3D Reconstruction Method of Core Image Microstructure Based on LSTM

Abstract

The present invention discloses a three-dimensional reconstruction method for the microscopic structure of core images based on LSTM, comprising the following steps: (1) collecting and producing a core binary image data set for the training and testing of the network; (2) designing a core image recurrent generation network model based on a recurrent neural network; (3) designing a three-dimensional core image generation module based on a recurrent neural network, an autoencoder, and a convolutional block attention network; (4) designing a cross-sectional loss function L based on the two-dimensional image morphological features sec ; (5) based on the above model and loss function, completing the training to obtain a core image recurrent generation network model; (6) based on the trained model, using the generation module therein to complete the three-dimensional reconstruction of digital core images; the reconstruction method proposed by the present invention can reconstruct core images with a larger field of view 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 microstructures, and particularly relates to a method for three-dimensional reconstruction of core image microstructures based on a Long-short term memory network (LSTM) model, belonging to the technical field of image processing. Background Art

[0002] The microstructures of cores determine their macroscopic properties (such as permeability, conductivity, etc.). By performing three-dimensional reconstruction on digital core images, various characteristics of formations can be quantitatively analyzed and simulated, solving problems that are difficult to overcome in traditional petrophysical experiments (such as difficulties in displacing low-porosity and low-permeability rocks, and difficulties in obtaining representative cores of fractured carbonate rocks). It provides important basic research data for the exploration and development of unconventional reservoir resources and geological scientific exploration, and is an important development direction at home and abroad in recent years.

[0003] Currently, there are mainly two ways to obtain core microstructures: one is the direct reconstruction method, that is, using imaging equipment (such as a CT scanner, etc.) to directly scan the cut core samples to obtain their three-dimensional core structures. Although this method can accurately obtain the three-dimensional structure of the core, there are still some problems that are difficult to overcome. For example, there is a contradiction between the resolution and the sample size in CT scanning imaging. In order to obtain a high-precision three-dimensional structure, high-resolution pore structure images need to be obtained, but the scanned sample size is restricted, making the representativeness of the rock samples somewhat lacking. Therefore, it is difficult to obtain a three-dimensional structure with high accuracy and good sample representativeness through the direct scanning method. The second digital core reconstruction method is the model reconstruction method. Different from the direct model reconstruction method, the model reconstruction method usually only uses limited two-dimensional images, and reconstructs its corresponding three-dimensional structure by learning its patterns and data distributions. That is to say, it usually completes three-dimensional reconstruction using the limited information contained in the two-dimensional images. Currently, the main three-dimensional reconstruction methods based on models are: reconstruction methods based on optimization, reconstruction methods based on multi-point statistics, reconstruction methods based on super-dimensions, and reconstruction methods based on machine learning. The first two types of methods can be classified as traditional model reconstruction methods. Such methods have problems such as low reconstruction efficiency, a geometric multiplication relationship between the reconstruction size and time complexity, and insufficient diversity of the generated samples, and are difficult to apply to actual industrial scenarios.

[0004] With the continuous breakthroughs in deep learning technologies, using deep learning algorithms to complete the reconstruction of core image microstructures has gradually become a research hotspot. Currently, the mainstream deep learning algorithm is the three-dimensional reconstruction model based on the generative adversarial network. However, the algorithms based on this model have problems that are difficult to overcome, such as limited model generation ability and small generated samples (64 3 or 128 3) The training is unstable and problems such as mode collapse are likely to occur. Therefore, the present invention proposes a three-dimensional reconstruction model of core images based on a recurrent generation network, and this algorithm can increase the size of the generated images to 512 3 . Summary of the Invention

[0005] The object of the present invention is to propose a three-dimensional reconstruction method for the microscopic structure of core images based on LSTM, and increase the scale of the generated samples to 512 3 .

[0006] The present invention realizes the above object through the following technical solutions:

[0007] (1) Collect and produce a core binary image dataset for the training and testing of the network;

[0008] (2) Design a core image recurrent generation network model based on a recurrent neural network;

[0009] (3) Design a three-dimensional core image generation model based on a long short-term memory network, an autoencoder, and an attention network for spatial\channel fusion;

[0010] (4) Design a cross-section loss function L based on the morphological features of two-dimensional images sec ;

[0011] (5) Based on the above model and loss function, complete the training to obtain a core image recurrent generation network model;

[0012] (6) Based on the trained model, use the generation module therein to complete the three-dimensional reconstruction of digital core images. Description of the Drawings

[0013] Figure 1 is a flowchart of the three-dimensional reconstruction method of core images based on the recurrent generation network model of the present invention;

[0014] Figure 2 is the three-dimensional reconstruction network structure of large field-of-view core images based on the recurrent generation network model proposed by the present invention;

[0015] Figure 3 is the generation module of the recurrent generation network model and the unfolded diagram when generating its recurrent sequence;

[0016] Figure 4 is the network structure diagram of the generation module in the recurrent generation network model;

[0017] Figure 5 is the structural block diagram of the cross-section loss;

[0018] Figure 6 Visual comparison diagram of the three-dimensional reconstruction of homogeneous core images;

[0019] Figure 7 Statistical function quantization comparison chart of homogeneous core images;

[0020] Figure 8 Three-dimensional reconstruction visual comparison chart of thin section images;

[0021] Figure 9 Statistical function quantization comparison chart of thin section images; Detailed implementation manners

[0022] The embodiments of the present invention will be 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 the following text, the present invention can be implemented in various forms and is not limited by the embodiments described in the accompanying drawings and the following text. The embodiments provided in the accompanying drawings and the following text are for enabling the present invention to be more completely and accurately understood by those skilled in the art.

[0023] Figure 1 Among them, a new three-dimensional reconstruction method for core images based on a recurrent neural network can be specifically divided into the following steps:

[0024] (1) Collect and produce a core binary image data set for the training and testing of the network;

[0025] (2) Design a core image recurrent generation network model based on a recurrent neural network;

[0026] (3) Design a three-dimensional core image generation module based on a recurrent neural network, an autoencoder, and a convolutional block attention network;

[0027] (4) Design a cross-section loss function L based on two-dimensional image morphological features sec ;

[0028] (5) Based on the above model and loss function, complete the training to obtain a core image recurrent generation network model;

[0029] (6) Based on the trained model, use the generation module therein to complete the three-dimensional reconstruction of digital core images.

[0030] In the step (2), the present invention proposes a digital core three-dimensional reconstruction network structure based on a recurrent neural network as shown in Figure 2 . The network consists of two parts, one is a generation module and the other is a reference module. The generation module consists of an encoder Encoder, a recurrent prediction network LSTM θ (recurrent neural network unit) and a decoder Decoder. The reference model consists of an encoder Encoder, a prior network LSTM φ (recurrent neural network unit), and an inference network LSTM ψ(Recurrent neural network unit), where the reference model and the generation model share an encoder. During the training phase, by randomly segmenting and sampling the training data set, the current training image sequence I is formed t：t+n ={I t , I t+1 , I t+2 …I t+n}, where in the embodiment of the present invention, the size of the training data set is 512×512×512, and the length of the segment n = 20, that is, 20 images are randomly and continuously sampled along the Z direction of the training samples each time, such as: {I0, I1, I2…I 19}. During the training phase, the encoder in the generation model learns the current frame image X t of the training image sequence and generates a feature hidden vector Z enc containing the current image features. At the same time, by sampling the data distribution learned by the prior network LSTM φ , a feature vector Z t+1 containing the next frame image is obtained. Z enc and Z t+1 are merged in the feature dimension to obtain the total feature vector as shown in Equation (1):

[0031] Z total =Z enc +Z t+1 (1)

[0032] Where in the embodiment of the present invention, the dimension of Z enc is 128×1×1, the dimension of Z t+1 is 32×1×1, then the dimension of the total feature vector noise Z total is 160×1×1. The recurrent prediction network LSTM φ generates a vector g total containing the sequence relationship of the feature vectors by recurrently learning the feature vector Z t . Finally, g t is input into the decoder Decoder, and finally the next frame of core image is generated and then the reconstruction of the entire generated image sequence is completed

[0033] During the training phase, the reference module learns the features of the current frame I t and the next frame of real image I t+1 of the training image sequence through the encoder Encoder, and correspondingly inputs them into the prior recurrent neural network LSTM φ and the inference recurrent neural network LSTM ψ to enable the corresponding network to learn the prior distribution based on the current recommendation and the inferred distribution P of the next frame image ψ (Z t+1 |I t+1 ). Using the KL divergence, continuously optimize the distance between these two distributions so that the prior long short-term memory (LSTM) neural network φ can learn the data distribution of the next frame image.

[0034]

[0035] In the step (3), design and construct an image generation model as shown in Figure 3 and Figure 4 . After the 3D reconstruction model shown in step (2) is completed training, the generation model completes learning the training image features and the information between image layers. As shown in Figure 3 , in the core image generation stage, only need to input the first reference image X t at the generation model end, and the encoder will encode this image into Z enc . At the same time, in order to enrich the diversity of the generated images, sample the noise vector Z t in the Gaussian space. Since the noise Z t has been constrained during the training stage, then jointly encode the noise Z enc and the Gaussian noise Z t and input them into the decoder, and a synthetic image with both accuracy and diversity can be generated.

[0036] In order to enable the generation model to fully learn the features of the 2D image, improve the traditional U-NET network structure. In order to increase the diversity of the generated images and overcome the direct copying of images caused by overfitting, add a convolutional block attention module and Gaussian noise with pixel-wise addition on the original U-NET network structure. As shown in Figure 5 , by inputting the feature map into the convolutional block attention module (CBAM), improve the model's learning ability for the feature map. By introducing Gaussian noise, the diversity of the generated images can be enhanced to a limited extent. In this embodiment, taking the first-layer feature map of the encoder as an example, when the 2D image X t with a scale of 1×512×512 passes through the first convolutional layer, the generated feature map F1 = 64×256×256. At this time, input this feature map into the convolutional block attention module to improve the feature expression ability of the feature map in both the image channel and spatial dimensions, and improve the feature extraction and morphological reconstruction ability of the generation module for the image. At the same time, introduce a standard Gaussian noise with a dimension of N1 = 64×256×256, and perform pixel-wise addition with the feature map processed by the convolutional block attention module to increase the generalization of the generation network.

[0037] In step (4), in order to improve the accuracy of the model in generating images and enhance the continuity of the images in the generation direction (Z direction), a sectional loss function based on the pre-trained VGG-16 network model is designed. Among them, the VGG-16 model is a neural network model trained by a research team at the University of Oxford on a large number of image datasets. This model can be publicly obtained and used. For a training data sequence and a generated data sequence, images of size 256×512 are extracted at a certain step length on their XZ planes and simultaneously input into the pre-trained VGG-16 model that has been trained. Define the real training data sequence I t：t+n ={I t , I t+1 , I t+2 …I t+n}, similarly, define the generated image sequence Define, then define the loss function L sec of the gradient image as

[0038]

[0039] wherein, the pre-trained VGG-16 network is represented by , represents the feature map processed by the j-th network module of the VGG-16 network. j represents the neural network module in the VGG-16 network. For the VGG-16 network, there are a total of 13 neural network modules. In the example of the present invention, j = 3. C j , H j , W j represent the number of channels, the length and width of the feature map of the j module. N is the number of cross-sections taken for the training data and the generated data on the XZ plane. In the example of the present invention, N = 3. represents the i-th image obtained from the training data sequence on the XZ plane. Similarly, represents the i-th image obtained from the generated data sequence on the XZ plane.

[0040] In addition, there is also a pixel reconstruction loss L rec in the training stage. The pixel reconstruction loss is to measure the pixel error between the generated image sequence and the real training image sequence I t：t+n ={I t , I t+1 , I t+2 …I t+n}, and its expression is:

[0041]

[0042] Through the constraint of the loss function, the reconstructed three-dimensional sample is closer to the real sample in statistical metrics. Then the total loss function L in the training stage total is as follows:

[0043] L total = λ rec L rec + λ kl L kl + λ sec L sec (5)

[0044] In the example of the present invention, λ rec , λ kl and λ sec are respectively taken as 1, 100, and 50.

[0045] In the step (5), based on the above-designed model and loss function, training is completed to obtain a cyclic generation three-dimensional reconstruction model.

[0046] In the step (6), after training is completed, only a reference image X t is input into the generation model, and then the network can cyclically generate a synthetic image That is, the layer-by-layer three-dimensional image reconstruction based on the recurrent neural network is completed. In the example of the present invention, taking the generation of a three-dimensional core image with a scale of 512×512×512 as an example, initially, a reference image I0 needs to be input into the generator, and through the cyclic iteration of the generator, the remaining 511 core images are generated Then the reference image and the generated image are superimposed to synthesize a core image with a scale of 512×512×512.

[0047] To prove the effectiveness of the method of the present invention, the present invention reconstructs homogeneous core images and optical thin section images, and determines the effectiveness of the algorithm through visual reconstruction effect comparison and statistical function quantification comparison. The relevant experimental results are as follows:

[0048] Figure 6 is the three-dimensional reconstruction result for the homogeneous core image. Among them, Figure (a) is the real reconstruction target, and Figure (b) is the reconstructed image. Through visual comparison, it can be seen that the three-dimensional structures in Figure (a) and Figure (b) have a high morphological similarity. By comparing the cross-sectional views of the three orthogonal sections of the three-dimensional structure in Figure (a) and Figure (b), it can be known that the generated image has a high similarity in morphology with the real image, and the multi-facetedness of the generated image is also good. Figure 7 Shows the quantification comparison of the statistical parameters after the three-dimensional reconstruction of the homogeneous core. Through the quantification comparison of the two-point correlation function, the linear path function, and the two-point cluster function, it can be known that the three-dimensional model reconstructed by the method of the present invention has a high matching degree with the real three-dimensional structure in various statistical features.

[0049] Figure 8 It is the three-dimensional reconstruction result of the optical thin-section core image, where Figure (a) is the original optical thin-section image with a size of 3000×2280. Figure (b) is the slice to be reconstructed with a size of 512×512 taken from the original optical thin-section image. (c) shows the reconstruction result of the slice to be reconstructed. By observing the reconstructed three-dimensional structure and comparing its two-dimensional form with the real image in Figure (b) through analysis, it can be seen that the generated two-dimensional image has similarity in form with the image to be reconstructed. In addition, by conducting statistical parameter comparisons, such as Figure 9 as shown. Through statistical parameter comparisons, it can be known that the statistical parameters of the generated image are in good agreement with those of the image to be reconstructed, which proves the effectiveness of the model.

[0050] Combined with the comparison and verification of the subjective visual effect and the objective statistical function, it can be seen that the method of the present invention has a good reconstruction effect on the core image. To sum up, the present invention is an effective method for three-dimensional reconstruction of core images. The invention can serve the field of petroleum geology, reduce the cost of core sample image acquisition, improve the accuracy of core image analysis, and has great value in practical applications such as oil and gas exploration and exploitation.

[0051] The above embodiments are only the preferred embodiments of the present invention, and do not limit the technical solutions described in the present invention. Any technical solution that can be achieved on the basis of the above embodiments without creative labor shall be regarded as falling within the protection scope of the content of the present invention.

Claims

1. Three-dimensional reconstruction method for microscopic structure of core images based on LSTM It is characterized by the following steps: (1) Collect and produce a core binary image dataset for the training and testing of the network; (2) Design a cyclic generation network model for core images based on LSTM. The model structure consists of a generation module and a reference module. The generation block is composed of an encoder Encoder, a cyclic prediction network LSTM θ and a decoder Decoder. The reference model consists of an encoder Encoder, a prior network LSTM φ , and an inference network LSTM ψ . Among them, the reference module and the generation module share an encoder. In the training stage, by randomly segmenting and sampling the training data set, the current training image sequence I t:t+n = {I t , I t+1 , I t+2 … I t+n} is formed. The encoder in the generation model generates a feature hidden vector Z t containing the current image features by learning the current frame image I enc of the training image sequence. At the same time, the reference module learns the features of the next frame real image I t+1 of the training image sequence through the encoder Encoder, generates the corresponding feature hidden vector, and inputs the feature hidden vector into the prior network LSTM φ , so that the recurrent neural network continuously learns the sequence relationship of the training sample data. Finally, sample the data distribution generated by LSTM φ to obtain a feature vector Z t+1 containing the next frame image. Connect and merge Z enc and Z t+1 to form a feature vector Z total . The cyclic prediction network LSTM φ generates a vector g total containing the sequence relationship of the feature vector sequence by cyclically learning the feature vector Z t . Finally, input g t into the decoder Decoder to finally generate the next frame of core image (3) Design a three-dimensional core image generation module based on long short-term memory network, autoencoder and convolutional block attention network. Different from the traditional U-NET network structure, in order to increase the diversity of the generated images and overcome the direct copying of images caused by overfitting, a convolutional block-based attention module and Gaussian noise added by pointwise pixel addition are added to the original U-NET network structure; through training, the generation model can learn the image features and the inter-layer relationship features between images in the training images; when the model training is completed, only a reference image I is input to the generation model. t Then the network can cyclically generate synthetic images according to the learned inter-layer information. That is, the layer-by-layer three-dimensional image reconstruction based on the recurrent neural network is completed. (4) Design the cross-section loss function L based on the morphological features of two-dimensional images sec ; (5) Based on the above model and loss function, complete the training to obtain a core image cyclic generation network model; (6) Based on the trained model, use the generation module therein to complete the 3D reconstruction of the digital core image.

2. The three-dimensional reconstruction method of the core image microstructure based on LSTM according to claim 1, characterized in that The cross-section loss function L based on two-dimensional image morphological features described in step (4) sec ; In order to improve the accuracy of the images generated by the model, a cross-section loss function based on the pre-trained VGG-16 network model is designed. For a training data sequence and a generated data sequence, images of size 256×512 are extracted at a certain step length on their XZ planes and simultaneously input into the pre-trained VGG-16 model that has been trained. Define the true training data sequence I t:t+n ={I t , I t+1 , I t+2 …I t+n}, Similarly, define the generated image sequence Define, then define the loss function L of the gradient image sec as Among them, the VGG-16 pre-trained network is denoted by ; represents the feature map processed by the j-th network module of the VGG-16 network; j represents the neural network module in the VGG-16 network; C j , H j , W j represent the number of channels, the length and width of the feature map of the j-th module; N is the number of cross-sections taken for the training data and the generated data on the XZ plane; denotes the i-th image obtained from the training data sequence on the XZ plane; similarly, denotes the i-th image obtained from the generated data sequence on the XZ plane.

Images