Seismic image super-resolution reconstruction method based on double-branch multiple attention network

Abstract

The invention discloses a seismic image super-resolution reconstruction method based on a double-branch multiple attention network, and belongs to the technical field of seismic exploration, and the method comprises the steps: constructing a data sample library for model training; adopting a super-resolution backbone network and a gradient branch auxiliary network to construct a double-branch multiple attention network model; and training the double-branch multiple attention network model by adopting the data sample library and the mixed loss function to obtain a network model for realizing seismic image super-resolution reconstruction and high-frequency texture information reconstruction, and performing reconstruction based on the trained network model to obtain a seismic image subjected to super-resolution reconstruction. According to the method, a gradient branch is introduced as an auxiliary network, so that the super-resolution reconstruction effect is enhanced. Meanwhile, a mixed loss function is adopted to train the network, and a loss function for constraining high-frequency information is particularly introduced, so that the common structural deformation problem in a super-resolution task is effectively solved, the generation of artifacts is prevented, and a new technical approach is provided for high-precision seismic interpretation.

Description

Technical Field

[0001] This application relates to the field of seismic exploration technology, and in particular to a method for super-resolution reconstruction of seismic images based on a bi-branch multi-attention network. Background Technology

[0002] With the continuous expansion of my country's oil and gas exploration, exploration targets have gradually shifted to concealed oil and gas reservoirs in structurally complex areas and stratigraphic lithological traps. Exploration depths have also extended from shallow to deep and ultra-deep layers. This trend places more stringent demands on seismic data processing technologies. Deep geological environments exhibit significant complexity, primarily manifested in: significant attenuation of effective seismic signal energy at depth, exacerbated spatial aliasing effects, and a sharp decrease in signal-to-noise ratio. Against this backdrop, improving the regularity, signal-to-noise ratio, resolution, and fidelity of seismic data can provide a reliable data foundation for subsequent key processes such as migration imaging, full-waveform inversion, and seismic interpretation, thereby effectively supporting the identification and evaluation of complex oil and gas reservoirs.

[0003] In the field of super-resolution reconstruction of seismic data, structural distortion and artifact suppression are core issues that urgently need to be addressed. Specifically, structural distortion can lead to geometrical distortion or edge blurring of key geological structures such as stratigraphic interfaces and fault systems; artifact interference is mainly manifested as abnormal enhancement of high-frequency noise, and false signals can severely mask the true seismic response, resulting in multiple interpretations of geological feature identification. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this application provides a seismic image super-resolution reconstruction method based on a bi-branch multi-attention network, which solves the problems of geometric distortion or edge blurring and artifact interference that are difficult to avoid in the process of seismic data super-resolution reconstruction.

[0005] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides a seismic image super-resolution reconstruction method based on a dual-branch multi-attention network, including: S1: Construct a data sample library for model training; S2: A dual-branch multi-attention network model is constructed using a super-resolution backbone network and a gradient branch auxiliary network; S3: The dual-branch multi-attention network model is trained using a data sample library and a hybrid loss function to obtain a network model for realizing super-resolution reconstruction of seismic images and reconstruction of high-frequency texture information. Based on the trained network model, super-resolution reconstruction of seismic images is performed to obtain the super-resolution reconstructed seismic images.

[0006] Furthermore, the dual-branch multi-attention network model includes a super-resolution backbone network, a gradient branch auxiliary network, and a feature fusion module; Both the super-resolution backbone network and the gradient branch auxiliary network include a shallow feature extraction module, a deep feature extraction module, an upsampling module, and a reconstruction layer; the super-resolution backbone network also includes a feature fusion module. The shallow feature extraction module uses convolutional layers to extract shallow features from the super-resolution backbone network and gradient branch auxiliary network. The calculation formula is as follows:

[0007]

[0008] In the formula, This refers to the shallow features extracted from the super-resolution backbone network. The shallow features of the gradient branch auxiliary network are extracted. For convolution operations, For low-resolution input images, This serves as the input to the gradient branch auxiliary network; The deep feature extraction module consists of multiple cascaded residual groups and a long skip connection. Each residual group contains multiple residual channel attention blocks. The deep feature extraction module extracts deep features from the super-resolution backbone network and the gradient branch auxiliary network, calculated using the following formula:

[0009]

[0010] In the formula, For the first super-resolution backbone network Deep features output from each residual set, For the gradient branch auxiliary network Deep features output from each residual set, and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals For the first super-resolution backbone network Input of one set of residuals, For the gradient branch auxiliary network Input of one set of residuals, For the first The output of the attention block structure for each residual channel in each residual group. For the first The output of the attention block structure for each residual channel in each residual group; The upsampling module employs subpixel convolution and fuses shallow and deep features through skip connections, outputting the sampled features:

[0011]

[0012] In the formula, and These are the sampled features of the super-resolution backbone network and the gradient branch auxiliary network, respectively. and These are the sampling operators on the super-resolution backbone network and the gradient branch auxiliary network, respectively; The reconstruction layer uses a convolutional layer to reconstruct the amplified features, as shown in the formula:

[0013]

[0014] In the formula, and These are the reconstructed and amplified features of the super-resolution backbone network and the gradient branch auxiliary network, respectively. and These are the convolution operators for the super-resolution backbone network and the gradient branch auxiliary network, respectively. and For super-resolution backbone network and gradient branch auxiliary network; The feature fusion module employs a weight-adjusted fusion method to fuse the reconstructed and amplified features from the super-resolution backbone network and the gradient branch auxiliary network, resulting in the final super-resolution reconstructed seismic image. The formula is as follows:

[0015] In the formula, To reconstruct fused features from seismic images at super-resolution resolution, For weight generation function, For element-wise multiplication, This is an element-wise addition.

[0016] Furthermore, the input to the gradient branch auxiliary network is the image gradient; The expression for the image gradient is:

[0017] in, For image gradient, Let be the partial derivative of the image in the horizontal direction. Let be the partial derivative of the image in the vertical direction. For earthquake images, In the horizontal direction, The vertical direction.

[0018] Furthermore, the hybrid loss function includes a pixel loss function, a perceptual loss function, and an adversarial loss function; The hybrid loss function The expression is:

[0019] In the formula, , and These represent the SR image perceptual loss weight coefficients, super-resolution image pixel loss weight coefficients, and super-resolution image gradient map loss weight coefficients, respectively. and The adversarial loss weight coefficients are the super-resolution image and the gradient map of the super-resolution image. The loss weights applied to the super-resolution gradient map output by the gradient branch auxiliary network. For the perceptual loss function, The loss of the super-resolution image in the super-resolution backbone network. The loss on the gradient map of the super-resolution image is given by the gradient branching auxiliary network. To provide adversarial loss for super-resolution backbone networks, For the adversarial loss applied to the gradient branch auxiliary network, The loss is applied to the SR gradient map output by the gradient branch auxiliary network; The formula for calculating the pixel loss function is as follows:

[0020]

[0021]

[0022] In the formula, To calculate the expected value of the loss for all generated super-resolution images SR, Calculate the expected value of the loss for all generated super-resolution gradient maps. The sum of absolute values Input images at high resolution; The expression for the perceptual loss function is:

[0023] In the formula, For Euclidean distance For the VGG feature extraction network, the first Feature map after layer extraction; The expression for the adversarial loss function is:

[0024]

[0025] In the formula, For the discriminator of the super-resolution backbone network, It serves as the discriminator for gradient branch auxiliary networks.

[0026] Furthermore, the data sample library includes low-resolution images and high-resolution images.

[0027] The beneficial effects of this application are: This application presents a seismic image super-resolution reconstruction method based on a bi-branch multi-attention network. By introducing a gradient branch as an auxiliary network, the method enhances the super-resolution reconstruction effect. Simultaneously, a hybrid loss function is used to train the network, with a loss function specifically introduced to constrain high-frequency information. This effectively addresses the structural deformation problem commonly encountered in super-resolution tasks, further preventing artifact generation and providing a new technical approach for high-precision seismic interpretation. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0029] Figure 1 This is a flowchart illustrating a seismic image super-resolution reconstruction method based on a dual-branch multi-attention network, as provided in an embodiment of this application.

[0030] Figure 2 This is a schematic diagram illustrating the overall algorithm flow of a seismic image super-resolution reconstruction method based on a dual-branch multi-attention network, provided in an embodiment of this application.

[0031] Figure 3 This is an overall structural diagram of a dual-branch multi-attention network provided in an embodiment of this application.

[0032] Figure 4 The diagram shows the RIR and RG structures provided in the embodiments of this application.

[0033] Figure 5 This is a schematic diagram of the RCAB structure provided in an embodiment of this application.

[0034] Figure 6 This is a schematic diagram of feature fusion provided for an embodiment of this application.

[0035] Figure 7 This is an example of seismic image edge information provided in this application.

[0036] Figure 8 A diagram illustrating the data modeling process provided in this application embodiment.

[0037] Figure 9 This is a data processing flowchart provided for an embodiment of this application.

[0038] Figure 10 A comparison diagram of the super-resolution reconstruction results of seismic images and their high-frequency information provided in the embodiments of this application.

[0039] Figure 11 A comparison diagram of amplitude characteristics provided for embodiments of this application.

[0040] Figure 12 A diagram showing the training data recording provided in an embodiment of this application.

[0041] Figure 13 A comparison chart of super-resolution reconstruction results provided in the embodiments of this application.

[0042] Figure 14 A comparison diagram of high-frequency information of seismic images provided in the embodiments of this application.

[0043] Figure 15 Comparison of earthquake image reconstruction effects provided in the embodiments of this application.

[0044] Figure 16 Two sets of field seismic images provided for embodiments of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0046] Example 1: The combination of super-resolution (SR) and multi-task learning (MTL) is an important research direction in computer vision, aiming to improve the performance of super-resolution models by jointly optimizing multiple related tasks. MTL can share useful information from multiple learning tasks, thereby helping each task achieve more accurate predictions. For example, it can extract common features from images by sharing convolutional layers, design specific branches for different tasks, and utilize common joint loss functions, such as introducing multi-task losses (e.g., ...). L 1 / L The weighted sum of 2 loss and perceptual loss is used to balance the optimization objectives of different tasks. High-frequency information recovery is a current challenge in SR processing, especially in SR tasks using seismic data. During the super-resolution reconstruction of seismic data, excessive enhancement of high-frequency information or misprocessing of noise by the algorithm may introduce artifacts, resulting in false anomalous signals, discontinuous stratigraphic interfaces, or distorted waveform features, thereby reducing the reliability of the reconstruction results. Complete high-frequency information is usually reflected in the edge information of the image, and extracting the edge information of the image can be used to assist the SR task of the original image. In seismic image data, edge information is an important component of geological models, which proves the rationality of using edge image reconstruction as an auxiliary branch of multi-task learning, and the gradient of the image reflects the edge, texture, and detail information of the image.

[0047] Based on this, embodiments of this application provide a method for super-resolution reconstruction of seismic images based on a dual-branch multi-attention network, which can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a seismic image super-resolution reconstruction method based on a dual-branch multi-attention network provided in this application, comprising: S1: Construct a data sample library for model training, which includes low-resolution and high-resolution images.

[0048] S2: A dual-branch multi-attention network model is constructed using a super-resolution backbone network and a gradient branch auxiliary network.

[0049] In one embodiment of this application, the network model consists of two parts: a super-resolution backbone network and a gradient branch auxiliary network. The backbone network focuses on the super-resolution reconstruction task, while the auxiliary gradient branch network is used to reconstruct high-frequency texture information to further improve the detail recovery capability of the reconstruction results. The overall algorithm is as follows: Figure 2 As shown.

[0050] The feature extraction of both the backbone and branch networks in this application is based on an improvement of RCAN. Both branches use the same network structure, and the parameters within the network structure remain consistent, such as the kernel size, padding, and stride in the convolutional layers. This is done to ensure that the receptive field size of the gradient branch and the main branch is essentially the same. The RCAN structure used in this application is a typical super-resolution reconstruction network. This network effectively fuses shallow and deep features. Its structure consists of multiple cascaded residual sets. The overall algorithm structure is as follows: Figure 3 As shown, it includes a shallow feature extraction module, a deep feature extraction module, an upsampling module, a reconstruction layer, and a feature fusion module.

[0051] First, shallow feature extraction is performed. This application uses a... Convolutional layers extract shallow features from the backbone network and auxiliary branches respectively, using... This represents the shallow feature extraction process of the backbone network. This represents the shallow feature extraction process of the auxiliary network. This represents the convolution operation. This indicates a low-resolution input image. This represents the input to the auxiliary branch network, through... This is achieved by extracting high-frequency information. The shallow feature extraction operation is shown in the following formula:

[0052]

[0053] Secondly, the backbone networks of the SR branch and the gradient branch are each composed of m cascaded residual groups (RGs), and a long skip connect (LSC) is added to fuse shallow and deep features. Fusion is achieved through pixel-level accumulation. Each RG consists of N residual channel attention blocks (RCABs). Deep feature extraction from the LR image and gradient map can be represented as:

[0054]

[0055] In the formula, For the first super-resolution backbone network Deep features output from each residual set, For the gradient branch auxiliary network Deep features output from each residual set, and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals For the first super-resolution backbone network Input of one set of residuals, For the gradient branch auxiliary network Input of one set of residuals, For the first The output of the attention block structure for each residual channel in each residual group. For the first The output of the attention block structure for each residual channel in each residual group.

[0056] like Figure 4 The diagram shows the RG structure, where each RG structure contains multiple RCAB modules. RCAB is a residual block structure with channel attention. In traditional CNN structures, features from different channels are assigned the same weights for training. However, the feature information of each channel is actually different, and it is desirable for the model to purposefully focus on information with larger weights to improve the feature extraction capability of the network. The channel attention mechanism helps the model learn feature representations better by assigning different weights to each channel. This type of network can be easily integrated into existing networks to improve network performance at a relatively low cost.

[0057] like Figure 5 The diagram shows the structure of RCAB. After the data stream enters, it goes through two skip connections. The first one goes directly to the output layer for addition, which is the same as a normal residual block. The second skip connection performs channel attention operation after convolution, downsampling the previous features through the learned channel attention. The specific channel attention operation is described below.

[0058] A1: Different channel characteristics It will go through a global pooling process to become Channel descriptor for each channel This process can be represented as:

[0059] in, This is a global average pooling operation. and The spatial dimensions of the feature map are height and width. and For the first The spatial location of each channel The pixel value at that location.

[0060] A2: Through a downsampling layer and upsampling layer This operation is performed through two The convolution implementation uses sigmoid and ReLU activation functions between scaling operations to increase its non-linearity, ultimately yielding channel statistics. :

[0061] in, It is the sigmoid activation function. It is the ReLU activation function. For all channels The channel description vector is composed of these vectors.

[0062] A3: Multiply the input raw channel data by the channel statistic to obtain the scaled channel information. :

[0063] in, This is the raw channel data.

[0064] Then, the two branches pass through and , and Skip connections are used for feature fusion between shallow and deep layers. The upsampling module employs sub-pixel convolution ESPCN, assuming... and These correspond to the SR branch and gradient branch upsampling operators, respectively, and the output sampled features are denoted as... and :

[0065]

[0066] This part obtains high-resolution feature maps through convolution and multi-channel recombination. It's 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 preserves more texture areas in the low-resolution space by upsampling only at the end of the model, achieving better reconstruction results. Finally, a convolutional layer reconstructs the amplified features. and This represents the convolution operator for the SR branch and gradient branch reconstruction layers. and These represent the SR branch and the gradient branch, respectively:

[0067]

[0068] Finally, to avoid domain conflicts caused by directly concatenating feature maps from the gradient domain and the image domain, and to further and more effectively fuse features, a weight-adjusted fusion method is adopted for feature fusion. This method uses two features as guidance, allowing the network to perform soft selection, and determines their respective weights through training, such as... Figure 6 As shown, X and Y are the outputs of the gradient branch and the backbone network after long skip connections, respectively. One output of the gradient branch is fused with the output of the backbone network to form the final backbone network output, corresponding to Z in the figure. The feature fusion formula is:

[0069] in, For weight generation function, For element-wise multiplication, This is an element-wise addition.

[0070] S3: The dual-branch multi-attention network model is trained using a data sample library and a hybrid loss function to obtain a network model for realizing super-resolution reconstruction of seismic images and reconstruction of high-frequency texture information. Based on the trained network model, super-resolution reconstruction of seismic images is performed to obtain the super-resolution reconstructed seismic images.

[0071] In one embodiment of this application, the loss is still presented as a fusion loss, and the network is trained by weighted summation of pixel loss, perceptual loss, and adversarial loss. This method enables the network to produce super-resolution reconstruction results that are more consistent with human visual perception. Specifically, the pixel loss in this application employs... Using this loss as a pixel loss ensures sharper edges in the reconstructed image and also guarantees better convergence during training. The formula for calculating the loss function is:

[0072]

[0073]

[0074] in, and These correspond to the SR image of the backbone network and the loss on the gradient map of the SR image, respectively. The loss is the effect of the SR gradient map output by the gradient branch, used... This indicates that the operation was performed via the network. This indicates the gradient calculation operation. To calculate the expected value of the loss for all generated super-resolution images SR, Calculate the expected value of the loss for all generated super-resolution gradient maps. for The paradigm, or the sum of absolute values, Given a high-resolution input image, it can be seen that the pixel objective function includes the pixel loss of the SR reconstruction result and the HR seismic image from the backbone network, the pixel loss of the gradient extracted from the SR reconstruction result and the gradient extracted from the HR image, and the pixel loss of the gradient of the gradient branch reconstruction and the gradient extracted from the HR image.

[0075] Perceptual loss is a concept proposed in SRGAN. It uses a pre-trained VGG network to extract features from reconstructed and high-resolution images, and defines the Euclidean distance between the extracted features as the perceptual loss.

[0076] in, For the perceptual loss function, for The paradigm, namely Euclidean distance, For the VGG feature extraction network, the first Feature map after layer extraction.

[0077] To ensure that the reconstruction results better match the characteristics of human visual perception, this application introduces an adversarial loss mechanism in the super-resolution task. Since the network architecture contains a dual-branch structure, the optimization objective involves two discriminators: one part is used for discriminative optimization of the super-resolution (SR) image generated by the backbone network, and this discriminator is denoted as... Another part is used for the discriminant optimization of gradient branch generation of gradient graphs, and this discriminator is denoted as... The corresponding adversarial loss functions can be expressed as follows:

[0078]

[0079]

[0080]

[0081] in, and It refers to the countermeasures against losses applied to the backbone network. and This is an adversarial loss applied to the gradient branch. To fully utilize gradient information, the multi-task learning dual-branch network designed in this application adopts a highly symmetrical structure, ensuring that the features extracted by the backbone network and the gradient branch at the same depth are consistent at the semantic level. This symmetrical design can serve as a high-frequency constraint term, enhancing the consistency of feature representations and improving the network's ability to recover high-frequency details, thereby significantly improving the quality of super-resolution reconstruction.

[0082] Final Mixed Loss Function The form is:

[0083] in, , and These represent the SR image perceptual loss weight coefficients, super-resolution image pixel loss weight coefficients, and super-resolution image gradient map loss weight coefficients of the super-resolution backbone network, with values ​​of 1, 0.01, and 0.01, respectively. and The adversarial loss weight coefficients for the super-resolution image and its gradient map are 0.005 and 0.005, respectively. The loss weight coefficient applied to the super-resolution gradient map output by the gradient branch auxiliary network is 0.5.

[0084] In one embodiment of this application, gradient information plays a crucial role in super-resolution learning, particularly in seismic data reconstruction. It significantly enhances the recovery of high-frequency details, strengthens structural consistency, and suppresses artifacts. By rationally designing gradient branches and loss functions, gradient information can be fully utilized to achieve higher-quality seismic data super-resolution reconstruction. In super-resolution learning, gradient information is an important auxiliary feature, reflecting the rate of change of pixel values ​​or amplitudes in an image or data. It is typically used to capture structural features such as edges, textures, and high-frequency details. In seismic data super-resolution reconstruction, gradient information can highlight high-frequency components in the data (such as edges and abrupt changes), helping the model better recover details such as stratigraphic interfaces, fault boundaries, and small-scale geological features. On one hand, gradient branches recover the gradient map of the super-resolution seismic image, providing additional structural priors for the SR reconstruction process. On the other hand, a gradient loss function is proposed, using the image's gradient information as additional supervision and constraint, equivalent to imposing a secondary constraint on the super-resolution image. Combined with the previous image space loss function, the gradient space objective helps the generative network focus more on the geometric structure. The model can more accurately reconstruct the details in the seismic data, avoid over-smoothing or distortion, maintain the structural consistency of the data, and prevent artifacts or structural deformation.

[0085] Assuming earthquake edge information is represented by a one-dimensional graph, such as... Figure 7 As shown, (a) represents the boundary information of the HR seismic image; (d) represents the gradient rendering state of the boundary portion of the HR seismic image. Common deep learning-based super-resolution reconstruction schemes usually only include tasks with pixel loss function as the objective function, and can only generate... Figure 7 The blurred boundary effect shown in (b) has gradient information as follows: Figure 7As shown in (e); to ensure that the reconstructed SR image contains clearer tonal slicing and sharper contour information, this application uses gradient information to constrain the optimization objective function, enabling the model to learn clearer local detail information from the gradient space. Therefore, the application will be as follows: Figure 7 The gradient extraction information shown in (f) is also used as input to the model for training, making the reconstruction result closer to... Figure 7 The presentation shown in (c) demonstrates that this multi-task learning approach can not only avoid... Figure 7 The transition smoothing effect shown in (b) can effectively avoid the problem of over-sharpening geometric deformation caused by adding GAN, thus generating a more realistic SR seismic image effect.

[0086] The goal of the gradient branch is to learn the mapping process from LR seismic images to HR seismic images. The gradient includes both magnitude and direction. In this design, the direction information of the gradient is not considered because the magnitude of the gradient is sufficient to represent local details in the region during reconstruction. Therefore, this application only selects the magnitude of the gradient as the gradient map input. Assume an image is represented as... The gradient is represented as Then the difference formula is:

[0087] in, The partial derivative of the image in the horizontal direction (horizontal gradient component) represents the rate of change of that pixel in the horizontal direction. The partial derivative of the image in the vertical direction (vertical gradient component) represents the rate of change of that pixel in the vertical direction. In the horizontal direction, The vertical direction.

[0088] As can be seen from the above formula, the image gradient refers to the rate of change of a pixel in the x and y directions (compared to neighboring pixels). It is a two-dimensional vector composed of two components: the change along the X-axis and the change along the Y-axis. The change along the X-axis is the difference between the pixel value to the right (X+1) and the pixel value to the left (X-1). The change along the Y-axis is the difference between the pixel value below (Y+1) and the pixel value above (Y-1). It can be seen that the image gradient can be calculated using a convolution kernel.

[0089] in, To calculate the vertical gradient, To calculate the horizontal gradient.

[0090] This gradient map can be used as input to the gradient branch and as the target of the auxiliary task, jointly trained with the main task (super-resolution reconstruction). By simultaneously optimizing the reconstruction results and gradient information, the model can learn richer feature representations, thereby improving overall performance. Meanwhile, gradient information is sensitive to noise; through reasonable network design and loss functions, gradient information can be used to distinguish noise from real signals, thus suppressing noise during reconstruction. The recovered gradients can be integrated into the super-resolution branch, providing structural priors for super-resolution. Furthermore, gradients can highlight areas requiring greater attention to sharpness and structure, thus explicitly guiding the generation of high-quality images. The gradient branch structure in this application connects several intermediate branches in the SR pathway, fusing gradient information with original data features as network input or intermediate features to enhance the model's ability to perceive details, ensuring that the SR backbone contains sufficient structural or boundary information.

[0091] Example 2: This application's embodiments conducted reconstruction experiments using synthetic seismic data. For the process of generating massive amounts of synthetic seismic data of different types, this application adopts the methods proposed by Wu (2017) 34 and Li et al. (2020). The data modeling process is as follows: Figure 8 As shown, the process includes reflection coefficient modeling, adding wrinkle construction, adding faults, and Ricker wavelet convolution. Using the training data generation method described above, a total of 3200 pairs of 2D training data were generated. Of these, 200 pairs were used to test the model's performance, and the remaining 3000 pairs were used as the training and validation sets in a 7:3 ratio. The validation set was used to update the PSNR performance at each epoch. To ensure better network convergence, the input data to the network underwent normalization preprocessing, as follows:

[0092] in, and For each 2D data slice, find the maximum and minimum values. Slice the data into 2D data. This is the data after normalization.

[0093] The data processing flowchart of the algorithm provided in this embodiment is as follows: Figure 9 As shown, the two-dimensional profile of the data is used as the training input HR and LR images. The gradient convolution kernel described above is used to extract the boundary information of LR and HR. The LR gradient map and HR gradient map in the corresponding image are shown in the image. At the same time, the SR image output by the network and the gradient information of the SR image are limited by pixel loss. Furthermore, perceptual loss and adversarial loss are introduced to better preserve the complete edge information.

[0094] Training uses the ADAM optimizer, where parameters are set. , , The batch size is 4, the maximum number of training epochs is 150, and the initial learning rate step size is [value missing]. Each interval The learning rate will decay in each iteration, decreasing to its original value. The network was tested on a validation set to visually observe the changes in PSNR. The final network structure was determined through experiments. The deep learning framework used in this embodiment is PyTorch, and the experiments were trained using an NVIDIA 3080 graphics card.

[0095] To ensure the model's effectiveness, its reconstruction performance was evaluated on a synthetic data test set. Specifically, low-resolution seismic slices were used as input, and the reconstruction results were observed. Figure 10 As shown, (a) is the original low-resolution seismic image, (b) is the seismic image processed by commercial software at high resolution, and (c) is the reconstructed super-resolution seismic image. Simultaneously, Sobel operations are performed on the images to extract simple boundary information, such as... Figure 10 As shown in (d), (e), and (f), (d) is the low-resolution high-frequency information map, (e) is the software-processed high-frequency information map, and (f) is the super-resolution reconstructed high-frequency information map. To verify the reliability of the model, 2D seismic image slices from different 3D seismic data models were specifically selected for testing. The results show that the model successfully removed noise while effectively preserving the high-frequency information of fault boundaries, resulting in a clearer and sharper appearance in areas such as the boundaries.

[0096] This application also compares the amplitude characteristics of the super-resolution reconstruction results of the synthetic seismic slices with those of the original high-resolution seismic slices. Amplitude characteristic plots for three samples are shown, as follows. Figure 11 The traces shown are as follows: the blue traces represent the amplitude curves of the original seismic image slices, while the orange traces represent the amplitude curves of the super-resolution reconstructed seismic image slices. Observation reveals that their characteristics are nearly identical across the entire sampling range.

[0097] The objective evaluation metric is to compare the average peak PSNR (Peak Signal to Noise Ratio, PSNR) and SSIM (Structure Similarity Image Measure, SSIM) of all test sets, as shown in the table. PSNR is an image quality assessment metric used to evaluate the difference between two images. The calculation process for PSNR values ​​in this application is as follows:

[0098] In the above formula, The maximum value for each set of earthquake data. Mean squared error (MSE) measures the squared average of the pixel-level differences between the reconstructed image and the original image. The smaller the value, the closer the two images are. For super-resolution image reconstruction at location Pixel value at that location, For the original high-resolution real image at location Pixel value at that location, and This indicates the width and height of the image.

[0099] SSIM values ​​are derived from brightness. Contrast ,structure The result is obtained by multiplying these three factors together. The calculation process is as follows:

[0100] In the formula, , These represent the mean values ​​of the super-resolution image and the high-resolution image, respectively. , Used to represent the standard deviation between images The covariance between images and It is a stability constant used to prevent instability when the denominator approaches zero.

[0101] like Figure 12 As shown, the fusion loss during training is visualized, and the change process of the fusion loss between the training set and the validation set during training is illustrated through... Figure 12 In the data for record 'a', the training set loss decreases significantly in the early stages of training, and then the validation set loss and training set loss gradually stabilize. The PSNR value of the validation set changes during training through... Figure 12 Record b in the middle.

[0102] Example 3: This application's embodiments conducted reconstruction experiments using actual data. Taking the seismic data of Block F3 as an example, F3 is a block in the Dutch part of the North Sea, used for exploring Upper Jurassic to Lower Cretaceous strata. Large-scale S-shaped bedding is very evident, exhibiting textbook-level lower, upper, lateral, and truncation structures. It consists of sediments from a large river delta system, with overall high porosity (20-33%) and some carbonate cemented bands (Alaudah et al., 2019). The original F3 dataset is quite noisy, allowing for the differentiation of several seismic facies: transparent, chaotic, linear, and scaly. The transparent facies consists of relatively homogeneous rocks, possibly sandstone or shale; the chaotic facies likely represents sliding sediments; and the scaly facies at the base of the slope has been shown to be composed of sandy turbidity currents.

[0103] This application uses the Dutch F3 block as the seismic image dataset. To demonstrate the effectiveness of the method using limited seismic data for training, the original seismic data is selected as the label HR data. Then, each training sample is downsampled by a factor of 2 to obtain low-resolution data as input LR data, forming training data pairs. To facilitate the description and discussion of the seismic data, the temporal inline and xline numbers based on the training data blocks are reassigned to 1-701 and 1-401, respectively, for 256 time sampling points.

[0104] To verify the effectiveness of the model, this application validated the reconstruction results on the F3 test set, using the original seismic data as HR data and the 2x downsampled data as LR data to form training data pairs. Figure 13 As shown, the images correspond to the super-resolution reconstruction results, where (a) is the LR seismic image, (b) is the original seismic image, and (c) is the reconstructed seismic image. To ensure the effectiveness of the model, the selected 2D seismic image slices are from 3D seismic data models of different regions. The reconstruction results show that the model not only removes noise but also effectively preserves high-frequency information of the boundaries.

[0105] For high-frequency information extraction, the Sobel operator is first applied to calculate the gradient information of pixels. To further eliminate interference from non-edge pixels, a non-edge information suppression operation is performed on the gradient image obtained using Sobel. Specifically, the gradient magnitude is compared for each point, and the local maximum value is retained. A minimum threshold and a maximum threshold are set. The edge values ​​less than the minimum threshold are assigned 0. Simultaneously, based on the minimum and maximum thresholds, the gradient information can be divided into strong edges, weak edges, and non-edges. Non-edges are set to 0, and the remaining strong and weak edges are connected to form complete edge information. Figure 14As shown, the leftmost image is the LR seismic image, the middle image is the original seismic image, and the rightmost image is the reconstructed seismic image. The algorithm in this application reconstructs more complete and richer high-frequency edge information. This high-frequency information further reflects the fold and fault information of the seismic profile data, and it can be clearly seen that the high-frequency edge information of the reconstruction result is sharper and clearer.

[0106] To further verify the effectiveness of the proposed method, this application not only compared various objective experimental metrics to comprehensively evaluate its performance advantages, but also intuitively compared the actual reconstruction results of real seismic images. The performance of the multi-task learning model in restoring seismic image details, edge preservation, and noise suppression demonstrates its superiority over traditional super-resolution methods. Figure 15 The image compares the reconstruction results of traditional super-resolution algorithms with those of this application. The left side shows the reconstruction performance of the proposed model (PSNR=30.87, SSIM=0.90), while the right side shows the reconstruction performance of the traditional Bicubic algorithm (PSNR=29.25, SSIM=0.83). As can be seen from the image, the reconstructed image on the right performs poorly in terms of detail reconstruction, effectively demonstrating the effectiveness of the proposed algorithm in reconstructing high-frequency information.

[0107] Furthermore, to further verify the effectiveness of the model in this application, a comparison of the spectral analysis before and after seismic data reconstruction was conducted, such as... Figure 16 As shown, the amplitude for each frequency is obtained by averaging all seismic traces in the two-dimensional profile. The blue and red curves represent the output seismic profile amplitude spectrum and the input profile amplitude spectrum, respectively. It can be seen that the frequency band corresponding to the reconstructed seismic image is wider than that of the original input seismic profile, especially in the high-frequency part, further verifying that the model in this application can effectively reconstruct the high-frequency information of seismic data.

[0108] In the process of super-resolution reconstruction of seismic data, excessive enhancement of high-frequency information or misprocessing of noise by the algorithm may lead to the introduction of artifacts, specifically manifested as false anomaly signals, discontinuities at stratigraphic interfaces, and distortion of waveform features, thereby reducing the geological reliability of the reconstruction results. To effectively suppress artifacts and improve the identification ability of thin interbedded reservoirs, this application proposes a seismic image super-resolution reconstruction method based on a dual-branch multi-attention network. In terms of network architecture, a dual-branch structure consisting of a super-resolution backbone network and a gradient auxiliary branch is designed. The super-resolution backbone network is an improved RCAN (Residual Channel Attention Network) architecture, which achieves efficient fusion of shallow and deep features through multi-level residual groups, focusing on the super-resolution reconstruction of seismic data. The gradient auxiliary branch network learns the high-frequency texture information of the seismic image and constrains the reconstruction process through gradient regularization to avoid excessive smoothing or structural distortion, thereby maintaining the edge sharpness and geological structural consistency of the seismic data. In addition, in terms of optimization strategy, a hybrid loss function that fuses pixel loss, perceptual loss, and adversarial loss is designed, in which the high-frequency constraint term is specifically used to suppress artifacts, achieving a balance between resolution improvement and structural fidelity. Experiments have shown that this method not only effectively enhances the identification of thin interlayers, but also avoids the structural distortion problem common in traditional methods through a multi-task collaborative optimization mechanism, providing a reliable technical means for the fine description of complex reservoirs.

[0109] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A method for super-resolution reconstruction of seismic images based on a dual-branch multi-attention network, characterized in that, include: S1: Construct a data sample library for model training; S2: A dual-branch multi-attention network model is constructed using a super-resolution backbone network and a gradient branch auxiliary network; S3: The dual-branch multi-attention network model is trained using a data sample library and a hybrid loss function to obtain a network model for realizing super-resolution reconstruction of seismic images and reconstruction of high-frequency texture information. Based on the trained network model, super-resolution reconstruction of seismic images is performed to obtain the super-resolution reconstructed seismic images.

2. The seismic image super-resolution reconstruction method based on a dual-branch multi-attention network according to claim 1, characterized in that, The dual-branch multi-attention network model includes a super-resolution backbone network, a gradient branch auxiliary network, and a feature fusion module. Both the super-resolution backbone network and the gradient branch auxiliary network include a shallow feature extraction module, a deep feature extraction module, an upsampling module, and a reconstruction layer; The shallow feature extraction module uses convolutional layers to extract shallow features from the super-resolution backbone network and gradient branch auxiliary network. The calculation formula is as follows: In the formula, This refers to the shallow features extracted from the super-resolution backbone network. The shallow features of the gradient branch auxiliary network are extracted. For convolution operations, For low-resolution input images, This serves as the input to the gradient branch auxiliary network; The deep feature extraction module consists of multiple cascaded residual groups and a long skip connection. Each residual group contains multiple residual channel attention blocks. The deep feature extraction module extracts deep features from the super-resolution backbone network and the gradient branch auxiliary network, calculated using the following formula: In the formula, For the first super-resolution backbone network Deep features output from each residual set, For the gradient branch auxiliary network Deep features output from each residual set, and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals and For the super-resolution backbone network and gradient branch auxiliary network, the first Functional representation of a set of residuals For the first super-resolution backbone network Input of one set of residuals, For the gradient branch auxiliary network Input of one set of residuals, For the first The output of the attention block structure for each residual channel in each residual group. For the first The output of the attention block structure for each residual channel in each residual group; The upsampling module employs subpixel convolution and fuses shallow and deep features through skip connections, outputting the sampled features: In the formula, and These are the sampled features of the super-resolution backbone network and the gradient branch auxiliary network, respectively. and These are the sampling operators on the super-resolution backbone network and the gradient branch auxiliary network, respectively; The reconstruction layer uses a convolutional layer to reconstruct the amplified features, as shown in the formula: In the formula, and These are the reconstructed and amplified features of the super-resolution backbone network and the gradient branch auxiliary network, respectively. and These are the convolution operators for the super-resolution backbone network and the gradient branch auxiliary network, respectively. and For super-resolution backbone network and gradient branch auxiliary network; The feature fusion module employs a weight-adjusted fusion method to fuse the reconstructed and amplified features from the super-resolution backbone network and the gradient branch auxiliary network, resulting in the final super-resolution reconstructed seismic image. The formula is as follows: In the formula, To reconstruct fused features from seismic images at super-resolution resolution. For weight generation function, For element-wise multiplication, This is an element-wise addition.

3. The seismic image super-resolution reconstruction method based on a dual-branch multi-attention network according to claim 2, characterized in that, The input to the gradient branch auxiliary network is the image gradient; The expression for the image gradient is: in, For image gradient, Let be the partial derivative of the image in the horizontal direction. Let be the partial derivative of the image in the vertical direction. For earthquake images, In the horizontal direction, The vertical direction.

4. The seismic image super-resolution reconstruction method based on a dual-branch multi-attention network according to claim 2, characterized in that, The hybrid loss function includes a pixel loss function, a perceptual loss function, and an adversarial loss function; The hybrid loss function The expression is: In the formula, , and These represent the SR image perceptual loss weight coefficients, super-resolution image pixel loss weight coefficients, and super-resolution image gradient map loss weight coefficients, respectively. and The adversarial loss weight coefficients are the super-resolution image and the gradient map of the super-resolution image. The loss weights applied to the super-resolution gradient map output by the gradient branch auxiliary network. For the perceptual loss function, The loss of the super-resolution image in the super-resolution backbone network. The loss on the gradient map of the super-resolution image is given by the gradient branching auxiliary network. To provide adversarial loss for super-resolution backbone networks, For the adversarial loss applied to the gradient branch auxiliary network, The loss is applied to the SR gradient map output by the gradient branch auxiliary network; The formula for calculating the pixel loss function is as follows: In the formula, To calculate the expected value of the loss for all generated super-resolution images SR, Calculate the expected value of the loss for all generated super-resolution gradient maps. The sum of absolute values Input images at high resolution; The expression for the perceptual loss function is: In the formula, For Euclidean distance For the VGG feature extraction network, the first Feature map after layer extraction; The expression for the adversarial loss function is: In the formula, For the discriminator of the super-resolution backbone network, It serves as the discriminator for gradient branch auxiliary networks.

5. The seismic image super-resolution reconstruction method based on a dual-branch multi-attention network according to claim 1, characterized in that, The data sample library includes low-resolution images and high-resolution images.

Images