Weakly cemented porous medium CT image crack identification method

Abstract

This invention belongs to the interdisciplinary fields of digital core analysis, computer vision, and geological engineering. Specifically, it relates to a method for crack identification in CT images of weakly cemented porous media. The method includes: acquiring and preprocessing the original dataset of three-dimensional CT images of weakly cemented porous media samples; automatically generating crack label maps from the images through data mining and feature analysis and pairing them with the original images to form a training set; constructing and training a deep learning recognition model based on alternating row-column scanning; and applying the trained model to crack segmentation in actual CT images of weakly cemented porous media. This invention automatically extracts labels through data mining and feature analysis, and uses a deep learning recognition model based on alternating row-column scanning for crack identification. The high-precision identification results can be directly used for subsequent quantitative statistics of crack set parameters and seepage-mechanical coupling numerical simulations, providing crucial foundational data for engineering stability assessment and disaster prediction of weakly cemented porous media.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of digital core analysis, computer vision and geological engineering, and specifically relates to a method for identifying cracks in CT images of weakly cemented porous media. Background Technology

[0002] Weakly cemented porous media, such as loose sandstone and hydrate-bearing sediments, are widely present in geological engineering. The evolution of their internal fracture networks is a key factor controlling the macroscopic mechanical response and permeability behavior of the media. X-ray computed tomography (CT) technology can non-destructively acquire three-dimensional images of the internal structure of sediments, providing an important data foundation for the observation and analysis of fractures.

[0003] However, accurately identifying crack structures from CT images presents a significant challenge:

[0004] (1) Scarcity of labeled data: The accurate labeling of cracks relies heavily on manual drawing by geological experts, which is time-consuming, labor-intensive, and highly subjective, making it difficult to obtain large-scale, high-quality labeled data for supervised learning.

[0005] (2) The crack features are complex and the data is highly heterogeneous: the cracks in weakly cemented porous media are discontinuous, low-contrast, and slender structures with different shapes. The gray-scale distribution, noise level and structural features of CT images under different porous media types and different strain stages are significantly different. Traditional algorithms and general models are prone to breakage, misjudgment or over-smoothing, resulting in insufficient generalization ability when facing data under different working conditions.

[0006] Therefore, there is an urgent need for a recognition method that can automatically generate high-quality crack annotations and use these annotations to train a deep learning model specifically optimized for crack features, so as to achieve robust, accurate and automated extraction of cracks in CT images of weakly cemented porous media. Summary of the Invention

[0007] This invention overcomes the above-mentioned defects and provides a method for identifying cracks in CT images of weakly cemented porous media. Through data-driven feature mining and multi-dimensional feature fusion, it achieves automatic and high-precision crack identification.

[0008] To achieve the above objectives, the present invention provides a method for identifying cracks in CT images of weakly cemented porous media, comprising the following steps:

[0009] S1. Acquire three-dimensional CT image data of a weakly cemented porous medium sample, and preprocess the CT image data to extract the region of interest of the sample.

[0010] S2. Through data mining and feature analysis, crack features are automatically extracted from the preprocessed images and crack label maps are generated. These maps are then paired with the original CT image dataset to form a training dataset.

[0011] S3. Construct a deep learning recognition model based on alternating row-column scanning, and use the training dataset described in S2 to train and optimize the model to obtain a trained recognition model;

[0012] S4. Input the CT image to be identified into the trained recognition model for prediction, and perform post-processing on the prediction results to obtain the final crack recognition result.

[0013] Furthermore, the preprocessing involves automatically dividing the entire three-dimensional region of interest into multiple three-dimensional core units of fixed size.

[0014] Furthermore, in step S2, the specific steps for automatically generating crack label maps through data mining and feature analysis are as follows:

[0015] S21. An adaptive threshold segmentation method is used to extract all pore phases within the three-dimensional core unit, and a morphological watershed algorithm is used to separate connected pores to obtain independent pore individuals.

[0016] S22. Quantitatively characterize each individual pore, calculate multiple feature parameters, and construct a high-dimensional pore feature space;

[0017] S23. The high-dimensional pore feature space is reduced in dimensionality by non-negative matrix decomposition, and its low-dimensional components are retained as feature vectors.

[0018] S24. Use a Gaussian mixture model to perform probabilistic clustering on the dimensionality-reduced pore individuals, divide them into two classes, and identify one of the classes as crack pores.

[0019] S25. Reassemble the clustering recognition results according to their original coordinates to generate a complete binary crack label map, with crack areas marked as 1 and non-crack areas marked as 0.

[0020] S26. Pair the crack label map with the corresponding original CT image to form a training dataset.

[0021] Furthermore, in step S23, the rank of the nonnegative matrix factorization is determined based on a threshold of 80% of the cumulative explained variance, and the low-dimensional features obtained from the factorization are used for subsequent cluster analysis.

[0022] Furthermore, the optimization objective of nonnegative matrix factorization is to minimize the reconstruction error, and its calculation formula is as follows:

[0023] ,

[0024] in, The original feature matrix, and The decomposed nonnegative matrix;

[0025] The log-likelihood function of the Gaussian mixture model is:

[0026] ,

[0027] in, , The mixing coefficient, and The first The mean and covariance matrix of each Gaussian component are used to solve for the model parameters using the expectation-maximization algorithm, and cracks and pores are identified based on the posterior probability.

[0028] Furthermore, in step S3, the deep learning recognition model based on alternating row-column scanning is based on an encoder-decoder structure, and the alternating row-column scanning feature extraction module is embedded in the downsampling stage of the encoding path, the upsampling stage of the decoding path, and the bottleneck layer, respectively.

[0029] Furthermore, the row-column alternating scanning feature extraction module reassembles the two-dimensional feature map into a one-dimensional sequence along the row and column directions respectively, and uses depthwise separable convolution to perform forward-backward alternating scanning; the module also has a dynamic weight adjustment mechanism inside.

[0030] Furthermore, the internal processing steps of the row-column alternating scan feature extraction module are as follows:

[0031] S31. Use a channel expansion layer with an expansion factor of 2 to map the input feature map to a higher-dimensional hidden space;

[0032] S32, Alternating forward and reverse scanning in the row direction: Input feature map... Remodeling For the first row sequence The formulas for its forward and reverse scan operations are as follows:

[0033] ,

[0034] ,

[0035] in, These are the weight parameters of depthwise separable convolution;

[0036] S33, Alternating forward and reverse scanning in the column direction: After transposing the input feature map, the same strategy as the row direction is used to scan in the column direction;

[0037] S34. Fuse the features obtained from the row and column scans to obtain global context features. with local details ;

[0038] S35. Introduce a dynamic weight adjustment mechanism to adaptively fuse global context features and local detail features. The fusion formula is as follows:

[0039] ,

[0040] in, It is a spatially adaptive gated weighted graph generated by the Sigmoid function. This represents element-wise multiplication. This represents a matrix of all 1s; when the weight values ​​are close to 1, local features are mainly used to maintain boundary accuracy; when the weight values ​​are close to 0, global context features are mainly used to maintain crack continuity.

[0041] S36. Through channel reduction layers and convolutional layers, and residual connections with the module input, the gradient flow and feature information are fully preserved.

[0042] This application provides a crack identification system for CT images of weakly cemented porous media, comprising:

[0043] The data acquisition and preprocessing module is used to acquire raw 3D CT image data and perform preprocessing.

[0044] The feature mining and label generation module is used to automatically process three-dimensional core elements and generate crack labels and training datasets.

[0045] A deep learning model library and training platform for storing, training, and evaluating the deep learning recognition model based on alternating row-column scanning;

[0046] The crack identification and post-processing module is used to load the trained model for prediction and output the crack identification results.

[0047] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for identifying cracks in CT images of weakly cemented porous media.

[0048] Compared with the prior art, the advantages of the present invention are as follows:

[0049] This invention automatically generates a large number of high-quality crack labels from unlabeled CT images through data-driven feature mining and probabilistic clustering analysis, eliminating the reliance on time-consuming, expensive, and subjective manual annotation, solving the bottleneck of labeled data, and providing a reliable data foundation for supervised learning.

[0050] The row-column alternating scanning deep learning model designed in this invention effectively captures long-distance spatial context information of cracks through forward-backward alternating scanning, and introduces a dynamic weight adjustment mechanism to optimize the fusion of local details and global features. This makes the model particularly sensitive to discontinuous, low-contrast slender cracks, and the continuity and completeness of the recognition results are significantly better than those of general models. Moreover, it has good robustness and generalization ability to image differences caused by different experimental conditions.

[0051] The high-precision crack identification results of this invention can be directly used for subsequent quantitative statistics of crack set parameters and seepage-mechanical coupling numerical simulation, providing key basic data for stability assessment and disaster prediction of weakly cemented porous media engineering. Attached Figure Description

[0052] Figure 1 This is a flowchart of a method for identifying cracks in CT images of weakly cemented porous media according to the present invention.

[0053] Figure 2 This is a flowchart of the feature mining and crack label generation process of the present invention;

[0054] Figure 3 This is a schematic diagram of the deep learning recognition model architecture based on alternating row-column scanning according to the present invention.

[0055] Figure 4 This is a schematic diagram of the internal structure of the row-column alternating scanning feature extraction module of the present invention and the forward-reverse alternating scanning operation;

[0056] Figure 5 This is a comparison chart of crack identification results in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0058] like Figure 1 This invention provides a method for identifying cracks in CT images of weakly cemented porous media, specifically including:

[0059] I. Data Acquisition and Preprocessing Stage. Raw datasets of three-dimensional CT images of weakly cemented porous media, such as loose sandstone and hydrate-bearing sediments, were simultaneously acquired during different triaxial shear experiments. The raw CT images were preprocessed using the following steps: a) Extracting the region of interest (ROI) of the sample using an image segmentation tool; b) Automatically dividing the entire three-dimensional ROI into multiple three-dimensional core units of fixed size.

[0060] The region of interest extracted by this invention can eliminate interference from the background and experimental setup. Moreover, the divided three-dimensional core units represent regions with relatively uniform local structures, which helps to reduce the overall data discrepancies in subsequent feature mining.

[0061] II. Feature mining and crack label generation, such as Figure 2 As shown, the purpose is to automatically generate high-quality crack labels from unlabeled CT data.

[0062] Specifically, firstly, pore extraction and separation are performed. For each three-dimensional core unit, an adaptive threshold segmentation method is used to extract all pore phases within the unit, and a morphological watershed algorithm is applied to separate connected pore volumes, obtaining independent pore individuals. Next, multi-dimensional feature extraction is performed. Each independent pore individual is quantitatively characterized using multi-dimensional features, extracting feature parameters covering three major categories: morphology, spatial distribution, and gray-level statistics, constructing a high-dimensional pore feature space. Morphological parameters include, but are not limited to: volume, surface area, three-dimensional length, three-dimensional width, three-dimensional thickness, sphericity, anisotropy, elongation, and flattening; spatial distribution parameters include, but are not limited to: centroid coordinates, feature vector direction, and spatial extremum coordinates; gray-level statistical parameters include, but are not limited to: gray-level weighted center, gray-level weighted direction, and gray-level quality. Subsequently, feature dimensionality reduction and clustering recognition are performed on the high-dimensional feature data. a) After standardizing the feature data, non-negative matrix factorization is used for dimensionality reduction. Let the original feature matrix be... ( The number of pores (where is the original feature dimension), it is approximately decomposed into the product of two low-rank nonnegative matrices. ,in , , Since all elements are non-negative, the optimization objective of the decomposition is to minimize the reconstruction error:

[0063] ,

[0064] in, The original feature matrix, and denoted as the non-negative matrix after decomposition; b) Gaussian mixture model is used to perform probabilistic clustering on all pore individuals after dimensionality reduction. The Gaussian mixture model assumes that all samples are a mixture of K Gaussian distributions, and its log-likelihood function is:

[0065] ,

[0066] in, , The mixing coefficient, and The first The mean and covariance matrix of each Gaussian component are used to solve the model parameters using the expectation-maximization algorithm, and the fracture pores are identified based on the posterior probability; c) Label reconstruction and dataset construction are performed, and the clustering identification results of all three-dimensional core units are recombined according to their original coordinates to generate a complete binary fracture label map, in which fracture areas are labeled as 1 and non-fracture areas are labeled as 0; d) The label map is paired with the corresponding original CT image to form a training dataset for subsequent deep learning model training.

[0067] This invention employs nonnegative matrix factorization combined with Gaussian mixture models. The low-dimensional components obtained from nonnegative matrix factorization are all nonnegatively constrained, better reflecting actual pore characteristics. The decomposed feature components are highly interpretable, avoiding the shortcomings of methods like principal component analysis that generate abstract mixed positive and negative components, contain meaningless negative values, and easily sever the intrinsic connections of pore structures. Simultaneously, it effectively preserves subtle local features of fractures, preventing critical fracture information from being obscured. Gaussian mixture models, being probabilistic clustering, can accurately distinguish transitional pore structures, adapting to the irregular distribution of pores and fractures. This effectively solves the shortcomings of hard clustering methods like K-means, which easily misclassify or omit critical pore features, better reflecting the spatial separation patterns of pores and fractures in real rocks. Furthermore, this invention automatically generates a large number of high-quality fracture labels from unlabeled CT images through feature mining and cluster analysis of image data, completely eliminating the reliance on manual annotation and providing a reliable data foundation for subsequent deep learning model training.

[0068] III. Construction, training, and optimization of deep learning models based on alternating row-column scanning, such as Figure 3 As shown, this step constructs and trains a deep learning recognition model based on alternating row-column scanning. This model is based on the classic encoder-decoder structure and embeds a core alternating row-column scanning feature extraction module after each downsampling stage in its encoding path and before each upsampling stage in its decoding path.

[0069] Specifically, the encoder comprises four progressive downsampling stages, each sequentially performing max pooling, double convolutional blocks, and incorporating an alternating row-column scan module to capture long-range spatial dependencies; the bottleneck layer contains a separate alternating row-column scan module; the decoder comprises four upsampling stages, each employing bilinear interpolation to restore spatial resolution and fusing local detail information from the corresponding encoder layer via skip connections. The first three upsampling stages include an alternating row-column scan module to optimize multi-scale feature fusion, while the last upsampling stage uses standard convolution to maintain computational efficiency; the model output layer uses 1×1 convolutions and a sigmoid activation function to output a pixel-level crack probability map. Figure 4 As shown, the internal processing flow of the row-column alternating scan feature extraction module is as follows:

[0070] a) Use a channel expansion layer with an expansion factor of 2 to map the input feature map to a higher-dimensional hidden space;

[0071] b) Alternate scanning in the forward and reverse directions of the row to scan the input feature map. Remodeling That is, treating each row as an independent 1D sequence, for the ... row sequence Its bidirectional sequence scan operation is implemented using the following formula:

[0072] ,

[0073] ,

[0074] in, These are the weight parameters of depthwise separable convolution.

[0075] c) Perform alternating forward and reverse scanning in the column direction. After transposing the input feature map, process it using the same strategy as in the row direction, that is, reshape it into a column vector sequence and perform alternating forward and reverse scanning.

[0076] d) Perform feature fusion and dynamic weight adjustment, merging the feature representations obtained from row and column scans to obtain global context features. with local details A dynamic weight adjustment mechanism is introduced to adaptively fuse global context features and local detail features. The fusion formula is as follows:

[0077] ,

[0078] in, It is a spatially adaptive gated weighted graph generated by the Sigmoid function. This represents element-wise multiplication. It is an all-one matrix. When the weight values ​​are close to 1, the location mainly uses local features to maintain boundary accuracy; when the weight values ​​are close to 0, the global context features are mainly used to maintain crack continuity.

[0079] e) The fused features are processed through channel reduction layers and convolutional layers, and residual connections are made with the original features to ensure gradient flow and the integrity of feature information.

[0080] This invention introduces a row-column alternating scanning module into the encoder-decoder model. By scanning in both forward and reverse directions, it effectively captures the long-distance spatial context information of the crack. A dynamic weight adjustment mechanism is introduced to optimize the fusion of local details and global context features obtained by merging row and column features, thereby enhancing the model's sensitivity to slender cracks and making the identified cracks have better continuity and integrity.

[0081] IV. Crack Identification, Prediction, and Post-processing Stage: The CT image to be identified is input into the trained model for prediction. For large-format CT images whose size exceeds the model input size, an overlapping sliding window prediction strategy is used for processing, as detailed below:

[0082] Given input image size block size Overlapping step size The number of blocks in the horizontal and vertical directions are respectively , The formula is as follows:

[0083] ,

[0084] ,

[0085] For each pixel location in the image Its final predicted probability By covering all of that location Prediction probability of each image patch The weighted average is obtained, and the calculation formula is:

[0086] ,

[0087] in, This represents the fusion weight for the corresponding position.

[0088] The fused full-image probability map is binarized using a threshold to obtain preliminary crack identification results. Morphological operations can then be applied to remove small noise points or connect minor fractures to obtain the final crack identification results.

[0089] This application also provides a system for implementing the above-mentioned method for identifying cracks in CT images of weakly cemented porous media, comprising:

[0090] The data acquisition and preprocessing module is used to acquire raw 3D CT image data and perform preprocessing.

[0091] The feature mining and label generation module is used to automatically process three-dimensional core elements and generate crack labels and training datasets.

[0092] A deep learning model library and training platform for storing, training, and evaluating the deep learning recognition model based on alternating row-column scanning;

[0093] The crack identification and post-processing module is used to load the trained model for prediction and output the crack identification results.

[0094] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for identifying cracks in CT images of weakly cemented porous media.

[0095] The method described in this embodiment is used to identify cracks in CT images of weakly cemented porous media, such as... Figure 5 As shown, from the perspective of the overall crack structure, the crack network identified by this method is spatially consistent with the crack orientation visible in the original CT image. For weak crack areas with low gray-level contrast and blurred boundaries in the image, this method can still accurately delineate their contours without over-expansion or contraction. In local crack images, the boundary between the crack and the surrounding porous medium is blurred, and traditional methods are prone to breakage or omission in such areas. The identification results of this method maintain the complete crack path in this area, and the crack width variation along the path is highly consistent with the faintly visible gray-level transition zone in the image. In particular, in the section with the weakest gray level in the middle of the crack, this method can still accurately delineate its contour without misjudgment or breakage due to weak signal.

[0096] 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 scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for identifying fractures in a CT image of a weakly cemented porous medium, characterized in that, Includes the following steps: S1. Acquire three-dimensional CT image data of a weakly cemented porous medium sample, and preprocess the CT image data to extract the region of interest of the sample. S2. Through data mining and feature analysis, crack features are automatically extracted from the preprocessed images and crack label maps are generated. These maps are then paired with the original CT image dataset to form a training dataset. S3. Construct a deep learning recognition model based on alternating row-column scanning, and use the training dataset described in S2 to train and optimize the model to obtain a trained recognition model; S4. Input the CT image to be identified into the trained recognition model for prediction, and perform post-processing on the prediction results to obtain the final crack recognition result.

2. The method for identifying cracks in CT images of weakly cemented porous media according to claim 1, characterized in that, The preprocessing involves automatically dividing the entire three-dimensional region of interest into multiple three-dimensional core units of fixed size.

3. The method for identifying cracks in CT images of weakly cemented porous media according to claim 2, characterized in that, In step S2, the specific steps for automatically generating crack label maps through data mining and feature analysis are as follows: S21. An adaptive threshold segmentation method is used to extract all pore phases within the three-dimensional core unit, and a morphological watershed algorithm is used to separate connected pores to obtain independent pore individuals. S22. Quantitatively characterize each individual pore, calculate multiple feature parameters, and construct a high-dimensional pore feature space; S23. The high-dimensional pore feature space is reduced in dimensionality by non-negative matrix decomposition, and its low-dimensional components are retained as feature vectors. S24. Use a Gaussian mixture model to perform probabilistic clustering on the dimensionality-reduced pore individuals, divide them into two classes, and identify one of the classes as crack pores. S25. Reassemble the clustering recognition results according to their original coordinates to generate a complete binary crack label map, with crack areas marked as 1 and non-crack areas marked as 0. S26. Pair the crack label map with the corresponding original CT image to form a training dataset.

4. The method for identifying cracks in CT images of weakly cemented porous media according to claim 3, characterized in that, In step S23, the rank of the nonnegative matrix factorization is determined based on a threshold of 80% of the cumulative explained variance, and the low-dimensional features obtained from the factorization are used for subsequent cluster analysis.

5. The method for identifying cracks in CT images of weakly cemented porous media according to claim 3, characterized in that, The optimization objective of nonnegative matrix factorization is to minimize the reconstruction error, and its calculation formula is as follows: ， in, The original feature matrix, and The decomposed nonnegative matrix; The log-likelihood function of the Gaussian mixture model is: ， in, , The mixing coefficient, and The first The mean and covariance matrix of each Gaussian component are used to solve for the model parameters using the expectation-maximization algorithm, and cracks and pores are identified based on the posterior probability.

6. The method for identifying cracks in CT images of weakly cemented porous media according to claim 1, characterized in that, In step S3, the deep learning recognition model based on alternating row-column scanning is based on an encoder-decoder structure, and the alternating row-column scanning feature extraction module is embedded in the downsampling stage of the encoding path, the upsampling stage of the decoding path, and the bottleneck layer, respectively.

7. The method for identifying cracks in CT images of weakly cemented porous media according to claim 6, characterized in that, The row-column alternating scanning feature extraction module reassembles the two-dimensional feature map into a one-dimensional sequence along the row and column directions respectively, and uses depthwise separable convolution to perform forward-backward alternating scanning; the module also has a dynamic weight adjustment mechanism inside.

8. The method for identifying cracks in CT images of weakly cemented porous media according to claim 7, characterized in that, The internal processing steps of the row-column alternating scan feature extraction module are as follows: S31. Use a channel expansion layer with an expansion factor of 2 to map the input feature map to a higher-dimensional hidden space; S32, Alternating forward and reverse scanning in the row direction: Input feature map... Remodeling For the first row sequence The formulas for its forward and reverse scan operations are as follows: ， ， in, These are the weight parameters of depthwise separable convolution; S33, Alternating forward and reverse scanning in the column direction: After transposing the input feature map, the same strategy as the row direction is used to scan in the column direction; S34. Fuse the features obtained from the row and column scans to obtain global context features. with local details ; S35. Introduce a dynamic weight adjustment mechanism to adaptively fuse global context features and local detail features. The fusion formula is as follows: ， in, It is a spatially adaptive gated weighted graph generated by the Sigmoid function. This represents element-wise multiplication. This represents a matrix of all 1s; when the weight values ​​are close to 1, local features are mainly used to maintain boundary accuracy; when the weight values ​​are close to 0, global context features are mainly used to maintain crack continuity. S36. Through channel reduction layers and convolutional layers, and residual connections with the module input, the gradient flow and feature information are fully preserved.

9. A crack identification system for CT images of weakly cemented porous media, used to implement the method described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw 3D CT image data and perform preprocessing. The feature mining and label generation module is used to automatically process three-dimensional core elements and generate crack labels and training datasets. A deep learning model library and training platform for storing, training, and evaluating the deep learning recognition model based on alternating row-column scanning; The crack identification and post-processing module is used to load the trained model for prediction and output the crack identification results.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images