Deep learning-based methods, software products, and equipment for identifying matrix materials from rock strata images.

Abstract

This invention discloses a method, program product, and device for matrix material identification from rock strata images based on deep learning. The method includes: acquiring a rock strata image to be tested, the rock strata image to be tested including rock strata images under various energy channels; forming a rock strata input image to a trained matrix material identification model based on the rock strata image to be tested, wherein the matrix material identification model includes a decoder and a matrix material prediction network, and each decoding layer in the decoder includes an attention layer and an upsampling layer; acquiring a multi-attention feature map of the rock strata output by the attention layer based on the rock strata input image; acquiring upsampled rock strata features output by the upsampling layer based on the rock strata input image; fusing the multi-attention feature map of the rock strata and the upsampled rock strata features to obtain the rock strata fusion feature corresponding to the decoding layer; and predicting the matrix material data in the rock strata image to be tested through the matrix material prediction network based on the rock strata fusion feature corresponding to the decoding layer.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for identifying the matrix material of rock strata images based on deep learning, a computer program product, and a computing device. Background Technology

[0002] As unconventional oil and gas exploration and development extend to deeper reservoirs, and as the demand for precise early warning in coal mine safety monitoring increases, microscale characterization technology has become a core means of understanding coal and rock properties. Coal and rock, as a complex porous medium, are composed of organic matter, various minerals (quartz, kaolinite, calcite, pyrite, etc.), and pores and fractures of varying shapes. The microscopic characteristics of this multi-dimensional system directly determine the physicochemical properties of coal and rock, significantly impacting coalbed methane development and mine safety. In coalbed methane development, pore connectivity and tortuosity control gas adsorption / desorption efficiency and seepage capacity. In mine safety, the dissolution and recrystallization of minerals caused by microporous water-rock interaction weaken the mechanical properties of coal and rock, significantly increasing the risk of coal and gas outbursts.

[0003] To study the interior of coal and rock, current methods include traditional CT scanning, deep learning models, and threshold segmentation to characterize the internal composition. However, traditional CT data is only used for simple material decomposition and does not deeply explore the energy-material decay correlation features, making it difficult to distinguish the composition of complex materials. Conventional convolutional neural network deep learning models have fixed receptive fields, making it difficult to capture micropore or macropore features. Threshold segmentation relies on manual threshold selection and only coarsely classifies pores and minerals, leading to biased results. These methods are insufficient to address the material distribution patterns and pore evolution mechanisms of coal and rock at the microscale. Furthermore, when classifying minerals, they treat the entire coal and rock mass as a single mineral, without considering its actual mineral composition. Therefore, more precise methods for identifying coal and rock mineral components and more accurate pore extraction methods are needed. Summary of the Invention

[0004] To address the existing technical problems, this invention provides a method, computer program product, and computing device for identifying the matrix material of rock strata based on deep learning. This method can enhance attention features in densely porous areas to preserve details, and enhance upsampling features in homogeneous rock strata to maintain continuity, thereby improving the accuracy of pore identification.

[0005] In a first aspect, a method for identifying the base material of rock strata images based on deep learning is provided, comprising: acquiring a rock strata image to be tested, wherein the rock strata image to be tested includes rock strata images under each energy channel;

[0006] Based on the image of the rock strata to be tested, a rock strata input image is formed to the trained matrix material recognition model, wherein the matrix material recognition model includes a decoder and a matrix material prediction network, and each decoding layer in the decoder includes an attention layer and an upsampling layer.

[0007] Based on the input image of the rock strata, obtain the multi-attention feature map of the rock strata output by the attention layer;

[0008] Based on the input image of the rock strata, the upsampled rock strata features output by the upsampled layer are obtained;

[0009] The multi-attention feature map of the rock strata and the upsampled rock strata features are fused to obtain the rock strata fused features corresponding to the decoding layer;

[0010] Based on the rock strata fusion features corresponding to the decoding layer, the matrix material prediction network is used to predict the matrix material data in the image of the rock strata to be tested.

[0011] In a second aspect, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the deep learning-based method for identifying the matrix material of rock strata images as described in any embodiment of this application.

[0012] Thirdly, a computing device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform a deep learning-based method for identifying the matrix material of rock strata images as described in any embodiment of this application.

[0013] This application configures an attention layer in the decoder to obtain a multi-attention feature map of the rock strata, which can provide precise location information of the base material. An upsampling layer in the decoder obtains upsampled rock strata features, which can provide high-level semantic understanding. By fusing these two features, attention features can be enhanced in densely porous regions, thus preserving details, while upsampled features are enhanced in homogeneous rock strata regions, maintaining continuity and thereby improving the accuracy of pore identification. Attached Figure Description

[0014] Figure 1 This is an application environment diagram of a deep learning-based rock strata image matrix material identification method in one embodiment;

[0015] Figure 2 This is a flowchart of a deep learning-based method for identifying the base material of rock strata images in one embodiment;

[0016] Figure 3 This is a schematic diagram of the network structure of a base substance identification model in one embodiment;

[0017] Figure 4This is a schematic diagram of the network structure of the attention layer in one embodiment;

[0018] Figure 5 This is a schematic diagram of the network structure of a feature enhancement network in one embodiment;

[0019] Figure 6 This is a schematic diagram of a training base material recognition model in one embodiment;

[0020] Figure 7 This is a schematic diagram of the network structure for training a substrate recognition model in one embodiment;

[0021] Figure 8 This is a schematic diagram of a rock stratum image base material identification device based on deep learning in one embodiment;

[0022] Figure 9 This is a schematic diagram of a computing device in one embodiment. Detailed Implementation

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0025] In the following description, the expression “some embodiments” refers to a subset of all possible embodiments. However, it should be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0026] See Figure 1This diagram illustrates the application environment of a deep learning-based method for identifying the matrix material of rock strata images in one embodiment. The deep learning-based method is applied to a computing device 10, which can acquire projection data of rock strata under various energy channels. The projection data is acquired as follows: X-ray tube voltages are set to 45, 55, and 65 kVp, with corresponding tube current parameters. Scanning is performed sequentially under different tube voltage settings, and X-ray multispectral attenuation data is acquired using a sequential scanning method. The acquired data is preprocessed, and a blind source separation algorithm based on non-negative matrix factorization (NMF) is used to solve the preprocessed data. By setting appropriate iteration counts and convergence thresholds, the separation effect is optimized to obtain narrow-spectrum projections, i.e., projection data of rock strata under multiple energy channels, such as [35,45) KeV, [45,55) KeV, and [55,65) KeV. A reconstruction algorithm is then used to reconstruct images of the rock strata under each energy channel after blind separation, generating images of the rock strata under different energy channels. During the reconstruction process, the reconstruction algorithm parameters are adjusted according to the requirements of the imaging target to improve the quality and detail resolution of the decomposed image and reduce subsequent experimental errors. The computing device 10 uses a trained matrix material recognition model to identify various matrix materials in the rock strata image to be tested. These matrix materials include, but are not limited to, pores and various mineral components, such as CaCO3, SiO2, etc. The deep learning-based matrix material recognition method for rock strata images provided in this application can be applied to various rock strata environments, such as coal seams.

[0027] Please see Figure 2 This is a flowchart illustrating a deep learning-based method for identifying the matrix material of rock strata images, provided in an embodiment of this application. The deep learning-based method for identifying the matrix material of rock strata images is applied in a computing device and includes the following steps:

[0028] S11. Obtain the image of the rock layer to be tested, which includes the rock layer image under each energy channel.

[0029] In this embodiment, the step of reconstructing the projection data under each energy channel to obtain the image of the rock layer to be measured can be performed in a computing device or in other devices, and can be directly imported into the computing device.

[0030] S12. Based on the image of the rock strata to be tested, form the input image of the rock strata to the trained base material recognition model.

[0031] In this embodiment, the base material identification model is trained based on the training dataset, such as... Figure 3As shown, the matrix material identification model includes an encoder, a decoder, and a matrix material prediction network. The encoder includes a feature extraction network and a feature enhancement network. The feature extraction network includes encoding layers at different scales, and each decoding layer in the decoder includes an attention layer and an upsampling layer. The encoder is used to extract features at different scales from the image of the rock strata under test through the encoding layers at different scales. The feature enhancement network can be a depth-separable, dilated convolutional pyramid structure, used to expand the receptive field of the matrix material features extracted by the encoding layers without losing spatial resolution or increasing the computational cost of the convolution kernel. This is crucial for obtaining the global distribution, connectivity, and differentiation of true and false pores of different sizes. In this embodiment, the image of the rock strata under test can be used as the input image, or the processed image of the rock strata under test can be used as the input image.

[0032] In this embodiment, the feature extraction network uses the Xception module for feature extraction. The Xception module mainly includes: a regular convolutional layer, a separable convolutional layer, a batch normalization (BN) layer, a ReLU activation function, and a max pooling layer. The feature map output by the Xception feature extraction network is introduced into the feature enhancement network to enhance the semantic information of the feature map.

[0033] S13. Based on the input image of the rock strata, obtain the multi-attention feature map of the rock strata output by the attention layer.

[0034] In this embodiment, by extracting encoded features at different directional positions through an attention layer, the expressive power of the position-sensitive features of the base material can be improved. For example... Figure 3 As shown, the feature data output by the encoding layer at the same scale is used as the input to the attention layer at the same scale.

[0035] S14. Based on the input rock layer image, obtain the upsampled rock layer features output by the upsampled layer.

[0036] In this embodiment, as Figure 3 As shown, the output of the previous decoding layer of the upsampling layer is used as the input of the upsampling layer. For example, the input of the second-scale upsampling layer is the output of the third-scale decoding layer. The upsampling layer outputs the upsampled rock strata features at each scale.

[0037] S15. The multi-attention feature map of the rock strata and the upsampled rock strata features are fused to obtain the rock strata fusion features corresponding to the decoding layer.

[0038] In this embodiment, after obtaining the multi-attention feature map and upsampled strata features of the same scale layer, the two are fused to obtain the fused strata feature. The multi-attention feature map can provide precise location information of the matrix material, while the upsampled strata features can provide high-level semantic understanding. By fusing the two, attention features can be enhanced in densely porous areas, thus preserving details, while upsampled features can be enhanced in homogeneous strata areas, maintaining continuity.

[0039] S16. Based on the rock strata fusion features corresponding to the decoding layer, the matrix material prediction network is used to predict the matrix material data in the rock strata image to be tested.

[0040] In this embodiment, when the decoding layer is not the last decoding layer connected to the matrix material prediction network, for example, when it is not the first-scale decoding layer, the rock strata fusion feature corresponding to the decoding layer is used as the input of the next upsampling layer until the rock strata fusion feature output by the last decoding layer is obtained. This is then used as the input of the matrix material prediction network, and the matrix material data in the rock strata image is output. The matrix material data includes, but is not limited to, porosity data, mineral composition data, etc.

[0041] In the above embodiments, by configuring an attention layer in the decoder, a multi-attention feature map of the rock strata is obtained, which can provide precise location information of the base material; an upsampling layer in the decoder obtains upsampled rock strata features, which can provide high-level semantic understanding. By fusing the two, attention features can be enhanced in densely porous areas, thus preserving details, while upsampled features can be enhanced in homogeneous rock strata areas, maintaining continuity, thereby improving the accuracy of pore identification.

[0042] In some embodiments, the base material identification model includes an encoder, the encoder including multiple coding layers, the coding layers being used to output a rock layer feature map, the rock layer feature map including feature maps under each energy channel, the attention layer including a channel attention layer and a coordinate attention layer, and obtaining the rock layer multi-attention feature map output by the attention layer based on the rock layer input image includes:

[0043] Obtain the rock strata feature map belonging to the same scale layer as the attention layer, and use it as the input of the attention layer. Then, output the channel attention features of each energy channel through the channel attention layer.

[0044] The channel attention features under the energy channel are used as the input of the coordinate attention layer. The horizontal pooling features and vertical pooling features are output by the pooling layer in the coordinate attention layer. The horizontal pooling features and vertical pooling features under the energy channel are spliced ​​together to obtain the rock strata pooling splicing features under the energy channel.

[0045] Based on the rock strata pooling and splicing features, the rock strata multi-attention feature map under the energy channel is output through the dynamic convolution kernel in the coordinate attention layer. Based on the rock strata multi-attention feature map under each energy channel, the rock strata multi-attention feature map is obtained.

[0046] In this embodiment, introducing coordinate channel attention can provide more accurate location encoding of pore features. Building upon the features of both channel attention and coordinate attention mechanisms, a dynamic kernel prediction module is added. Compared to traditional attention modules, this module captures more local and global feature information across channels, resulting in a wider and more precise feature field of view. This allows the model to more accurately locate and identify targets with minute structures. Figure 4 As shown, Figure 4 This is a schematic diagram of the network structure of the attention layer in one embodiment. The attention layer includes a channel attention layer and a coordinate attention layer. Channel attention features are input to the horizontal average pooling layer and the vertical average pooling layer, respectively. The horizontal average pooling layer outputs horizontal pooling features, and the vertical average pooling layer outputs vertical pooling features. The dynamic convolution kernel generator generates adaptive convolution kernels based on the rock strata pooling and stitching features. This results in larger convolution kernels for large pores, capturing the overall morphology, and thinner convolution kernels for micro-cracks, enhancing linear features. Figure 4 As can be seen, the rock layer feature maps are output to the maximum pooling layer and the average pooling layer for processing. The outputs of the maximum pooling layer and the average pooling layer are input to the shared MLP. The data output by the shared MLP is then input to the first activation layer and the second activation layer for processing. The outputs of the first activation layer and the second activation layer are concatenated to obtain the channel attention features.

[0047] Optionally, the encoder includes a feature extraction network and a feature enhancement network, and the step of acquiring the rock strata feature map belonging to the same scale layer as the attention layer includes:

[0048] Based on the input image of the rock strata, the feature extraction network outputs rock strata feature maps corresponding to each coding layer, and the rock strata feature maps include feature maps under each energy channel;

[0049] The rock layer feature map output by the coding layer connected to the feature enhancement network is used as the input of the feature enhancement network, and the rock layer enhanced features are output through the feature enhancement network.

[0050] For example, such as Figure 3 As shown, three encoding layers of different scales are configured. The first-scale encoding layer, the second-scale encoding layer, and the third-scale encoding layer each output their corresponding rock layer feature maps. For example, if a rock layer image with three energy channels is input, the rock layer feature map will also include feature maps with all three energy channels.

[0051] Optionally, the step of outputting rock layer enhancement features through the feature enhancement network includes:

[0052] The inputs of each energy channel in the feature enhancement network are convolved using dilated convolution with different dilation ratios to obtain dilated convolution features for each energy channel.

[0053] Based on the void convolution features under each energy channel, point convolution processing is performed to obtain the rock layer enhancement features.

[0054] In this embodiment, the feature enhancement network depth includes a separable dilated convolution pyramid composed of two parts: depthwise atrous convolution and pointwise convolution, as shown in the diagram below. Figure 5 As shown, dilated convolution expands the receptive field of feature extraction without sacrificing spatial resolution or increasing the computational cost of the convolution kernel. This is crucial for obtaining the global distribution, connectivity, and differentiation of true and false pores of different sizes. Adding this module improves the extraction of micropore features while maintaining the model's computational efficiency. This module performs dilated convolution on the rock layer feature map under each input energy channel using different dilation rates. The convolution results are then concatenated and followed by a 1×1 convolution to obtain the final feature map. Point convolution is a 1×1 convolution, unifying the pore representation under different energy channels, while simultaneously performing channel fusion on the feature map output by dilated convolution. Different dilation rates are used for the rock layer feature maps under different energy channels to calculate the dilated convolution features for each energy channel. This ensures that micropores correspond to small dilation rates, preserving more details of micropores and preventing feature omissions due to uniform small dilation rates; and that macropores correspond to large dilation rates, achieving global perception and avoiding computational waste due to uniform large dilation rates.

[0055] In some embodiments, Figure 6 This is a schematic diagram of training a base material recognition model in one embodiment; the method further includes:

[0056] S61. Obtain the actual sample dataset.

[0057] In this embodiment, each actual sample in the actual sample dataset includes a sample rock layer image corresponding to the sample rock layer. The sample rock layer image includes sample images under each energy channel, and the sample images are reconstructed based on the projection data of the corresponding energy channel. The actual samples are reconstructed based on projection data collected from actual rock layers.

[0058] The CT X-ray tube voltages were configured with three different energies, such as 45, 55, and 65 kVp, and corresponding tube current parameters. Scans were performed sequentially at these different voltage settings to acquire X-ray multispectral attenuation data of the actual samples. The acquired data was preprocessed, and a blind source separation algorithm based on non-negative matrix factorization (NMF) was used to solve multiple sets of preprocessed data. By setting appropriate iteration counts and convergence thresholds, the separation effect was optimized to obtain narrow-spectrum projections, i.e., projection data of the actual rock strata in the [35,45) KeV, [45,55) KeV, and [55,65) KeV energy channels. Reconstruction algorithms were then used to reconstruct images of the projection data in each energy channel after blind separation, generating sample images for each energy channel. During reconstruction, the reconstruction algorithm parameters were adjusted according to the requirements of the imaging target to improve the quality and detail resolution of the decomposed images and reduce subsequent experimental errors.

[0059] S62. Obtain the simulation sample dataset.

[0060] In this embodiment, the simulation sample dataset is generated based on a three-dimensional physical model representing the relationship between rock strata and matrix material. Each simulation sample in the simulation sample dataset includes a simulated rock strata image and matrix material label data in the simulated rock strata image. The matrix material label data includes, but is not limited to, pore label data, various mineral label data, etc.

[0061] Determining the composition of rock strata involves using X-ray diffraction to measure the mineral composition (e.g., CaCO3, SiO2) of rock samples. The presence and absence of minerals are then separated to determine the location of pores within the rock strata. A three-dimensional physical model is then constructed based on the rock strata's composition and porosity to establish the relationship between the material attenuation coefficient and energy channels. The rock strata can be coal seams, shale seams, etc. For example, a simulation sample dataset can be obtained using simulation software based on the three-dimensional physical model.

[0062] S63. In the first round of iteration, a base material identification model is trained based on the simulation sample dataset to obtain a pre-trained base material identification model.

[0063] In this embodiment, training is first performed based on data labeled with matrix materials. This allows the matrix material identification model to learn the rock strata features corresponding to the matrix material labels, thereby accelerating the convergence of the model. During the first iteration, the model is iterated a preset number of times before ending. In this first iteration, the matrix material prediction loss value is calculated for each iteration. Based on this loss value, the gradient for parameter adjustment is calculated via backpropagation, and the parameters in the matrix material identification model are adjusted according to this gradient.

[0064] S64. In the second round of iteration, based on the actual sample dataset and the simulation sample dataset, calculate the total loss value for each iteration in the second round of iteration, and train the base material recognition model based on the total loss value.

[0065] In this embodiment, as Figure 7 As shown, the network structures of the encoder and decoder are similar to... Figure 3 The network structures shown are identical, except one is a pre-trained network and the other is an untrained network. The training network for the base material recognition model includes the encoder, the decoder connected to the encoder, an adversarial training network, and a total loss calculation network. The adversarial training network is used to learn the differences in base material features between the actual sample dataset and the simulated sample dataset through adversarial training. The adversarial training network includes a domain adversarial discriminator. The encoder includes a feature extraction network and a feature enhancement network. The feature extraction network outputs a rock layer sample feature map corresponding to the input sample image, and the feature enhancement network outputs enhanced rock layer sample features corresponding to the input sample image based on the rock layer sample feature map.

[0066] During network training, the actual sample dataset is an unlabeled sample set. To reduce the deviation between real and simulated data, which leads to decreased cross-domain performance and low accuracy, an adversarial training network (e.g., a feature-level adaptive domain adversarial discriminator (DANN)) is added after the feature enhancement network output. This optimizes the network's task loss (mineral classification, pore extraction) and domain adversarial loss in parallel. Simultaneously, data augmentation (image rotation, Gaussian noise addition) is applied to both the simulated and actual sample datasets, achieving domain invariance learning of the feature space between the simulated and actual sample datasets of the rock strata. First, the domain discriminator is frozen, and the network is pre-trained on the simulated sample dataset, i.e., the first iteration is performed. Then, the actual sample dataset is added for adversarial fine-tuning. This involves fixing the feature extractor and training the domain adversarial discriminator. The domain adversarial discriminator is fixed, and a gradient reversal layer (GRL) is used to multiply the gradient by a negative coefficient during backpropagation, making the optimization objectives of the feature extractor and the domain adversarial discriminator opposite. This optimizes the feature extractor and dynamically adjusts the parameters.

[0067] In this embodiment, during the second iteration, various losses are calculated, including but not limited to discriminant loss and base material prediction loss. In this embodiment, discriminant loss represents the core supervisory signal driving the domain discriminator in domain adversarial training to distinguish whether features originate from the actual sample dataset or the simulated sample dataset; its essence is binary cross-entropy loss. The role of discriminant loss is to: optimize the domain discriminator: by minimizing the loss, it forces the discriminator to accurately identify whether features come from the simulated sample dataset (e.g., laboratory simulated rock formations) or the actual sample dataset (e.g., real downhole data); guide the feature extractor: by inverting the loss gradient through a gradient inversion layer (GRL) and backpropagating it, it drives the feature extractor to generate domain-invariant features (e.g., removing equipment noise and retaining consistent pore structures across domains), thereby reducing the distribution difference between simulated and real data and improving the model's generalization ability in real-world scenarios. Base material prediction loss is the core supervisory signal in dual-energy CT rock formation analysis; its role is to constrain the neural network to accurately decompose the material components (e.g., minerals, pore fluids) in CT images. The matrix material prediction loss calculates the pixel-level difference between the predicted matrix material density map (such as calcite, quartz, and porosity) and the actual material distribution (commonly using L1 / L2 loss). This drives the model to learn the energy absorption characteristics of materials to X-rays, thereby accurately quantifying the mineral composition and porosity of the rock formation and providing key physical property parameters for reservoir evaluation.

[0068] Optionally, in the second round of iterations, calculating the total loss value for each iteration based on the actual sample dataset and the simulated sample dataset includes:

[0069] The feature maps of the rock strata samples are used as inputs to the decoder and the adversarial training network, respectively.

[0070] Based on the rock stratum sample feature map, the discrimination result corresponding to the rock stratum sample feature map is output through the domain adversarial discriminator, and the discrimination loss under the current iteration is calculated based on the discrimination result;

[0071] Based on the feature map of the rock strata sample, the decoder outputs the prediction result of the base material corresponding to the feature map of the rock strata sample. Based on the prediction result of the base material and the base material label data corresponding to the feature map of the rock strata sample, the prediction loss value of the base material under the current iteration is calculated.

[0072] Based on the discrimination loss and the predicted loss value of the base material, the total loss value under the current iteration is calculated through the total loss calculation network.

[0073] Optionally, the discrimination loss includes base material sample domain loss and base material feature confusion loss. The step of outputting the discrimination result corresponding to the rock layer sample feature map through the domain adversarial discriminator based on the rock layer sample feature map, and calculating the discrimination loss for the current iteration based on the discrimination result, includes:

[0074] Based on the rock stratum sample feature map, the domain adversarial discriminator outputs a first probability that the rock stratum sample feature map belongs to the simulated sample dataset and a second probability that the rock stratum sample feature map belongs to the actual sample dataset.

[0075] Based on the first probability and the second probability, the sample domain loss of the base material is calculated;

[0076] Based on the second probability, the feature confusion loss of the base material is calculated;

[0077] The discrimination loss is obtained by weighting the sample domain loss of the base material and the feature confusion loss of the base material.

[0078] In this embodiment, the base material prediction results include decomposed images of various base materials, such as pore images, CaCO3 images, and SiO2 images. The total loss value L... TOTAL =λ1×L TASK +λ2×L DANN Where λ1 and λ2 are weight parameters, L TASK Predicted loss value of base material, L DANN This indicates the loss to be identified. The formula for calculating the predicted loss value of the base material is as follows:

[0079] L TASK =L MSE +αL CE

[0080]

[0081] Where L MSE T represents the loss of the base material classification. k Let T′ represent the predicted category for the k-th decomposed image. k This represents the base material label data corresponding to the k-th decomposed image, where K represents the total number of decomposed images. α represents the coefficient. LCE represents the pixel loss, calculated as follows:

[0082]

[0083] T k,j Let J represent the pixel value of the j-th pixel in the k-th decomposed image of the base material label data, and let T′ represent the total number of pixels in the k-th decomposed image. k,j This represents the pixel value of the j-th pixel in the predicted k-th decomposed image.

[0084] The formula for calculating the loss is as follows:

[0085] L DANN =L disc +βL adv

[0086] L disc L represents the sample domain loss of the base material. adv β represents the feature confusion loss of the base material, and β represents the adversarial loss weight. The formula for calculating the base material sample domain loss is as follows:

[0087]

[0088] Where, n s n represents the number of input simulation samples in the simulation sample dataset. T This represents the total number of actual samples input into the actual sample dataset. This represents the i-th sample in the input simulation samples. The i-th sample in the actual input sample. Indicates discrimination The probability that it belongs to the actual sample. Indicates discrimination The probability that it belongs to the actual sample.

[0089] In some embodiments, after predicting the matrix material data in the image of the rock strata to be tested, i.e., each matrix material image, the topological analysis, porosity calculation, and pore size analysis of the pore image are of great significance to research and practice in fields such as energy, materials, and geology, and play a key role in many practical scenarios. Connectivity analysis is performed on the decomposed pores, converting the resulting pore image into a binary image, where pore regions are assigned a value of 1 and matrix regions are assigned a value of 0. A connected component labeling algorithm is used to label the connected components of the pore regions in the binary image.

[0090] Porosity Φ is an important indicator for measuring the degree of porosity development in coal and rock, defined as the ratio of pore volume to the total volume of coal and rock. In two-dimensional image analysis, porosity can be calculated through pixel statistics, using the following formula:

[0091]

[0092] Where, N p N represents the number of pixels in the pore region. total This represents the total number of pixels in the image.

[0093] One or more embodiments of this application have the following beneficial effects:

[0094] (1) This application uses conventional microfocus CT to acquire multi-energy projection data. Through multiple tube voltages, X-ray multi-spectral CT imaging based on multi-energy projection blind separation can calculate and separate multi-energy projections in blind scenarios where the X-ray energy spectrum is unknown, thereby obtaining narrow energy spectrum projections of multiple energy channels under different voltages.

[0095] (2) This application utilizes a depthwise separable dilated convolutional pyramid network model to decompose the components of multi-energy coal and rock CT images. Advanced deep learning methods, such as depthwise separable dilated convolution and an improved attention mechanism, are employed to extract image features, expand the receptive field, and obtain a wider and more accurate feature field, thereby improving the precision and accuracy of the decomposition of the base material and micropores. Simultaneously, a feature-level domain adversarial discriminator and adversarial loss are introduced during data training and testing, enabling the model to continuously migrate and optimize to better fit real-world data and generalize to actual datasets.

[0096] (3) Perform topological analysis and calculate porosity of the extracted pores to realize the research and practical application of pores in coal and rock in the fields of energy, materials, and geology.

[0097] (4) This application can realize the leap from "fuzzy recognition" to "precise quantification" in the micro-characterization of coal and rock minerals and pores. This technology has significant academic value and industrial application potential in the field of energy spectrum CT material decomposition, and is used in the study of some complex porous media.

[0098] In another aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based method for identifying the matrix material of rock strata images as described in any embodiment of this application.

[0099] In the computer program product, the optional implementation form of the program module architecture of the computer program that implements each step of the target recognition method can be a base material recognition device based on deep learning rock strata images.

[0100] Please see Figure 8 This application provides a deep learning-based matrix material identification device for rock strata images, comprising: an acquisition module 80 for acquiring a rock strata image to be tested, the rock strata image to be tested including rock strata images under various energy channels; a prediction module 81 for forming a rock strata input image to a trained matrix material identification model based on the rock strata image to be tested, wherein the matrix material identification model includes a decoder and a matrix material prediction network, each decoding layer in the decoder including an attention layer and an upsampling layer; the prediction module 81 is further configured to acquire a multi-attention feature map of the rock strata output by the attention layer based on the rock strata input image; the prediction module 81 is further configured to acquire upsampled rock strata features output by the upsampling layer based on the rock strata input image; the prediction module 81 is further configured to fuse the multi-attention feature map of the rock strata and the upsampled rock strata features to obtain a rock strata fusion feature corresponding to the decoding layer; and the prediction module 81 is further configured to predict matrix material data in the rock strata image to be tested based on the rock strata fusion feature corresponding to the decoding layer through the matrix material prediction network.

[0101] Optionally, the matrix material identification model includes an encoder, which includes multiple coding layers. The coding layers are used to output a rock layer feature map, which includes feature maps under each energy channel. The attention layer includes a channel attention layer and a coordinate attention layer. The prediction module 81 is further used for:

[0102] Obtain the rock strata feature map belonging to the same scale layer as the attention layer, and use it as the input of the attention layer. Then, output the channel attention features of each energy channel through the channel attention layer.

[0103] The channel attention features under the energy channel are used as the input of the coordinate attention layer. The horizontal pooling features and vertical pooling features are output by the pooling layer in the coordinate attention layer. The horizontal pooling features and vertical pooling features under the energy channel are spliced ​​together to obtain the rock strata pooling splicing features under the energy channel.

[0104] Based on the rock strata pooling and splicing features, the rock strata multi-attention feature map under the energy channel is output through the dynamic convolution kernel in the coordinate attention layer. Based on the rock strata multi-attention feature map under each energy channel, the rock strata multi-attention feature map is obtained.

[0105] Optionally, the encoder includes a feature extraction network and a feature enhancement network, and the prediction module 81 is further used for:

[0106] Based on the input image of the rock strata, the feature extraction network outputs rock strata feature maps corresponding to each coding layer, and the rock strata feature maps include feature maps under each energy channel;

[0107] The rock layer feature map output by the coding layer connected to the feature enhancement network is used as the input of the feature enhancement network, and the rock layer enhanced features are output through the feature enhancement network.

[0108] Optionally, the prediction module 81 is also used for:

[0109] The inputs of each energy channel in the feature enhancement network are convolved using dilated convolution with different dilation ratios to obtain dilated convolution features for each energy channel.

[0110] Based on the void convolution features under each energy channel, point convolution processing is performed to obtain the rock layer enhancement features.

[0111] Optionally, training module 82 is also used for:

[0112] Obtain an actual sample dataset, wherein each actual sample in the actual sample dataset includes a sample rock layer image corresponding to the sample rock layer, wherein the sample rock layer image includes sample images under each energy channel, and the sample images are reconstructed based on the projection data of the corresponding energy channel;

[0113] A simulation sample dataset is obtained, which is generated based on a three-dimensional physical model representing the relationship between rock strata and matrix material. Each simulation sample in the simulation sample dataset includes a simulated rock strata image and matrix material label data in the simulated rock strata image.

[0114] In the first round of iteration, the matrix material identification model is trained based on the simulation sample dataset to obtain a pre-trained matrix material identification model;

[0115] In the second round of iterations, based on the pre-trained base substance recognition model, and using the actual sample dataset and the simulation sample dataset, the total loss value for each iteration in the second round of iterations is calculated, and the base substance recognition model is trained again based on the total loss value.

[0116] Optionally, the training network of the matrix material recognition model includes the encoder, the decoder and adversarial training network connected to the encoder, and the total loss calculation network. The adversarial training network is used to learn the differences in matrix material features between the actual sample dataset and the simulated sample dataset through adversarial training. The adversarial training network includes a domain adversarial discriminator. The encoder includes a feature extraction network and a feature enhancement network. The feature extraction network is used to output a rock layer sample feature map corresponding to the input sample image. The feature enhancement network is used to output enhanced rock layer sample features corresponding to the input sample image based on the rock layer sample feature map.

[0117] Optionally, training module 82 is also used for:

[0118] The feature maps of the rock strata samples are used as inputs to the decoder and the adversarial training network, respectively.

[0119] Based on the rock stratum sample feature map, the discrimination result corresponding to the rock stratum sample feature map is output through the domain adversarial discriminator, and the discrimination loss under the current iteration is calculated based on the discrimination result;

[0120] Based on the feature map of the rock strata sample, the decoder outputs the prediction result of the base material corresponding to the feature map of the rock strata sample. Based on the prediction result of the base material and the base material label data corresponding to the feature map of the rock strata sample, the prediction loss value of the base material under the current iteration is calculated.

[0121] Based on the discrimination loss and the predicted loss value of the base material, the total loss value under the current iteration is calculated through the total loss calculation network.

[0122] Optionally, the discrimination loss includes the base material sample domain loss and the base material feature confusion loss, and the training module 82 is further used for:

[0123] Based on the rock stratum sample feature map, the domain adversarial discriminator outputs a first probability that the rock stratum sample feature map belongs to the simulated sample dataset and a second probability that the rock stratum sample feature map belongs to the actual sample dataset.

[0124] Based on the first probability and the second probability, the sample domain loss of the base material is calculated;

[0125] Based on the second probability, the feature confusion loss of the base material is calculated;

[0126] The discrimination loss is obtained by weighting the sample domain loss of the base material and the feature confusion loss of the base material.

[0127] Please see Figure 9 In another aspect of this application, a computing device 10 is also provided, including a memory 3011 and a processor 3012. The memory 3011 stores a computer program, and when the computer program is executed by the processor, the processor 3012 performs the steps of the deep learning-based rock strata image matrix identification method provided in any of the above embodiments of this application. The computing device 10 can be (e.g., a desktop computer, laptop computer, tablet computer, handheld computer, smart speaker, server, etc.), a mobile phone (e.g., a smartphone, cordless phone, etc.), a wearable device (e.g., a pair of smart glasses or a smartwatch), or a similar device.

[0128] The processor 3012 is the control center, connecting various parts of the computer device via various interfaces and lines. It executes software programs and / or modules stored in the memory 3011, and calls data stored in the memory 3011 to perform various functions and process data. Optionally, the processor 3012 may include one or more processing cores; preferably, the processor 3012 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user page, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 3012.

[0129] The memory 3011 can be used to store software programs and modules. The processor 3012 executes various functional applications and data processing by running the software programs and modules stored in the memory 3011. The memory 3011 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 3011 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 3011 may also include a memory controller to provide the processor 3012 with access to the memory 3011.

[0130] In another aspect, this application also provides a storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the deep learning-based rock strata image matrix identification method provided in any of the above embodiments of this application.

[0131] Those skilled in the art will understand that all or part of the processes in the methods provided in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0132] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying the matrix material of rock strata images based on deep learning, characterized in that, The method comprises: obtaining a to-be-detected rock stratum image, the to-be-detected rock stratum image comprising rock stratum images under respective energy channels; forming a rock stratum input image to a trained base substance recognition model based on the to-be-detected rock stratum image, wherein the base substance recognition model comprises a decoder and a base substance prediction network, each decoding layer in the decoder comprising an attention layer and an up-sampling layer; obtaining a rock stratum multi-attention feature map output by the attention layer based on the rock stratum input image; obtaining an up-sampled rock stratum feature output by the up-sampling layer based on the rock stratum input image; fusing the rock stratum multi-attention feature map and the up-sampled rock stratum feature to obtain a rock stratum fusion feature corresponding to the decoding layer; predicting base substance data in the to-be-detected rock stratum image through the base substance prediction network based on the rock stratum fusion feature corresponding to the decoding layer, the base substance recognition model comprising an encoder, the encoder comprising a plurality of encoding layers, the encoding layers being configured to output a rock stratum feature map, the rock stratum feature map comprising feature maps under respective energy channels, the attention layer comprising a channel attention layer and a coordinate attention layer, and the obtaining of the rock stratum multi-attention feature map output by the attention layer based on the rock stratum input image comprising: obtaining a rock stratum feature map belonging to a same scale layer as the attention layer as input of the attention layer, outputting channel attention features under respective energy channels through the channel attention layer; taking the channel attention features under the energy channels as input of the coordinate attention layer, respectively outputting horizontal pooling features and vertical pooling features through a pooling layer in the coordinate attention layer, splicing the horizontal pooling features and the vertical pooling features under the energy channels to obtain rock stratum pooling splicing features under the energy channels; outputting the rock stratum multi-attention feature map under the energy channels through a dynamic convolution kernel in the coordinate attention layer based on the rock stratum pooling splicing features, and obtaining the rock stratum multi-attention feature map based on the rock stratum multi-attention feature map under each of the energy channels. 2.The method of claim 1, wherein, The encoder comprises a feature extraction network and a feature enhancement network, and the obtaining of the rock stratum feature map belonging to the same scale layer as the attention layer comprises: outputting rock stratum feature maps corresponding to respective encoding layers through the feature extraction network based on the rock stratum input image, the rock stratum feature maps comprising feature maps under respective energy channels; taking the rock stratum feature maps output by the encoding layers connected with the feature enhancement network as input of the feature enhancement network, and outputting rock stratum enhanced features through the feature enhancement network. 3.The method of claim 2, wherein, The outputting of the rock stratum enhanced features through the feature enhancement network comprises: performing convolution calculation on inputs under respective energy channels in the feature enhancement network through different hole rates of hole convolution to obtain hole convolution features under the respective energy channels; performing point convolution processing based on the hole convolution features under the respective energy channels to obtain the rock stratum enhanced features. 4.The method of claim 1, wherein, The method further comprises: obtain an actual sample dataset, each actual sample in the actual sample dataset comprising a sample rock layer corresponding sample rock layer image, wherein the sample rock layer image comprises sample images under respective energy channels, the sample images being reconstructed based on projection data of the corresponding energy channels; obtain a simulation sample dataset, the simulation sample dataset being generated based on a three-dimensional physical model representing a rock layer and a base material, each simulation sample in the simulation sample dataset comprising a simulation rock layer image and base material label data in the simulation rock layer image; in a first iteration process, train the base material recognition model based on the simulation sample dataset to obtain a pre-trained base material recognition model; in a second iteration process, based on the actual sample dataset and the simulation sample dataset, calculate a total loss value of each iteration in the second iteration process, and continue to train the base material recognition model based on the total loss value. 5.The method of claim 4, wherein, The training network of the base material recognition model comprises the decoder, the adversarial training network, and the total loss calculation network connected with the encoder respectively, wherein the adversarial training network is used to learn the difference of base material features between the actual sample dataset and the simulation sample dataset through adversarial training, the adversarial training network comprises a domain adversarial discriminator, the encoder comprises a feature extraction network and a feature enhancement network, the feature extraction network is used to output a rock layer sample feature map corresponding to an input sample image, and the feature enhancement network is used to output a rock layer enhanced sample feature corresponding to the input sample image based on the rock layer sample feature map. 6.The method of claim 5, wherein, The total loss value of each iteration in the second iteration process is calculated based on the actual sample dataset and the simulation sample dataset in the second iteration process, which comprises: inputting the rock layer sample feature map into the decoder and the adversarial training network respectively; outputting a discrimination result corresponding to the rock layer sample feature map through the domain adversarial discriminator based on the rock layer sample feature map, and calculating a discrimination loss under the current iteration based on the discrimination result; outputting a base material prediction result corresponding to the rock layer sample feature map through the decoder based on the rock layer sample feature map, calculating a base material prediction loss value under the current iteration based on the base material prediction result and base material label data corresponding to the rock layer sample feature map; calculating a total loss value under the current iteration through the total loss calculation network based on the discrimination loss and the base material prediction loss value. 7.The method of claim 6, wherein, The discrimination loss comprises a base material sample domain loss and a base material feature confusion loss, and the discrimination loss under the current iteration is calculated based on the discrimination result, which comprises: outputting a first probability that the rock layer sample feature map belongs to the simulation sample dataset and a second probability that the rock layer sample feature map belongs to the actual sample dataset through the domain adversarial discriminator based on the rock layer sample feature map; calculating the base material sample domain loss based on the first probability and the second probability; calculating the base material feature confusion loss based on the second probability; weighting the base material sample domain loss and the base material feature confusion loss to obtain the discrimination loss.

8. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the base material identification method for rock stratum images based on deep learning according to any one of claims 1 to 7.

9. A computing device, comprising: A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute the base material identification method for rock stratum images based on deep learning according to any one of claims 1 to 7.

Images