Electric imaging logging blank area intelligent repairing method based on graph adversarial transfer learning

Abstract

The invention discloses an electric imaging logging blank area intelligent restoration method based on graph adversarial transfer learning. The method comprises the steps that a first training data set is established according to public image data; constructing an initial image restoration model based on the graph neural network and the generative adversarial network, and training and testing the initial image restoration model by using the first training data set to obtain a pre-trained image restoration model; acquiring conductivity logging data, and establishing a second training data set according to the conductivity logging data; carrying out migration training on the pre-trained image restoration model by utilizing the second training data set to obtain an optimized image restoration model; and inputting a to-be-repaired electric imaging logging image into the optimized image repair model to complete the repair of the electric imaging logging image. The method can effectively improve the restoration precision of the image of the electric imaging logging blank area, fills up the image missing area caused by the instrument problem, and also has remarkable advantages in the aspects of processing speed and adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical imaging logging, and in particular, to an intelligent repair method for blank areas of electrical imaging logging based on graph adversarial transfer learning. Background Art

[0002] Electrical imaging logging measures the potential difference between rocks and fluids around the wellbore by applying a high-voltage direct current or alternating current electric field. These potential differences reflect physical properties such as the resistivity and dielectric constant of rocks and fluids. Combining signal processing algorithms, electrical imaging logging can generate high-resolution two-dimensional images of the distribution of underground rocks and fluids, revealing the structural and attribute characteristics of geological bodies such as the sedimentary structure, fractures, and pores of rocks. However, due to the limitations of the logging instrument size and the measurement wellbore environment, the wellbore coverage rate of the data collected by the resistivity imaging instrument cannot reach 100%, and there are gaps between the electrodes, resulting in blank strip areas in the electrical imaging logging image, which affects the accurate interpretation of the reservoir. In order to improve the repair accuracy of the blank area and enhance the image quality, it is usually necessary to repair the electrical imaging logging image.

[0003] Most of the existing methods are based on the grayscale images of electrical imaging logging, and the missing areas in the images are intelligently repaired through digital image processing, traditional machine learning methods, and deep learning methods, such as interpolation algorithms, partial differential equation (PDE) algorithms, texture synthesis methods, least squares methods, and the Filtersim algorithm of multi-point geostatistics. However, the existing methods all have some significant drawbacks, mainly manifested in that when processing electrical imaging images with complex geological changes, the details in the images cannot be restored, and the repair effect is not natural, with image details being lost or becoming blurred. With the improvement of the accuracy requirements for oil and gas exploration, the repair of electrical imaging logging images needs to be significantly improved, and the current repair methods and repair effects are not ideal. The main technical problems in repair are insufficient repair accuracy and poor real-time performance in large-scale and irregular missing areas; in addition, the existing repair technologies do not completely repair the image geological features when facing complex geological features. At the same time, the scarcity problem of electrical imaging logging images also needs to be solved.

[0004] Therefore, the method of the present invention proposes an intelligent repair method for blank areas of electrical imaging logging based on graph adversarial transfer learning. Based on the deep fusion algorithm in deep learning, it effectively fills the blank areas in the image caused by instrument problems, improves the repair accuracy of the blank areas and enhances the image quality, providing help for the picking of formation dip angles and the interpretation of geological features. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent repair method for blank areas in electrical imaging logging based on graph adversarial transfer learning to solve the technical problems that most of the current digital image repair methods are based on simple mathematical models and are difficult to cope with complex geological changes and geological feature details in electrical imaging logging images, as well as large-scale repair requirements.

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

[0007] In a first aspect, the present invention provides an intelligent repair method for blank areas in electrical imaging logging based on graph adversarial transfer learning, including:

[0008] Establish a first training data set according to public image data;

[0009] Construct an initial image repair model based on a graph neural network and a generative adversarial network, and use the first training data set to train and test the initial image repair model to obtain a pre-trained image repair model;

[0010] Obtain conductivity logging data, and establish a second training data set according to the conductivity logging data;

[0011] Use the second training data set to perform transfer training on the pre-trained image repair model to obtain an optimized image repair model;

[0012] Input the actually collected electrical imaging logging image to be repaired into the optimized image repair model to complete the repair of the missing area of the electrical imaging logging image.

[0013] Further, the construction of the initial image repair model based on the graph neural network and the generative adversarial network includes:

[0014] The image repair model includes a graph neural network with a gated aggregation mechanism and an improved Pix2Pix network designed based on the generative adversarial network, which is connected to the graph neural network;

[0015] The graph neural network with a gated aggregation mechanism includes multiple graph convolutional layers, each graph convolutional layer is connected to multiple gated aggregation layers, the output end of the last gated aggregation layer is connected to a pooling layer, and the output end of the pooling layer is connected to a deconvolutional layer; the graph neural network is used to perform information aggregation and graph data feature matrix propagation on the input graph data, and is also used to adaptively weight the information of the target node and neighbor nodes by using the gated aggregation mechanism to preliminarily repair the image corresponding to the graph data;

[0016] The generator of the improved Pix2Pix network has a U-net structure and is used to take the input image as a conditional input and generate a target image that matches the input image through adversarial training.

[0017] Furthermore, the graph convolutional layer completes the aggregation of input information and the propagation of the graph data feature matrix by performing weighted summation on the adjacency matrix and the node feature matrix in the graph data, which is expressed by the formula:

[0018]

[0019] where, H l represents the node feature matrix of the l-th layer; represents the normalized adjacency matrix, represents the normalization of the adjacency matrix, represents the normalization of the degree matrix; W l is the weight matrix of the l-th layer, and σ is the activation function used to introduce non-linear transformation.

[0020] Furthermore, the gated aggregation mechanism of the graph neural network is expressed by the formula:

[0021]

[0022] where, W f and W g are learnable weight matrices, (I is the identity matrix), is the node degree matrix of Ai, and σ r and σ g are the ReLU and Sigmoid activation functions respectively.

[0023] Furthermore, the loss function of the graph neural network adopts the mean squared error loss function, which is used to calculate the difference between the restored image and the original image. The loss function is expressed by the formula:

[0024]

[0025] where, I pred (i) represents the restored value of the i-th pixel point, and I gt (i) is the true value of the i-th pixel point, and N is the total number of pixels in the image.

[0026] Furthermore, in the improved Pix2Pix network, the loss function of the generator consists of two parts: the generative adversarial loss and the L1 reconstruction loss function, which is expressed by the formula:

[0027] L D (G,D) = L GAN (G,D) + λL L1 (G),

[0028]

[0029]

[0030] Among them, λ is a weight parameter used to control the weight of L1; ||·||1 represents the L1 norm, which is used to measure the pixel difference between the generated image and the real image; D(I in , I gt ) represents the output of the discriminator for the real image (I in , I gt ), and G(I in ) represents the inpainted image output by the generator.

[0031] Furthermore, a first training dataset is established based on the common image data, including:

[0032] Randomly generate an irregular mask image from the common image data according to a preset mask processing method, and use the mask image to mark the inpainting area of the common image data;

[0033] The common image data is segmented into irregular pixel blocks by using the superpixel segmentation method, and different pixel blocks are converted into graph data to obtain a graph dataset; among them, each irregular pixel block is used as a node of the graph data, and the spatial adjacency relationship between nodes is determined according to node similarity;

[0034] A first training dataset is obtained according to the graph dataset and the corresponding original data, and the path information of the common image data in the dataset is recorded.

[0035] Furthermore, conductivity logging data is obtained, and a second training dataset is established based on the conductivity logging data, including:

[0036] Map the conductivity data obtained by the logging instrument into a grayscale image by means of grayscale mapping, and generate a random mask image of the corresponding size;

[0037] A second training dataset is obtained based on the random mask image and the grayscale image.

[0038] Furthermore, transfer training is performed on the pre-trained image inpainting model by using the second training dataset to obtain an optimized image inpainting model, including:

[0039] During transfer training, freeze the input layer and some graph convolutional layers of the graph convolutional neural network in the pre-trained image inpainting model, and retain the output layer to adapt to the topological structure characteristics of the electrical imaging image;

[0040] Train the network parameters of the last two layers of the generator and the discriminator to make the model adapt to the inpainting task of the electrical imaging image features.

[0041] On the other hand, the present invention also provides an intelligent repair system for blank areas of electrical imaging logging based on graph adversarial transfer learning, including:

[0042] A first acquisition module, configured to establish a first training data set according to public image data;

[0043] A model construction module, configured to construct an initial image repair model based on a graph neural network and a generative adversarial network;

[0044] A pre-training module, configured to train and test the initial image repair model using the first training data set to obtain a pre-trained image repair model;

[0045] A second acquisition module, configured to acquire conductivity logging data and establish a second training data set according to the conductivity logging data;

[0046] A transfer training module, configured to perform transfer training on the pre-trained image repair model using the second training data set to obtain an optimized image repair model;

[0047] An image repair module, configured to input the actually acquired electrical imaging logging image to be repaired into the optimized image repair model to complete the repair of the missing area of the electrical imaging logging image.

[0048] Compared with the prior art, the advantages provided by the present invention are as follows: By combining public image data with images in the field of electrical imaging logging, an image repair model is constructed based on a graph neural network and a generative adversarial network, making the image repair process more accurate and highly professional. Through transfer learning, the model can be optimized in multiple stages, not only improving the image repair quality, but also better solving the blank area problem in the repair of electrical imaging logging images, ultimately enhancing the usability and accuracy of electrical imaging logging images. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic flow chart of the intelligent repair method for blank areas of electrical imaging logging based on graph adversarial transfer learning provided by the present invention;

[0050] Figure 2 is a schematic mask diagram of the public data set provided by the present invention;

[0051] Figure 3 is a schematic diagram of converting image data into graph data through a feature pyramid provided by the present invention;

[0052] Figure 4 is a structural and data processing flow chart of the graph neural network model provided by the present invention;

[0053] Figure 5 is an architecture diagram of the improved Pix2Pix network provided by the present invention;

[0054] Figure 6 Schematic diagram for converting conductivity data into grayscale images and improving resolution provided by the present invention;

[0055] Figure 7 Mask schematic diagram of the electrical imaging logging data provided by the present invention;

[0056] Figure 8 Schematic diagram of the restoration effect of the image restoration model provided by the present invention;

[0057] Figure 9 Schematic diagram of the structure of the intelligent restoration system for blank areas of electrical imaging logging based on graph adversarial transfer learning provided by the present invention. Detailed implementation manners

[0058] The following will specifically describe the preferred embodiments of the present invention with reference to the accompanying drawings, where the accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0059] Please refer to Figure 1 , this embodiment provides an intelligent restoration method for blank areas of electrical imaging logging based on graph adversarial transfer learning, including:

[0060] Step S101: Establish a first training data set according to the common image data;

[0061] Step S102: Construct an initial image restoration model based on a graph neural network and a generative adversarial network, and use the first training data set to train and test the image restoration model to obtain a pre-trained image restoration model;

[0062] Step S103: Obtain historical conductivity data, and establish a second training data set according to the historical conductivity data;

[0063] Step S104: Use the second training data set to perform transfer training on the pre-trained image restoration model to obtain an optimized image restoration model for electrical imaging logging image restoration;

[0064] Step S105: Input the actually collected electrical imaging logging image to be restored into the optimized image restoration model to complete the restoration of the blank area of the image.

[0065] The method of this embodiment collects and processes electrical imaging logging image data, establishes an electrical imaging logging image data set, and establishes a graph adversarial neural network model and a transfer learning method to repair the blank areas of electrical imaging, improving the performance and repair accuracy of the model, realizing the intelligent repair of various irregular and large-area missing electrical imaging image data, providing a feasible basis for accurately interpreting the reservoir in oil and gas exploration, and having great engineering significance.

[0066] Since the electrical imaging logging images are usually generated by professional equipment, the acquisition of images is restricted by various factors such as equipment and environment, and it is difficult to collect enough samples for training. Therefore, this method adopts the method of transfer learning, which well solves the problem of data scarcity in repairing electrical imaging logging images using neural networks, and can accelerate the training process, enhance the generalization ability, avoid overfitting at the same time, and improve the actual performance of the image repair model.

[0067] As a specific embodiment, in step S101, in order to ensure the scale and diversity of the dataset, the public image data selects the Places2 and CelebA datasets containing a large number of natural scene picture data. CelebA is a large-scale facial attribute dataset containing more than 200,000 celebrity images. The images in the dataset cover a large range of pose changes and background clutter, and are suitable for the algorithm training of image repair tasks; the Places dataset contains more than 10 million images, including more than 400 unique scene categories. The Places dataset can learn deep scene features for various scene recognition tasks, and the goal is to establish new state-of-the-art performance on scene-centered benchmarks.

[0068] The size of each image in the above two types of datasets is 256×256 pixels. We use the Enhanced Deep Residual Network (EDSR) to increase the resolution of each image to 360×360 pixels, and the format of each image is PNG. Through this processing method, the details and clarity of the image can be improved, blurring can be reduced, and lossless quality can be maintained, which can provide higher-quality input data for subsequent deep learning analysis. Especially in geological structure analysis that requires detailed processing and high fidelity, using a public dataset with a higher resolution can bring more reliable and accurate results when pre-training the model.

[0069] As a preferred embodiment, in step S101, a first training dataset is established according to the public image data, including:

[0070] Step S201: Randomly generate an irregular mask image from the public image data according to a preset mask processing method, and use the mask image to mark the area to be repaired of the public image data;

[0071] Step S202: Use the superpixel segmentation method to segment the public image data into irregular pixel blocks, and convert different pixel blocks into graph data to obtain a graph dataset; wherein, each irregular pixel block is used as a node of the graph data, and the spatial adjacency relationship between nodes is determined according to node similarity;

[0072] Step S203: Divide the graph dataset into a training set, a validation set, and a test set to obtain a first training dataset, and record the path information of the common image data in the dataset.

[0073] As a specific embodiment, in step S201, the preset masking method can be a random rectangle / square mask, or a random circle / ellipse mask / random dot mask / random stripe mask or an irregular polygon mask, etc. Select a preset mask image to cover or set a corresponding area in the image to a special value (such as black, white, or other marker values).

[0074] As Figure 2 shown, for the Places2 and CelebA datasets, a unified mask image is used. The mask images account for 20%, 45%, 70% of the entire image respectively, and randomly generated mask images (black indicates the area to be repaired, Figure 2 the values in

[0075] As a specific embodiment, in step S202, please refer to Figure 3 , Figure 3 which shows the complete process of converting the images in the natural scene of the common image dataset into graph data.

[0076] First, perform superpixel clustering segmentation through the simple linear iterative clustering (SLIC) method. By decomposing the image into smaller regions with similar colors and spatial positions, the specific steps are shown in Table 1:

[0077] Table 1 Simple linear iterative clustering (SLIC) process

[0078]

[0079] Secondly, map the image after each superpixel segmentation to the feature pyramid. In this embodiment, the feature pyramid is designed with 5 layers, and the image resolution size of each layer is n represents the number of layers of the feature pyramid.

[0080] Thirdly, map each segmented pixel into graph data. Among them, each irregular pixel block is used as a node of the graph data, and the edge represents the spatial adjacency relationship of the superpixel blocks. The spatial adjacency relationship between nodes is determined according to node similarity, thereby constructing a large-scale graph dataset.

[0081] Finally, convert the graph data of each layer into an adjacency matrix A, which represents the node information to be transmitted. At the same time, each node contains the feature information X of the pixel block.

[0082] Through the above steps, the image data is converted into matrix data that can be input into the image inpainting model. The image dataset in the natural scene is divided into a training set, a validation set, and a test set according to the ratio of 8:1:1, and a CSV file is created to record the path information of the pictures in the dataset. When reading the dataset, the dataset can be imported into the network model by directly reading the CSV file.

[0083] As a preferred embodiment, in step S102, constructing the initial image inpainting model based on the graph neural network and the generative adversarial network includes:

[0084] The image inpainting model includes a graph neural network with a gated aggregation mechanism and an improved Pix2Pix network designed based on the generative adversarial network and connected to the graph neural network;

[0085] The graph neural network with a gated aggregation mechanism includes multiple graph convolutional layers, each graph convolutional layer is connected with multiple gated aggregation layers, the output end of the last gated aggregation layer is connected with a pooling layer, and the output end of the pooling layer is connected with a transposed convolutional layer; the graph neural network is used for information aggregation and propagation of the graph data feature matrix of the input graph data, and is also used for adaptively weighting the information of the target node and the neighbor nodes by using the gated aggregation mechanism to preliminarily repair the image corresponding to the graph data;

[0086] The generator of the improved Pix2Pix network has a U-net structure and is used to take the input image as a conditional input and generate a target image matching the input image through adversarial training.

[0087] As a specific embodiment, as Figure 4 shown, Figure 4 shows the structure and data processing flow of the graph neural network. The following will be described in detail in combination with Figure 4 for details.

[0088] In the image dataset in the natural scene, the values of the adjacency matrix A and the feature matrix X are used to aggregate the neighborhood information through the graph convolutional layer, and the feature matrix X is gradually propagated.

[0089] In this model structure, the graph convolutional layer performs weighted summation on the node feature matrix X through the adjacency matrix A, which is expressed by the formula:

[0090]

[0091] where H l represents the node feature matrix of the l-th layer; represents the normalized adjacency matrix, using represents the normalization of the adjacency matrix, represents the normalization of the degree matrix. Wl is the weight matrix of the l-th layer, and σ is the activation function used to introduce non-linear transformation.

[0092] After performing weighted summation on the adjacency matrix and the node matrix X through preliminary graph convolution, gated aggregation is then carried out to further aggregate the information of the missing feature matrix. Specifically, after each layer of graph convolution, a gating mechanism is used to control the influence (weight) of different neighbor nodes on the target node. It should be noted that the network is composed of several gated aggregation blocks with skip connections, and these modules perform feature aggregation iteratively in the graph G′ i to obtain high-dimensional feature vectors extracted at a certain depth from the input feature matrix F′ i and the adjacency matrix A i

[0093] The formula for calculating the gated aggregation weight g ij is as follows:

[0094] g ij = σ(W g [h i || h j )

[0095] where h i and h j are the feature vectors of node i and node j respectively, || represents the concatenation operation, W g is the weight matrix of the gating mechanism, and σ is the Sigmoid activation function.

[0096] Calculate the weights of the features of the missing nodes through the gated aggregation weights to obtain the dynamic weights g ij of the missing nodes. The specific gated weighted aggregation is that the gated weight g ij performs weighted aggregation on the node features, and the calculation formula is as follows:

[0097]

[0098] where N(i) represents the neighbor set of node i, W l represents the weight matrix of this layer, b l is the bias term, and the updated node feature matrix is

[0099] Through the operation of the graph convolution algorithm, the data information of the missing nodes can be obtained. After multiple layers of graph convolution and gated aggregation operations, the model will obtain the result of image inpainting. The last layer of the model is the activation function Sigmoid, and the missing node information I pred after passing through the activation function is:

[0100] I pred = σ(W​out H l +b out )

[0101] Among them, W out and b out are the weights and bias terms of the output layer, and H l is the data of the node feature matrix of the last layer. Through the above graph convolution design, the rough aggregation information of the missing nodes will be obtained. Then, the up and down adoption mechanism of the feature pyramid is used to map the graph structure data into the final rough repaired image I out′ .

[0102] In the above training, the Adma optimizer is selected, the initial value of the learning rate is 0.001, the batch training size is 16, the number of training times is set to 200, and the mean square error is used as the loss function. In the activation function, Relu is used in the graph convolution layer, and Sigmoid is used in the final output layer. The design of the loss function is crucial for the convergence and optimization of the graph convolution model. To train this model, we use the mean square error loss function to calculate the difference between the repaired image and the original image, which is expressed by the formula:

[0103]

[0104] Among them, I pred (i) represents the repaired value of the i-th pixel, and I gt (i) is the true value of the i-th pixel, and N is the total number of pixels in the image.

[0105] Through the graph convolution and gated aggregation mechanism, the model can effectively process the missing nodes in the image (such as the missing areas or noises in the image) and recover the missing image information. The gated mechanism plays a key role in this process. It can dynamically adjust the influence of each neighbor node, making the model more accurate in repairing the missing areas. Finally, through multiple graph convolution operations and gated aggregation, the model can generate the repaired image, restoring the details of the original image as much as possible and completing the preliminary repair of the image.

[0106] Next, the merging operation of the feature data is performed, and the aggregated graph data is merged into a rough repaired effect diagram through the feature pyramid. To finely repair the rough repaired image of the graph neural network, the present invention adopts an algorithm based on conditional adversarial neural network (improved Pix2Pix GAN).

[0107] As a specific embodiment, as Figure 5 shown,[[]]END]] Figure 5Shows the structural diagram and data processing flow of the improved Pix2Pix network. The improved Pix2Pix network is mainly used for image-to-image conversion. In image inpainting, the input image is the preliminarily inpainted image, and the output image is the complete inpainted image. Like other adversarial generative networks, this network model mainly includes a generator and a discriminator, and the main functions of the two components are as follows:

[0108] The goal of the generator (G) is to generate the inpainted image I according to the input image I out′ using the U-Net architecture, as shown in out For the generator of the Pix2Pix network, by extracting image features layer by layer and combining skip connections to retain high-resolution detail information. Figure 5 Shown, by extracting image features layer by layer and combining skip connections to retain high-resolution detail information.

[0109] The goal of the discriminator (D) is to help the generator optimize the generation process by distinguishing whether the input images (I out′ , I out ) come from the true data distribution.

[0110] The loss function of the improved Pix2Pix consists of two parts: the generative adversarial loss and the L1 reconstruction loss function. First, the generator G of the generative adversarial loss makes it difficult for the discriminator D to distinguish between the generated image and the real image, so that the generated image is as close as possible to the real image. The mathematical expression is as follows:

[0111]

[0112] where D(I in , I gt ) represents the output of the discriminator for the real images (I in , I gt ), and G(I in ) represents the inpainted image output by the generator. Through adversarial training, the effect of the generated image can be gradually optimized.

[0113] To maintain the pixel-level similarity between the generated image and the real image, in addition to using the generative adversarial loss, this embodiment also introduces the L1 reconstruction loss in the loss function to measure the difference between the generated image and the real image. The specific mathematical principle is as follows:

[0114]

[0115] where ||·||1 represents the L1 norm, which is used to measure the pixel difference between the generated image and the real image. In the present invention, the total loss function form is adopted for model training, that is, the generative adversarial loss and the L1 reconstruction loss are weighted and summed.

[0116] L D (G, D) = LGAN (G, D) + λL L1 (G)

[0117] Where λ is the weight parameter that controls the weight of L1 and is set to 100.

[0118] The above loss function design is mainly for the loss function of the generator.

[0119] The objective of the discriminator's loss function is to distinguish between real images and generated images as much as possible, maximize the probability of discriminating real images, and at the same time minimize the probability of discriminating generated images, so as to improve the discrimination ability between generated images and real images. The specific mathematical expression of the discriminator's loss function is as follows:

[0120]

[0121] Input the image data (I out′ ) initially repaired by graph convolution and the corresponding mask image into Pix2Pix GAN, and output the finely repaired image I out . At the same time, design and optimize the loss function of Pix2Pix GAN, and finally output the finely repaired image.

[0122] Through the above steps, the process of pre-training the graph neural network and the improved Pix2Pix network natural scene image pair model using images in the natural scene is carried out, and a finely repaired model for images in the natural scene is obtained. However, this model cannot be directly used for the missing repair of resistivity image logging. In order to repair the resistivity image logging model, and considering the scarcity of resistivity images at the same time, the present invention uses transfer learning to fine-tune and train the repair model again, so that the model can adapt to the repair task of resistivity image logging.

[0123] Before fine-tuning training, it is first necessary to obtain historical resistivity image logging images.

[0124] As a preferred embodiment, in step S103, the resistivity image logging image dataset is a conductivity dataset collected by Schlumberger's logging instrument Fullbore formation microimager (FMI). In order to establish a second training dataset based on conductivity logging data, the following operations need to be carried out:

[0125] First, convert the obtained conductivity data into the.CSV data format, select a data block size of 45×45, and use Python to write code to match the 45×45 data block. If there are no missing values in the data block, save the data block; otherwise, move 40 data lengths to the right each time until the entire CSV data is traversed, obtaining the final complete 45×45 dataset. Map the 45×45 dataset to grayscale to convert it into a grayscale image with a resolution of 45×45.

[0126] Next, use the Enhanced Deep Residual Network (EDSR) for super-resolution improvement, upgrading the 45×45 grayscale image to a resolution of 360×360, as Figure 6 shown. Figure 6 It shows the process of converting conductivity data into a grayscale image and improving the resolution. Use a mask image unified with the Places2 and CelebA datasets for the grayscale image, and use the mask image to mark the area to be repaired, as Figure 7 shown, in the PNG format. The image data used during transfer learning is the small amount of electrical imaging grayscale images generated at this time.

[0127] As a preferred embodiment, use the second training dataset to perform transfer learning on the pre-trained image inpainting model to obtain an optimized image inpainting model, including:

[0128] During transfer learning, freeze the input layer and some graph convolutional layers of the graph convolutional neural network in the pre-trained image inpainting model, and retain the output layer to adapt to the topological structure characteristics of electrical imaging images;

[0129] Train the network parameters of the last two layers of the generator and discriminator to make the model adapt to the inpainting task of electrical imaging image features.

[0130] As a specific embodiment, the learning rate during transfer learning is less than the learning rate during pre-training, and the model is fine-tuned. By reducing the learning rate, it is possible to avoid damaging the model structure in the electrical imaging logging image inpainting task and ensure the image inpainting ability of the original model. Through the method of transfer learning, the pre-inpainting model is fine-tuned, enabling the model to adapt to the blank area inpainting task of electrical imaging logging images and obtaining the final model for electrical imaging logging image inpainting.

[0131] Furthermore, after obtaining the optimized electrical imaging logging image inpainting model, use the test set to test the model and comprehensively evaluate the inpainting results. Calculate the inpainting accuracy of the model for various types of images to evaluate the actual generalization performance of the model. When both the training and test results of the model meet the expectations, save the trained model. During application, directly load the saved model and input the missing electrical imaging data and the corresponding mask image to achieve intelligent inpainting of the blank area of electrical imaging logging images. AsFigure 8 As shown Figure 8 The figure shows a comparison diagram of using the optimized model in this time to repair natural scene images and missing electrical imaging images respectively. As can be seen Figure 8 from it, the model of the present invention can accurately repair large-area and irregular image missing areas.

[0132] Such as Figure 9 shown, this embodiment also provides an intelligent repair system 900 for blank areas of electrical imaging logging based on graph adversarial transfer learning, including:

[0133] A first acquisition module 901, configured to establish a first training data set according to public image data;

[0134] A model construction module 902, configured to construct an initial image repair model based on a graph neural network and a generative adversarial network;

[0135] A pre-training module 903, configured to train and test the initial image repair model using the first training data set to obtain a pre-trained image repair model;

[0136] A second acquisition module 904, configured to acquire conductivity logging data and establish a second training data set according to the conductivity logging data;

[0137] A transfer training module 905, configured to perform transfer training on the pre-trained image repair model using the second training data set to obtain an optimized image repair model;

[0138] An image repair module 906, configured to input the actually acquired electrical imaging logging image to be repaired into the optimized image repair model to complete the repair of the missing area of the electrical imaging logging image.

[0139] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent repair method for blank areas in electrical imaging logging based on graph adversarial transfer learning, characterized in that: include: Establishing a first training data set based on public image data; Building an initial image restoration model based on a graph neural network and a generative adversarial network, and training and testing the initial image restoration model using a first training data set to obtain a pre-trained image restoration model; Acquire conductivity logging data, and establish a second training data set based on the conductivity logging data; Using the second training data set to perform migration training on the pre-trained image restoration model to obtain an optimized image restoration model; The actually acquired electrical imaging logging image to be repaired is input into the optimized image repair model to complete the repair of the missing area of ​​the electrical imaging logging image.

2. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 1 is characterized in that: The initial image restoration model is constructed based on the graph neural network and the generative adversarial network, including: The image restoration model includes a graph neural network with a gated aggregation mechanism, and an improved Pix2Pix network based on a generative adversarial network design connected to the graph neural network; The graph neural network with a gated aggregation mechanism includes a plurality of graph convolutional layers, each of which is connected to a plurality of gated aggregation layers, the output end of the gated aggregation layer at the end is connected to the pooling layer, and the output end of the pooling layer is connected to the deconvolution layer; the graph neural network is used to perform information aggregation and graph data feature matrix propagation on the input graph data, and is also used to adaptively weight the information of the target node and the neighboring node using the gated aggregation mechanism to perform preliminary repair on the image corresponding to the graph data; The generator of the improved Pix2Pix network has a U-net structure, which is used to take an input image as a conditional input and generate a target image matching the input image through adversarial training.

3. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 2 is characterized in that: The graph convolution layer aggregates the input information and propagates the graph data feature matrix by performing weighted summation on the adjacency matrix and node feature matrix in the graph data, which can be expressed as: Among them, H l Represents the node feature matrix of the lth layer; represents the normalized adjacency matrix, represents the normalization of the adjacency matrix, represents the normalization of the degree matrix; W l is the weight matrix of the lth layer, and σ is the activation function used to introduce nonlinear transformation.

4. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 3 is characterized in that: The gated aggregation mechanism of the graph neural network is expressed as follows: Among them, W f and W g is a learnable weight matrix, (I is the identity matrix), is the node degree matrix of Ai, σ r and σ g They are ReLU and Sigmoid activation functions respectively.

5. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 2 is characterized in that: The loss function of the graph neural network adopts the mean square error loss function, which is used to calculate the difference between the restored image and the original image. The loss function is expressed as follows: Among them, I pred (i) represents the repair value of the i-th pixel, I gt (i) is the true value of the i-th pixel, and N is the total number of pixels in the image.

6. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 2 is characterized in that: In the improved Pix2Pix network, the generator loss function consists of two parts: generative adversarial loss and L1 reconstruction loss function, which can be expressed as follows: L D (G,D)=L GAN (G,D)+λL L1 (G), Among them, λ is a weight parameter used to control the weight of L1; ||·||1 represents the L1 norm, which is used to measure the pixel difference between the generated image and the real image; D(I in ,I gt ) represents the discriminator's response to the real image (I in ,I gt ) output, G(I in ) represents the repaired image output by the generator.

7. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 1 is characterized in that: A first training data set is established based on public image data, including: Randomly generate an irregular mask image from the public image data according to a preset mask processing method, and use the mask image to mark the area to be repaired in the public image data; The public image data is segmented into irregular pixel blocks by using the superpixel segmentation method, and different pixel blocks are converted into graph data to obtain a graph data set; wherein each irregular pixel block is used as a node of the graph data, and the spatial adjacency relationship between nodes is determined according to the node similarity; A first training data set is obtained according to the image data set and the corresponding original data, and path information of common image data in the data set is recorded.

8. The method for intelligent repair of blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 1 is characterized in that: Obtaining conductivity logging data, and establishing a second training data set based on the conductivity logging data, including: The conductivity data obtained by the logging instrument is mapped into a grayscale image by grayscale mapping, and a random mask image of a corresponding size is generated; A second training data set is obtained based on the random mask image and the grayscale image.

9. The method for intelligently repairing blank areas in electrical imaging logging based on graph adversarial transfer learning according to claim 2, characterized in that: The pre-trained image restoration model is transferred and trained using the second training data set to obtain an optimized image restoration model, including: During transfer training, the input layer and part of the graph convolutional layer of the graph convolutional neural network in the pre-trained image restoration model are frozen, and the output layer is retained to adapt to the topological structure characteristics of the electrophysiological image; The network parameters of the last two layers of the generator and discriminator are trained to adapt the model to the task of restoring electro-imaging image features.

10. An intelligent repair system for blank areas in electrical imaging logging based on graph adversarial transfer learning, characterized in that: include: A first acquisition module, used to establish a first training data set based on public image data; Model building module, used to build an initial image restoration model based on graph neural network and generative adversarial network; A pre-training module, used to train and test the initial image restoration model using the first training data set to obtain a pre-trained image restoration model; A second acquisition module is used to acquire conductivity logging data and establish a second training data set according to the conductivity logging data; A migration training module, used to perform migration training on the pre-trained image restoration model using the second training data set to obtain an optimized image restoration model; The image restoration module is used to input the actually acquired electrical imaging logging image to be restored into the optimized image restoration model to complete the restoration of the missing area of ​​the electrical imaging logging image.

Images