Rock core CT image three-dimensional super-division reconstruction method and device based on convolutional neural network

Abstract

The invention discloses a core CT image three-dimensional super-division reconstruction method and device based on a convolutional neural network. The method comprises the following steps: inputting a pre-obtained preprocessed core CT sequence slice into a trained particle segmentation reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-segmentation image; the particle segmentation three-dimensional reconstruction model comprises a Transform and a 3D-CNN (Convolutional Neural Network); inputting the preprocessed core CT sequence slices into the trained pore segmentation reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-segmentation image; the pore segmentation three-dimensional reconstruction model comprises a Transform and a 3D-CNN (Convolutional Neural Network); and fusing the particle segmentation longitudinal super-division image and the pore segmentation longitudinal super-division image to obtain a rock core CT three-dimensional super-division reconstructed image. According to the method, the clearer core CT three-dimensional high-resolution particle image providing more detail information can be obtained, and the volume thickness is increased, so that the basic detail information of the sample is fully displayed, effective data guarantee is provided for scientific research and technical personnel, and the working efficiency is greatly improved.

Description

Technical Field

[0001] The present application relates to a method and device for three-dimensional super-resolution reconstruction of core CT images based on convolutional neural networks. Background Art

[0002] Core analysis is the gold standard for oil and gas exploration and development. Image analysis is the primary tool for core analysis. Core images include meter-scale core images, centimeter-scale thin-section images, and micro- and nanoscale CT and electron microscopy images. CT technology not only observes structural changes within the core but also simulates core imaging and fluid movement, making it indispensable in oil exploration and development.

[0003] Currently, CT technology is widely used in oil exploration and development. This requires meticulous geological surveys and core analysis, and CT technology can accurately reconstruct three-dimensional models of core samples. Since oil is typically stored in the pores and fractures of rock, understanding the structural composition of the rock and the structure of its pores is crucial for assessing oil reserves and developing appropriate extraction plans. CT technology can also be used for quantitative analysis of rock samples, such as determining parameters like density and porosity, and providing additional information. This helps oil explorers and developers better understand the underground geological structure and reservoir characteristics.

[0004] CT technology in oil and gas exploration and development often relies on mature software, such as Avizo. This software offers excellent application and visualization capabilities for particle segmentation and pore extraction. However, it can be sensitive to noise and artifacts in images, which can easily interfere with the results and lead to inaccurate segmentation results. Furthermore, threshold-based segmentation fails to incorporate semantic and geological information between images, resulting in low accuracy in particle classification. Furthermore, interpolation is often used for 3D reconstruction of serial CT core slices.

[0005] Interpolation is an important means of improving the quality of three-dimensional reconstructed images. Currently, commonly used interpolation methods are divided into grayscale-based interpolation, shape-based interpolation, and wavelet-based interpolation. Grayscale-based interpolation includes linear interpolation and cubic spline interpolation. These methods are computationally inefficient and easy to implement, but they can cause severe boundary blurring. To address the shortcomings of the above methods, some researchers have proposed an interpolation method based on corresponding point matching. This method finds the best matching point pair in the upper and lower tomographic images based on certain criteria and uses the grayscale values ​​of the best point pair to interpolate a new image. Although these studies have solved the problem of insufficient accuracy to a certain extent, further improvement is still needed. Summary of the Invention

[0006] In order to better achieve the segmentation and reconstruction of core CT images, the embodiment of the present application provides a three-dimensional super-resolution reconstruction method and device for core CT images based on convolutional neural networks.

[0007] In a first aspect, an embodiment of the present application provides a method for three-dimensional super-resolution reconstruction of core CT images based on a convolutional neural network, the method comprising:

[0008] Inputting the pre-processed core CT serial slices into the trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; the particle segmentation 3D reconstruction model includes Transformer and 3D-CNN;

[0009] Inputting the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes a Transformer and a 3D-CNN;

[0010] The particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image are fused to obtain a core CT three-dimensional super-resolution reconstructed image.

[0011] In an optional implementation of the embodiment of the present application, the trained particle segmentation and reconstruction model is obtained by the following method:

[0012] Dividing the pre-constructed first core CT serial slice sample set into a first training set, a first validation set, and a first test set according to a preset ratio;

[0013] Inputting the first training set into a pre-built particle segmentation 3D reconstruction model for training, outputting a multi-slice-based particle segmentation longitudinal super-resolution image, and obtaining a trained particle segmentation 3D reconstruction model;

[0014] Inputting the first verification set into the trained particle segmentation 3D reconstruction model to obtain a verification result, and comparing it with the corresponding real labeled image to update the trained particle segmentation 3D reconstruction model;

[0015] Iterative training is performed until the difference between the verification result and the corresponding true label image meets a preset requirement, and the first test set is used for testing to obtain a trained particle segmentation 3D reconstruction model.

[0016] In an optional implementation of the embodiment of the present application, the first training set is input into a pre-built particle segmentation 3D reconstruction model for training, and a multi-slice-based particle segmentation longitudinal super-resolution image is output to obtain a trained particle segmentation 3D reconstruction model, including:

[0017] Based on the Transformer, perform in-image feature extraction on each core CT serial slice in the first training set, and output a single-slice particle segmentation transverse super-resolution image;

[0018] Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based particle segmentation transverse super-resolution image, and a multi-slice-based particle segmentation longitudinal super-resolution image is output to obtain a trained particle segmentation three-dimensional reconstruction model.

[0019] In an optional implementation of the embodiment of the present application, the first core CT serial slice sample set is constructed in the following manner:

[0020] The pre-processed multiple sets of historical core CT image serial slices are annotated with particle contours and categories to obtain a core CT serial slice label set;

[0021] A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set;

[0022] The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the first core CT sequence slice sample set.

[0023] In an optional implementation of the embodiment of the present application, a trained pore segmentation 3D reconstruction model is obtained by the following method:

[0024] Dividing the pre-constructed second core CT sequence slice sample set into a second training set, a second validation set, and a second test set according to a preset ratio;

[0025] Inputting the second training set into a pre-built pore segmentation 3D reconstruction model for training, outputting a multi-slice-based pore segmentation longitudinal super-resolution image, and obtaining a trained pore segmentation 3D reconstruction model;

[0026] Inputting the second verification set into the trained pore segmentation 3D reconstruction model to obtain a verification result, and comparing it with the corresponding real label image to update the trained pore segmentation 3D reconstruction model;

[0027] Iterative training is performed until the difference between the verification result and the corresponding true label image meets the preset requirements, and the second test set is used for testing to obtain a trained pore segmentation 3D reconstruction model.

[0028] In an optional implementation of the embodiment of the present application, the second training set is input into a pre-built pore segmentation 3D reconstruction model for training, and a multi-slice-based pore segmentation longitudinal super-resolution image is output to obtain a trained pore segmentation 3D reconstruction model, including:

[0029] Based on the Transformer, extract in-image features from each core CT serial slice in the second training set, and output a transverse super-resolution image of pore segmentation based on a single slice;

[0030] Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based pore segmentation transverse super-resolution image, and a multi-slice-based pore segmentation longitudinal super-resolution image is output to obtain a trained pore segmentation three-dimensional reconstruction model.

[0031] In an optional implementation of the embodiment of the present application, the second core CT serial slice sample set is constructed in the following manner:

[0032] The pore contours of multiple sets of pre-processed historical core CT image sequence slices are annotated to obtain a core CT sequence slice label set;

[0033] A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set;

[0034] The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the second core CT sequence slice sample set.

[0035] In a second aspect, an embodiment of the present application provides a device for three-dimensional super-resolution reconstruction of core CT images based on a convolutional neural network, the device comprising:

[0036] A particle segmentation module is used to input the pre-processed core CT serial slices obtained in advance into a trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; the particle segmentation 3D reconstruction model includes a Transformer and a 3D-CNN;

[0037] A pore segmentation module is used to input the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes a Transformer and a 3D-CNN;

[0038] A fusion module is used to fuse the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image to obtain a core CT three-dimensional super-resolution reconstructed image.

[0039] In a third aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned convolutional neural network-based three-dimensional super-resolution reconstruction method for core CT images.

[0040] In a fourth aspect, an embodiment of the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the three-dimensional super-resolution reconstruction method of core CT images based on a convolutional neural network as described above is implemented.

[0041] In a fifth aspect, an embodiment of the present application provides a computer program product comprising instructions. When the computer program product is run on a computer device, the computer device executes the above-mentioned convolutional neural network-based three-dimensional super-resolution reconstruction method for core CT images.

[0042] In a sixth aspect, an embodiment of the present application provides a chip, which includes a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run a computer program or instruction to implement the above-mentioned convolutional neural network-based three-dimensional super-resolution reconstruction method of core CT images.

[0043] The beneficial effects of the above technical solutions provided by the embodiments of the present application include at least:

[0044] The embodiment of the present application provides a three-dimensional super-resolution reconstruction method for core CT images based on convolutional neural networks. By training a particle segmentation three-dimensional reconstruction model and a pore segmentation three-dimensional reconstruction model based on Transformer and 3D-CNN, the core CT images are segmented and reconstructed for particles and pores respectively. Finally, the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image output by the fusion model are combined to obtain a core CT three-dimensional super-resolution reconstruction image. Compared with the high-precision hardware performance of CT image reconstruction in the prior art, this method is low-cost and easy to promote. It can be applied to sample images collected by most CT and direct media interface (DMI) devices. At the same time, the method can exclude the influence of the environment, human factors and sample image storage process of the CT device, obtain a clearer and more detailed core CT three-dimensional high-resolution particle image, and increase the volume thickness to avoid the error caused by reconstruction through interlayer interpolation, thereby fully displaying the basic detail information of the sample, providing effective data protection for scientific research and technical personnel, and greatly improving work efficiency.

[0045] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0046] The technical solution of the present application is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are used to provide a further understanding of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings:

[0048] Figure 1 A schematic diagram of the steps of a convolutional neural network-based three-dimensional super-resolution reconstruction method for core CT images provided in an embodiment of the present application;

[0049] Figure 2 A schematic diagram of the structure of the pore segmentation three-dimensional reconstruction model provided in an embodiment of the present application;

[0050] Figure 3 A schematic diagram of the 3D U-net network architecture provided in an embodiment of the present application;

[0051] Figure 4 A schematic diagram of an example of sub-pixel convolution layer processing provided in an embodiment of the present application;

[0052] Figure 5 A schematic diagram of the results of an experiment using the method of the present invention provided in an embodiment of the present application;

[0053] Figure 6 The three-dimensional particle size distribution diagram of the core CT provided in the embodiment of this application;

[0054] Figure 7 A schematic diagram of the structure of a three-dimensional super-resolution reconstruction device for core CT images based on a convolutional neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0056] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0057] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0058] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0059] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0060] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0061] It should be understood that the size of the serial numbers of the steps in the following embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0062] In order to illustrate the technical solution of the present application, specific embodiments are provided below.

[0063] The inventors discovered that existing CT technology, due to limitations such as the inability to address low-dose CT scanning, leads to increased quantum noise and decreased overall resolution, which impacts the accuracy and reliability of core analysis. Therefore, a new CT technology is needed to address these issues.

[0064] At present, convolutional neural networks (CNN) are the most widely used in high-resolution reconstruction deep learning tasks. Common models include RCAN, HAN, IGNN and NLSA. Most of them focus on sophisticated architectural design, such as residual learning and dense connections. Although the performance is much better than that of methods based on traditional models, there are still significant limitations in CT three-dimensional reconstruction. CT three-dimensional reconstruction requires not only high resolution in two-dimensional image imaging, but also high-precision reconstruction of interlayer images in the spatial direction. Based on this, the inventors have made this application after further research and development, providing a method and device for three-dimensional super-resolution reconstruction of core CT images based on convolutional neural networks.

[0065] Example 1

[0066] The present invention provides a method for 3D super-resolution reconstruction of core CT images based on convolutional neural network. Figure 1 As shown, the method includes the following steps S101 to S103:

[0067] S101: Inputting the pre-processed core CT serial slices obtained in advance into a trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; the particle segmentation 3D reconstruction model includes a Transformer and a 3D-CNN.

[0068] S102: Inputting the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes Transformer and 3D-CNN.

[0069] S103: Fusing the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image to obtain a core CT three-dimensional super-resolution reconstructed image.

[0070] In the above step S101, after obtaining the low-dose and large-interval core CT serial slices that need to be processed for 3D super-resolution reconstruction, the core CT serial slices can be preprocessed, such as resizing, and then the preprocessed core CT serial slices are input into the trained particle segmentation and reconstruction model.

[0071] In the embodiment of the present application, a trained particle segmentation and reconstruction model can be obtained through the following steps S1011 to S1014:

[0072] S1011: Divide the pre-constructed first core CT serial slice sample set into a first training set, a first validation set, and a first test set according to a preset ratio.

[0073] In the above step S1011, the first core CT serial slice sample set may be constructed in the following manner:

[0074] The pre-processed multiple sets of historical core CT image serial slices are annotated with particle contours and categories to obtain a core CT serial slice label set;

[0075] A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set;

[0076] The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the first core CT sequence slice sample set.

[0077] In an embodiment of the present application, during the training of the particle segmentation and reconstruction model, multiple sets of historical core CT image sequence slices are collected and preprocessed, such as resizing. Due to the limitation of server computing power, low-quality core CT images with excessive resolution can be cropped according to the original aspect ratio; core CT images can also be processed by data enhancement to achieve the purpose of increasing the data set.

[0078] The preprocessed sets of historical core CT image sequence slices were annotated with particle outlines and categories to obtain a core CT sequence slice label set. Furthermore, several images were randomly intermittently extracted from each set of historical core CT image sequence slices and noise was added to simulate low-dose CT scan images to form training images, thus obtaining the core CT sequence slice training set. The training images in the core CT sequence slice training set and the corresponding labeled images in the core CT sequence slice label set were combined in the [training image, labeled image] format to create the first core CT sequence slice sample set.

[0079] The first core CT serial slice sample set is divided into a first training set, a first validation set, and a first test set according to a preset ratio. For example, 80% of the training images and corresponding labeled images can be randomly selected to form the training set, 10% of the training images and corresponding labeled images can be randomly selected to form the validation set, and the rest can be the test set.

[0080] S1012: Input the first training set into a pre-built particle segmentation 3D reconstruction model for training, output a multi-slice-based particle segmentation longitudinal super-resolution image, and obtain a trained particle segmentation 3D reconstruction model.

[0081] In the above step S1012, the first training set is input into a pre-built particle segmentation 3D reconstruction model for training, and a multi-slice-based particle segmentation longitudinal super-resolution image is output to obtain a trained particle segmentation 3D reconstruction model, which specifically includes:

[0082] Based on the Transformer, perform in-image feature extraction on each core CT serial slice in the first training set, and output a single-slice particle segmentation transverse super-resolution image;

[0083] Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based particle segmentation transverse super-resolution image, and a multi-slice-based particle segmentation longitudinal super-resolution image is output to obtain a trained particle segmentation three-dimensional reconstruction model.

[0084] In the examples of this application, the concept of transfer learning is employed in the construction of a particle segmentation and 3D reconstruction model. This approach leverages the ability to identify similarities between existing and new knowledge, facilitating the learning of more general features. This allows for better model training results even on small datasets, while also improving the speed of network training and saving time. The inventors used a 3D-CNN model trained on a dataset of brain CT slice images from patients at Kaiyuan Hospital as a foundation and fine-tuned the model using transfer learning on a dataset of core CT scan images.

[0085] The inventors constructed a core CT scan image model (i.e., a particle segmentation 3D reconstruction model and a pore segmentation 3D reconstruction model) based on Transformer+3D-CNN. Figure 3 Figure 2 shows the schematic diagram of the pore segmentation and 3D reconstruction model. The model uses Transformer to extract intra-image features from CT sequence slices to obtain a transverse super-resolution image based on a single slice. 3D-CNN then performs intra-image feature extraction and inter-layer feature extraction to obtain a longitudinal super-resolution image based on multiple slices. The model architecture has a sub-convolutional module r = 2.

[0086] In the embodiment of this application, in the particle segmentation and 3D reconstruction model built based on Transformer + 3D-CNN, the Transformer uses a residual convolutional neural network to extract in-image features, obtaining shallow details such as edges, color, and brightness. The specific operation of in-image feature extraction is as follows:

[0087] Take the low-quality blurred core CT scan image (LQ) as input image (I LQ ) is the input of the particle segmentation 3D reconstruction model. During the training process, the first training set is used as the input of the particle segmentation 3D reconstruction model. LQ ∈R H×W×D Where H is the height of the core CT scan image, W is the width of the core CT scan image, D is the number of core CT scan image slices, and R is the data set.

[0088] Extract shallow features through convolution (F o ), F o∈R H×W×C , where H is the height of the core CT scan image, W is the width of the core CT scan image, C is the number of output feature channels, and R is the data set.

[0089] After the above operations, the Transformer extracts in-image features, obtains shallow features of the core CT scan image, and outputs a single-slice, particle-segmented, lateral super-resolution image. Applying the Transformer to early visual processing helps stabilize training and achieve better results.

[0090] In the embodiment of the present application, in the particle segmentation and 3D reconstruction model based on Transformer+3D-CNN, the 3D-CNN convolution module can use a 3D U-net network. The network structure of 3D U-net is very similar to that of the standard U-net, consisting of three parts: encoder downsampling, decoder upsampling and skip connection; the difference is that 3D U-net takes 3D volume as input and performs corresponding 3D operations, followed by 3D convolution, 3D dilated convolution layer, 3D maximum pooling and 3D upconvolution layer. The 3D U-net network architecture is as follows: Figure 4 As shown, its architecture is as follows:

[0091] 3D convolution layer: It is mainly composed of multiple feature planes. Each feature plane is extracted through a 3×3×3 convolution kernel. As the network depth increases, the extracted features change from simple to complex.

[0092] 3D dilated convolution layer: It consists of multiple feature planes, each of which is extracted using a 5×5×5 dilated convolution kernel. As the network depth increases, the extracted features also change from simple to complex.

[0093] Activation function: You can use the activation function Relu. For Relu, if the input x>0, the output is equal to the input, otherwise the output is 0. Relu can filter out extremely subtle features, which not only improves the generalization ability of the model, but also further overcomes the gradient vanishing problem and speeds up the convergence speed. The formula is as follows:

[0094] f(x)=max(0,x) Formula 1;

[0095] Where f(x) is the output and x is the input.

[0096] 3D pooling layer: Consists of multiple feature planes, each of which uniquely corresponds to a feature plane in the convolutional layer. Therefore, the number of feature planes in this layer corresponds to the number of convolutional layers and does not change. The pooling layer can be considered to have a secondary feature extraction function, obtaining spatially invariant features by reducing the resolution of the feature planes.

[0097] 3D convolution layer: gradually restores the details and size of the image through interpolation method, and further obtains the final prediction result through convolution processing. An example of sub-pixel convolution layer processing in 3D convolution layer is Figure 5 As shown in the figure, an example of a low-resolution input activation map and a corresponding high-resolution output activation map is given by a three-dimensional convolutional layer for magnification in one depth direction. For a scaling factor of r = 2 in one direction, the three-dimensional convolutional layer requires r = 2 activation maps as input.

[0098] Loss function: MAE is used as the loss function, as shown below:

[0099]

[0100] Where, loss is the mean absolute error; Y i is a real CT image; pCT i is the predicted image; N is the number of real CT images or predicted images, i = 1, 2, ..., N. The error obtained by the loss function is back-propagated using the Adam random optimization function to further optimize the parameters.

[0101] After the above operations, 3D-CNN performs intra-image feature extraction and inter-layer feature extraction, outputs a multi-slice-based particle segmentation longitudinal super-resolution image, and obtains a trained particle segmentation 3D reconstruction model.

[0102] S1013: Input the first verification set into the trained particle segmentation 3D reconstruction model to obtain a verification result, compare it with the corresponding real label image, and update the trained particle segmentation 3D reconstruction model.

[0103] S1014: Repeat the above steps S1012 to S1013, iteratively train until the difference between the verification result and the corresponding real label image meets the preset requirements, use the first test set to test, and obtain a trained particle segmentation 3D reconstruction model.

[0104] During the iterative training process, the loss function is set to minimize the difference between the label image and the output image of the particle segmentation 3D reconstruction model. The objective function is typically the mean absolute difference (MAD). Given a low-resolution input image I and the corresponding ground truth high-resolution image O, the MAD loss is formally defined as follows:

[0105]

[0106] Where L(θ,I,O) is the loss value based on the mean absolute value; θ is the parameter (weight) of 3D-CNN; I is the low-resolution input image; O is the real-value high-resolution image corresponding to I; f is the transformation function learned by 3D-CNN during training; D is the width of the real image; w is the height of the real image; h is the depth of the real image.

[0107] In addition, during iterative training, the Adam optimizer can be used, which converges faster than stochastic gradient descent.

[0108] In the embodiment of the present application, the pre-processed core CT serial slices are input into the particle segmentation and reconstruction model trained through the above process to obtain a multi-slice based particle segmentation longitudinal super-resolution image.

[0109] In the embodiment of the present application, the trained pore segmentation longitudinal super-resolution image can be obtained through the following steps S1021 to S1024:

[0110] S1021: Divide the pre-constructed second core CT sequence slice sample set into a second training set, a second validation set, and a second test set according to a preset ratio.

[0111] In the above step S1021, the second core CT serial slice sample set may be constructed in the following manner:

[0112] The pore contours of multiple sets of pre-processed historical core CT image sequence slices are annotated to obtain a core CT sequence slice label set;

[0113] A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set;

[0114] The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the second core CT sequence slice sample set.

[0115] S1022: Input the second training set into a pre-built pore segmentation 3D reconstruction model for training, output a multi-slice-based pore segmentation longitudinal super-resolution image, and obtain a trained pore segmentation 3D reconstruction model.

[0116] In the above step S1022, the second training set is input into the pre-built pore segmentation 3D reconstruction model for training, and a multi-slice-based pore segmentation longitudinal super-resolution image is output to obtain a trained pore segmentation 3D reconstruction model, including:

[0117] Based on the Transformer, extract in-image features from each core CT serial slice in the second training set, and output a transverse super-resolution image of pore segmentation based on a single slice;

[0118] Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based pore segmentation transverse super-resolution image, and a multi-slice-based pore segmentation longitudinal super-resolution image is output to obtain a trained pore segmentation three-dimensional reconstruction model.

[0119] S1023: Input the second verification set into the trained pore segmentation 3D reconstruction model to obtain a verification result, and compare it with the corresponding real label image to update the trained pore segmentation 3D reconstruction model.

[0120] S1024: Repeat the above steps S1022 to S1023, iteratively train until the difference between the verification result and the corresponding real label image meets the preset requirements, use the second test set for testing, and obtain a trained pore segmentation 3D reconstruction model.

[0121] In the embodiment of the present application, the specific implementation process of the above steps S1021 to S1024 can refer to the process of training the particle segmentation and reconstruction model in S1011 to S1014, and will not be repeated here.

[0122] In an embodiment of the present application, the preprocessed core CT serial slices are segmented and reconstructed into particles and pores respectively through a particle segmentation reconstruction model and a pore segmentation reconstruction model, and the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image output by the two models are fused to obtain a core CT three-dimensional super-resolution reconstructed image corresponding to the preprocessed core CT serial slices, wherein the fusion method can be a feature addition and merging method.

[0123] In the embodiments of the present application, based on the particle segmentation and reconstruction model and the pore segmentation and reconstruction model, advanced performance is achieved in the 3D serial slice image reconstruction task of core CT. The image quality evaluation criteria Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) can be used as evaluation indicators for comparative experiments.

[0124] PSNR is used to measure the degree of image distortion. The image to be evaluated is considered as a superposition of the original image and noise, and the ratio is calculated to achieve the specific value. The specific calculation formula of PSNR is as follows: Formula 4:

[0125]

[0126] Where PSNR is the peak signal-to-noise ratio; MAX is the largest pixel in the image; MSE is the mean square error of the image; I is a clean image of size m×n; and K is the noise.

[0127] SSIM evaluates the similarity between two images from three aspects: brightness, contrast, and structure. The specific calculation method is to calculate the local SSIM through sliding windows of different sizes and take the average as the global evaluation result. The brightness is estimated by the mean, the contrast is calculated by the standard deviation, and the structural similarity is measured by covariance. The specific calculation formula of SSIM is as follows: Formula 5:

[0128]

[0129] Where x and y are two samples; l is brightness; c is contrast; s is structure; α, β, and γ represent the proportion of different features in the SSIM measurement, usually α = β = γ = 1; μ is the mean; σ is the variance; c1 and c2 are equal to 0.01 and 0.03 multiplied by the range of pixel values, respectively.

[0130] In addition, for the three-dimensional reconstruction of rock particles and the three-dimensional reconstruction of pore structure, AP (average precision) is used as the main evaluation indicator of the model in the present invention. For a given input image, the model designed by the present invention outputs a set of prediction masks for each individual particle and pore, and compares each individual prediction with all real particle and pore instances. If the intersection over union (IOU) of the predicted particle segmentation mask and the actual mask is greater than a set threshold t, it is regarded as a true positive (TP); if it is less than t, it is regarded as a false positive (FP). Finally, a precision-recall curve is created by calculating the precision and recall rate of different thresholds t, and the area under the curve is used as the average precision (AP). Among them, the formula for the IOU intersection over union ratio is as follows: Formula 6:

[0131]

[0132] Where IOU(A,B) is the intersection-over-union ratio of the predicted particle segmentation mask and the actual mask; A is the predicted particle segmentation mask; B is the actual mask.

[0133] In a specific embodiment, a core CT scan image with an image resolution of 1956*2008 is used. According to the convolutional neural network-based 3D super-resolution reconstruction method for core CT images provided by the present invention, the image is first unified to a size of 1280*960. Then, relevant experts from the research institute perform data annotation. The core CT 3D super-resolution reconstruction image is obtained through the particle segmentation reconstruction model and the pore segmentation reconstruction model. The experimental results are shown in Figure 2. Figure 6 As shown, the top row of images are core CT scan images, the second row of images are pore segmentation longitudinal super-resolution images, and the third row are particle segmentation longitudinal super-resolution images.

[0134] The experimental results are shown in Tables 1 and 2. Table 1 compares the 3D reconstruction results of core CT image sequences, including the PSNR and SSIM calculation results of the 3D reconstruction results of core CT image sequences obtained by linear interpolation, wavelet interpolation, and the method of the present invention. As can be seen from Table 1, the 3D reconstruction results of the method of the present invention are superior to those of currently used traditional 3D reconstruction schemes such as linear interpolation and wavelet interpolation.

[0135] Table 1 Three-dimensional reconstruction results of core CT sequence images

[0136] algorithm PSNR SSIM Linear interpolation 32.64 0.9008 Wavelet interpolation 32.68 0.9006 Method of the present invention 33.51 0.9240

[0137] Table 2 Pore and particle segmentation results of core CT sequence images

[0138] algorithm AP AP50 AP75 Threshold segmentation 0.35 0.53 0.32 Pore ​​segmentation 0.4 0.58 0.39 Particle segmentation 0.44 0.61 0.42

[0139] Table 2 also shows the pore and particle segmentation results for core CT image sequences. These include the average precision (AP) of the pore and particle segmentation results obtained using threshold segmentation, as well as the average precision of the pore and particle segmentation results obtained using the method of the present invention. AP50 is the AP value when IOU = 0.5, and AP75 is the AP value when IOU = 0.75. Table 2 shows that the 3D reconstruction of rock particle segmentation and pore structure segmentation using the method of the present invention also outperforms traditional threshold segmentation methods.

[0140] In addition, core CT images are often used to calculate rock particle size distribution, which is useful information for geological analysis and engineering applications (such as sand control). Figure 7 The figure shows a 3D particle size distribution diagram of a core CT image. It includes the 3D particle size distributions of the core CT image reconstructed using threshold interpolation and the method of the present invention, as well as the corresponding true values. The figure shows that the particle size distribution obtained by the convolutional neural network-based 3D super-resolution reconstruction method for core CT images provided by the present invention more closely matches the true values. Therefore, the 3D reconstruction of rock particle segmentation using the method of the present invention is superior to the traditional threshold interpolation reconstruction method.

[0141] The embodiment of the present application provides a three-dimensional super-resolution reconstruction method for core CT images based on convolutional neural networks. By training a particle segmentation three-dimensional reconstruction model and a pore segmentation three-dimensional reconstruction model based on Transformer and 3D-CNN, the core CT images are segmented and reconstructed for particles and pores respectively. Finally, the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image output by the fusion model are combined to obtain a core CT three-dimensional super-resolution reconstruction image. Compared with the high-precision hardware performance of CT image reconstruction in the prior art, this method is low-cost and easy to promote. It can be applied to sample images collected by most CT and direct media interface (DMI) devices. At the same time, the method can exclude the influence of the environment, human factors and sample image storage process of the CT device, obtain a clearer and more detailed core CT three-dimensional high-resolution particle image, and increase the volume thickness to avoid the error caused by reconstruction through interlayer interpolation, thereby fully displaying the basic detail information of the sample, providing effective data protection for scientific research and technical personnel, and greatly improving work efficiency.

[0142] Example 2

[0143] Based on the same inventive concept, the present application also provides a three-dimensional super-resolution reconstruction device for core CT images based on a convolutional neural network, referring to Figure 7 As shown, the device includes:

[0144] The particle segmentation module 101 is used to input the pre-processed core CT serial slices obtained in advance into the trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; the particle segmentation 3D reconstruction model includes Transformer and 3D-CNN;

[0145] The pore segmentation module 102 is used to input the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes a Transformer and a 3D-CNN;

[0146] The fusion module 103 is configured to fuse the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image to obtain a core CT three-dimensional super-resolution reconstructed image.

[0147] In the second embodiment, the above modules perform the convolutional neural network-based three-dimensional super-resolution reconstruction process of core CT images in steps S101 to S103 of the first embodiment, and the repeated parts are not repeated here.

[0148] Example 3

[0149] Based on the same inventive concept, an embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements the three-dimensional super-resolution reconstruction method of core CT images based on convolutional neural network as described in the above embodiment 1.

[0150] Example 4

[0151] Based on the same inventive concept, an embodiment of the present application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the three-dimensional super-resolution reconstruction method of core CT images based on convolutional neural network as described in the above embodiment 1.

[0152] Example 5

[0153] Based on the same inventive concept, an embodiment of the present application also provides a computer program product comprising instructions. When the computer program product is run on a computer device, the computer device executes the three-dimensional super-resolution reconstruction method of core CT images based on convolutional neural networks as described in the above-mentioned embodiment 1.

[0154] Example 6

[0155] Based on the same inventive concept, an embodiment of the present application also provides a chip, which includes a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run computer programs or instructions to implement the three-dimensional super-resolution reconstruction method of core CT images based on convolutional neural networks as described in the above embodiment one.

[0156] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) that contain computer-usable program code.

[0157] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0158] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0159] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0160] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A three-dimensional super-resolution reconstruction method for core CT images based on convolutional neural networks, characterized in that: include: The pre-processed core CT serial slices are input into the trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; The particle segmentation 3D reconstruction model includes Transformer and 3D-CNN; Inputting the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes a Transformer and a 3D-CNN; The particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image are fused to obtain a core CT three-dimensional super-resolution reconstructed image.

2. The method according to claim 1, wherein The trained particle segmentation and reconstruction model is obtained by the following method: Dividing the pre-constructed first core CT serial slice sample set into a first training set, a first validation set, and a first test set according to a preset ratio; Inputting the first training set into a pre-built particle segmentation 3D reconstruction model for training, outputting a multi-slice-based particle segmentation longitudinal super-resolution image, and obtaining a trained particle segmentation 3D reconstruction model; Inputting the first verification set into the trained particle segmentation 3D reconstruction model to obtain a verification result, and comparing it with the corresponding real labeled image to update the trained particle segmentation 3D reconstruction model; Iterative training is performed until the difference between the verification result and the corresponding true label image meets a preset requirement, and the first test set is used for testing to obtain a trained particle segmentation 3D reconstruction model.

3. The method according to claim 2, wherein The step of inputting the first training set into a pre-built particle segmentation three-dimensional reconstruction model for training, outputting a multi-slice-based particle segmentation longitudinal super-resolution image, and obtaining a trained particle segmentation three-dimensional reconstruction model comprises: Based on the Transformer, perform in-image feature extraction on each core CT serial slice in the first training set, and output a single-slice particle segmentation transverse super-resolution image; Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based particle segmentation transverse super-resolution image, and a multi-slice-based particle segmentation longitudinal super-resolution image is output to obtain a trained particle segmentation three-dimensional reconstruction model.

4. The method according to claim 2, wherein The first core CT serial slice sample set is constructed in the following manner: The pre-processed multiple sets of historical core CT image serial slices are annotated with particle contours and categories to obtain a core CT serial slice label set; A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set; The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the first core CT sequence slice sample set.

5. The method according to claim 1, wherein The trained pore segmentation 3D reconstruction model is obtained by the following method: Dividing the pre-constructed second core CT sequence slice sample set into a second training set, a second validation set, and a second test set according to a preset ratio; Inputting the second training set into a pre-built pore segmentation 3D reconstruction model for training, outputting a multi-slice-based pore segmentation longitudinal super-resolution image, and obtaining a trained pore segmentation 3D reconstruction model; Inputting the second verification set into the trained pore segmentation 3D reconstruction model to obtain a verification result, and comparing it with the corresponding real label image to update the trained pore segmentation 3D reconstruction model; Iterative training is performed until the difference between the verification result and the corresponding true label image meets the preset requirements, and the second test set is used for testing to obtain a trained pore segmentation 3D reconstruction model.

6. The method according to claim 5, wherein The step of inputting the second training set into a pre-built pore segmentation three-dimensional reconstruction model for training, outputting a multi-slice-based pore segmentation longitudinal super-resolution image, and obtaining a trained pore segmentation three-dimensional reconstruction model comprises: Based on the Transformer, extract in-image features from each core CT serial slice in the second training set, and output a transverse super-resolution image of pore segmentation based on a single slice; Based on the 3D-CNN, intra-image feature extraction and inter-layer feature extraction are performed on each core CT serial slice in the single-slice-based pore segmentation transverse super-resolution image, and a multi-slice-based pore segmentation longitudinal super-resolution image is output to obtain a trained pore segmentation three-dimensional reconstruction model.

7. The method according to claim 5, wherein The second core CT serial slice sample set is constructed in the following manner: The pore contours of multiple sets of pre-processed historical core CT image sequence slices are annotated to obtain a core CT sequence slice label set; A number of slices are randomly and discontinuously extracted from each group of the historical core CT image sequence slices and noise is added to obtain a core CT sequence slice training set; The core CT sequence slice label set is matched with the core CT sequence slice training set to obtain the second core CT sequence slice sample set.

8. A three-dimensional super-resolution reconstruction device for core CT images based on convolutional neural networks, characterized in that: include: A particle segmentation module is used to input the pre-processed core CT serial slices obtained in advance into a trained particle segmentation and reconstruction model to obtain a multi-slice-based particle segmentation longitudinal super-resolution image; the particle segmentation 3D reconstruction model includes a Transformer and a 3D-CNN; A pore segmentation module is used to input the pre-processed core CT serial slices into a trained pore segmentation and reconstruction model to obtain a multi-slice-based pore segmentation longitudinal super-resolution image; the pore segmentation 3D reconstruction model includes a Transformer and a 3D-CNN; A fusion module is used to fuse the particle segmentation longitudinal super-resolution image and the pore segmentation longitudinal super-resolution image to obtain a core CT three-dimensional super-resolution reconstructed image.

9. A computer-readable storage medium having a computer program stored therein, wherein when the program is executed by a processor, the processor executes the convolutional neural network-based three-dimensional super-resolution reconstruction method for core CT images as described in any one of claims 1 to 7.

10. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for three-dimensional super-resolution reconstruction of core CT images based on a convolutional neural network as described in any one of claims 1 to 7 is implemented.

11. A computer program product comprising instructions, which, when executed on a computer device, causes the computer device to execute the method for three-dimensional super-resolution reconstruction of core CT images based on a convolutional neural network according to any one of claims 1 to 7.

12. A chip comprising a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is configured to execute a computer program or instruction to implement the convolutional neural network-based three-dimensional super-resolution reconstruction method for core CT images according to any one of claims 1 to 7.

Images