res2Net module with a squeeze-and-excitation module, and an upsampling layer.

Images