A multi-scale digital core image construction method based on deep learning

Abstract

The application discloses a kind of multi-scale digital core image construction methods based on deep learning.The present application utilizes core sample preparation plunger sample and scanning, batch acquisition different resolution core scanning image and adjust gray value, based on SIFT feature to each core scanning image is carried out image matching, based on deep learning method establishes double generative adversarial network model, using medium, high resolution core scanning image trains first generative adversarial network model, generates double scale digital core image, again using double scale digital core image and image matched low resolution core scanning image trains second generative adversarial network model, and using the second generative adversarial network model after training generates the three-dimensional digital core image of core sample, determines the pore boundary and micropore structure of core sample.The present application solves the problem that different scale core scanning image information cannot be fused, realizes the construction of multi-scale digital core image, and improves the resolution of digital core image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of oil and gas field exploration and development, and particularly relates to a multi-scale digital core image construction method based on deep learning. BACKGROUND

[0002] With the development of digital core technology, CT scanning, FIB-SEM scanning, QEMSCAN scanning and other technologies have been widely used in the construction of multi-component digital core models reflecting the real state of the reservoir, and deep learning technology has also played an important role in digital core image reconstruction and digital core analysis. The traditional deep learning super-resolution method usually needs to use high-resolution digital core images as training data sets. Although this method can effectively improve the resolution of the core image, it is difficult to complete the missing image information in the core image.

[0003] Due to the difficulty in fusing image information of different scale core images, the multi-resolution core images obtained in the digital core experiment are difficult to be fully utilized. Therefore, it is urgent to propose a multi-scale digital core construction method to fully integrate the accuracy of high-resolution core images and the breadth of low-resolution core images, and realize the fusion of image information of different resolution core images. SUMMARY

[0004] The present application proposes a multi-scale digital core image construction method based on deep learning to solve the problem of difficulty in fusing different scale digital core images. This method fully utilizes the accuracy of high-resolution core scanning images and the breadth of low-resolution core scanning images, and realizes the construction of multi-scale digital core images by fusing different scale digital core images, solves the problem of missing core image information caused by the inability to fuse different scale core scanning images, and improves the resolution of digital core images.

[0005] The present application adopts the following technical solutions:

[0006] A multi-scale digital core image construction method based on deep learning, specifically comprising the following steps:

[0007] Step 1: Obtain core samples according to the coring data of the study area, drill multiple plunger samples at the same position of the core sample, and obtain core scanning images of different resolutions by scanning the plunger samples multiple times;

[0008] Step 2: Scale the image gray value of each core scanning image to a unified gray range according to the gray histogram of each core scanning image, perform image matching on the core scanning images of each plunger sample based on SIFT features, and obtain low-resolution core scanning images, medium-resolution core scanning images and high-resolution core scanning images of the core sample after image matching;

[0009] Step 3, a first generative adversarial network model for constructing a double-scale digital core image is established, the first generative adversarial network model is trained based on deep learning by using the matched medium-resolution core scanning image and high-resolution core scanning image of the core sample image as training images, and the matched medium-resolution core scanning image of the core sample image is input into the trained first generative adversarial network model, and the first generative adversarial network model is used to generate a double-scale digital core image of the core sample;

[0010] Step 4, a second generative adversarial network model for constructing a multi-scale digital core image is established, the second generative adversarial network model is trained based on deep learning by using the double-scale digital core image generated in step 3 and the matched low-resolution core scanning image of the core sample image as training images, and the matched low-resolution core scanning image of the core sample image is input into the trained second generative adversarial network model, and the second generative adversarial network model is used to generate a three-dimensional digital core image of the core sample;

[0011] Step 5, according to the three-dimensional digital core image of the core sample, the pore boundary and micro-pore structure of the core sample are determined.

[0012] Preferably, in step 1, the high-resolution core scanning image, the medium-resolution core scanning image and the low-resolution core scanning image are obtained based on CT scanning or FIB-SEM scanning or QEMSCAN scanning of the plug sample.

[0013] Preferably, the resolution of the high-resolution core scanning image is set to 1 μm, the resolution of the medium-resolution core scanning image is set to 4.16 μm, and the resolution of the low-resolution core scanning image is set to 15.4 μm.

[0014] Preferably, in step 2, the image gray scale range of the core scanning image is 0-255.

[0015] Preferably, in step 3, the following steps are included:

[0016] Step 3.1, a first generative adversarial network model for constructing a double-scale digital core image is established, the first generative adversarial network model is initialized, the parameters of the generator and the discriminator are set, and the matched medium-resolution core scanning image and high-resolution core scanning image of the core sample image are used as training images;

[0017] Step 3.2, input the training image into the first generative adversarial network model, and generate a double-scale digital core image containing double image information of the medium-resolution core scanning image and the high-resolution core scanning image by using the generator in the first generative adversarial network model;

[0018] Step 3.3, keep the parameters of the discriminator in the first generative adversarial network model unchanged, input the double-scale digital core image and the high-resolution core scanning image matched with the core sample image into the discriminator, and the discriminator judges by extracting the image information in the double-scale digital core image and the high-resolution core scanning image; if the discriminator can judge that the double-scale digital core image generated by the generator is false, the parameters of the generator are adjusted, and the generator continues to be trained in step 3.2; if the discriminator cannot judge that the double-scale digital core image generated by the generator is false, step 3.4 is entered;

[0019] Step 3.4, keep the parameters of the generator in the first generative adversarial network model unchanged, input the double-scale digital core image generated by the generator and the high-resolution core scanning image matched with the core sample image into the discriminator, and the discriminator judges by extracting the image information in the double-scale digital core image and the high-resolution core scanning image; if the discriminator cannot judge that the double-scale digital core image generated by the generator is false, the parameters of the discriminator are adjusted, and the discriminator continues to be trained in step 3.4; if the discriminator can judge that the double-scale digital core image generated by the generator is false, step 3.5 is entered;

[0020] Step 3.5, analyze the loss functions of the discriminator and the generator; if the loss functions in the discriminator and the generator do not reach Nash equilibrium, return to step 3.2 to continue training the first generative adversarial network model; if the loss functions in the discriminator and the generator reach Nash equilibrium, the training of the first generative adversarial network model is completed, and step 3.6 is entered;

[0021] Step 3.6, input the medium-resolution core scanning image matched with the core sample image into the trained first generative adversarial network model, and generate a double-scale digital core image of the core sample by using the first generative adversarial network model.

[0022] Preferably, in step 4, the following steps are specifically included:

[0023] Step 4.1, a second generative adversarial network model for constructing a multi-scale digital core image is established, the second generative adversarial network model is provided with a generator and a discriminator, the second generative adversarial network model is initialized, the parameters of the generator and the discriminator are set, and the double-scale digital core image generated in step 3 and the low-resolution core scanning image matched with the core sample image are used as training images.

[0024] Step 4.2, input the training image into the second generative adversarial network model, and generate a multi-scale digital core image containing multiple image information of low-resolution core scanning images, medium-resolution core scanning images and high-resolution core scanning images by using the generator in the second generative adversarial network model;

[0025] Step 4.3, keeping the parameters of the discriminator in the second generative adversarial network model unchanged, inputting the multi-scale digital core image and the double-scale digital core image generated in step 3 into the discriminator, and the discriminator judges by extracting the image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, the parameters of the generator are adjusted, and the generator is returned to step 4.2 for continuous training, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, step 4.4 is entered;

[0026] Step 4.4, keeping the parameters of the generator in the second generative adversarial network model unchanged, inputting the multi-scale digital core image and the double-scale digital core image generated by the generator into the discriminator, and the discriminator judges by extracting the image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, the parameters of the discriminator are adjusted, and the discriminator is returned to step 4.4 for continuous training, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, step 4.5 is entered;

[0027] Step 4.5, analyzing the loss function of the discriminator and the loss function of the generator, if the loss functions in the discriminator and the generator do not reach Nash equilibrium, step 4.2 is returned for continuous training of the second generative adversarial network model, if the loss functions in the discriminator and the generator all reach Nash equilibrium, the training of the second generative adversarial network model is completed, and step 4.6 is entered;

[0028] Step 4.6, inputting the low-resolution core scanning image matched with the core sample image into the trained second generative adversarial network model, and generating a multi-scale digital core image of the core sample by using the second generative adversarial network model.

[0029] The present application has the following beneficial effects:

[0030] The application trains a first generative adversarial network model for constructing a double-scale digital core image and a second generative adversarial network model for constructing a multi-scale digital core image based on deep learning, generates a multi-scale digital core image by using the trained second generative adversarial network model, fully utilizes the accuracy of a high-resolution core scanning image and the breadth of a low-resolution core scanning image, realizes the fusion of multi-resolution digital core image information, solves the problem of missing core image information caused by the inability of different scale core scanning images to be fused, improves the resolution of the digital rock image, lays a foundation for guiding the exploration and development of various strata, and has a wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a flowchart of a multi-scale digital core image construction method based on deep learning.

[0032] Figure 2 It is a matched core sample image. Figure 2 In the figure, (a) is a low-resolution core scanning image matched with the core sample image, (b) is a medium-resolution core scanning image matched with the core sample image, and (c) is a high-resolution core scanning image matched with the core sample image.

[0033] Figure 3 It is a structural diagram of a generative adversarial network.

[0034] Figure 4 It is a schematic diagram of a multi-scale digital core image construction process.

[0035] Figure 5 It is a double-scale digital core image of a core sample.

[0036] Figure 6 It is a comparison diagram of a multi-scale digital core image of a core sample. Figure 6 In the figure, (a) is a low-resolution core scanning image matched with the core sample image, and (b) is a multi-scale digital core image of the core sample. DETAILED DESCRIPTION

[0037] The specific embodiments of the application will be further described below in combination with the drawings and an oilfield as an example:

[0038] Taking an oilfield as a research area, a multi-scale digital core image is constructed by using a multi-scale digital core image construction method based on deep learning proposed by the application, as shown in the figure, which specifically includes the following steps: Figure 1

[0039] ​Step 1, obtaining core samples according to coring data of a study area, drilling multiple plug samples at the same position of the core sample, and obtaining core scanning images of different resolutions by scanning the plug samples based on CT scanning or FIB-SEM scanning or QEMSCAN scanning.

[0040] In the process of preparing the plug samples by multistage drilling in the embodiment, a first plug sample with a diameter of 2.54 cm is first drilled, a low-resolution core scanning image with a resolution of 15.4 μm is obtained by CT scanning the first plug sample, a second plug sample with a diameter of 4 mm is drilled in the first plug sample, a medium-resolution core scanning image with a resolution of 4.16 μm is obtained by CT scanning the second plug sample, and finally, a third plug sample with a diameter of 1 mm is drilled in the second plug sample, and a low-resolution core scanning image with a resolution of 1 μm is obtained by CT scanning the third plug sample.

[0041] Step 2, scaling the image gray value of each core scanning image to the gray scale range of 0-255 according to the gray histogram of each core scanning image, and performing image matching on the core scanning images of each plug sample based on SIFT features to obtain low-resolution core scanning images (image resolution is 15.4 μm, image size is 162x162x162 voxels, as shown in (a)), medium-resolution core scanning images (image resolution is 4.16 μm, image size is 144x144x144 voxels, as shown in (b)) and high-resolution core scanning images (image resolution is 1 μm, image size is 600x600x600 voxels, as shown in (c)) of the core sample after image matching. Figure 2 Figure 2 Figure 2

[0042] The image matching based on SIFT features is prior art, and the image matching using SIFT features registers the high-resolution core scanning image with the medium-resolution core scanning image, and registers the medium-resolution core scanning image with the low-resolution core scanning image, that is, restores the core scanning image of the next higher resolution to the position of the core scanning image of the first resolution, so as to match the core scanning images of different resolutions at the same position of the core sample.

[0043] Step 3, establishing a first generative adversarial network model for constructing a dual-scale digital core image, as shown in (a) and (b). Figure 3 Figure 4 ​​​​As shown, based on deep learning, the medium resolution scan image and the high resolution scan image matched with the core sample image are used as training images to train the first generative adversarial network model, and the medium resolution scan image matched with the core sample image is input into the trained first generative adversarial network model, and the first generative adversarial network model is used to generate the dual-scale digital core image of the core sample, specifically including the following steps:

[0044] Step 3.1, a first generative adversarial network model for constructing a dual-scale digital core image is established, a generator and a discriminator are arranged in the first generative adversarial network model, the first generative adversarial network model is initialized, parameters of the generator and the discriminator are set, and the medium resolution scan image and the high resolution scan image matched with the core sample image are used as training images.

[0045] In this embodiment, the loss function of the generator is:

[0046] l G =λ mse l MSE +λ perc l X +λ gen l Gen (1)

[0047] Wherein, l MSE is the reconstruction loss, l X is the content loss, and l Gen is the adversarial loss.

[0048]

[0049] Wherein, W i,j and H i,j are the dimensions of each feature map in the VGG network, D θD (G θG (I LR )) is the probability of the reconstructed image, and G θG (I LR ) is the probability of the high-level resolution image.

[0050] The loss function of the discriminator is:

[0051]

[0052] Step 3.2, input the training image into the first generative adversarial network model, and use the generator in the first generative adversarial network model to generate a dual-scale digital core image, which contains dual image information of the medium resolution core scan image and the high resolution core scan image.

[0053] Step 3.3, keeping the parameters of the discriminator in the first generative adversarial network model unchanged, inputting the high-resolution scanned image matched with the core sample image into the discriminator, and the discriminator judging by extracting the image information in the dual-scale digital core image and the high-resolution scanned image. If the discriminator can judge that the dual-scale digital core image generated by the generator is false, the parameters of the generator are adjusted, and the generator is continued to be trained in step 3.2. If the discriminator cannot judge that the dual-scale digital core image generated by the generator is false, step 3.4 is entered.

[0054] Step 3.4, keeping the parameters of the generator in the first generative adversarial network model unchanged, inputting the dual-scale digital core image generated by the generator and the high-resolution scanned image matched with the core sample image into the discriminator, and the discriminator judging by extracting the image information in the dual-scale digital core image and the high-resolution scanned image. If the discriminator cannot judge that the dual-scale digital core image generated by the generator is false, the parameters of the discriminator are adjusted, and the discriminator is continued to be trained in step 3.4. If the discriminator can judge that the dual-scale digital core image generated by the generator is false, step 3.5 is entered.

[0055] Step 3.5, analyzing the loss functions of the discriminator and the generator. If the loss functions in the discriminator and the generator do not reach Nash equilibrium, the first generative adversarial network model is continued to be trained in step 3.2. If the loss functions in the discriminator and the generator reach Nash equilibrium, the training of the first generative adversarial network model is completed, and step 3.6 is entered.

[0056] Step 3.6, inputting the medium-resolution scanned image matched with the core sample image into the trained first generative adversarial network model, and generating the dual-scale digital core image of the core sample by the first generative adversarial network model (the image size is 600x600x600 voxels, as shown in FIG. 8). The dual-scale digital core image contains the image information of the high-resolution core scanned image. Figure 5

[0057] Step 4, establishing a second generative adversarial network model for constructing a multi-scale digital core image, training the second generative adversarial network model based on deep learning using the dual-scale digital core image generated in step 3 and the low-resolution scanned image matched with the core sample image as training images, and inputting the low-resolution scanned image matched with the core sample image into the trained second generative adversarial network model, and generating a three-dimensional digital core image of the core sample by the second generative adversarial network model. Specifically, the following steps are included:

[0058] ​Step 4.1, a second generative adversarial network model for constructing a multi-scale digital core image is established, a generator and a discriminator are arranged in the second generative adversarial network model, the second generative adversarial network model is initialized, parameters of the generator and the discriminator are set, loss functions of the generator and the discriminator in the second generative adversarial network model are the same as those of the first generative adversarial network model, and the low-resolution scanning image matched with the double-scale digital core image and the core sample image generated in step 3 is taken as a training image.

[0059] Step 4.2, the training image is input into the second generative adversarial network model, the generator in the second generative adversarial network model is used to generate a multi-scale digital core image, and the multi-scale digital core image contains multiple image information of a low-resolution core scanning image, a medium-resolution core scanning image and a high-resolution core scanning image.

[0060] Step 4.3, parameters of the discriminator in the second generative adversarial network model are kept unchanged, the multi-scale digital core image and the double-scale digital core image generated in step 3 are input into the discriminator, the discriminator judges by extracting image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, the parameters of the generator are adjusted, and the generator is continuously trained in step 4.2, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, step 4.4 is entered.

[0061] Step 4.4, parameters of the generator in the second generative adversarial network model are kept unchanged, the multi-scale digital core image generated by the generator and the double-scale digital core image are input into the discriminator, the discriminator judges by extracting image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, the parameters of the discriminator are adjusted, and the discriminator is continuously trained in step 4.4, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, step 4.5 is entered.

[0062] Step 4.5, the loss functions of the discriminator and the generator are analyzed, if the loss functions in the discriminator and the generator do not reach Nash equilibrium, step 4.2 is returned to continuously train the second generative adversarial network model, if the loss functions in the discriminator and the generator both reach Nash equilibrium, the training of the second generative adversarial network model is completed, and step 4.6 is entered.

[0063] Step 4.6, the low-resolution scanning image of the core sample after image matching (image size: 600x600x600 voxels) is input into the trained second generative adversarial network model, and the second generative adversarial network model is used to generate a multi-scale digital core image (image size: 600x600x600 voxels) of the core sample.

[0064] Step 5, according to the three-dimensional digital core image of the core sample, by comparing the low-resolution core scanning image of the core sample after image matching and the multi-scale digital core image, the pore boundary and micro-pore structure of the core sample are determined as shown in Figure 6 Figure 6 (a) and Figure 6 (b) can be obtained, the multi-scale digital core image constructed by the method has fused image information of the low-resolution core scanning image, the medium-resolution core scanning image and the high-resolution core scanning image, has clearer pore boundary and micro-pore structure, and lays a foundation for guiding exploration and development of various strata.

[0065] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples, and changes, modifications, additions or replacements made by those skilled in the art within the essential scope of the present application shall also belong to the protection scope of the present application.​

Claims

1. A deep learning-based multi-scale digital core image construction method, characterized in that, Specifically comprising the following steps: Step 1, obtaining a core sample according to coring data of a research area, drilling multiple piston samples at the same position of the core sample, and obtaining core scanning images of different resolutions by scanning the piston samples multiple times; Step 2, scaling the image gray value of each core scanning image to a unified gray scale range according to the gray histogram of each core scanning image, performing image matching on the core scanning images of each piston sample based on SIFT features, and obtaining low-resolution core scanning images, medium-resolution core scanning images and high-resolution core scanning images of the core sample after image matching; Step 3, establishing a first generative adversarial network model for constructing a dual-scale digital core image, training the first generative adversarial network model based on deep learning by using the medium-resolution core scanning images and the high-resolution core scanning images after image matching of the core sample as training images, inputting the medium-resolution core scanning images after image matching of the core sample into the trained first generative adversarial network model, and generating a dual-scale digital core image of the core sample by using the first generative adversarial network model; Step 4, establishing a second generative adversarial network model for constructing a multi-scale digital core image, training the second generative adversarial network model based on deep learning by using the dual-scale digital core image generated in step 3 and the low-resolution core scanning images after image matching of the core sample as training images, inputting the low-resolution core scanning images after image matching of the core sample into the trained second generative adversarial network model, and generating a three-dimensional digital core image of the core sample by using the second generative adversarial network model; Step 5, determining the pore boundary and micro-pore structure of the core sample according to the three-dimensional digital core image of the core sample.

2. The method of claim 1, wherein, In the step 1, the high-resolution core scanning image, the medium-resolution core scanning image and the low-resolution core scanning image are obtained by scanning the piston sample based on CT scanning or FIB-SEM scanning or QEMSCAN scanning.

3. The method of claim 2, wherein, The resolution of the high-resolution core scanning image is set to 1 μm, the resolution of the medium-resolution core scanning image is set to 4.16 μm, and the resolution of the low-resolution core scanning image is set to 15.4 μm.

4. The method of claim 1, wherein, In the step 2, the image gray scale range of the core scanning image is 0-255.

5. The method of claim 1, wherein, In the step 3, specifically comprising the following steps: Step 3.1, establishing a first generative adversarial network model for constructing a dual-scale digital core image, setting a generator and a discriminator in the first generative adversarial network model, initializing the first generative adversarial network model, setting parameters of the generator and the discriminator, and taking the medium-resolution core scanning images and the high-resolution core scanning images after image matching of the core sample as training images; Step 3.2, inputting the training images into the first generative adversarial network model, generating a dual-scale digital core image by using the generator in the first generative adversarial network model, and the dual-scale digital core image containing dual image information of the medium-resolution core scanning image and the high-resolution core scanning image; Step 3.3, keeping the parameters of the discriminator in the first generative adversarial network model unchanged, inputting the high-resolution core scanning image matched with the dual-scale digital core image and the core sample image into the discriminator, and the discriminator judging by extracting the image information in the dual-scale digital core image and the high-resolution core scanning image; if the discriminator can judge that the dual-scale digital core image generated by the generator is false, adjusting the parameters of the generator, returning to step 3.2 to continue training the generator; if the discriminator cannot judge that the dual-scale digital core image generated by the generator is false, entering step 3.4; Step 3.4, keeping the parameters of the generator in the first generative adversarial network model unchanged, inputting the high-resolution core scanning image matched with the dual-scale digital core image generated by the generator and the core sample image into the discriminator, and the discriminator judging by extracting the image information in the dual-scale digital core image and the high-resolution core scanning image; if the discriminator cannot judge that the dual-scale digital core image generated by the generator is false, adjusting the parameters of the discriminator, returning to step 3.4 to continue training the discriminator; if the discriminator can judge that the dual-scale digital core image generated by the generator is false, entering step 3.5; Step 3.5, analyzing the loss functions of the discriminator and the generator; if the loss functions in the discriminator and the generator do not reach Nash equilibrium, returning to step 3.2 to continue training the first generative adversarial network model; if the loss functions in the discriminator and the generator reach Nash equilibrium, completing the training of the first generative adversarial network model, and entering step 3.6; Step 3.6, inputting the core sample image matched with the medium-resolution core scanning image into the trained first generative adversarial network model, and generating the dual-scale digital core image of the core sample by the first generative adversarial network model.

6. The method of claim 1, wherein, In step 4, the following steps are specifically included: Step 4.1, establishing a second generative adversarial network model for constructing a multi-scale digital core image, setting a generator and a discriminator in the second generative adversarial network model, initializing the second generative adversarial network model, setting the parameters of the generator and the discriminator, and taking the low-resolution core scanning image matched with the dual-scale digital core image and the core sample image generated in step 3 as a training image; Step 4.2, inputting the training image into the second generative adversarial network model, and generating a multi-scale digital core image by the generator in the second generative adversarial network model, the multi-scale digital core image containing multiple image information of low-resolution core scanning images, medium-resolution core scanning images and high-resolution core scanning images; Step 4.3, keeping the parameters of the discriminator in the second generative adversarial network model unchanged, inputting the multi-scale digital core image and the double-scale digital core image generated in step 3 into the discriminator, and the discriminator judging by extracting image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, adjusting the parameters of the generator, returning to step 4.2 to continue training the generator, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, entering step 4.4; Step 4.4, keeping the parameters of the generator in the second generative adversarial network model unchanged, inputting the multi-scale digital core image generated by the generator and the double-scale digital core image into the discriminator, and the discriminator judging by extracting image information in the multi-scale digital core image and the double-scale digital core image, if the discriminator cannot judge that the multi-scale digital core image generated by the generator is false, adjusting the parameters of the discriminator, returning to step 4.4 to continue training the discriminator, if the discriminator can judge that the multi-scale digital core image generated by the generator is false, entering step 4.5; Step 4.5, analyzing the loss function of the discriminator and the loss function of the generator, if the loss functions in the discriminator and the generator do not reach Nash equilibrium, returning to step 4.2 to continue training the second generative adversarial network model, if the loss functions in the discriminator and the generator all reach Nash equilibrium, completing the training of the second generative adversarial network model, entering step 4.6; Step 4.6, inputting the low-resolution core scanning image matched with the core sample image into the trained second generative adversarial network model, and generating the multi-scale digital core image of the core sample by using the second generative adversarial network model.

Images