Multi-source digital rock image fusion method based on ensemble learning

Abstract

The invention discloses a multi-source digital rock image fusion method based on ensemble learning, and relates to the technical field of oil and gas resource geological exploration and development evaluation, and the method comprises the steps: S1, scanning and preprocessing a multi-source rock image; s2, generator optimization of the cyclic consistent generative adversarial network; s3, optimizing three types of weak discriminators; and S4, constructing a strong discriminator based on ensemble learning and reconstructing the multi-source digital rock. According to the method, multi-source digital rock image fusion is realized, the advantages of a large-scale image and a high-resolution image are effectively fused, the problem that a multi-scale three-dimensional pore structure cannot be reconstructed by using a single image scanning mode is solved, and compared with a traditional cyclic consistent generative adversarial network method, the method has the advantages that the method is simple and convenient to implement. The relative error of the constructed digital rock physical parameters is remarkably reduced, the accuracy of rock physical numerical simulation is improved, the method is easy to apply and popularize in different oil and gas fields, and a foundation is laid for fine evaluation of the oil and gas reservoir rock physical parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of geological exploration and development evaluation of oil and gas resources, and in particular to a multi-source digital rock image fusion method based on ensemble learning. Background Art

[0002] Three-dimensional digital rock uses computer technology to represent the microstructure of rocks and simulates the physical properties of rocks through mathematical modeling. Compared with traditional laboratory methods, three-dimensional digital rock technology has several advantages. Once a three-dimensional digital rock model is established, it can be used for multiple physical simulations of the same rock sample. In addition, it avoids the use of chemical reagents and is an environmentally friendly and non-polluting method. Therefore, the reconstruction of three-dimensional digital rocks has attracted more and more attention from scholars in recent years. There are two main methods for reconstructing three-dimensional digital rocks: physical methods and numerical reconstruction methods. Physical methods mainly involve using X-ray micro-computed tomography (micro-CT) or high-precision equipment such as scanning electron microscopes (SEM) to capture two-dimensional images of different cross-sections of rocks. These two-dimensional images are used to reconstruct three-dimensional digital rocks through specific algorithms. Numerical reconstruction methods rely on a limited number of two-dimensional digital rock images to extract the required information. Through this smaller data set, the reconstruction algorithm is used to construct a complete three-dimensional digital rock.

[0003] In the reconstruction of three-dimensional digital rock, the inherent multi-scale characteristics of rocks and the limitations of single scanning methods have caused a constant conflict between field of view and resolution. Low-resolution images can capture representative volume elements (RVEs) and provide a broader view of the overall structure of rocks, but cannot resolve fine microstructures; on the other hand, high-resolution images can reveal detailed microstructures but cannot capture enough representative volume elements. To solve this problem, a method to strike a balance between field of view and resolution is needed. A promising approach is to fuse CT and SEM images. CT provides a large field of view and captures the overall structure of rocks, while SEM provides high-resolution images of fine microstructures. By integrating these multi-source digital rock images, a three-dimensional digital rock model containing cross-scale information can be constructed. This fusion method can significantly improve the accuracy of rock physics numerical simulation because it provides a more complete representation of rock properties at multiple scales, thereby better predicting rock physical properties. However, the current multi-source digital rock image fusion method constructs low-precision three-dimensional digital rocks. For example, the traditional cycle-consistent generative adversarial network method, although it works well for rocks with good homogeneity, when the homogeneity of the rock deteriorates, the digital rock constructed by this method has low accuracy and unreliable rock physics numerical simulation results.

[0004] Therefore, it is necessary to optimize the generator and multi-class weak discriminator of the cycle-consistent generative adversarial network, realize the fusion of multi-source digital rock images based on the ensemble learning strong discriminator, reconstruct multi-source digital rock, effectively integrate the advantages of large-scale images and high-resolution images, overcome the problem that the multi-scale three-dimensional pore structure cannot be reconstructed using a single image scanning method, and lay the foundation for the detailed evaluation of the petrophysical parameters of oil and gas reservoirs. Summary of the invention

[0005] In order to solve the above technical problems, the present invention discloses a multi-source digital rock image fusion method based on ensemble learning.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A multi-source digital rock image fusion method based on ensemble learning includes the following steps:

[0008] S1. Multi-source rock image scanning and preprocessing;

[0009] S2. Generator optimization of cycle-consistent generative adversarial networks;

[0010] S3. Optimization of three types of weak discriminators;

[0011] S4. Strong discriminator construction and multi-source digital rock reconstruction based on ensemble learning.

[0012] Optionally, in step S1, multi-source rock image scanning and preprocessing specifically include:

[0013] Collect rock CT scan and scanning electron microscope (SEM) image data, crop the area without rock data around the image, and only keep the image part with rock data. Then convert the CT scan image and SEM image into 8-bit grayscale format images with distribution in the [0,255] interval, extract the SEM sub-image and downsample it. According to the image fusion requirements, adjust the resolution to 1 / 2, 1 / 4, 1 / 8 or 1 / 16 times the CT scan image resolution, and randomly select N downsampled SEM images and N CT scan images as training sets.

[0014] Optionally, in step S2, the generator optimization of the cycle-consistent generative adversarial network specifically includes:

[0015] S2.1. Based on the ResNet generator of the cycle-consistent generative adversarial network, the optimization is performed. The Batch Normalization (BN) in the original network structure is changed to the Instance Normalization (IN) layer, and the transposed convolution upsampling module is changed to Resize-Convolution. The upsampling and convolution operations are separated. The image is first enlarged by 2 times using the nearest neighbor interpolation, and then the feature map is reduced through convolution and the checkerboard artifacts are reduced. A jump structure is added before and after the CBAM module to promote the fusion of features at different depths.

[0016] S2.2. Add the Cutmix enhancement strategy. During the training process of the discriminator, the fake images generated by the generator are randomly cropped and spliced ​​with the real images, and used as new training samples to train the discriminator, guiding the discriminator to focus on the structural differences between the generated images and the real images, thereby enhancing the discriminator's ability to identify real and fake images.

[0017] Optionally, in step S3, three types of weak discriminators are optimized, specifically including:

[0018] S3.1. Optimize the weak discriminator Patch Discriminator, correspond each value in the input mapping matrix to the features extracted from each patch in the original image, and finally calculate the average of the extracted features. In the Patch Discriminator of the low-resolution image, a smaller convolution kernel (2×2 convolution kernel) and feature map number are used. The number of feature maps is half of that in the Patch Discriminator of the high-resolution image, reducing the amount of parameter calculations.

[0019] S3.2. Using Attn U-Net as a weak discriminator, we added Attention Gate to the traditional U-Net jump structure to suppress redundant information in the encoding process and splice it to the corresponding decoding channel to improve the discriminator accuracy.

[0020] S3.3. Optimize the weak discriminator VGG16 Discriminator, use transfer learning to freeze the pre-trained weights of the VGG16 feature layer, and redesign a new classifier. Add a Sigmoid layer after the fully connected layer to control the output results within the range of [0,1] to reduce the time and performance of training the model.

[0021] Optionally, in step S4, the strong discriminator construction based on ensemble learning and multi-source digital rock reconstruction specifically includes:

[0022] S4.1. Generator loss function L G By adversarial loss L adv , cycle consistency loss L cycand identity loss L identity The calculation process is as follows:

[0023]

[0024] In the formula, X represents a low-resolution image, Y represents a high-resolution image, G1 represents the generator from LR to HR, G2 represents the generator from HR to LR, and D X Patch Discriminator representing a low-resolution image, Patch Discriminator representing a high-resolution image, Attn U-Net Discriminator representing high-resolution images, VGG16 Discriminator representing high-resolution images, α, β, γ are weight parameters;

[0025]

[0026] In the formula, it is necessary to satisfy λ1+λ2+λ3=1;

[0027]

[0028] In the formula, 1 a×b Represents a matrix of size a×b with all elements set to 1, where the values ​​of a and b should be consistent with the size of the low-resolution image X;

[0029]

[0030] In the formula, y reszie Indicates the size of the high-resolution image to be resampled to the low-resolution image, x reszie Indicates the size of resampling a low-resolution image to a high-resolution image;

[0031] S4.2. By adjusting different weak discriminator weights α i (i=1,2,3,4), a strong discriminator is constructed by linear combination, and the weight parameter λ in step S4.1 is i The parameters are adjusted from (0,1) at intervals of 0.1 to reconstruct multi-source digital rocks with different parameter combinations. The entropy weight method is used to assign different weights to the four indicators, namely porosity, permeability, FID index and grayscale histogram similarity, to obtain the final index F:

[0032] F=α1|R por -X por |+α2|R per -X per |+α3X fid +α4(1-XHis )

[0033] Where, X por , X per , X fid , X His Respectively represent the porosity, permeability, FID index and grayscale histogram similarity of multi-source digital rocks reconstructed with different parameter combinations, R por and R per represent the rock porosity and permeability measured in real experiments, respectively;

[0034] S4.3. Reconstruct the index F of multi-source digital rock with different parameter combinations through step S4.2, and select the maximum value of F as the weight parameter λ i The optimal solution of the multi-source digital rock physical property parameters reconstructed under this weight parameter combination is closest to the true value.

[0035] The beneficial effect of the present invention is that the present invention realizes the fusion of multi-source digital rock images, effectively integrates the advantages of large-scale images and high-resolution images, overcomes the problem that a single image scanning method cannot reconstruct a multi-scale three-dimensional pore structure, and compared with the traditional cyclic consistent generative adversarial network method, the relative error of the digital rock physical property parameters constructed by the present invention is significantly reduced, the accuracy of rock physics numerical simulation is improved, and it is easy to apply and promote in different oil and gas fields, laying a foundation for the detailed evaluation of rock physical parameters in oil and gas reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a multi-source digital rock image fusion method based on ensemble learning in the present invention;

[0037] Figure 2 A comparison diagram of the digital rock constructed by the present invention and the digital rock constructed by CT scanning, wherein (a) is a three-dimensional visualization diagram of the digital rock constructed by CT scanning, (b) is a three-dimensional visualization diagram of the digital rock constructed by the present invention, (c) is a three-dimensional visualization diagram of the pores of the digital rock constructed by CT scanning, and (d) is a three-dimensional visualization diagram of the pores of the digital rock constructed by the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0039] A multi-source digital rock image fusion method based on ensemble learning includes the following steps:

[0040] S1. Multi-source rock image scanning and preprocessing: Collect image data obtained by rock CT scanning and scanning electron microscope (SEM), crop the area without rock data around the image, and only retain the image part with rock data. Then convert the CT scan image and SEM image into 8-bit grayscale format images distributed in the [0,255] interval, extract the SEM sub-image and downsample it. According to the image fusion requirements, adjust the resolution to 1 / 2, 1 / 4, 1 / 8 or 1 / 16 times the CT scan image resolution, and randomly select N downsampled SEM images and N CT scan images as training sets.

[0041] S2. Generator optimization of cycle-consistent generative adversarial networks, including:

[0042] S2.1. Based on the ResNet generator of the cycle-consistent generative adversarial network, the optimization is performed. The Batch Normalization (BN) in the original network structure is changed to the Instance Normalization (IN) layer, and the transposed convolution upsampling module is changed to Resize-Convolution. The upsampling and convolution operations are separated. The image is first enlarged by 2 times using the nearest neighbor interpolation, and then the feature map is reduced through convolution and the checkerboard artifacts are reduced. A jump structure is added before and after the CBAM module to promote the fusion of features at different depths.

[0043] S2.2. Add the Cutmix enhancement strategy. During the training process of the discriminator, the fake images generated by the generator are randomly cropped and spliced ​​with the real images, and used as new training samples to train the discriminator, guiding the discriminator to focus on the structural differences between the generated images and the real images, thereby enhancing the discriminator's ability to identify real and fake images.

[0044] S3. Optimization of three types of weak discriminators, including:

[0045] S3.1. Optimize the weak discriminator Patch Discriminator, correspond each value in the input mapping matrix to the features extracted from each patch in the original image, and finally calculate the average of the extracted features. In the Patch Discriminator of the low-resolution image, a 2×2 convolution kernel and the number of feature maps are used, and the number of feature maps is half of that in the Patch Discriminator of the high-resolution image, reducing the amount of parameter calculation;

[0046] S3.2. Using Attn U-Net as a weak discriminator, we added Attention Gate to the traditional U-Net jump structure to suppress redundant information in the encoding process and splice it to the corresponding decoding channel to improve the discriminator accuracy.

[0047] S3.3. Optimize the weak discriminator VGG16 Discriminator, use transfer learning to freeze the pre-trained weights of the VGG16 feature layer, and redesign a new classifier. Add a Sigmoid layer after the fully connected layer to control the output results within the range of [0,1] to reduce the time and performance of training the model.

[0048] S4. Construction of strong discriminator based on ensemble learning and reconstruction of multi-source digital rock, including:

[0049] S4.1. Generator loss function L G By adversarial loss L adv , cycle consistency loss L cyc and identity loss L identity The calculation process is as follows:

[0050]

[0051] In the formula, X represents a low-resolution image, Y represents a high-resolution image, G1 represents the generator from LR to HR, G2 represents the generator from HR to LR, and D X Patch Discriminator representing low-resolution images, D Y Patch Patch Discriminator representing high-resolution images, D Y U Attn U-Net Discriminator representing high-resolution images, D Y VGG VGG16 Discriminator representing high-resolution images, α, β, γ are weight parameters;

[0052]

[0053] In the formula, it is necessary to satisfy λ1+λ2+λ3=1;

[0054]

[0055] In the formula, 1 a×b Represents a matrix of size a×b with all elements set to 1, where the values ​​of a and b should be consistent with the size of the low-resolution image X;

[0056]

[0057] In the formula, y reszie Indicates the size of the high-resolution image to be resampled to the low-resolution image, x reszie Indicates the size of resampling a low-resolution image to a high-resolution image;

[0058] S4.2. By adjusting different weak discriminator weights α i (i=1,2,3,4), a strong discriminator is constructed by linear combination, and the weight parameter λ in step S4.1 is i The parameters are adjusted from (0,1) at intervals of 0.1 to reconstruct multi-source digital rocks with different parameter combinations. The entropy weight method is used to assign different weights to the four indicators, namely porosity, permeability, FID index and grayscale histogram similarity, to obtain the final index F:

[0059] F=α1|R por -X por |+α2|R per -X per |+α3X fid +α4(1-X His )

[0060] Where, X por , X per , X fid , X His Respectively represent the porosity, permeability, FID index and grayscale histogram similarity of multi-source digital rocks reconstructed with different parameter combinations, R por and R per represent the rock porosity and permeability measured in real experiments, respectively;

[0061] S4.3. Reconstruct the index F of multi-source digital rock with different parameter combinations through step S4.2, and select the maximum value of F as the weight parameter λ i The optimal solution of the multi-source digital rock physical property parameters reconstructed under this weight parameter combination is closest to the true value.

[0062] Example

[0063] Taking rock A coring from an oil field in China as an example, the multi-source digital rock image fusion method based on ensemble learning is presented in the present invention. Figure 1 As shown, the following steps are included:

[0064] S1. Multi-source rock image scanning and preprocessing

[0065] Image data obtained from rock CT scans (resolution of 8.0 μm) and SEM (resolution of 0.5 μm) were collected, and the areas without rock data around the images were cropped, leaving only the image parts with rock data. The CT scan images and SEM images were then converted into 8-bit grayscale formats distributed in the [0,255] interval. The SEM sub-images were extracted and downsampled. In order to achieve higher accuracy in rock physics simulation and larger rock size, their resolution was adjusted to 1 / 4 of the CT scan image resolution, i.e., 2.0 μm. 1088 downsampled SEM sub-images and 1088 CT scan images were randomly selected as training sets.

[0066] S2. Generator Optimization for Cycle-Consistent Generative Adversarial Networks

[0067] S2.1. Based on the ResNet generator of the cycle-consistent generative adversarial network, the optimization is performed. The Batch Normalization (BN) in the original network structure is changed to the Instance Normalization (IN) layer, and the transposed convolution upsampling module is changed to Resize-Convolution. The upsampling and convolution operations are separated. The image is first enlarged by 2 times using the nearest neighbor interpolation, and then the feature map is reduced through convolution and the checkerboard artifacts are reduced. A jump structure is added before and after the CBAM module to promote the fusion of features at different depths.

[0068] S2.2. Add the Cutmix enhancement strategy. During the training process of the discriminator, the fake images generated by the generator are randomly cropped and spliced ​​with the real images, and used as new training samples to train the discriminator, guiding the discriminator to pay more attention to the structural differences between the generated images and the real images, thereby enhancing the discriminator's ability to identify true and false images.

[0069] S3. Optimization of three types of weak discriminators

[0070] S3.1. Optimize the weak discriminator Patch Discriminator, correspond each value in the input mapping matrix to the features extracted from each patch in the original image, and finally calculate the average of the extracted features. In the Patch Discriminator of the low-resolution image, a smaller convolution kernel and feature map number are used. The number of feature maps is only half of that of the Patch Discriminator of the high-resolution image, reducing the amount of parameter calculation;

[0071] S3.2. Attn U-Net is used as a weak discriminator. It adds Attention Gate to the traditional U-Net jump structure, suppresses redundant information in the encoding process and splices it to the corresponding decoding channel to improve the discriminator accuracy.

[0072] S3.3. Optimize the weak discriminator VGG16 Discriminator, use transfer learning to freeze the pre-trained weights of the VGG16 feature layer, and redesign a new classifier. Add a Sigmoid layer after the fully connected layer to control the output results within the range of [0,1] to reduce the time and performance of training the model.

[0073] S4. Construction of strong discriminator based on ensemble learning and multi-source digital rock reconstruction

[0074] S4.1. Generator loss function L G Mainly due to the adversarial loss L adv , cycle consistency loss L cyc and identity loss L identity The specific calculation process is as follows:

[0075]

[0076] In the formula, X represents a low-resolution image, Y represents a high-resolution image, G1 represents the generator from LR to HR, G2 represents the generator from HR to LR, and D X Patch Discriminator representing a low-resolution image, Patch Discriminator representing a high-resolution image, Attn U-Net Discriminator representing high-resolution images, VGG16 Discriminator representing high-resolution images, α, β, γ are weight parameters;

[0077]

[0078] In the formula, it is necessary to satisfy λ1+λ2+λ3=1;

[0079]

[0080] In the formula, 1 a×b Represents a matrix of size a×b with all elements set to 1, where the values ​​of a and b should be consistent with the size of the low-resolution image X;

[0081]

[0082] In the formula, y reszie Indicates the size of the high-resolution image to be resampled to the low-resolution image, x reszie Indicates the size of resampling a low-resolution image to a high-resolution image;

[0083] S4.2. By adjusting different weak discriminator weights α i (i=1,2,3,4), a strong discriminator is constructed by linear combination, and the weight parameter λ in step S4.1 is i (i=1,2,3) are adjusted from (0,1) at intervals of 0.1, and multi-source digital rocks with different parameter combinations are reconstructed. The porosity, permeability, FID index and grayscale histogram similarity are given different weights by the entropy weight method, and the final index F is obtained:

[0084] F=α1|17.3-X por |+α2|109.85-X per |+α3X fid +α4(1-X His );

[0085] S4.3. Reconstruct the index F of multi-source digital rock with different parameter combinations through step S4.2, and select the maximum value of F (0.877) as the weight parameter λ i The optimal solution is λ1=0.6,λ2=0.2,λ3=0.2. The multi-source digital rock reconstructed under this weight parameter combination (such as Figure 2 shown). Figure 2 (a) and Figure 2 (b) a three-dimensional visualization of digital rocks constructed by CT scanning and a three-dimensional visualization of digital rocks constructed by the present invention, respectively; Figure 2 (c) and Figure 2(d) The three-dimensional visualization of the digital rock pores constructed by CT scanning and the three-dimensional visualization of the digital rock pores constructed by the present invention are respectively. It can be seen from the comparison diagram that the digital rock constructed by the present invention has a finer pore structure, which makes up for the problem of low porosity of the digital rock constructed by CT scanning due to insufficient resolution. The porosity of the digital rock constructed by the present invention is 12.35%, and the permeability is 61.334mD, which are closest to the true values ​​of rock physics experiments. Compared with the digital rock constructed by the traditional cycle-consistent generative adversarial network method, the relative errors of porosity and permeability are reduced by 10.75% and 44.30%, respectively.

[0086] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A multi-source digital rock image fusion method based on ensemble learning, characterized in that: The following steps are involved: S1. Multi-source rock image scanning and preprocessing; S2. Generator optimization for cycle-consistent generative adversarial networks; S3. Optimization of three types of weak discriminators; S4. Strong discriminator construction and multi-source digital rock reconstruction based on ensemble learning.

2. The multi-source digital rock image fusion method based on ensemble learning according to claim 1, characterized in that: In step S1, multi-source rock image scanning and preprocessing specifically include: Collect rock CT scan and SEM image data, crop the area without rock data around the image, retain the image part with rock data, convert the CT scan image and SEM image into 8-bit grayscale format images with distribution in the [0,255] interval, extract the SEM sub-image and downsample it, adjust the resolution to 1 / 2, 1 / 4, 1 / 8 or 1 / 16 times the CT scan image resolution according to the image fusion requirements, and randomly select N downsampled SEM images and N CT scan images as training sets.

3. The multi-source digital rock image fusion method based on ensemble learning as claimed in claim 1, characterized in that: In step S2, the generator optimization of the cycle-consistent generative adversarial network specifically includes: S2.

1. Based on the ResNet generator of the cycle-consistent generative adversarial network, the Batch Normalization in the original network structure is modified to the Instance Normalization layer, the transposed convolution upsampling module is modified to Resize-Convolution, the upsampling and convolution operations are separated, the image is first enlarged by 2 times using the nearest neighbor interpolation, and then the feature map is reduced through convolution and the checkerboard artifacts are reduced. Jump structures are added before and after the CBAM module to promote the fusion of features at different depths; S2.

2. Add the Cutmix enhancement strategy. During the training process of the discriminator, the fake images generated by the generator are randomly cropped and spliced ​​with the real images, and used as new training samples to train the discriminator, guiding the discriminator to focus on the structural differences between the generated images and the real images, thereby enhancing the discriminator's ability to identify real and fake images.

4. The multi-source digital rock image fusion method based on ensemble learning as claimed in claim 1, characterized in that: In step S3, three types of weak discriminators are optimized, specifically including: S3.

1. Optimize the weak discriminator Patch Discriminator, correspond each value in the input mapping matrix to the features extracted from each patch in the original image, and finally calculate the average of the extracted features. In the PatchDiscriminator of the low-resolution image, a 2×2 convolution kernel and the number of feature maps are used. The number of feature maps is half of that in the PatchDiscriminator of the high-resolution image, which reduces the amount of parameter calculations. S3.

2. Using Attn U-Net as a weak discriminator, AttentionGate is added to the traditional U-Net skip structure to suppress the redundant information in the encoding process and splice it to the corresponding decoding channel to improve the discriminator accuracy; S3.

3. Optimize the weak discriminator VGG16 Discriminator, use transfer learning to freeze the pre-trained weights of the VGG16 feature layer, and redesign a new classifier, adding a Sigmoid layer after the fully connected layer to control the output results within the range of [0,1].

5. The multi-source digital rock image fusion method based on ensemble learning according to claim 1, characterized in that: In step S4, the strong discriminator construction based on ensemble learning and multi-source digital rock reconstruction specifically include: S4.

1. Generator loss function L G By adversarial loss L adv , cycle consistency loss L cyc and identity loss L identity The calculation process is as follows: In the formula, X represents the low-resolution LR image, Y represents the high-resolution HR image, G1 represents the LR→HR generator, G2 represents the HR→LR generator, and D X Patch Discriminator representing a low-resolution image, Patch Discriminator representing a high-resolution image, Attn U-Net Discriminator representing high-resolution images, VGG16 Discriminator for high-resolution images, λ i ,i=1,2,3 are weight parameters, which need to satisfy λ1+λ2+λ3=1; S4.

2. By adjusting different weak discriminator weights α i , i = 1, 2, 3, 4, a strong discriminator is constructed by linear combination, and the weight parameter λ in step S4.1 is i The parameters are adjusted from (0,1) at intervals of 0.1 to reconstruct multi-source digital rocks with different parameter combinations. The entropy weight method is used to assign different weights to the four indicators, namely porosity, permeability, FID index and grayscale histogram similarity, to obtain the final index F: F=α1|R por -X por |+α2|R per -X per |+α3X fid +α4(1-X His ) Where, X por , X per , X fid , X His Respectively represent the porosity, permeability, FID index and grayscale histogram similarity of multi-source digital rocks reconstructed with different parameter combinations, R por and R per represent the rock porosity and permeability measured in real experiments, respectively; S4.

3. Reconstruct the index F of multi-source digital rock with different parameter combinations through step S4.2, and select the maximum value of F as the weight parameter λ i The optimal solution of .

Images