Method and equipment for identifying pores in rock image based on deep learning

Abstract

The invention discloses a method and equipment for identifying pores in a rock image based on deep learning. The method comprises the following steps: acquiring a three-dimensional CT (Computed Tomography) slice image acquired from a rock; based on the three-dimensional CT slice image, input image data of a pre-trained rock pore recognition model is formed, and the rock pore recognition model comprises a multi-scale pore feature extraction network and a pore recognition network; extracting multi-scale pore feature data from the input image data by using the multi-scale pore feature extraction network; and taking the pore feature data of each scale as the input of the pore identification network, and identifying to obtain each pore in the three-dimensional CT slice image and the three-dimensional position area data of each pore.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and computing device for identifying pores in rock images based on deep learning. Background Art

[0002] As resources continue to advance deeper, the development of rock layers (such as coalbed methane) has received increasing attention. The key to studying how gas migrates in rocks is to obtain the distribution of pores. However, due to the limitations of the research scale of coal rocks, it is impossible to obtain a detailed pore distribution. Existing research shows that the fractal characteristics of pores are scale-invariant. The pore identification methods in the existing technology have low pore identification accuracy and it is difficult to distinguish between noise and pores. When pore identification is inaccurate, it is impossible to obtain an accurate pore distribution. Summary of the Invention

[0003] In order to solve the existing technical problems, the present invention provides a pore identification method and computing device in rock images based on deep learning, which can improve the accuracy of pore identification.

[0004] In a first aspect, a method for identifying pores in rock images based on deep learning is provided, comprising: obtaining three-dimensional CT slice images collected from rocks; forming input image data of a pre-trained rock pore identification model based on the three-dimensional CT slice images, wherein the rock pore identification model includes a multi-scale pore feature extraction network and a pore identification network; utilizing the multi-scale pore feature extraction network to extract pore feature data of multiple scales from the input image data; and using the pore feature data of each scale as input to the pore identification network to identify each pore in the three-dimensional CT slice image and obtain three-dimensional position area data of each pore.

[0005] In a second aspect, a computing device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the deep learning-based pore identification method in rock images provided in the first aspect of the present application.

[0006] This application can directly identify pores in three-dimensional CT slice images through a deep learning network, and can improve the accuracy of pore identification by performing identification based on pore characteristic data at different scales. Moreover, after identifying the three-dimensional position area data of the pores, the pore shape factor is directly output through a pore shape factor calculation network, facilitating the subsequent acquisition of more accurate pore distribution data in the rock image. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 FIG2 is a diagram showing an application environment of a method for identifying pores in rock images based on deep learning in one embodiment;

[0008] Figure 2 Flowchart of a method for identifying pores in rock images based on deep learning in one embodiment;

[0009] Figure 3 FIG1 is a network structure diagram of a rock pore identification model according to an embodiment of the present invention;

[0010] Figure 4 A flowchart of training a rock pore identification model in one embodiment;

[0011] Figure 5 This is a network structure diagram of an example training rock pore recognition model in one embodiment;

[0012] Figure 6 A flowchart of training a rock pore identification model in another embodiment;

[0013] Figure 7 Schematic diagram of a device for identifying pores in rock images based on deep learning in one embodiment;

[0014] Figure 8 is a schematic diagram of a computing device in one embodiment. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the scope of protection of the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] In the following description, reference is made to “some embodiments” which describe a subset of all possible embodiments, but it should be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0018] See Figure 1, is a diagram illustrating the application environment of a deep learning-based pore identification method in rock images in one embodiment. The deep learning-based pore identification method in rock images is applied to a computing device 10, which can acquire 3D computed tomography (CT) slice images. A 3D CT slice image is a 3D slice image within a 3D image reconstructed from a 3D measurement projection map, where a 3D slice image is a 3D slice image within a 3D image. 3D measurement projection maps can be acquired by performing 3D CT scanning on the rock using a detection device. The radiation source in the detection device rotates around the rock, acquiring 3D measurement projection maps of the rock from different angles. This projection data is then used to reconstruct a 3D image of the scanned area. The acquisition of the measurement projection map involves the attenuation of the radiation source. Tissues of different densities absorb the radiation to varying degrees, and this difference is used to distinguish different tissue structures. During scanning, a tube emits a narrow beam of radiation, and a detector opposite the tube measures the radiation attenuation. The data collected by the detection device consists of ray projection and profile data. The computing device 10 can reconstruct the 3D measurement projection map to obtain a 3D image. In some embodiments, the process of reconstructing the three-dimensional measurement projection map to obtain a three-dimensional image can also be calculated in other devices. The computing device 10 directly obtains the reconstructed three-dimensional image, and then segments the three-dimensional image to obtain a three-dimensional CT slice image.

[0019] See also Figure 2 , is a flow chart of a method for identifying pores in rock images based on deep learning according to an embodiment of the present application. The method for identifying pores in rock images based on deep learning is applied to a computing device and includes the following steps:

[0020] S11. Acquire a three-dimensional CT slice image of the rock.

[0021] S12. Based on the three-dimensional CT slice image, forming input image data of a pre-trained rock pore recognition model, wherein the rock pore recognition model includes a multi-scale pore feature extraction network and a pore recognition network.

[0022] In this embodiment, the rock pore identification model is trained based on a training dataset. The rock pore identification model includes a multi-scale pore feature extraction network and a pore identification network. The multi-scale pore feature extraction network is used to extract features of pores of different sizes. Since pores in rock formations vary in size, extracting features of pores of different scales facilitates the subsequent pore identification network's ability to identify pores of varying sizes.

[0023] S13. Utilize a multi-scale pore feature extraction network to extract pore feature data of multiple scales from the input image data.

[0024] In this embodiment, the multi-scale pore feature extraction network can be a multi-scale convolutional neural network (CNN), which includes multiple convolutional layers of different sizes, connected together. Each convolutional layer uses a convolution kernel of different sizes (e.g., 1×1, 3×3, and 5×5) to capture local features at different scales. By outputting the corresponding pore feature data at different convolutional layers, pore feature data at multiple scales is obtained.

[0025] For example Figure 3 As shown, Figure 3 Figure 1 shows the network structure of a rock pore identification model used in an embodiment. Three different scales of pore characteristic data are extracted from 3D CT slice images: first-scale pore characteristic data, second-scale pore characteristic data, and third-scale pore characteristic data. These three scales of pore characteristic data represent different pore sizes.

[0026] S14. Using the pore characteristic data of each scale as input to the pore recognition network, each pore in the three-dimensional CT slice image and the three-dimensional position area data of each pore are identified.

[0027] In this embodiment, the pore identification network is primarily used to detect pores and their three-dimensional location area data based on input pore feature data at different scales. The pore identification network includes a detection head, which identifies pore boundaries and pore centers. The three-dimensional location area data includes pore boundary data and the three-dimensional coordinates of the pore center. The detection head includes, but is not limited to, convolutional layers, fully connected layers, activation layers, pooling layers, and the like. For example, the detection head can be the detection head network structure used in the Faster R-CNN model or the network structure used in the YOLO detection head.

[0028] For example Figure 3 As shown, the first-scale pore characteristic data, the second-scale pore characteristic data and the third-scale pore characteristic data are respectively used as inputs of the pore identification network to obtain the first pore identification result, the second pore identification result and the third pore identification result respectively; then the first pore identification result, the second pore identification result and the third pore identification result are fused to obtain the final pore identification result, so that pores of different sizes can be included in the final pore identification result.

[0029] In the above embodiment, the three-dimensional CT slice image collected from the rock is input into a pre-trained rock pore identification model, and the pore feature data of different scales can be directly extracted through the multi-scale pore feature extraction network. The pore feature data of different scales are respectively used as input to the pore identification network to obtain multiple pore identification results of different scales. The pore identification results of different scales are fused to obtain the pores in the three-dimensional CT slice image and the three-dimensional position area data of each pore. The present application can directly identify the pores in the three-dimensional CT slice image through the deep learning network, and the identification is based on the pore feature data of different scales, which can improve the accuracy of pore identification, thereby obtaining more accurate pore distribution data in the rock image.

[0030] In some embodiments, the rock pore identification model includes a pore shape factor calculation network, and the method further includes:

[0031] The three-dimensional position area data of each pore is used as the input of the pore shape factor calculation network to calculate the shape factor of each pore.

[0032] In this embodiment, the pore shape factor is a dimensionless parameter that describes the geometry of the pores. For example, in petroleum engineering, the pore shape factor of a rock affects the permeability of oil and gas. When the pore shape is regular (large shape factor), the flow path of fluids such as oil and gas in the rock pores is smoother and the permeability is higher. Therefore, when detecting rock pores, it is also necessary to accurately detect the pore shape factor. The pore shape factor calculation network is used to construct three-dimensional structural data of the pores based on the three-dimensional position area data of the pores and calculate the pore shape factor using the formula of the principal axis length method. The pore shape factor calculation network includes but is not limited to convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layer is used to capture local boundary information and subtle shape changes of the pores. The main function of the pooling layer is to downsample the features extracted by the convolutional layer. The fully connected layer is located at the end of the network and globally integrates and classifies the features extracted by the previous convolutional and pooling layers. It maps the input feature vector to the output space to predict the final value of the pore shape factor. The output layer outputs the pore shape factor.

[0033] In the above embodiment, the present application can directly identify pores in three-dimensional CT slice images through a deep learning network, and the identification is based on pore characteristic data of different scales, which can improve the accuracy of pore identification; and after identifying the three-dimensional position area data of the pore, the pore shape factor is directly output through the pore shape factor calculation network, so that more accurate pore distribution data can be obtained in the rock image later.

[0034] In some embodiments, the method further comprises:

[0035] The rock pore identification model is trained.

[0036] like Figure 4 As shown in Figure 2, the process steps for training the rock pore recognition model include the following:

[0037] S41. Obtain a training data set.

[0038] In this embodiment, each training sample in the training dataset includes a 3D CT slice sample image collected from a rock sample and a sample label for the 3D CT slice sample image. The sample label includes a 3D position label region of the sample pores and a shape factor label of the sample pores. The 3D position label region indicates the 3D position of the pores in the slice sample image. The shape factor label indicates the parameter value of the pore geometry.

[0039] In this embodiment, for each training sample, the porosity, maximum pore radius, minimum pore radius, and shape factor of the training sample can be calculated in the region where the three-dimensional position label of the sample pore is known. The shape factor ω = 4πA / L 2 , A is the pore surface area of ​​the sample, and L is the pore length.

[0040] S42. Construct an initial rock pore identification model.

[0041] In this embodiment, the initial rock pore identification model includes an initial multi-scale pore feature extraction network, an initial pore identification network, and an initial pore shape factor calculation network.

[0042] S43. Obtain training samples from the training data set as input training samples, and iteratively train the initial rock pore identification model based on the input training samples until a training termination condition is met, thereby obtaining a pre-trained rock pore identification model.

[0043] Optionally, during each iterative training process, S43 specifically includes the following steps:

[0044] Extracting pore sample features at multiple scales from the input training samples using a multi-scale pore feature extraction network in a current iteration;

[0045] Using the pore sample features at each scale as input to the pore recognition network in the current iteration, respectively, to identify the sample pores and the three-dimensional position area data of the sample pores in the input training samples, and calculating the loss value between the three-dimensional position area data of the sample pores and the three-dimensional position label area of ​​the sample pores based on a first loss function to obtain a first loss value;

[0046] Using the three-dimensional position area data of the sample pore as input to the pore shape factor calculation network in the current iteration, calculating the shape factor corresponding to the sample pore, and calculating the loss value between the shape factor corresponding to the sample pore and the shape factor label of the sample pore based on a second loss function to obtain a second loss value;

[0047] Obtaining a total loss value in a current iteration based on the first loss value and the second loss value;

[0048] When the total loss value in the current iteration does not meet the training termination condition, continue to obtain training samples from the training data set as input training samples, continue iterative training until the training termination condition is met, and use the rock pore recognition model after meeting the training termination condition as the pre-trained rock pore recognition model.

[0049] In this embodiment, the network structure of the rock pore identification model under training is identical to that of the trained rock pore identification model, except that the model parameters in the rock pore identification model under training have not yet been solidified, requiring iterative training to find the optimal model parameters. The first loss function and the second loss function can be the same or different, for example, they can be a mean square error loss function, a cross entropy loss function, etc. When there are multiple sample pores in an input training sample, for each sample pore, the loss value between the 3D position region data corresponding to each sample pore and the 3D position label region is calculated. The loss values ​​between the 3D position region data corresponding to each sample pore and the 3D position label region are accumulated to obtain a first loss value. Similarly, for each sample pore, the loss value between the shape factor corresponding to each sample pore and the shape factor label is calculated. The loss values ​​between the shape factor corresponding to each sample pore and the shape factor label are accumulated to obtain a second loss value. The first and second loss values ​​can be weighted to obtain a total loss value.

[0050] For example, Figure 5This is a network diagram illustrating the training of a rock pore recognition model in an embodiment. The input training samples form the input to a multi-scale pore feature extraction network. The multi-scale pore feature extraction network outputs first-scale pore sample features, second-scale pore sample features, and third-scale pore sample features, which serve as inputs to the pore recognition network. These results are then integrated to obtain the sample pores and their 3D location area data for the input training samples. A first loss value and a second loss value are then calculated based on the 3D location area data of the sample pores obtained in this iteration, ultimately yielding a total loss value. When the total loss value indicates the need for further iterative training, further training samples are obtained for training. During the training process, the rock pore recognition model continuously learns the location and shape features of sample pores at different scales from the input training samples. The location and boundary feature data of the sample pores are trained using the 3D location label areas of the sample pores as training targets, and the shape features of the sample pores are trained using the shape factor labels of the sample pores as training targets.

[0051] In the above embodiment, during the training process, the combination of two different losses can enable the rock pore identification model to learn more features and accelerate the convergence speed of the rock pore identification model, thereby improving the pore identification accuracy of the rock pore identification model.

[0052] In some embodiments, as Figure 6 As shown, S43 may further include the following steps:

[0053] S431, obtaining the three-dimensional position area data of the current sample pore in each input training sample output by the rock pore recognition model that meets the iterative training conditions.

[0054] In this embodiment, the iterative training condition includes the number of iterations reaching one round of iterations, for example, setting 1000 times as the number of iterations. After one round of iterations is completed, the rock pore identification model after the iteration can be obtained, and part or all of the training samples can be obtained from the training data set as input training samples. The three-dimensional position area data of the current sample pores in each input training sample is output through the rock pore identification model after the iteration.

[0055] S432: Based on the three-dimensional position area data of the current sample pore in each input training sample, a target input training sample that meets the screening condition is obtained.

[0056] Optionally, the step of obtaining target input training samples meeting the screening conditions based on the three-dimensional position area data of the current sample pore in each input training sample includes:

[0057] Calculate the porosity corresponding to each input training sample based on the three-dimensional position area data of the current sample pore in each input training sample;

[0058] The input training samples with porosity less than the preset porosity are selected as target input training samples.

[0059] In this embodiment, porosity represents the ratio of pore volume to total volume. For any input training sample, after obtaining the three-dimensional location data of the pores in the current sample within the input training sample, the volume of all pores can be calculated, thereby calculating the porosity of each input training sample.

[0060] Due to the pore volume effect, some pores are difficult to effectively identify due to noise interference. These pores are characterized by having a lower attenuation coefficient than the surrounding image neighborhood (i.e., they generally appear as darker areas in the pore image). Furthermore, pore fractal features are scale-invariant, so the porosity obtained based on the 3D location data of the current sample pores is generally lower than the true porosity of the input training sample. Therefore, pore reconstruction is necessary. This involves adding pore areas to the target area that meets the requirements to enhance the characteristics of this area. This means enhancing the dataset for this area so that the rock pore recognition model under training can learn more features from these darker areas. This results in a porosity closer to the true porosity obtained based on the 3D location data of the identified sample pores.

[0061] S433 , based on the three-dimensional position area data of the current sample pore in the target input training sample, generate a polyhedron corresponding to each current sample pore in the target input training sample, obtain multiple polyhedrons, and calculate the average voxel value of each polyhedron.

[0062] In this embodiment, when there are multiple current sample pores for the target input training sample, a polyhedron is generated corresponding to each current sample pore. The process of generating the polyhedron can be based on the three-dimensional boundary data of the current sample pore.

[0063] S434: Filter target polyhedrons that meet the requirements based on the average voxel value of each polyhedron.

[0064] In this embodiment, since the pore dataset in the darker area needs to be enhanced, a polyhedron having an average voxel value lower than a preset voxel value needs to be selected as a target polyhedron.

[0065] S435, in the target polyhedron, generating target sample pores that meet the pore conditions and generating sample labels corresponding to the target sample pores, updating the target input training samples based on the target sample pores and the generated sample labels corresponding to the target sample pores, and updating the updated target input training samples to the training data set to continue training the rock pore recognition model.

[0066] Optionally, S435 may further include:

[0067] Obtaining a first parameter and a second parameter of the target input training sample, wherein the first parameter represents a ratio of characteristic sizes of a fractal structure at two adjacent scales, and the second parameter represents a ratio of the number of primitives covering the fractal structure at adjacent scales;

[0068] Obtaining the maximum pore radius in the target input training sample based on the three-dimensional position area data of the current sample pore in the target input training sample;

[0069] updating the maximum pore radius according to the first parameter to obtain an updated maximum pore radius;

[0070] Using the updated maximum pore radius as the radius of the target sample pore and determining the number of the target sample pores according to the second parameter;

[0071] The target sample pores are generated in the target polyhedron according to the number of the target sample pores and the radius of the target sample pores.

[0072] In this embodiment, for the target input training sample, the first and second parameters can be obtained based on the three-dimensional position label area of ​​the sample pores in the target input training sample and using multi-scale coverage and fractal scaling law analysis. The calculation of this part is existing technology and will not be repeated here. The maximum pore radius is the maximum distance from the center to the boundary of the pore. Therefore, the number of target sample pores and the radius of the target sample pores can be used to randomly generate pore regions that meet the conditions in the target polyhedron. The geometric distribution of the pores conforms to scale invariance, which is specifically reflected in the first and second parameters. The first parameter is the scaling ratio of the fractal pore size, which represents the ratio of each pore size to the previous pore size in the fractal iteration. In the fractal iteration, each iteration will produce smaller pores. Updating the maximum pore radius based on the first parameter can ensure the self-similarity of the pore structure at different iteration levels. At the same time, by dynamically adjusting the maximum pore radius through the first parameter, the model can better adapt to different types of coal matrices and different geological conditions.

[0073] The second parameter is a scaling override for the number of fractal pores. It represents the ratio of the total number of pores in each iteration to the total number of pores in the previous iteration. As the fractal iterations progress, more pores are generated with each iteration. The scaling factor determines the number of new pores after each iteration to ensure consistent self-similarity in the pore distribution.

[0074] By updating the maximum pore radius and determining the number of pores through the first and second parameters in fractal topography theory, the pore structure of rocks can be characterized more accurately and the self-similarity and universality of the model can be improved.

[0075] Optionally, the updated maximum pore radius is equal to the ratio of the maximum pore radius to the first parameter.

[0076] Optionally, the number of target sample pores is equal to the product of the second parameter and the number of current sample pores in the target input training sample.

[0077] In the above embodiment, a target input training sample having a porosity generally lower than the true porosity of the input training sample is selected, and then a polyhedron corresponding to each current sample pore is generated based on the three-dimensional position area data of the current sample pore of the target input training sample, and a target polyhedron of a darker area is selected, and a data set of the pore area is enhanced in the target polyhedron. The updated target input training sample is obtained and updated to the training data set to continue training the rock pore recognition model, so that the rock pore recognition model under training can learn more features of such darker areas. In this way, the porosity obtained based on the three-dimensional position area data of the identified sample pores will be closer to the true porosity, so that the rock pore recognition model after training can more accurately identify the pores in such darker areas, thereby improving the accuracy of pore recognition.

[0078] On the other hand, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the deep learning-based pore identification method in rock images described in any embodiment of the present application.

[0079] Among them, in the computer program product, an optional implementation form of the program module architecture of the computer program that implements each step of the target identification method can be a pore identification device in rock images based on deep learning.

[0080] See also Figure 7 An embodiment of the present application provides a pore identification device in rock images based on deep learning, including: an acquisition module 71, used to obtain three-dimensional CT slice images collected from rocks; an identification module 72, used to form input image data of a pre-trained rock pore identification model based on the three-dimensional CT slice images, wherein the rock pore identification model includes a multi-scale pore feature extraction network and a pore identification network; the identification module 72 is also used to use the multi-scale pore feature extraction network to extract pore feature data of multiple scales from the input image data; the identification module 72 is also used to use the pore feature data of each scale as input to the pore identification network, and identify each pore in the three-dimensional CT slice image and the three-dimensional position area data of each pore.

[0081] Optionally, the identification module 72 is further configured to:

[0082] The three-dimensional position area data of each pore is used as the input of the pore shape factor calculation network to calculate the shape factor of each pore.

[0083] Optionally, the deep learning-based pore recognition device in rock images further includes a training module 73 for:

[0084] training the rock pore identification model;

[0085] The training of the rock pore identification model comprises:

[0086] Acquire a training data set, wherein each training sample in the training data set includes: a three-dimensional CT slice sample image collected from a rock sample and a sample label of the three-dimensional CT slice sample image, wherein the sample label includes a three-dimensional position label area of ​​the sample pore and a shape factor label of the sample pore;

[0087] Construct an initial rock pore identification model;

[0088] A training sample is obtained from the training data set as an input training sample, and an initial rock pore identification model is iteratively trained based on the input training sample until a training termination condition is met, thereby obtaining a pre-trained rock pore identification model.

[0089] Optionally, the training module 73 is further configured to:

[0090] Extracting pore sample features at multiple scales from the input training samples using a multi-scale pore feature extraction network in a current iteration;

[0091] Using the pore sample features at each scale as input to the pore recognition network in the current iteration, respectively, to identify the sample pores and the three-dimensional position area data of the sample pores in the input training samples, and calculating the loss value between the three-dimensional position area data of the sample pores and the three-dimensional position label area of ​​the sample pores based on a first loss function to obtain a first loss value;

[0092] Using the three-dimensional position area data of the sample pore as input to the pore shape factor calculation network in the current iteration, calculating the shape factor corresponding to the sample pore, and calculating the loss value between the shape factor corresponding to the sample pore and the shape factor label of the sample pore based on a second loss function to obtain a second loss value;

[0093] Obtaining a total loss value in a current iteration based on the first loss value and the second loss value;

[0094] When the total loss value in the current iteration does not meet the training termination condition, continue to obtain training samples from the training data set as input training samples, continue iterative training until the training termination condition is met, and use the rock pore recognition model after meeting the training termination condition as the pre-trained rock pore recognition model.

[0095] Optionally, the training module 73 is further configured to:

[0096] Obtaining the three-dimensional position area data of the current sample pores in each input training sample output by the rock pore recognition model that meets the iterative training conditions;

[0097] Based on the three-dimensional position area data of the current sample pore in each input training sample, a target input training sample that meets the screening conditions is obtained;

[0098] Based on the three-dimensional position area data of the current sample pore in the target input training sample, a polyhedron corresponding to each current sample pore in the target input training sample is generated to obtain a plurality of polyhedrons, and an average voxel value of each polyhedron is calculated;

[0099] According to the average voxel value of each polyhedron, the target polyhedron that meets the conditions is selected;

[0100] In the target polyhedron, target sample pores that meet the pore conditions and sample labels corresponding to the target sample pores are generated. Based on the target sample pores and the sample labels corresponding to the target sample pores, the target input training samples are updated. The updated target input training samples are updated to the training data set to continue training the rock pore recognition model.

[0101] Optionally, the training module 73 is further configured to:

[0102] Calculate the porosity corresponding to each input training sample based on the three-dimensional position area data of the current sample pore in each input training sample;

[0103] The input training samples with porosity less than the preset porosity are selected as target input training samples.

[0104] Optionally, the training module 73 is further configured to:

[0105] Obtaining a first parameter and a second parameter of the target input training sample, wherein the first parameter represents a ratio of characteristic sizes of a fractal structure at two adjacent scales, and the second parameter represents a ratio of the number of primitives covering the fractal structure at adjacent scales;

[0106] Obtaining the maximum pore radius in the target input training sample based on the three-dimensional position area data of the current sample pore in the target input training sample;

[0107] updating the maximum pore radius according to the first parameter to obtain an updated maximum pore radius;

[0108] Using the updated maximum pore radius as the radius of the target sample pore and determining the number of the target sample pores according to the second parameter;

[0109] The target sample pores are generated in the target polyhedron according to the number of the target sample pores and the radius of the target sample pores.

[0110] Optionally, the updated maximum pore radius is equal to the ratio of the maximum pore radius to the first parameter.

[0111] Optionally, the number of target sample pores is equal to the product of the second parameter and the number of current sample pores in the target input training sample.

[0112] See also Figure 8 In another aspect of an embodiment of the present application, a computing device 10 is provided, comprising a memory 3011 and a processor 3012. The memory 3011 stores a computer program. When the computer program is executed by the processor, the processor 3012 performs the steps of the deep learning-based pore identification method in rock images provided in any of the above embodiments of the present application. The computing device 10 is, for example, a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc., a mobile phone (e.g., a smartphone, a wireless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smart watch), or a similar device.

[0113] The processor 3012 is the control center, connecting the various components of the entire computer device using various interfaces and lines. It executes the various functions of the computer device and processes data by running or executing software programs and / or modules stored in the memory 3011 and accessing data stored in the memory 3011. Optionally, the processor 3012 may include one or more processing cores. Preferably, the processor 3012 may integrate an application processor and a modem processor, wherein the application processor primarily processes the operating system, user interfaces, and application programs, while the modem processor primarily handles wireless communications. It is understood that the modem processor may not be integrated into the processor 3012.

[0114] The memory 3011 can be used to store software programs and modules. The processor 3012 executes various functional applications and data processing by running the software programs and modules stored in the memory 3011. The memory 3011 may mainly include a program storage area and a data storage area. The program storage area may store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created based on the use of the computer device, etc. In addition, the memory 3011 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 3011 may also include a memory controller to provide the processor 3012 with access to the memory 3011.

[0115] On the other hand, an embodiment of the present application further provides a storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the deep learning-based pore identification method in rock images provided in any of the above embodiments of the present application.

[0116] Those skilled in the art will appreciate that all or part of the processes in the methods provided in the above embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. The scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for identifying pores in rock images based on deep learning, characterized in that: include: Acquire three-dimensional CT slice images of rock samples; Based on the three-dimensional CT slice image, forming input image data of a pre-trained rock pore identification model, wherein the rock pore identification model includes a multi-scale pore feature extraction network and a pore identification network; Extracting pore feature data of multiple scales from the input image data using the multi-scale pore feature extraction network; The pore characteristic data of each scale is respectively used as the input of the pore recognition network to identify each pore in the three-dimensional CT slice image and obtain the three-dimensional position area data of each pore.

2. The method for identifying pores in rock images based on deep learning according to claim 1, wherein: The rock pore identification model includes a pore shape factor calculation network, and the method further includes: The three-dimensional position area data of each pore is used as the input of the pore shape factor calculation network to calculate the shape factor of each pore.

3. The method for identifying pores in rock images based on deep learning according to claim 1, wherein: The method further comprises: training the rock pore identification model; The training of the rock pore identification model comprises: Acquire a training data set, wherein each training sample in the training data set includes: a three-dimensional CT slice sample image collected from a rock sample and a sample label of the three-dimensional CT slice sample image, wherein the sample label includes a three-dimensional position label area of ​​the sample pore and a shape factor label of the sample pore; Construct an initial rock pore identification model; A training sample is obtained from the training data set as an input training sample, and an initial rock pore identification model is iteratively trained based on the input training sample until a training termination condition is met, thereby obtaining a pre-trained rock pore identification model.

4. The method for identifying pores in rock images based on deep learning according to claim 3, wherein: The method of obtaining a training sample from the training data set as an input training sample, iteratively training an initial rock pore identification model based on the input training sample until a training termination condition is met, and obtaining a pre-trained rock pore identification model, includes: Extracting pore sample features at multiple scales from the input training samples using a multi-scale pore feature extraction network in a current iteration; Using the pore sample features at each scale as input to the pore recognition network in the current iteration, respectively, to identify the sample pores and the three-dimensional position area data of the sample pores in the input training samples, and calculating the loss value between the three-dimensional position area data of the sample pores and the three-dimensional position label area of ​​the sample pores based on a first loss function to obtain a first loss value; Using the three-dimensional position area data of the sample pore as input to the pore shape factor calculation network in the current iteration, calculating the shape factor corresponding to the sample pore, and calculating the loss value between the shape factor corresponding to the sample pore and the shape factor label of the sample pore based on a second loss function to obtain a second loss value; Obtaining a total loss value in a current iteration based on the first loss value and the second loss value; When the total loss value in the current iteration does not meet the training termination condition, continue to obtain training samples from the training data set as input training samples, continue iterative training until the training termination condition is met, and use the rock pore recognition model after meeting the training termination condition as the pre-trained rock pore recognition model.

5. The method for identifying pores in rock images based on deep learning according to claim 3, wherein: The method of obtaining a training sample from the training data set as an input training sample, iteratively training an initial rock pore identification model based on the input training sample until a training termination condition is met, and obtaining a pre-trained rock pore identification model, includes: Obtaining the three-dimensional position area data of the current sample pores in each input training sample output by the rock pore recognition model that meets the iterative training conditions; Based on the three-dimensional position area data of the current sample pore in each input training sample, a target input training sample that meets the screening conditions is obtained; Based on the three-dimensional position area data of the current sample pore in the target input training sample, a polyhedron corresponding to each current sample pore in the target input training sample is generated to obtain a plurality of polyhedrons, and an average voxel value of each polyhedron is calculated; According to the average voxel value of each polyhedron, the target polyhedron that meets the conditions is selected; In the target polyhedron, target sample pores that meet the pore conditions and sample labels corresponding to the target sample pores are generated. Based on the target sample pores and the sample labels corresponding to the target sample pores, the target input training samples are updated. The updated target input training samples are updated to the training data set to continue training the rock pore recognition model.

6. The method for identifying pores in rock images based on deep learning according to claim 5, wherein: The step of obtaining target input training samples that meet the screening conditions based on the three-dimensional position area data of the current sample pores in each input training sample includes: Calculate the porosity corresponding to each input training sample based on the three-dimensional position area data of the current sample pore in each input training sample; The input training samples with porosity less than the preset porosity are selected as target input training samples.

7. The method for identifying pores in rock images based on deep learning according to claim 5, wherein: The pore condition includes the number of target sample pores and the radius of the target sample pores, and is characterized in that generating target sample pores that meet the pore condition in the target polyhedron includes: Obtaining a first parameter and a second parameter of the target input training sample, wherein the first parameter represents a ratio of characteristic sizes of a fractal structure at two adjacent scales, and the second parameter represents a ratio of the number of primitives covering the fractal structure at adjacent scales; Obtaining the maximum pore radius in the target input training sample based on the three-dimensional position area data of the current sample pore in the target input training sample; updating the maximum pore radius according to the first parameter to obtain an updated maximum pore radius; Using the updated maximum pore radius as the radius of the target sample pore and determining the number of the target sample pores according to the second parameter; The target sample pores are generated in the target polyhedron according to the number of the target sample pores and the radius of the target sample pores.

8. The method for identifying pores in rock images based on deep learning according to claim 7, wherein: The updated maximum pore radius is equal to the ratio of the maximum pore radius to the first parameter.

9. The method for identifying pores in rock images based on deep learning according to claim 7, wherein: The number of target sample pores is equal to the product of the second parameter and the number of current sample pores in the target input training sample.

10. A computing device, characterized in that The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the pore identification method in rock images based on deep learning as described in any one of claims 1 to 9.

Images