Super-resolution reconstruction method of seismic images based on deep residual attention mechanism

Abstract

The present invention discloses a method for super-resolution reconstruction of seismic images based on a deep residual attention mechanism, which relates to the field of seismic exploration technology and includes the following steps: constructing a data sample library for model training; constructing a super-resolution reconstruction network based on a residual attention mechanism; using the data sample library and a mixed loss function to train the super-resolution reconstruction network, obtaining a network model for realizing super-resolution reconstruction and denoising of seismic images, and completing the method for super-resolution reconstruction of seismic images based on a deep residual attention mechanism. The present invention solves the problem that the prior art regards denoising and super-resolution reconstruction as two independent parts, which increases the time complexity of the algorithm.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration technology, and in particular to a seismic image super-resolution reconstruction method based on a deep residual attention mechanism. Background Art

[0002] Seismic exploration is one of the important means of developing oil and gas reservoir resources. Compared with other geophysical exploration methods (gravity, magnetic, and electrical), seismic exploration can more accurately and detailedly characterize the structural information of the entire stratum. With the continuous exploitation of oil and gas resources, the targets of oil and gas seismic exploration have gradually shifted to complex, concealed, deep, and unconventional oil and gas reservoirs, which places higher requirements on the quality of seismic data. Affected by factors such as the complex natural environment, seismic data collected in the field will inevitably be interfered by a large amount of regular and irregular noise, which seriously affects the signal-to-noise ratio and resolution of seismic data, and thus affects the imaging of underground structures and the interpretation of seismic data. Therefore, suppressing the noise in seismic data and improving the resolution of seismic data are of great significance in the field of seismic data processing.

[0003] Current research regards denoising and super-resolution reconstruction as two independent parts, which increases the time complexity of the algorithm. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the seismic image super-resolution reconstruction method based on the deep residual attention mechanism provided by the present invention solves the problem that the prior art regards denoising and super-resolution reconstruction as two independent parts, which increases the time complexity of the algorithm.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a seismic image super-resolution reconstruction method based on deep residual attention mechanism, comprising the following steps:

[0006] S1: Build a data sample library for model training;

[0007] S2: Construct a super-resolution reconstruction network based on residual attention mechanism;

[0008] S3: The super-resolution reconstruction network is trained using a data sample library and a mixed loss function to obtain a network model for super-resolution reconstruction and denoising of seismic images, and to complete super-resolution reconstruction of seismic images based on a deep residual attention mechanism.

[0009] The beneficial effect of the above scheme is: the present invention proposes a super-resolution reconstruction network based on the residual attention mechanism, which is used to achieve random noise suppression and seismic profile resolution improvement of the seismic profile, and at the same time adopts a mixed loss function algorithm to train the network, so as to further improve the perceptual quality of the reconstructed image and obtain sharper seismic image edge information, thereby solving the problem that the prior art treats denoising and super-resolution reconstruction as two independent parts, which increases the time complexity of the algorithm.

[0010] Furthermore, the data sample library in S1 includes low-resolution images and high-resolution images.

[0011] The beneficial effect of the above further solution is: in order to solve the problem that network training depends on a large number of training data pairs, the network is trained by obtaining training image pairs with the same structure but different resolutions.

[0012] Furthermore, S1 includes the following sub-steps:

[0013] S2-1: Establish an initial reflectivity model. To simulate the fold and fault structures in real seismic wave images, add fold structures and plane fault structures to the initial reflectivity model.

[0014] S2-2: Convolve the reflectivity model with fold structure and plane fault structure with the Ricker wavelet, and highlight the high-frequency information to obtain the seismic data model;

[0015] S2-3: Perform 2D sectioning on the seismic data model to obtain high-resolution images as training input;

[0016] S2-4: Based on high-resolution images, a high-order degradation model is constructed to obtain low-resolution images with different degradation factors and a data sample library for model training.

[0017] The beneficial effects of the above further scheme are as follows: firstly, an initial reflectivity model is established, and the fold structure is simulated by a two-dimensional normal distribution function, and then plane faults are added to the model, each of which contains different dips and strikes to form fault-related folds. Finally, the generated reflectivity model is convolved with the Ricker wavelet to obtain a seismic data model, and a two-dimensional profile of the data is used as a high-resolution image for training input, and a high-order data degradation model is established to obtain rich low-resolution seismic data with different degradation factors, so as to obtain the data sample library required for training.

[0018] Furthermore, the wrinkle structure in S2-1 is simulated by a two-dimensional normal distribution function, and the formula is:

[0019]

[0020] Among them, S1(·) is the wrinkle structure, a is the position of vertical shear along the z direction, N z is the number of sampling points, N g is the number of control points for constructing the fold structure, b k is the size of the vertical fold structure, c k is the position of the wrinkle structure in the x direction, d k is the position of the wrinkle structure in the y direction, σ k is the radius of the fold structure, and k is the number of fold structures.

[0021] The beneficial effect of the above further scheme is: in order to simulate the folds and fault structures in the real seismic wave image, it is necessary to add the fold structure to the model, and the specific spatial fold structure is simulated by a two-dimensional normal distribution function.

[0022] Furthermore, the high-order degradation model in S2-4 is traversed twice, and each traversal performs blur processing, noise processing and downsampling processing in sequence. The blur processing is: the blur kernel is processed on the image in the form of convolution, and the formula is:

[0023]

[0024] U=[i,j] T

[0025]

[0026] Where k(·) is the blur kernel, i and j are the sliding window ranges, N is the normalization constant, U is the spatial coordinate, superscript T is the transpose of the matrix, Σ is the covariance matrix, R is the rotation matrix, σ1 and σ2 are the eigenvalues ​​of the covariance matrix, and θ is the degree of rotation.

[0027] The beneficial effect of the above further scheme is: in order to simulate some complex degradation problems (unknown noise and complex distortion), the classic degradation models are arranged and combined to construct a new high-order degradation model. The entire degradation process will perform the above three degradation kernels twice (here is because of the balance between simplicity and effectiveness) to construct a low-resolution data set.

[0028] Furthermore, the generator subnetwork RIR in the super-resolution reconstruction network in S2 includes a shallow feature extraction module, a deep feature extraction module, an upsampling module and a reconstruction layer;

[0029] The shallow feature extraction module uses a convolutional layer to extract shallow features F0, and the formula is:

[0030] F0=conv(I LR )

[0031] Among them, conv(·) is the convolution operation, I LR For input;

[0032] The deep feature extraction module is formed by cascading multiple residual groups RG, and a long skip connection LSC is added as a whole. The residual group RG is formed by cascading multiple residual channel attention blocks RCAB. The formula is:

[0033] F DF =H RIR (F0)

[0034] Among them, H RIR (·) is the entire composite function for deep feature extraction, F DF is the deep feature;

[0035] F deep =F0+conv(F g )=F0+conv g (H g (H g-1 (…(H1(F0))…)))

[0036] F g,b =H g,b (F g,b-1 )

[0037] F g =F g-1 +conv g,b (F g,b )

[0038] Among them, F deep For shallow and deep feature fusion, conv g (·) is the convolutional layer after the last RG, H g (·) is the function of the g-th RG, H g-1 (·) is the function of the g-1th RG, H1(·) is the function of the first RG, F g,b is the input and output of the bth RCAB in the gth RG, H g,b (·) is F g,b The corresponding function, F g,b-1 is the input and output of the b-1th RCAB in the gth RG, F g is the output of the g-th RG, F g-1 is the output of the g-1th RG, conv g,b (·) is the convolutional layer after the last RCAB in the g-th RG, F g,b is the output of the bth RCAB in the gth RG;

[0039] The upsampling module uses sub-pixel convolution ESPCN to output the sampled feature F UP for:

[0040] F UP =H UP (F DF )

[0041] Among them, H UP (·) is the upsampling module function;

[0042] The reconstruction layer uses the convolutional layer to reconstruct the enlarged features, and the formula is:

[0043] I SR =H REC (F UP )

[0044] Among them, I SR For RCAB output, H REC (·) is the convolution operator of the reconstruction layer.

[0045] The beneficial effect of the above further scheme is: the generator network in this scheme adopts the RIR structure to construct a deeper network to further extract high-level feature information. At the same time, the channel attention (CA) mechanism is introduced to adaptively learn the weights of different feature channels and automatically extract the degree of correlation dependency between each channel. The model can purposefully focus on information with larger weights, effectively improving the feature extraction capability in the network.

[0046] Furthermore, the hybrid loss function L in S3 SR Including pixel level loss L1, perceptual loss and relative GAN loss The formula is:

[0047]

[0048] Among them, η, λ and α are hyperparameters for balancing different loss functions;

[0049]

[0050] Among them, W and H are the sizes of RCAB output, C is the upsampling multiple, is the L1 loss of high-resolution image pixels, L1 loss for reconstructed image pixels;

[0051]

[0052] Among them, W i,j and H i,j is the dimension of each feature map in the feature extraction network, φ i,j(·) is the feature map extracted before the i-th maximum pooling layer and after the j-th convolutional layer is activated, I HR For high-resolution images, To reconstruct the image;

[0053]

[0054] in, is the mean operation of the real data, is the mean operation of the reconstructed data, D Ra (·) is the relative probability predicted by the discriminator, x r is the real image, x f It is a fake image.

[0055] The beneficial effect of the above further scheme is that the loss function in this scheme exists in a synthetic form. In addition to the pixel-level loss defined by L1 loss, perceptual loss and adversarial loss are also introduced. The final loss is obtained by balancing and accumulating the weight coefficients of different losses. The common super-resolution effect with pixel-level loss as the objective function is too smooth and not sharp enough in boundary areas such as faults and folds. This problem is greatly improved after adding perceptual loss and GAN. The GAN loss function can be adjusted and optimized for different tasks and data. By introducing an adversarial training mechanism, the generator and discriminator are jointly optimized in competition and cooperation, which helps the generator better capture the true distribution of the data and generate more realistic samples.

[0056] Furthermore, relative to GAN loss Obtained through a relative discriminator, the relative discriminator is:

[0057]

[0058] Among them, sigmoid(·) is the activation function and C(·) is the unconverted discriminator output.

[0059] The beneficial effect of the above further scheme is that: this scheme also designs a relative discriminator, which predicts that the true image has a higher probability of being true than the false image. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Flowchart of a seismic image super-resolution reconstruction method based on deep residual attention mechanism.

[0061] Figure 2 This is the structural diagram of the high-order degradation model.

[0062] Figure 3 Flowchart of the architecture for training data.

[0063] Figure 4This is the structure diagram of the generator subnetwork.

[0064] Figure 5 This is a diagram of the RIR structure.

[0065] Figure 6 This is the RCAB structure diagram.

[0066] Figure 7 Flowchart of the channel attention mechanism operation.

[0067] Figure 8 This is the image reconstruction effect diagram.

[0068] Fig. 9 This is a graph showing the changes in the PSNR value of the validation set during data training.

[0069] Fig.10 Comparison chart of image reconstruction effects with different resolutions. DETAILED DESCRIPTION

[0070] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0071] like Figure 1 As shown, a seismic image super-resolution reconstruction method based on a deep residual attention mechanism comprises the following steps:

[0072] S1: Build a data sample library for model training;

[0073] S2: Construct a super-resolution reconstruction network based on residual attention mechanism;

[0074] S3: The super-resolution reconstruction network is trained using a data sample library and a mixed loss function to obtain a network model for super-resolution reconstruction and denoising of seismic images, and to complete super-resolution reconstruction of seismic images based on a deep residual attention mechanism.

[0075] The data sample library in S1 includes low-resolution images and high-resolution images.

[0076] S1 includes the following sub-steps:

[0077] S2-1: Establish an initial reflectivity model. To simulate the fold and fault structures in real seismic wave images, add fold structures and plane fault structures to the initial reflectivity model.

[0078] S2-2: Convolve the reflectivity model with fold structure and plane fault structure with the Ricker wavelet, and highlight the high-frequency information to obtain the seismic data model;

[0079] S2-3: Perform 2D sectioning on the seismic data model to obtain high-resolution images as training input;

[0080] S2-4: Based on high-resolution images, a high-order degradation model is constructed to obtain low-resolution images with different degradation factors and a data sample library for model training.

[0081] The wrinkle structure in S2-1 is simulated by a two-dimensional normal distribution function, and the formula is:

[0082]

[0083] Among them, S1(·) is the wrinkle structure, a is the position of vertical shear along the z direction, N z is the number of sampling points, N g is the number of control points for constructing the fold structure, b k is the size of the vertical fold structure, c k is the position of the wrinkle structure in the x direction, d k is the position of the wrinkle structure in the y direction, σ k is the radius of the fold structure, and k is the number of fold structures.

[0084] The high-order degradation model in S2-4 is traversed twice, and each traversal performs blur processing, noise processing and downsampling processing in sequence. The blur processing is: the blur kernel is used to process the image in the form of convolution, and the formula is:

[0085]

[0086] U=[i,j] T

[0087]

[0088] Where k(·) is the blur kernel, i and j are the sliding window ranges, N is the normalization constant, U is the spatial coordinate, superscript T is the transpose of the matrix, Σ is the covariance matrix, R is the rotation matrix, σ1 and σ2 are the eigenvalues ​​of the covariance matrix, and θ is the degree of rotation.

[0089] In one embodiment of the present invention, in order to simulate some complex degradation problems (unknown noise and complex distortion), the classic degradation models are arranged and combined to construct a new high-order degradation model. The degradation process of this solution includes three types of processing: Blur processing, Noise processing, and Resize processing, as follows Figure 2 As shown, the Blur scheme and the Resize interpolation method are selected by random probability, the noise distribution randomly selects Gaussian noise or random noise, and the noise form selects a single channel.

[0090] In this embodiment, the downsampling ratio is 2, the size of the input high-resolution seismic data set is 256×256, and the size of the low-resolution seismic data set is 128×128. According to the above data generation process, a total of 3200 pairs of 2D training data pairs are generated, of which 200 pairs are used to test the model effect, and the remaining 3000 pairs are used as training sets and validation sets in a ratio of 7:3. The training set contains 2100 pairs of 2D data slices. In order to prevent the problem of memory explosion during training, the training data will be cropped into small blocks and sent to the network. Through the comparison of experimental results, the low-resolution input block size is balanced to adopt a size of 96×96. Similarly, the high-resolution input block size is 192×192. In addition, simple geometric transformations are used to increase the diversity of the data, such as common data enhancement methods such as random flipping. The overall training process is as follows Figure 3 shown.

[0091] like Figure 4 As shown, the generator subnetwork RIR in the super-resolution reconstruction network in S2 includes a shallow feature extraction module, a deep feature extraction module, an upsampling module and a reconstruction layer;

[0092] The shallow feature extraction module uses a convolutional layer to extract shallow features F0, and the formula is:

[0093] F0=conv(I LR )

[0094] Among them, conv(·) is the convolution operation, I LR For input;

[0095] The deep feature extraction module is formed by cascading multiple residual groups RG, and a long skip connection LSC is added as a whole. The residual group RG is formed by cascading multiple residual channel attention blocks RCAB. The formula is:

[0096] F DF =H RIR (F0)

[0097] Among them, H RIR (·) is the entire composite function for deep feature extraction, F DF is the deep feature;

[0098] F deep =F0+conv(F g )=F0+conv g (H g (H g-1 (…(H1(F0))…)))

[0099] F g,b =H g,b (F g,b-1 )

[0100] F g =F g-1 +conv g,b (F g,b )

[0101] Among them, F deep For shallow and deep feature fusion, conv g (·) is the convolutional layer after the last RG, H g (·) is the function of the g-th RG, H g-1 (·) is the function of the g-1th RG, H1(·) is the function of the first RG, F g,b is the input and output of the bth RCAB in the gth RG, H g,b (·) is F g,b The corresponding function, F g,b-1 is the input and output of the b-1th RCAB in the gth RG, F g is the output of the g-th RG, F g-1 is the output of the g-1th RG, conv g,b (·) is the convolutional layer after the last RCAB in the g-th RG, F g,b is the output of the bth RCAB in the gth RG;

[0102] Each RG is composed of multiple residual channel attention blocks (RCAB), which can obtain a larger receptive field. Because different channels contain different feature information, some channels' feature maps contain important information, while some channels may contain invalid or even unfavorable feature maps for image reconstruction. Therefore, by adding a channel attention mechanism to each residual block, different reduction factors are given to feature maps of different channels, so that the feature maps of different channels in the entire feature map have different effects on the output.

[0103] The upsampling module uses sub-pixel convolution ESPCN to output the sampled feature F UP for:

[0104] F UP =H UP (F DF )

[0105] Among them, H UP (·) is the upsampling module function;

[0106] The reconstruction layer uses the convolutional layer to reconstruct the enlarged features, and the formula is:

[0107] ISR =H REC (F UP )

[0108] Among them, I SR For RCAB output, H REC (·) is the convolution operator of the reconstruction layer.

[0109] This part obtains a high-resolution feature map through convolution and reorganization between multiple channels. It is an efficient, fast, and parameter-free pixel rearrangement upsampling method. Choosing this upsampling method not only reduces training time and saves GPU memory, but also its upsampling method only at the end of the model retains more texture areas in the low-resolution space, achieving better reconstruction effects.

[0110] The RIR (Residual in Residual) structure is used in the network structure of this solution, which adds low-level feature information to high-level feature information through the LSC structure. RIR is a structure in which the residual group RG contains the residual channel attention block RCAB. From a macroscopic point of view, the overall structure is composed of G RG modules cascaded, and finally a long jump connection is used to fuse the shallow and deep features, such as Figure 5 The lower part shown is a structural diagram of RG composition.

[0111] The channel attention mechanism helps the model learn feature expressions better by assigning different weights to each channel. This network can be easily integrated into the existing network to improve network performance at a low cost. Figure 6 The figure shows the structure of RCAB. After the data stream enters, it will go through two skip connections. The first one is directly added to the output layer, which is the same as the ordinary Residual Block. The second skip connection is to perform channel attention operation after convolution, and rescale the previous features through the learned channel attention sub- Figure 7 As shown, the specific channel attention operation is described as follows:

[0112] First, different channel features x c It will go through a global pooling to become a channel description operator z with c channels c , the process can be expressed as:

[0113]

[0114] Then pass through a downsampling layer W D and upsampling layer W U , which is implemented by two l×l convolutions, using sigmoid and ReLu activation functions to increase its nonlinearity between scaling, and finally obtaining the channel statistics s:

[0115] s=f(W U δ(W D z))

[0116] Multiply the input original channel data by the channel statistics to get the scaled channel information for:

[0117]

[0118] Where f(·) is the S-type gate, δ(·) is the ReLU activation function, z is the channel statistic, s c is the scaling factor of the C channel.

[0119] Mixed loss function L in S3 SR Including pixel level loss L1, perceptual loss and relative GAN loss The formula is:

[0120]

[0121] Among them, η, λ and α are hyperparameters for balancing different loss functions;

[0122]

[0123] Among them, W and H are the sizes of RCAB output, C is the upsampling multiple, is the L1 loss of high-resolution image pixels, L1 loss for reconstructed image pixels;

[0124]

[0125] Among them, W i,j and H i,j is the dimension of each feature map in the feature extraction network, φ i,j (·) is the feature map extracted before the i-th maximum pooling layer and after the j-th convolutional layer is activated, I HR For high-resolution images, To reconstruct the image;

[0126]

[0127] in, is the mean operation of the real data, To average the data of each mini-batch, D Ra (·) is the relative probability predicted by the discriminator, x r is the real image, x f It is a fake image.

[0128] Relative GAN loss Obtained through a relative discriminator, the relative discriminator is:

[0129]

[0130] Among them, sigmoid(·) is the activation function and C(·) is the unconverted discriminator output.

[0131] In this embodiment, Figure 8 As shown in Figure 1, (a) is the reconstruction result containing only L1 loss, and (b) is the weighted synthetic loss result of this scheme. Ablation analysis is performed on the reconstruction result containing only L1 loss, the addition of GAN without introducing perceptual loss, and the mixed loss of this scheme. The objective evaluation indicators are compared, and the peak PSNR average values ​​of all test sets are taken for comparison, as shown in Table 1.

[0132] Table 1 Evaluation indicators of the same dataset under different loss combinations

[0133]

[0134] It can be seen from the above table that the objective evaluation index PSNR using the mixed loss form of this scheme reaches the optimal value.

[0135] Using the above training data generation method, a total of 3200 pairs of 2D training data pairs were generated, of which 200 pairs were used to test the model effect, and the remaining 3000 pairs were used as training sets and validation sets in a ratio of 7:3. The validation set was used to update the PSNR performance in each epoch. In order to ensure better convergence of the network, the data input to the network was normalized and preprocessed. Then the data enhancement processing mentioned above was performed; the training used the ADAM optimizer, in which the parameters β1 = 0.9, β2 = 0.999, ε = 10 -8 , the initial learning rate step is 1e-4, the batchsize value is 4, and the maximum number of training epochs is 150, such as Fig. 9 As shown, the change process of loss during training is Fig. 9 In (a), the loss of the training set decreases significantly in the early stage of training, and the subsequent validation set loss and training set loss gradually stabilize. The PSNR value of the validation set changes during the training process. Fig. 9 (b) The experiment was trained using an NVIDIA 3080 graphics card.

[0136] In order to ensure the effectiveness of the model, the reconstruction effect test was carried out on an independent test set, and the low-resolution seismic slices were used as input to obtain the reconstruction effect. Fig.10As shown in the figure, (a) is the corresponding low-resolution seismic slice image, (b) is the original high-resolution image, and (c) is the reconstructed seismic image. In order to ensure the effectiveness of the model, we selected 2D seismic image slices from different 3D seismic data models. It can be seen that the noise is effectively removed, and the fault boundary effectively retains high-frequency information, and the boundary and other areas have achieved a sharper effect.

[0137] The present invention targets the characteristics of seismic data. The generator network in the algorithm adopts the RIR structure to extract higher-level feature information of deeper seismic data, and introduces the residual block structure RCAB with channel attention, so that the network weights tend to channels with more stratigraphic features, thereby improving the reconstruction effect of the generator. At the same time, we use a mixed loss function algorithm to train the network, introduce a perceptual loss function to further improve the perceptual quality of the reconstructed image, and use the adversarial loss RaGAN to obtain sharper seismic image edge information. In view of the problem that network training depends on a large number of training data pairs, synthetic high-resolution seismic data is used, and a high-order data degradation model is established to obtain low-resolution seismic data with different degradation factors. The analysis of the experimental processing results of synthetic seismic data and actual data shows that the proposed workflow can significantly improve the quality perception of the original data, obtain an ideal signal-to-noise ratio, further improve the resolution effect of the reconstructed profile image, and construct a more accurate structural and stratigraphic feature map than the original seismic image.

[0138] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the invention.

Claims

1. A super-resolution reconstruction method for seismic images based on deep residual attention mechanism, characterized in that: The following steps are involved: S1: Build a data sample library for model training; S2: Construct a super-resolution reconstruction network based on residual attention mechanism; S3: The super-resolution reconstruction network is trained using a data sample library and a mixed loss function to obtain a network model for realizing super-resolution reconstruction and denoising of seismic images, and to complete super-resolution reconstruction of seismic images based on a deep residual attention mechanism; The S1 includes the following sub-steps: S1-1: Establish an initial reflectivity model. To simulate the fold and fault structures in real seismic wave images, add fold structures and plane fault structures to the initial reflectivity model. S1-2: Convolve the reflectivity model with fold structure and plane fault structure with the Ricker wavelet, and highlight the high-frequency information to obtain the seismic data model; S1-3: Perform 2D sectioning of the seismic data model to obtain high-resolution images as training input; S1-4: Based on high-resolution images, a high-order degradation model is constructed to obtain low-resolution images with different degradation factors and a data sample library for model training; The high-order degradation model in S1-4 is traversed twice, and each traversal performs blur processing, noise processing and downsampling processing in sequence. The blur processing is: the blur kernel is processed on the image in the form of convolution, and the formula is: U=[i1,j1] T Where k(·) is the blur kernel, i1 and j1 are the sliding window ranges, N is the normalization constant, U is the spatial coordinate, superscript T is the transpose of the matrix, Σ is the covariance matrix, R is the rotation matrix, σ1 and σ2 are the eigenvalues ​​of the covariance matrix, and θ is the degree of rotation; The generator subnetwork RIR in the super-resolution reconstruction network in S2 includes a shallow feature extraction module, a deep feature extraction module, an upsampling module and a reconstruction layer; The shallow feature extraction module uses a convolutional layer to extract shallow features F0, and the formula is: F0=conv(I LR ) Among them, conv(·) is the convolution operation, I LR is the input; The deep feature extraction module is formed by cascading multiple residual groups RG, and a long skip connection LSC is added as a whole. The residual group RG is formed by cascading multiple residual channel attention blocks RCAB. The formula is: F DF =H RIR (F0) Among them, H RIR (·) is the entire composite function for deep feature extraction, F DF is the deep feature; F deep =F0+conv(F g )=F0+conv g (H g (H g-1 (…(H1(F0))…))) F g,b =H g,b (F g,b-1 ) F g =F g-1 +conv g,b (F g,b ) Among them, F deep For the fusion of shallow and deep features, conv g (·) is the convolutional layer after the last RG, H g (·) is the function of the g-th RG, H g-1 (·) is the function of the g-1th RG, H1(·) is the function of the first RG, F g,b is the input and output of the bth RCAB in the gth RG, H g,b (·) is F g,b The corresponding function, F g,b-1 is the input and output of the b-1th RCAB in the gth RG, F g is the output of the g-th RG, F g-1 is the output of the g-1th RG, conv g,b (·) is the convolutional layer after the last RCAB in the g-th RG; The upsampling module uses sub-pixel convolution ESPCN to output the sampled feature F UP for: F UP =H UP (F DF ) Among them, H UP (·) is the upsampling module function; The reconstruction layer uses the convolutional layer to reconstruct the enlarged features, and the formula is: I SR =H REC (F UP ) Among them, I SR For RCAB output, H REC (·) is the convolution operator of the reconstruction layer.

2. The method for super-resolution reconstruction of seismic images based on deep residual attention mechanism according to claim 1, characterized in that: The data sample library in S1 includes low-resolution images and high-resolution images.

3. The method for super-resolution reconstruction of seismic images based on deep residual attention mechanism according to claim 1, characterized in that: The wrinkle structure in S1-1 is simulated by a two-dimensional normal distribution function, and the formula is: Among them, S1(·) is the wrinkle structure, a is the position of vertical shear along the z direction, N z is the number of sampling points, N g is the number of control points for constructing the fold structure, b k is the size of the vertical fold structure, c k is the position of the wrinkle structure in the x direction, d k is the position of the wrinkle structure in the y direction, σ k is the radius of the fold structure, and k is the number of fold structures.

4. The method for super-resolution reconstruction of seismic images based on deep residual attention mechanism according to claim 1, characterized in that: The mixed loss function L in S3 SR Including pixel level loss L1, perceptual loss and relative GAN loss The formula is: Among them, η, λ and α are hyperparameters for balancing different loss functions; Among them, W and H are the sizes of RCAB output, C is the upsampling multiple, is the L1 loss of high-resolution image pixels, L1 loss for reconstructed image pixels; Among them, W i,j and H i,j is the dimension of each feature map in the feature extraction network, φ i,j (·) is the feature map extracted before the i-th maximum pooling layer and after the j-th convolutional layer is activated, I HR For high-resolution images, To reconstruct the image; in, is the mean operation of the real data, To average the data of each mini-batch, D Ra (·) is the relative probability predicted by the discriminator, x r is the real image, x f It is a fake image.

5. The method for super-resolution reconstruction of seismic images based on deep residual attention mechanism according to claim 4, characterized in that: The relative GAN loss Obtained through a relative discriminator, the relative discriminator is: Among them, sigmoid(·) is the activation function and C(·) is the unconverted discriminator output.

Images