A digital core reconstruction method

Abstract

The present invention discloses a digital core reconstruction method, comprising the following steps: S1: converting a three-dimensional digital core image into a core sequence slice; S2: converting the core sequence slice into a digital image matrix; S3: using the digital image matrix as input to a generative adversarial network to train the generative adversarial network; S4: using the core sequence slice as input to an improved long short-term memory neural network to train the improved long short-term memory neural network and obtain temporal features; S5: inputting the temporal features into the trained generative adversarial network to generate a predicted core image, converting the predicted core image into the latest digital image matrix, and continuing to reconstruct the next three-dimensional digital core image. The present invention utilizes the long short-term memory neural network to accurately obtain the temporal features of the sample sequence images, and the generative adversarial network to generate clear digital core slices. Combining the two can predict and generate clear digital core sequence images.

Description

Technical Field

[0001] The present invention belongs to the technical field of core reconstruction, and in particular relates to a digital core reconstruction method. Background Art

[0002] As a non-renewable energy source, the efficient and rational extraction and utilization of oil and natural gas resources is crucial for advancing technological development in the petroleum industry. Digital core technology, a recently emerging and effective method for core analysis, has been widely and successfully applied in core analysis of sandstone, carbonate, and shale rock types. Early digital cores were based on physical experimental methods such as two-dimensional scanning electron microscope images or three-dimensional electron microscope scanning images. These images were captured or scanned using laboratory instruments to obtain a large number of two-dimensional images of the core. Numerical modeling, on the other hand, is based on a small number of two-dimensional thin-section images obtained through physical experimental methods. It analyzes the information contained in these images and uses a series of mathematical methods to reconstruct digital cores.

[0003] At present, the methods for reconstructing digital rock cores are mainly divided into three categories: (1) Physical experimental methods, which mainly use highly precise instruments such as scanning electron microscopes, high-power optical microscopes or CT imagers to obtain two-dimensional plane images of the rock core. The digital rock core can be obtained by three-dimensional reconstruction of the two-dimensional plane images. However, this method will damage the sample itself when performing tomographic scanning on the sample, thereby affecting the accuracy of the observation. (2) Numerical reconstruction methods, which mainly use a small amount of data such as rock core plane images to analyze the image and extract modeling information, and then use corresponding mathematical methods to establish the digital rock core. The numerical reconstruction method has many methods, a large calculation volume, and makes targeted treatments for the reconstruction of different digital rock core images. (3) Hybrid modeling method, which combines the advantages of physical experimental methods and numerical reconstruction methods to reconstruct digital rock cores.

[0004] Numerical reconstruction methods primarily include process methods, simulated annealing, and Markov Chain Monte Carlo methods. Process methods simulate the formation of real cores, providing a good description of rock deposition, compaction, and diagenesis. The reconstructed core images exhibit good connectivity, but they are computationally complex, difficult, and expensive. Simulated annealing methods offer advantages such as simple description, ease of use, flexibility, smaller size, and fewer initial condition constraints. However, they can easily reconstruct chaotic pore structures. Markov Chain Monte Carlo methods offer advantages for core reconstruction, such as fast computational speed, wide applicability, and the ability to capture core anisotropy. However, the porosity of the reconstructed core model sections differs significantly from that of the original image sections. Hybrid modeling methods were proposed to overcome the limitations of single modeling methods in digital core modeling. They combine numerical reconstruction with physical experimental methods to improve the accuracy and efficiency of core reconstruction.

[0005] In recent years, image generation technology using generative adversarial networks (GANs) has become a mainstream research area. Leveraging the concept of a zero-sum game, GANs use noise to generate new image samples, significantly outperforming other technologies in the field of image generation. Recurrent neural networks are well-suited for solving sequential problems, but they can suffer from exploding or vanishing gradients when the network is too deep or when there is too much time series data. Long-short-term memory (LSTM) neural networks utilize a gate structure to accept recently useful information, selectively forget long-term, unused information, and determine output based on the current state. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a digital core reconstruction method.

[0007] The technical solution of the present invention is: a digital core reconstruction method comprises the following steps:

[0008] S1: Acquire 3D digital core images and convert them into core serial slices;

[0009] S2: pre-process the core sequence slices and convert them into digital image matrices;

[0010] S3: Using the digital image matrix of the 3D digital core image as the input of the generative adversarial network to train the generative adversarial network;

[0011] S4: Using the core sequence slices as input to the improved long short-term memory neural network, training the improved long short-term memory neural network, and obtaining time series features;

[0012] S5: Input the time series features into the trained generative adversarial network to generate a predicted core image, and convert the predicted core image into the latest digital image matrix to continue reconstructing the next three-dimensional digital core image.

[0013] Furthermore, step S1 includes the following sub-steps:

[0014] S11: collecting a three-dimensional digital core image, and performing median filtering on the three-dimensional digital core image to obtain a core grayscale image;

[0015] S12: performing binarization processing on the core grayscale image to obtain a core binary image;

[0016] S13: Determine the reconstructed size of the core according to the porosity of the core, and convert the core binary image into core serial slices.

[0017] Furthermore, in step S3, the generative adversarial network includes a generative model and a discriminative model.

[0018] Furthermore, the generative model includes a first-layer generative network, a second-layer generative network, a third-layer generative network, a fourth-layer generative network, a fifth-layer generative network, a sixth-layer generative network, a seventh-layer generative network, and an eighth-layer generative network; the first-layer generative network, the second-layer generative network, the third-layer generative network, the fourth-layer generative network, the fifth-layer generative network, the sixth-layer generative network, and the seventh-layer generative network have the same structure, and all include a deconvolution layer, a normalization layer, and a ReLu activation function layer; the eighth-layer generative network includes a deconvolution layer and a Tanh activation function layer;

[0019] The first-layer generation network is used to receive the digital image matrix and output image features with a resolution of 4*4; the second-layer generation network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 8*8; the third-layer generation network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 16*16; the fourth-layer generation network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 32*32; the fifth-layer generation network is used to receive image features with a resolution of 32*32 and output image features with a resolution of 64*64; the sixth-layer generation network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 128*128; the seventh-layer generation network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 256*256; the eighth-layer generation network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 512*512.

[0020] Furthermore, the discriminant model includes a first discriminant network layer, a second discriminant network layer, a third discriminant network layer, a fourth discriminant network layer, a fifth discriminant network layer, a sixth discriminant network layer, a seventh discriminant network layer and an eighth discriminant network layer; the first discriminant network layer, the second discriminant network layer, the third discriminant network layer, the fourth discriminant network layer, the fifth discriminant network layer, the sixth discriminant network layer and the seventh discriminant network layer have the same structure, and all include a convolution layer, a normalization layer and a LeakyReLu activation function layer; the eighth discriminant network layer includes a convolution layer and a sigmoid activation function layer;

[0021] The first discriminant network layer is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second discriminant network layer is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third discriminant network layer is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth discriminant network layer is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth discriminant network layer is used to receive image features with a resolution of 32*32 and output image features with a resolution of 16*16; the sixth discriminant network layer is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh discriminant network layer is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; the eighth discriminant network layer is used to receive image features with a resolution of 4*4 and output the probability that the 3D digital core image is a real image.

[0022] Furthermore, in step S4, the specific method for training the improved long short-term memory neural network is: setting a sliding window with a step size of 20, and using the sliding window to sequentially use the 20 core sequence slices as inputs of the improved long short-term memory neural network, and training them sequentially.

[0023] Furthermore, the improved long short-term memory neural network includes a first layer of memory neural network, a second layer of memory neural network, a third layer of memory neural network, a fourth layer of memory neural network, a fifth layer of memory neural network, a sixth layer of memory neural network, a seventh layer of memory neural network, an eighth layer of memory neural network, a ninth layer of memory neural network and a tenth layer of memory neural network; the first layer of memory neural network, the second layer of memory neural network, the third layer of memory neural network, the fourth layer of memory neural network, the fifth layer of memory neural network, the sixth layer of memory neural network, the seventh layer of memory neural network, the eighth layer of memory neural network and the ninth layer of memory neural network have the same structure and all include a convolution layer, a normalization layer and a LeakyReLu activation function layer; the tenth layer of memory neural network includes a fully connected layer and an LSTM layer;

[0024] The first layer of memory neural network is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second layer of memory neural network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third layer of memory neural network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth layer of memory neural network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth layer of memory neural network is used to receive image features with a resolution of 32*32 and output 1 The sixth layer of memory neural network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh layer of memory neural network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; the eighth layer of memory neural network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 2*2; the ninth layer of memory neural network is used to receive image features with a resolution of 2*2 and output image features with a resolution of 1*1; the tenth layer of memory neural network is used to receive image features with a resolution of 2*2 and output time series features with a vector size of 1024.

[0025] The beneficial effects of the present invention are:

[0026] (1) This digital core reconstruction method does not require large-scale sampling of samples, but only requires a small number of sample slices, which maintains the integrity of the sample and does not damage the sample during the collection process, reducing experimental loss costs and greatly improving sample collection speed and accuracy;

[0027] (2) The digital core reconstruction method has high reconstruction efficiency and takes a short time. The numerical simulation method calculates the sample formation process and the random process, which greatly improves the reconstruction efficiency of the core. The present invention only needs a small amount of time and pre-training process to accurately predict the subsequent reconstruction, reducing the large-scale simulation calculations brought by the numerical simulation method and solving the problem of the generated image not being consistent with the original image.

[0028] (3) The digital core reconstruction method has high generalization ability and does not require special solutions to special problems. During the training process, the characteristic distribution of different types of samples will be learned to generate image types of specific samples. The long short-term memory neural network can accurately obtain the temporal characteristics of the sample sequence images, and the generative adversarial network can generate clear digital core slices. Combining the two can predict and generate clear digital core sequence images. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Flowchart of the digital core reconstruction method;

[0030] Figure 2 Schematic diagram of the predicted structure. DETAILED DESCRIPTION

[0031] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, the present invention provides a digital core reconstruction method, comprising the following steps:

[0033] S1: Acquire 3D digital core images and convert them into core serial slices;

[0034] S2: pre-process the core sequence slices and convert them into digital image matrices;

[0035] S3: Using the digital image matrix of the 3D digital core image as the input of the generative adversarial network to train the generative adversarial network;

[0036] S4: Using the core sequence slices as input to the improved long short-term memory neural network, training the improved long short-term memory neural network, and obtaining time series features;

[0037] S5: Input the time series features into the trained generative adversarial network to generate a predicted core image, and convert the predicted core image into the latest digital image matrix to continue reconstructing the next three-dimensional digital core image.

[0038] In this embodiment of the present invention, step S1 includes the following sub-steps:

[0039] S11: collecting a three-dimensional digital core image, and performing median filtering on the three-dimensional digital core image to obtain a core grayscale image;

[0040] S12: performing binarization processing on the core grayscale image to obtain a core binary image;

[0041] S13: Determine the reconstructed size of the core according to the porosity of the core, and convert the core binary image into core serial slices.

[0042] In an embodiment of the present invention, in step S3, the generative adversarial network includes a generative model and a discriminative model.

[0043] In the embodiment of the present invention, the generative model in the generative adversarial network is to fix the discriminant model in the network so that the generated image is discriminated as 1 as much as possible, so that the generated image makes it impossible for the discriminant model to distinguish the source sample of the model. The calculation formula is: The generative model includes a first-layer generative network, a second-layer generative network, a third-layer generative network, a fourth-layer generative network, a fifth-layer generative network, a sixth-layer generative network, a seventh-layer generative network, and an eighth-layer generative network; the first-layer generative network, the second-layer generative network, the third-layer generative network, the fourth-layer generative network, the fifth-layer generative network, the sixth-layer generative network, and the seventh-layer generative network have the same structure, and all include a deconvolution layer, a normalization layer, and a ReLu activation function layer; the eighth-layer generative network includes a deconvolution layer and a Tanh activation function layer;

[0044] The first-layer generation network is used to receive the digital image matrix and output image features with a resolution of 4*4; the second-layer generation network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 8*8; the third-layer generation network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 16*16; the fourth-layer generation network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 32*32; the fifth-layer generation network is used to receive image features with a resolution of 32*32 and output image features with a resolution of 64*64; the sixth-layer generation network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 128*128; the seventh-layer generation network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 256*256; the eighth-layer generation network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 512*512.

[0045] In the embodiment of the present invention, the discriminant model in the generative adversarial network is to fix the generative model in the network so that the input generated image is discriminated as 0 as much as possible, and the input training image is discriminated as 1 as much as possible, so that the generated image is closer to the sample image. The calculation formula is: The discriminant model includes a first-layer discriminant network, a second-layer discriminant network, a third-layer discriminant network, a fourth-layer discriminant network, a fifth-layer discriminant network, a sixth-layer discriminant network, a seventh-layer discriminant network, and an eighth-layer discriminant network; the first-layer discriminant network, the second-layer discriminant network, the third-layer discriminant network, the fourth-layer discriminant network, the fifth-layer discriminant network, the sixth-layer discriminant network, and the seventh-layer discriminant network have the same structure, and all include a convolution layer, a normalization layer, and a LeakyReLu activation function layer; the eighth-layer discriminant network includes a convolution layer and a sigmoid activation function layer;

[0046] The first discriminant network layer is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second discriminant network layer is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third discriminant network layer is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth discriminant network layer is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth discriminant network layer is used to receive image features with a resolution of 32*32 and output image features with a resolution of 16*16; the sixth discriminant network layer is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh discriminant network layer is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; the eighth discriminant network layer is used to receive image features with a resolution of 4*4 and output the probability that the 3D digital core image is a real image. The closer the probability is to 1, the closer the generated image is to the real image, and the closer the output is to 0, the less likely it is to be a real image. This distinguishes the images generated by the generator from the real images, forming a dynamic game process between the generator and the discriminator.

[0047] In an embodiment of the present invention, in step S4, the specific method for training the improved long short-term memory neural network is: setting a sliding window with a step size of 20, and using the sliding window to sequentially use 20 core sequence slices as inputs of the improved long short-term memory neural network, and training them sequentially.

[0048] The present invention sets a sliding window with a step size of 20. Each time training, 20 core sequence images arranged in strict order are input as training images for a single training. The last image of the 20 training images is used as the predicted image label for verification. After each training is completed, the sliding window is slid back one unit. The calculation formula is:

[0049] In an embodiment of the present invention, the present invention uses a single slice core image of size 512*512. In order to match the core image used in the present invention, several layers of convolution layers, normalization layers, and LeakyReLu activation function layers are added to the traditional long short-term memory neural network to use the training image as input, perform feature extraction on the training image to reduce the image scale, and then input the extracted feature image into the LSTM layer to obtain time series features. The improved long short-term memory neural network includes a first layer of memory neural network, a second layer of memory neural network, a third layer of memory neural network, a fourth layer of memory neural network, a fifth layer of memory neural network, a sixth layer of memory neural network, a seventh layer of memory neural network, an eighth layer of memory neural network, a ninth layer of memory neural network, and a tenth layer of memory neural network; the first layer of memory neural network, the second layer of memory neural network, the third layer of memory neural network, the fourth layer of memory neural network, the fifth layer of memory neural network, the sixth layer of memory neural network, the seventh layer of memory neural network, the eighth layer of memory neural network, and the ninth layer of memory neural network have the same structure, all including convolution layers, normalization layers, and LeakyReLu activation function layers; the tenth layer of memory neural network includes a fully connected layer and an LSTM layer;

[0050] The first layer of memory neural network is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second layer of memory neural network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third layer of memory neural network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth layer of memory neural network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth layer of memory neural network is used to receive image features with a resolution of 32*32 and output 1 The sixth layer of memory neural network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh layer of memory neural network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; the eighth layer of memory neural network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 2*2; the ninth layer of memory neural network is used to receive image features with a resolution of 2*2 and output image features with a resolution of 1*1; the tenth layer of memory neural network is used to receive image features with a resolution of 2*2 and output time series features with a vector size of 1024.

[0051] The working principle and process of this invention are as follows: This invention proposes using a generative adversarial network (GAN) combined with an improved long-short-term memory (LSTM) neural network to accurately predict and reconstruct subsequent core slices within a certain range of digital cores. This digital core slice sequence is represented as a spatiotemporal sequence problem and solved through machine learning. This invention extracts feature temporal information using the improved LSTM neural network and, leveraging the powerful reconstruction capabilities of the GAN, inputs the extracted temporal information vector into the GAN's generative model, thereby predicting and reconstructing subsequent slices of the sample sequence.

[0052] like Figure 2 As shown in the figure, during the neural network training process, the long short-term memory neural network is pre-trained strictly according to the digital core image sequence as input, and the temporal features of the predicted image are output. The predicted temporal features are input into the generative adversarial network for image generation, and the generated image samples and the real image samples are input into the discriminator of the generative adversarial network. The probability that the generated image is a real image is output and returned to the network in the form of a loss function, so that the generator and the discriminator of the generative adversarial network compete with each other, so that the generator can achieve the purpose of generating images as real as possible.

[0053] The beneficial effects of the present invention are:

[0054] (1) This digital core reconstruction method does not require large-scale sampling of samples, but only requires a small number of sample slices, which maintains the integrity of the sample and does not damage the sample during the collection process, reducing experimental loss costs and greatly improving sample collection speed and accuracy;

[0055] (2) The digital core reconstruction method has high reconstruction efficiency and takes a short time. The numerical simulation method calculates the sample formation process and the random process, which greatly improves the reconstruction efficiency of the core. The present invention only needs a small amount of time and pre-training process to accurately predict the subsequent reconstruction, reducing the large-scale simulation calculations brought by the numerical simulation method and solving the problem of the generated image not being consistent with the original image.

[0056] (3) The digital core reconstruction method has high generalization ability and does not require special solutions to special problems. During the training process, the characteristic distribution of different types of samples will be learned to generate image types of specific samples. The long short-term memory neural network can accurately obtain the temporal characteristics of the sample sequence images, and the generative adversarial network can generate clear digital core slices. Combining the two can predict and generate clear digital core sequence images.

[0057] 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 it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A digital core reconstruction method, characterized in that: The following steps are involved: S1: Acquire 3D digital core images and convert them into core serial slices; S2: pre-process the core sequence slices and convert them into digital image matrices; S3: Using the digital image matrix of the 3D digital core image as the input of the generative adversarial network to train the generative adversarial network; S4: Using the core sequence slices as input to the improved long short-term memory neural network, training the improved long short-term memory neural network, and obtaining time series features; S5: Input the time series features into the trained generative adversarial network to generate a predicted core image, convert the predicted core image into the latest digital image matrix, and continue to reconstruct the next 3D digital core image; In step S4, the specific method for training the improved long short-term memory neural network is as follows: setting a sliding window with a step size of 20, using the sliding window to sequentially use the 20 core sequence slices as inputs of the improved long short-term memory neural network, and training them sequentially; The improved long short-term memory neural network includes a first layer of memory neural network, a second layer of memory neural network, a third layer of memory neural network, a fourth layer of memory neural network, a fifth layer of memory neural network, a sixth layer of memory neural network, a seventh layer of memory neural network, an eighth layer of memory neural network, a ninth layer of memory neural network and a tenth layer of memory neural network; the first layer of memory neural network, the second layer of memory neural network, the third layer of memory neural network, the fourth layer of memory neural network, the fifth layer of memory neural network, the sixth layer of memory neural network, the seventh layer of memory neural network, the eighth layer of memory neural network and the ninth layer of memory neural network have the same structure and all include a convolution layer, a normalization layer and a LeakyReLu activation function layer; the tenth layer of memory neural network includes a fully connected layer and an LSTM layer; The first layer of memory neural network is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second layer of memory neural network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third layer of memory neural network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth layer of memory neural network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth layer of memory neural network is used to receive image features with a resolution of 32*32 and output image features with a resolution of 1 The sixth layer of the memory neural network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh layer of the memory neural network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; the eighth layer of the memory neural network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 2*2; the ninth layer of the memory neural network is used to receive image features with a resolution of 2*2 and output image features with a resolution of 1*1; the tenth layer of the memory neural network is used to receive image features with a resolution of 2*2 and output time series features with a vector size of 1024.

2. The digital core reconstruction method according to claim 1, characterized in that: The step S1 includes the following sub-steps: S11: collecting a three-dimensional digital core image, and performing median filtering on the three-dimensional digital core image to obtain a core grayscale image; S12: performing binarization processing on the core grayscale image to obtain a core binary image; S13: Determine the reconstructed size of the core according to the porosity of the core, and convert the core binary image into core serial slices.

3. The digital core reconstruction method according to claim 1, characterized in that: In step S3, the generative adversarial network includes a generative model and a discriminative model.

4. The digital core reconstruction method according to claim 3, characterized in that: The generative model includes a first-layer generative network, a second-layer generative network, a third-layer generative network, a fourth-layer generative network, a fifth-layer generative network, a sixth-layer generative network, a seventh-layer generative network and an eighth-layer generative network; the first-layer generative network, the second-layer generative network, the third-layer generative network, the fourth-layer generative network, the fifth-layer generative network, the sixth-layer generative network and the seventh-layer generative network have the same structure and all include a deconvolution layer, a normalization layer and a ReLu activation function layer; the eighth-layer generative network includes a deconvolution layer and a Tanh activation function layer; The first-layer generation network is used to receive a digital image matrix and output image features with a resolution of 4*4; the second-layer generation network is used to receive image features with a resolution of 4*4 and output image features with a resolution of 8*8; the third-layer generation network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 16*16; the fourth-layer generation network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 32*32; the fifth-layer generation network is used to receive image features with a resolution of 32*32 and output image features with a resolution of 64*64; the sixth-layer generation network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 128*128; the seventh-layer generation network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 256*256; The eighth layer generation network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 512*512.

5. The digital core reconstruction method according to claim 3, characterized in that: The discriminant model includes a first discriminant network layer, a second discriminant network layer, a third discriminant network layer, a fourth discriminant network layer, a fifth discriminant network layer, a sixth discriminant network layer, a seventh discriminant network layer and an eighth discriminant network layer; the first discriminant network layer, the second discriminant network layer, the third discriminant network layer, the fourth discriminant network layer, the fifth discriminant network layer, the sixth discriminant network layer and the seventh discriminant network layer have the same structure and all include a convolution layer, a normalization layer and a LeakyReLu activation function layer; the eighth discriminant network layer includes a convolution layer and a sigmoid activation function layer; The first layer of the discriminant network is used to receive image features with a resolution of 512*512 and output image features with a resolution of 256*256; the second layer of the discriminant network is used to receive image features with a resolution of 256*256 and output image features with a resolution of 128*128; the third layer of the discriminant network is used to receive image features with a resolution of 128*128 and output image features with a resolution of 64*64; the fourth layer of the discriminant network is used to receive image features with a resolution of 64*64 and output image features with a resolution of 32*32; the fifth layer of the discriminant network is used to receive image features with a resolution of 32*32 and output image features with a resolution of 16*16; the sixth layer of the discriminant network is used to receive image features with a resolution of 16*16 and output image features with a resolution of 8*8; the seventh layer of the discriminant network is used to receive image features with a resolution of 8*8 and output image features with a resolution of 4*4; The eighth layer of the discriminant network is used to receive image features with a 4*4 resolution and output the probability that the three-dimensional digital core image belongs to a real image.

Images