Rock core image microstructure multi-scale fusion reconstruction method based on NNTM

Abstract

The invention discloses a core image microstructure multi-scale fusion reconstruction method based on NNTM. The core image microstructure multi-scale fusion reconstruction method comprises the following steps: (1) collecting a large-vision-field low-resolution three-dimensional image and a small-vision-field high-resolution two-dimensional image of a core; (2) designing a core image microstructure multi-scale fusion reconstruction algorithm based on neural network transformation mapping; (3) designing a network structure of the mapping network; (4) designing a structure loss function # imgabs0 # for controlling and generating a structure pore trunk; (5) designing a detail loss function # imgabs1 # for controlling and generating structure pore details; and (6) carrying out network optimization based on the model and the loss functions to realize multi-scale fusion reconstruction. The multi-scale fusion reconstruction method provided by the invention can combine a small-vision-field high-resolution two-dimensional core image and a large-vision-field low-resolution three-dimensional image to model a large-vision-field high-resolution three-dimensional structure of the core, and has important application value in the field of petroleum geology.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional reconstruction of core microstructure, and specifically relates to a multi-scale fusion reconstruction method of core image microstructure based on Neural Network-based Transformation Mapping (NNTM), which belongs to the technical field of image processing. Background Art

[0002] Microstructure modeling is an important way to study the pore space inside the core. As a place for transmission, diffusion and reaction, the pore space plays a vital role in determining the macroscopic physical characteristics and migration phenomena of the core. The key to studying the microstructure of the core is the connection between its microscopic properties and macroscopic characteristics, which can provide important basic research data for resource exploration and development of unconventional reservoirs, geological scientific exploration, geological modeling and other research.

[0003] With the development of imaging technology (such as micro-computed tomography (CT)), digital images of core sample materials can be directly obtained and three-dimensional microstructure models can be established. For the three-dimensional microstructure of the core modeled, the following conditions are usually required: 1) there must be a large enough field of view to make the structure representative; 2) the resolution is high enough to provide key details. However, due to the limitations of the imaging principle, traditional CT needs to make a trade-off between resolution and field of view, and it is difficult to obtain high-resolution three-dimensional structures with a large field of view through direct imaging. Generally, the expansion of the field of view will lead to a decrease in resolution and the inability to capture detailed features (such as tiny pores, subtle features at the edges of pores, and narrow channels). To obtain higher-resolution images, it is necessary to cut out smaller parts from the sample for scanning and imaging. At this time, although the local detailed features can be captured, the local sample lacks representativeness.

[0004] Multiscale fusion reconstruction is a feasible method to obtain high-resolution three-dimensional microstructure with a large field of view. Multiscale reconstruction combines low-resolution images with a small field of view of the same core to model high-resolution microstructure with a large field of view. In recent years, a variety of methods have been developed to achieve multiscale fusion reconstruction, including methods based on hybrid superposition, methods based on pattern matching, methods based on statistical reconstruction, and methods based on neural networks.

[0005] Hybrid superposition-based methods aim to match images of different resolutions (or scales) through feature matching methods, and then design superposition rules to construct multi-scale models. Hybrid superposition-based methods do not need to assume the self-similarity or statistical properties of the pore space, so they are applicable to homogeneous and heterogeneous porous materials. However, limited by the accuracy of the three-dimensional structural feature matching method, this type of method usually performs well when processing relatively uncomplicated two-dimensional images, but it is difficult to avoid the overlap of different types of pores (or components) when processing complex three-dimensional structures. Pattern matching-based methods usually use templates to directly match high-resolution and low-resolution patterns, or establish pattern dictionaries / sets to modify low-resolution images or three-dimensional pore structures. Pattern matching-based methods avoid pore overlap to a certain extent and focus on the modification of pore geometry and pore surface morphology. However, the limitation of template size may introduce noise into the reconstructed structure, causing additional discrete pores and / or erroneous connections in the reconstructed microstructure. Statistical reconstruction-based methods rely on the uniformity assumption and focus on the reconstruction of statistical features, so the reconstruction results can follow the structural characteristics of the reference image, but often ignore the pore geometry and pore surface morphology. In addition, due to the difficulties in characterizing heterogeneous structures, statistical reconstruction-based methods often give unsatisfactory results when applied to reconstruct such structures.

[0006] The application of deep learning provides new possibilities for exploring more effective methods in this field. In recent years, a variety of multi-scale fusion reconstruction methods for cores based on neural networks have emerged. These methods use objective functions to guide the fusion of semantic features of images in high-dimensional feature space, thereby avoiding the noise and overlap problems introduced in traditional methods, and can generate reconstruction results with similar physical and topological properties (such as porosity, local porosity distribution and permeability, etc.) to the real three-dimensional pore structure. However, these deep learning methods still have some difficult problems to solve. On the one hand, the generalization of the model depends on the provided data set. For complex and diverse core structures, it is impossible to obtain a sufficiently complete data set to train a general model suitable for various types of cores. On the other hand, the size of the data set and the device video memory also limit the acquisition of general models. To this end, the present invention proposes a multi-scale fusion reconstruction method for core images based on transformation mapping of neural networks. This method does not need to be trained on a data set. It only provides a small field of view high-resolution two-dimensional image and a large field of view low-resolution three-dimensional image, and can model a large field of view high-resolution fine microstructure. Summary of the invention

[0007] The purpose of the present invention is to propose a multi-scale fusion reconstruction method of core image microstructure based on NNTM, which can model a large-viewing-area high-resolution microstructure (Rec) using only a small-viewing-area high-resolution two-dimensional image (HRI) and a large-viewing-area low-resolution three-dimensional image (LRI).

[0008] The present invention achieves the above object through the following technical solutions:

[0009] (1) Collecting low-resolution 3D images with a large field of view and high-resolution 2D images with a small field of view of the core;

[0010] (2) Design a multi-scale fusion reconstruction algorithm for core image microstructure based on neural network transformation mapping;

[0011] (3) Design the network structure of the mapping network;

[0012] (4) Designing a structural loss function to control the generation of structural pore backbones

[0013] (5) Design a detail loss function to control the generated structural pore details

[0014] (6) Based on the above model and loss function, network optimization is performed to achieve multi-scale fusion reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the multi-scale fusion reconstruction method of core images based on neural network transformation mapping of the present invention;

[0016] Figure 2 It is a block diagram of the algorithm for multi-scale fusion reconstruction based on neural network transformation mapping proposed by the invention;

[0017] Figure 3 is the network structure of the mapping network proposed by the present invention;

[0018] Figure 4 It is the calculation process of the structure loss and detail loss proposed by the present invention;

[0019] Figure 5 This is a visual comparison of the multi-scale fusion reconstruction of the core image of the Fontainebleau sandstone;

[0020] Figure 6 This is a quantitative comparison of the multi-scale fusion reconstruction of the core image of the Fontainebleau sandstone;

[0021] Figure 7 This is a visual comparison of the multi-scale fusion reconstruction of the core image of the Estaillades carbonate rock;

[0022] Figure 8 This is a quantitative comparison of the multi-scale fusion reconstruction of the core image of the Estaillades carbonate rock; DETAILED DESCRIPTION

[0023] The embodiments of the present invention are described in more detail below with reference to the accompanying drawings. Although the embodiments of the present invention are given in the accompanying drawings and below, the present invention can be implemented in many forms and is not limited by the embodiments described in the accompanying drawings and below. The embodiments described in the accompanying drawings and below are provided to enable the present invention to be more completely and accurately understood by those skilled in the art.

[0024] Figure 1 In this paper, a multi-scale fusion reconstruction method of core image microstructure based on neural network transformation mapping is proposed, which can be divided into the following steps:

[0025] (1) Collecting low-resolution 3D images with a large field of view and high-resolution 2D images with a small field of view of the core;

[0026] (2) Design a multi-scale fusion reconstruction algorithm for core image microstructure based on neural network transformation mapping;

[0027] (3) Design the network structure of the mapping network;

[0028] (4) Designing a structural loss function to control the generation of structural pore backbones

[0029] (5) Design a detail loss function to control the generated structural pore details

[0030] (6) Based on the above model and loss function, network optimization is performed to achieve multi-scale fusion reconstruction.

[0031] For step (1), since the multi-scale fusion reconstruction method of core image microstructure based on neural network transformation mapping of the present invention only requires a large-viewing-area low-resolution three-dimensional image and a small-viewing-area high-resolution two-dimensional image, a large-viewing-area high-resolution three-dimensional structure can be generated without the need to prepare a data set.

[0032] In step (2), the algorithm block diagram of the multi-scale fusion reconstruction based on the neural network transformation mapping proposed in the present invention is shown in Figure 2. The multi-scale fusion reconstruction algorithm of the neural network transformation mapping proposed in the present invention regards the multi-scale fusion reconstruction as an image fusion problem, adopts a generative paradigm, and uses the neural network G(θ) to fit the transformation mapping of the noise z to the reconstructed structure Rec Then there is G(z;θ) = Rec. The pore space backbone of the reconstructed structure is obtained by the structural loss function Control, the pore details are determined by the detail loss function Control, total loss The pore space backbone of the reconstructed structure should be consistent with the large-viewing-area low-resolution image, and the pore space details of the reconstructed structure should be similar to the small-viewing-area high-resolution image.

[0033] Specifically, in step (3), the network structure of the mapping network proposed by the present invention is as follows: Figure 3 As shown. The fusion of HRI and LRI involves the fusion of features of different scales of the two images, mapping Decompose into multiple sub-maps, so that the network can better fuse features at the corresponding scale, mapping It can be expressed as follows:

[0034]

[0035] N in formula (1) represents the number of feature scales, which is determined according to the point size ratio of HRI and LRI. The specific calculation formula is as follows:

[0036]

[0037] P in formula (2) LRI and P HRI Represent the point sizes of HRI and LRI respectively. According to formula (3), the number of layers of the mapping network can be determined. The network consists of N nonlinear mapping units G and 1 feature integration module C. The network G can be expressed as:

[0038] G=G 1 G 2 …G n …G N C (3)

[0039] Among them G n Represents the nonlinear mapping unit of the nth feature scale. C represents the feature integration module, which integrates features of different scales and generates reconstruction results. This design enables the network to process features of different scales layer by layer, thereby better capturing the relationship between the pore space skeleton and the pore details. At the same time, it can also make the network flexibly applicable to multi-scale fusion reconstruction of HRI and LRI of different resolutions. In order to improve the reconstruction efficiency without affecting the reconstruction accuracy, the network should be designed as a lightweight structure. Reducing the network complexity can speed up the network reasoning and reconstruction process. Nonlinear mapping unit G n The module is mainly composed of convolutional layers, activation functions and normalization layers. The nonlinear ability of the module is achieved by combining convolutional layers and activation functions:

[0040] Y=h(θ W X+θ B ) (4)

[0041] In formula (4), X represents the input of the convolutional layer; Y is the output of the convolutional layer processed by the activation function h; θ W and θ BRepresent the weight and bias of the convolution layer respectively. The convolution layer uses 3D convolution, the activation function uses LeakReLU activation function, and the normalization layer uses Batch Normalization. The first N-1 nonlinear mapping units (G 1 ,G 2 ,…,G n ,…,G N-1 ) is expanded in size through an upsampling operation, so that the network G generates a reconstructed structure by gradually upsampling, which ensures that each nonlinear mapping unit focuses on optimizing a specific feature scale. Each nonlinear mapping unit G n Both fit the transformation sub-map f of the current feature scale n , the noise z of the current scale n Transformed into the output g of the current layer n The network input is multi-scale noise, and the number of noise scales is determined according to formula (3). Assuming the size of LRI is l×w×h, the initialization noise z N The size is 2 N l×2 N w×2 N h, noise of other scales is obtained by N Downsample sequentially to obtain the noise z of the nth scale n The size is 2 n l×2 n w×2 n h. In addition, the lower-level feature maps (output g from the previous module) n-1 ) Noise z n (Current feature scale) is connected in the channel dimension through the "concatenate" operation, and then the result of the connection is input into G n These low-level feature maps are gradually propagated to higher feature scales as prior information, providing additional constraints for the reconstruction process, improving optimization speed and reconstruction accuracy. Each nonlinear mapping unit G n Both fit the transformation sub-map f of the current feature scale n , the process can be expressed as:

[0042]

[0043] The structural loss function described in step (4) and the detail loss function described in step (5) The calculation process is as follows Figure 4 As shown. The reconstructed structure Rec is downsampled to the same size as the LRI structure Rec / 2 N-1 , then calculate Rec / 2 N-1 The L2 distance between LRI and the LRI is used as the structural loss, as shown in formula (6):

[0044]

[0045] Where l, w, h represent the length, width, and height of the LRI respectively; (x, y, z) is the coordinate index. The structural loss function controls the pore backbone of the reconstructed structure to be consistent with the LRI. Random slices S are taken from the x, y, and z orthogonal directions of the reconstructed structure. xy , S yz , S zx , then calculate S xy , S yz , S zx The VGG loss between S and HRI is used as the detail loss. The VGG loss first uses a CNN model (such as VGG-16) pre-trained on a large dataset to obtain the xy , S yz , S zx and HRI to extract feature maps of multiple scales. Then the Gram matrix of each feature map is calculated to measure the similarity of their feature distributions, so as to constrain the detail similarity of the reconstructed structure with HRI. In order to calculate the Gram matrix, the feature map is vectorized using the "flatten" operation. The calculation of the Gram matrix is ​​shown in formula (7):

[0046]

[0047] In formula (7) represents the inner product of the i-th and j-th vectorized feature maps in the output of the m-th layer of the pre-trained model. The VGG loss of the m-th layer is shown in formula (9):

[0048]

[0049] In formula (8), I m is the number of feature maps output by layer m; J m is the number of elements in each feature map. The total VGG loss is shown in formula (10):

[0050]

[0051] In formula (9), β m represents the weight of the VGG loss of the mth layer, which is usually set to 1. The calculation of the detail loss is shown in formula (11):

[0052]

[0053] In step (6), based on the above model and loss function, network optimization is performed to achieve multi-scale fusion reconstruction. Update the network parameters θ to search for an optimal network to generate a reconstruction structure G(z; θ) with the minimum loss. This process is shown in formula (11):

[0054]

[0055] To prove the effectiveness of the method of the present invention, the present invention uses the core image of Fontainebleau sandstone and the core image of Estaillades carbonate rock to reconstruct, which have different pore morphologies. The effectiveness of the algorithm is determined by comparing the visual reconstruction effect and the quantitative comparison of statistical functions. The relevant experimental results are as follows:

[0056] Figure 5 It is the result of multi-scale fusion reconstruction of the core image of Fontainebleau sandstone. LRI is a low-resolution structure obtained by downsampling Original HRI. The Overall pore space column shows the three-dimensional structure of LRI, Original HRI and reconstructed structure Rec. The Layer column shows the two-dimensional images of the 1st, 128th and 256th slices of Original HRI and reconstructed structure Rec, as well as the two-dimensional images of the 1st, 32nd and 64th slices of LRI. The pore trunk of the reconstructed structure Rec is basically consistent with LRI and Original HRI. The pore geometry and pore edge details of Rec are very similar to those of Original HRI, and some pores that are isolated in LRI due to limited resolution are reconnected.

[0057] Figure 6 The quantitative comparison of statistical parameters of multi-scale fusion reconstruction of core images of Fontainebleau sandstone is shown. Through the comparison of pore size distribution, pore volume distribution, pore shape factor distribution and pore shape factor-pore size joint distribution, it can be seen that the three-dimensional structure obtained by multi-scale fusion reconstruction by the method of the present invention is highly similar to the original HRI in all indicators.

[0058] Figure 7It is the result of multi-scale fusion reconstruction of the core image of the Estaillades carbonate rock. LRI is a low-resolution structure obtained by downsampling the Original HRI. The Overall pore space column shows the three-dimensional structure of LRI, OriginalHRI and the reconstructed structure Rec. The Layer column shows the two-dimensional images of the 1st, 256th, and 512th slices of the Original HRI and the reconstructed structure Rec, as well as the two-dimensional images of the 1st, 64th, and 128th slices of the LRI. The pore trunk of the reconstructed structure Rec is basically consistent with LRI and Original HRI. The pore geometry and pore edge details of Rec are very similar to those of Original HRI, and some pores that are isolated in LRI due to limited resolution are reconnected. In addition, Figure 8 The quantitative comparison of statistical parameters of multi-scale fusion reconstruction of Estaillades carbonate core images is shown. Through the comparison of statistical parameters, it can be seen that the three-dimensional structure obtained by the reconstruction result is very consistent with the original HRI in all indicators, which proves the effectiveness of the method.

[0059] Combining the comparison and verification of subjective visual effects and objective statistical functions, it can be seen that the method of the present invention has a good multi-scale fusion reconstruction effect for core images. In summary, the present invention is an effective multi-scale fusion reconstruction method for core images. The invention can serve the field of petroleum geology, model large-viewing-area high-resolution three-dimensional structures, improve the accuracy of core image analysis, and has great value in practical applications such as oil and gas exploration and mining.

[0060] The above embodiments are only preferred implementation cases of the present invention and are not limitations of the technical solutions described in the present invention. Any technical solution that can be implemented on the basis of the above implementation cases without creative work should be deemed to fall within the protection scope of the present invention.

Claims

1. Multi-scale fusion reconstruction method of core image microstructure based on NNTM, It is characterized by the following steps: (1) Collecting low-resolution 3D images with a large field of view and high-resolution 2D images with a small field of view of the core; (2) Design a multi-scale fusion reconstruction algorithm for core image microstructure based on neural network transformation mapping; (3) Design the network structure of the mapping network; (4) Designing a structural loss function to control the generation of structural pore backbones (5) Design a detail loss function to control the generated structural pore details (6) Based on the above model and loss function, network optimization is performed to achieve multi-scale fusion reconstruction.

2. The NNTM-based multi-scale fusion reconstruction method for core image microstructure according to claim 1 is characterized in that Step (1) Design of the multi-scale fusion reconstruction algorithm based on neural network transformation mapping described in step (2); The multi-scale fusion reconstruction algorithm based on neural network transformation mapping proposed in the present invention regards multi-scale fusion reconstruction as an image fusion problem, adopts a generative paradigm, and uses the neural network G(θ) to fit the transformation mapping from noise z to the reconstructed structure Rec By combining only a large-viewing-area low-resolution 3D image and a small-viewing-area 2D image, a large-viewing-area high-resolution 3D structure can be modeled.

3. The NNTM-based multi-scale fusion reconstruction method of core image microstructure according to claim 1 is characterized in that The mapping network design described in step (3) is as follows: In order to make the network better able to fuse features at the corresponding scale, the mapping Decomposed into several sub-maps: N in formula (1) represents the number of feature scales, which is determined according to the point size ratio of HRI and LRI. The specific calculation formula is as follows: P in formula (2) LRI and P HRI Represent the point sizes of HRI and LRI respectively; According to formula (3), the number of layers of the mapping network can be determined, and the network G can be expressed as: G=G1G2…G n …G N C (3) The network G consists of N nonlinear mapping units G and 1 feature integration module C.

4. The NNTM-based multi-scale fusion reconstruction method of core image microstructure according to claim 1 is characterized in that Design of the structural loss function and detail loss function described in step (4) and step (5); In order to constrain the reconstructed structure Rec to be consistent with the pore trunk of LRI, the structural loss function designed by the present invention is shown in formula (4): Where l, w, h represent the length, width, and height of the LRI respectively; (x, y, z) is the coordinate index; Rec / 2 N-1 The reconstructed structure is downsampled to a structure of the same size as LRI. In order to control the detail similarity between the reconstructed structure and HRI, the detail loss function designed by the present invention is shown in formula (5): Where S xy , S yz , S zx Random slices in the x, y, and z orthogonal directions to reconstruct the structure; is vgg loss; I represents the inner product of the i-th and j-th vectorized feature maps in the output of the m-th layer of the pre-trained model; m is the number of feature maps output by layer m; J m is the number of elements in each feature map; β m Represents the weight of the m-th layer vgg loss.

Images