Sandstone slice image semi-supervised classification and segmentation method based on GL-SLIC

Abstract

The invention discloses a sandstone slice image component identification method based on a GL-SLIC algorithm and semi-supervised learning. The method comprises the following steps: firstly, acquiring a sandstone slice image amplified by 200 times under single polarization, converting an RGB image into a CIE-Lab color space, and extracting a GLBP texture feature vector; then a GL-SLIC superpixel segmentation algorithm is constructed, and adaptive segmentation based on texture and color features is realized in combination with a Gabor filter and a local binary pattern; then implementing a region merging algorithm to generate complete mineral particles and pore fragments; finally, a classifier based on VGG16 and a discriminator based on ResNet18 are constructed, a semi-supervised self-training framework is adopted, a model is initialized by using about 6% of manual labeling samples, a high-confidence-coefficient pseudo-label extension training data set is generated through iteration, and automatic recognition of sandstone components such as quartz, pores, kaolinite, rock debris and a matrix is achieved. According to the method, the recognition accuracy of 96.3% on a test set is achieved, compared with a traditional method, the segmentation precision and the recognition accuracy are remarkably improved, the data labeling cost is greatly reduced, and an efficient technical scheme is provided for automatic analysis of geological images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological exploration and rock analysis, and particularly relates to a sandstone thin section image composition automatic identification technology based on a semi-supervised learning framework and an improved GL-SLIC superpixel segmentation algorithm. BACKGROUND

[0002] In the field of geological engineering, sandstone thin section composition identification plays a crucial role in geological exploration and mineral analysis, and is essential in oil exploration and mineral development. Traditional methods usually involve manual observation and analysis of the composition and structure of sandstone thin sections, with manual marking of the boundaries between mineral grains and pores to identify the mineral and pore characteristics of the target rock. However, this manual process is not only time-consuming, but also susceptible to subjective bias, limiting the quantification and reliability of the identification results.

[0003] In recent years, with the development of deep learning technology, automated thin section analysis has received extensive attention. Ren et al. introduced a multi-channel attention transformer model, significantly improving the instance segmentation accuracy in sandstone images, surpassing previous methods in terms of mean average precision (mAP). Liu et al. applied artificial intelligence technology for automatic identification of sandstone thin sections, improving accuracy and efficiency. Shebl et al. demonstrated the application of cognitive image recognition in automated thin section image description, providing significant improvements for carbonate rock analysis. Yalamanchi et al. used machine learning algorithms and scanning electron microscope (SEM) images to predict the pore structure and permeability of carbonate reservoirs.

[0004] In terms of image segmentation, traditional superpixel segmentation methods such as SLIC and Watershed have problems with inaccurate boundary positioning when dealing with complex geological images. Dabek et al. proposed a superpixel grain segmentation method specifically for sandstone thin sections, significantly improving the accuracy of grain boundary identification. However, existing methods still have the following problems when processing sandstone thin section images:

[0005] Inaccurate boundary identification: traditional segmentation algorithms have difficulty accurately capturing the boundary features of complex mineral grains;

[0006] Strong dependence on data labeling: deep learning methods require a large amount of manually labeled data, and the diversity and complexity of geological images make the manual labeling process labor-intensive and prone to errors;

[0007] Limited feature extraction capability: existing methods have limited ability to extract unique texture patterns and structural complexity of sandstone thin sections.

[0008] Therefore, it is necessary to develop a sandstone thin section image composition identification method that effectively combines texture feature extraction and semi-supervised learning to solve the above technical problems. SUMMARY

[0009] In order to solve the problems of large amount of labeled data, inaccurate boundary identification and insufficient generalization ability in the prior art, the present application provides a sandstone thin section image component recognition method based on semi-supervised learning and GL-SLIC algorithm. The method realizes accurate image pre-segmentation by combining the GL-SLIC superpixel segmentation algorithm of Gabor filter and local binary pattern, and uses a semi-supervised learning framework to generate high-confidence pseudo-labels to expand the training set using a small amount of labeled data, thereby realizing high-precision automatic identification of quartz, pores, kaolinite, matrix, debris and other sandstone components. The technical scheme adopted by the present application is:

[0010] 1. A sandstone thin section image component recognition method based on GL-SLIC algorithm and semi-supervised learning, characterized in that it comprises the following steps:

[0011] 1) Collecting sandstone thin section image data to obtain resolution images taken at 200 times magnification under single polarized light;

[0012] 2) Preprocessing the sandstone image, converting the RGB image to CIE-Lab color space, and constructing a texture feature vector through GLBP feature extraction, while constructing a labeled data set and an unlabeled data set for semi-supervised learning;

[0013] 3) Constructing a GL-SLIC superpixel segmentation algorithm, combining Gabor filter and local binary pattern to extract GLBP features, and realizing adaptive superpixel segmentation based on texture and color features;

[0014] 4) Implementing a region merging algorithm based on the color feature similarity of adjacent superpixels to generate complete mineral grain and pore fragments;

[0015] 5) Constructing a classifier model based on VGG16 and a discriminator model based on ResNet18, and using a semi-supervised self-training framework to train the model;

[0016] 6) Initializing the model using a small amount of manually labeled samples, generating high-confidence pseudo-labels through iterative adversarial training, and expanding the training data set;

[0017] 7) Using the trained model to perform real-time component recognition on sandstone thin section images, and outputting the recognition results of quartz, pores, kaolinite, debris, matrix and other categories, thereby providing an automatic solution for geological analysis.

[0018] 2. The method of claim 1, wherein the sandstone thin section image data in step 1) comprises:

[0019] The tight sandstone reservoir image has an average porosity of 9.66% and a permeability of 0.05 mD, and mainly contains quartz, kaolinite and clay minerals.

[0020] 3. The method of claim 1, wherein the GLBP feature extraction in step 2) comprises the following steps:

[0021] 2.1) Apply 36 Gabor filters with different scales and directions to the original sandstone thin section image, including 6 different scales and 6 directions.

[0022] 2.2) Calculate the mean value of the 6-direction Gabor features at each scale to obtain 6 mean value feature images, and the calculation formula is:

[0023]

[0024] wherein GM li is the mean value feature of the i-th scale, is the Gabor feature of the i-th scale and the j-th direction.

[0025] 2.3) Apply LBP operator to encode the 6 mean value feature images respectively to emphasize local texture patterns.

[0026] 2.4) Perform principal component analysis (PCA) dimension reduction on the 6 LBP encoded image matrices to extract the most important features to form the GLBP feature vector.

[0027] 4. The method of claim 1, wherein the GL-SLIC algorithm in step 3) comprises the following steps:

[0028] 3.1) Divide the image into K superpixel grids and initialize the cluster centers C k = [l k ,a k ,b k ,x k ,y k ,GL k ], wherein l k ,a k , b k are CIE-Lab color features, x k ,y k are spatial positions, and GL k is the GLBP feature value.

[0029] 3.2) Optimize the initial cluster center position and move it to the area with the smallest gradient in the neighborhood, and the gradient calculation formula is:

[0030]

[0031] 3.3) Calculate the integrated distance of pixels and cluster centers, distance metrics include color feature distance, spatial feature distance and texture feature distance:

[0032]

[0033] where d c is the color distance, d s is the spatial distance, and d gl is the texture distance;

[0034] 3.4) Iteratively update the cluster centers until convergence, complete the superpixel segmentation.

[0035] 5. The method of claim 1, wherein the region merging algorithm in step 4) comprises:

[0036] 4.1) Construct an adjacency matrix based on the average color features of superpixel blocks;

[0037] 4.2) Calculate the Euclidean distance between adjacent regions as a dissimilarity measure;

[0038] 4.3) Set the merging threshold Adjacent superpixels that meet the following conditions are merged:

[0039]

[0040] where w ij is the weight between superpixels i and j;

[0041] 4.4) Perform graph-based segmentation merging using a breadth-first search algorithm.

[0042] 6. The method of claim 1, wherein the semi-supervised learning framework in steps 5) and 6) comprises:

[0043] 5.1) Initialize the VGG16 classifier and ResNet18 discriminator using about 6% manually labeled samples;

[0044] 5.2) Classify unlabeled data and select samples with confidence greater than 94% as high-confidence pseudo-labels;

[0045] 5.3) The discriminator model verifies the effectiveness of the pseudo-labels;

[0046] 5.4) Combine the original labeled samples and high-confidence pseudo-label samples for model fine-tuning;

[0047] 5.5) Repeat steps 5.2 to 5.4 until the termination condition is met, and realize the automatic expansion of the dataset.

[0048] 7. The method of claim 6, wherein the classifier employs a modified VGG16 architecture comprising:

[0049] with an input layer size of (3, 32, 32), passing through the first 14 layers of a VGG16 network with batch normalization, containing two convolutional stages, each stage containing two convolutional blocks, each block including convolutional operations, batch normalization, ReLU activation, and max pooling layers, and finally outputting a probability distribution of 6 classes through fully connected layers and a softmax function.

[0050] 8. A sandstone image component recognition system implementing the method of any one of claims 1-7, comprising:

[0051] an image preprocessing module for performing image acquisition, CIE-Lab color space conversion, and GLBP feature extraction; a GL-SLIC segmentation module for implementing adaptive superpixel segmentation based on texture and color features; a region merging module for superpixel merging based on similarity; a semi-supervised learning module containing a VGG16 classifier and a ResNet18 discriminator to implement a self-training framework; and a result output module for displaying segmentation results, recognition results, and performance evaluation indicators.

[0052] The advantages of the present application over the prior art mainly include:

[0053] (1) High data efficiency: Through the semi-supervised learning framework, only about 6% of the labeled data is needed to achieve high-precision recognition, greatly reducing the data labeling cost and workload, and solving the problem of difficulty in obtaining labeled data in geological image analysis.

[0054] (2) High segmentation accuracy: The GL-SLIC algorithm combined with Gabor filters and LBP features can more accurately identify the complex boundaries of mineral particles in sandstone thin sections, and has significantly improved the under-segmentation error rate, boundary recall rate, and precision rate compared to the traditional SLIC algorithm.

[0055] (3) Strong recognition accuracy: The final model achieves a recognition accuracy of 96.3% on the test set, which is significantly higher than the 72.12% of the UNet semantic segmentation method, achieving a comprehensive accuracy of 89.33%.

[0056] (4) Good generalization ability: The semi-supervised learning framework expands the diversity of training data through the pseudo-label mechanism, improving the model's adaptability to different geological conditions and imaging conditions.

[0057] (5) Strong practicality: This method provides an efficient and automated tool for sandstone reservoir evaluation, porosity and permeability analysis, and other geological applications, which helps to improve the efficiency and objectivity of geological analysis.

[0058] In summary, the sandstone thin section image component recognition method based on semi-supervised learning and GL-SLIC algorithm provided by the application successfully solves the key technical problems such as large data labeling requirement and inaccurate boundary recognition in traditional methods through the organic combination of innovative feature extraction, superpixel segmentation, region merging and semi-supervised learning technology, provides an efficient and accurate solution for automatic analysis of geological images, and has important scientific significance and application value. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below.

[0060] Figure 1 The overall workflow diagram of the method of the application is shown in Figure 1.

[0061] Figure 2 The mineral recognition workflow diagram of GL-SLIC segmentation and semi-supervised training is shown in Figure 2.

[0062] Figure 3 The GLBP feature extraction workflow diagram is shown in Figure 3.

[0063] Figure 4 The semi-supervised self-training process diagram is shown in Figure 4.

[0064] Figure 5 The modified VGG16 classifier architecture diagram is shown in Figure 5.

[0065] Figure 6 The discriminator model architecture diagram is shown in Figure 6.

[0066] Figure 7 The training and testing accuracy change curve diagram of the main model is shown in Figure 7.

[0067] Figure 8 The iteration model training and data enhancement process diagram is shown in Figure 8.

[0068] Figure 9 The confusion matrix diagram of the final model evaluation is shown in Figure 9.

[0069] Figure 10 The component recognition result comparison diagram is shown in Figure 10. DETAILED DESCRIPTION

[0070] Embodiment 1:

[0071] Referring to Figure 1 and Figure 2 The sandstone thin section image component recognition method based on semi-supervised learning and GL-SLIC algorithm of the application comprises the following steps:

[0072] Step 1: Data acquisition and preprocessing

[0073] The images of sandstone thin sections were collected from a tight sandstone reservoir with an average porosity of 9.66% and a permeability of 0.05 mD. The samples were collected perpendicular to the bedding, and the main constituents include quartz, kaolinite, and clay minerals, with carbonate, siliceous, and iron oxide cements. The grain sizes vary from fine to medium, with point and long contacts.

[0074] The images of sandstone thin sections were collected from a tight sandstone reservoir with an average porosity of 9.66% and a permeability of 0.05 mD. The samples were collected perpendicular to the bedding, and the main constituents include quartz, kaolinite, and clay minerals, with carbonate, siliceous, and iron oxide cements. The grain sizes vary from fine to medium, with point and long contacts.

[0075] Step 2: GL-SLIC superpixel segmentation

[0076] As shown in Figure 3 the GLBP feature extraction process includes the following sub-steps:

[0077] 2.1) Gabor filter feature extraction: Apply Gabor filters to the original sandstone thin section images, with 6 different scales and 6 different orientations, creating a total of 36 different Gabor filters. These filters systematically extract features from the images, resulting in 36 Gabor feature vectors.

[0078] 2.2) Average feature calculation: The Gabor features extracted for each different direction and scale are averaged to calculate the average features for each corresponding setting. The calculation formula is:

[0079]

[0080] where i = 1, 2,..., 6 represents different scale parameters.

[0081] As shown in Figure 4 the Gabor filter transformation effectively extracts unique texture features of sandstone thin sections at different scales and orientations.

[0082] 2.3) LBP encoding: The average feature images obtained from Gabor filters are encoded using the LBP operator. This operator emphasizes local texture patterns by comparing each pixel with its neighborhood.

[0083] 2.4) GLBP feature construction: Perform principal component analysis (PCA) on the 6 image matrices generated by LBP encoding. This step reduces the dimensionality and extracts the most important features, forming a feature vector that succinctly represents the texture features of the sandstone images.

[0084] 2.5) Improved SLIC segmentation: Perform superpixel segmentation using the GL-SLIC algorithm, which includes the following steps:

[0085] Convert the image from the RGB color space to the CIE-Lab color space;

[0086] Initialize cluster centers, with each cluster center corresponding to a superpixel center C. k =[l k ,a k ,b k ,x k ,y k ,GL k ];

[0087] Optimize the initial cluster centers by adjusting them to the regions with the lowest gradients in their neighborhoods;

[0088] Pixel classification is performed using a comprehensive distance metric.

[0089]

[0090] in:

[0091] Color feature distance:

[0092] Spatial feature distance:

[0093] Texture feature distance:

[0094] Step 3: Region Merging Optimization

[0095] 3.1) Constructing the adjacency matrix: Construct an adjacency matrix based on the average color features of each superpixel block, and measure the dissimilarity between adjacent regions using Euclidean distance.

[0096] 3.2) Set the merging threshold: Use the normalized difference threshold formula:

[0097]

[0098] 3.3) Perform region merging: Use a breadth-first traversal algorithm to merge superpixels, merging regions with low similarity.

[0099] Experimental results show that the merged superpixels mainly represent complete mineral particles, and the algorithm effectively merges broken mineral particles and slender porous structures into coherent units.

[0100] Step 4: Construction of a Semi-Supervised Learning Framework

[0101] like Figure 4 As shown, a semi-supervised self-training framework is constructed:

[0102] 4.1) Main Classifier Design: A modified VGG16 architecture is adopted, such as...Figure 5 As shown, the first 14 layers and batch normalization are included. The network is divided into two segments, each containing two convolutional blocks. The input tensor size is (3, 32, 32), and the output is a feature tensor of (128, 8, 8) after processing, flattened to (1, 8192) and passed through two fully connected layers, and finally a softmax function outputs the probability distribution of N classes.

[0103] 4.2) Discriminator Design: A subset of ResNet18 architecture is adopted, as shown in Figure 6 From the initial convolution, followed by batch normalization and max pooling. The subsequent layers include Layer1 to Layer3, each composed of two basic block types. The output tensor size is (256, 3, 3), flattened to (1, 2304) and passed through the final fully connected layer.

[0104] Step 5: Model Training and Optimization

[0105] The model training process includes:

[0106] 5.1) Initial Training: The main classifier (VGG16) is initially trained using 600 hand-labeled samples. This basic training establishes initial learning parameters that guide the classification of unlabeled samples.

[0107] 5.2) Unlabeled Sample Classification: The main model processes the remaining 7920 unlabeled samples, classifying them based on features learned during initial training. Each sample receives a preliminary label and a confidence score.

[0108] 5.3) Pseudo-Label Generation: The main model assigns pseudo-labels to the unlabeled samples. Only samples with a confidence score greater than 94% are considered. These high-confidence pseudo-labels indicate that the model is very certain about their classification.

[0109] 5.4) High-Confidence Sample Selection: Samples that meet the high-confidence threshold are selected for further processing. The discriminator model (ResNet18) reviews these pseudo-labeled samples to confirm their validity.

[0110] 5.5) Enhanced Data Fine-Tuning: The model is refined using both original labeled samples and newly identified high-confidence samples, repeating steps 5.2 to 5.4 to gradually expand the dataset of high-confidence pseudo-labeled samples.

[0111] By the end, the data for each of the 5 classes (quartz, pore, detritus, matrix, kaolinite) reaches approximately 1300 images. After multiple experiments, the model converges around the 120th iteration.

[0112] Step 6: Component Identification and Evaluation

[0113] 6.1) Segmentation Quality Evaluation: The following evaluation metrics are used:

[0114] Undersegmentation error (UE):

[0115]

[0116] Boundary recall (BR):

[0117]

[0118] Precision:

[0119]

[0120] 6.2) Classification performance evaluation: accuracy, precision, recall, F1 score, etc. Final model implementation: training accuracy 93.7%, test accuracy 96.3%, F1 score 96.4%.

[0121] 6.3) Comparative experiment: compared with UNet semantic segmentation method, the method of the application achieves an accuracy of 89.33% in the comparison of manual annotation results, while the semantic segmentation method only achieves 72.12%.

[0122] The experimental results show that the method proposed by the application has significant advantages in sandstone thin section image component recognition, can effectively improve the recognition accuracy, reduce the data labeling demand, and provides an effective technical scheme for automatic analysis of geological images.

[0123] Example 2

[0124] This embodiment verifies the superiority of the GL-SLIC algorithm compared with other superpixel segmentation algorithms.

[0125] As Figure 7 shown, GL-SLIC algorithm is compared with mainstream superpixel segmentation algorithms LSC, SLIC, FH, QS, SEEDS and Watershed. The average results on the 50 sandstone thin section image data sets are as follows:

[0126]

[0127] The GL-SLIC algorithm is superior to other algorithms in all three indicators, showing its excellent segmentation accuracy and fidelity to the edges of real mineral particles.

[0128] Example 3

[0129] This embodiment details the effectiveness of the region merging algorithm.

[0130] The region merging algorithm based on GL-SLIC segmentation algorithm is used to process 50 sandstone thin section images, and the average results before and after merging are as follows:

[0131]

[0132] The merged indicators are significantly improved, both precision and boundary recall are improved, indicating that the merged superpixels are more consistent with the true boundaries of mineral particles, producing a more accurate representation of the actual geological structure.

Claims

1. A sandstone thin section image component identification method based on GL-SLIC algorithm and semi-supervised learning, characterized in that, The method comprises the following steps: 1) Collecting sandstone thin section image data to obtain resolution images taken under single polarization at 200 times magnification; 2) Preprocessing the sandstone images, converting the RGB images to CIE-Lab color space, and constructing texture feature vectors through GLBP feature extraction, while constructing a labeled dataset and an unlabeled dataset for semi-supervised learning; 3) Constructing a GL-SLIC superpixel segmentation algorithm, combining Gabor filters and local binary pattern to extract GLBP features, and realizing adaptive superpixel segmentation based on texture and color features; 4) Implementing a region merging algorithm based on the color feature similarity of adjacent superpixels to generate complete mineral grain and pore fragments; 5) Constructing a VGG16-based classifier model and a ResNet18-based discriminator model, and using a semi-supervised self-training framework for model training; 6) Initializing the model using a small number of manually labeled samples, generating high-confidence pseudo-labels through iterative adversarial training, and expanding the training dataset; 7) Using the trained model to perform real-time component recognition on sandstone thin section images, and outputting the recognition results of quartz, pores, kaolinite, detritus, matrix, and other categories to provide an automated solution for geological analysis.

2. The method of claim 1, wherein, In the step 1), the sandstone thin section image data includes: The dense sandstone reservoir image has an average porosity of 9.66% and a permeability of 0.05 millidarcy, and the main components include quartz, kaolinite, and clay minerals.

3. The method of claim 1, wherein, The specific process of GLBP feature extraction in the step 2) includes: 2.1) Applying 36 Gabor filters of different scales and directions to the original sandstone thin section image, including 6 different scales and 6 directions; 2.2) Calculating the mean value of the 6-direction Gabor features at each scale to obtain 6 mean feature images, and the calculation formula is: wherein GM li is the mean feature of the i-th scale, is the Gabor feature of the i-th scale and j-th direction. 2.3) Applying LBP operators to the 6 mean feature images for encoding to emphasize local texture patterns; 2.4) Performing principal component analysis PCA dimensionality reduction on the 6 LBP encoded image matrices to extract the most important features to form the GLBP feature vector.

4. The method of claim 1, wherein, The specific process of the GL-SLIC algorithm in the step 3) includes: 3.1) Divide the image into K superpixel grids, initialize clustering centers C k = [l k ,a k ,b k ,x k ,y k ,GL k ], where l k ,a k , b k are CIE-Lab color features, x k ,y k are spatial locations, GL k is GLBP feature value; 3.2) Optimizing the initial cluster center position by moving it to the area with the smallest gradient in the neighborhood, and the gradient calculation formula is: 3.3) Calculating the comprehensive distance between the pixel and the cluster center, and the distance metric includes color feature distance, spatial feature distance, and texture feature distance: where d c is the color distance, d s is the spatial distance, d gl is the texture distance; 3.4) Iteratively updating the cluster center until convergence to complete superpixel segmentation.

5. The method of claim 1, wherein, The region merging algorithm in the step 4) includes: 4.1) Constructing an adjacency matrix based on the average color features of superpixel blocks; 4.2) Calculating the Euclidean distance between adjacent regions as a dissimilarity measure; 4.3) Set the merge threshold Adjacent superpixels that meet the following conditions are merged: where w ij is the weight between superpixels i and j; 4.4) Using a breadth-first traversal algorithm to perform graph-based segmentation merging.

6. The method of claim 1, wherein, The semi-supervised learning framework in the steps 5) and 6) includes: 5.1) Initializing the VGG16 classifier and the ResNet18 discriminator using approximately 6% of the manually labeled samples; 5.2) Classify unlabeled data, select samples with confidence greater than 94% as high-confidence pseudo-labels; 5.3) Discriminator model verifies the effectiveness of pseudo-labels; 5.4) Combine original labeled samples and high-confidence pseudo-label samples for model fine-tuning; 5.5) Repeat steps 5.2 to 5.4 until the termination condition is met, and realize the automatic expansion of the dataset.

7. The method of claim 6, wherein, The classifier uses a modified VGG16 architecture, including: The input layer size is (3, 32, 32), and the VGG16 network is normalized by the first 14 layers, containing two convolutional sections, each containing two convolutional blocks, each including convolution operation, batch normalization, ReLU activation and maximum pooling layer, finally outputting the probability distribution of 6 categories through fully connected layer and softmax function.

8. A sandstone image component recognition system implementing the method of any one of claims 1 to 7, characterized by Including: Image preprocessing module, used for image acquisition, CIE-Lab color space conversion and GLBP feature extraction; GL-SLIC segmentation module, realizing adaptive superpixel segmentation based on texture and color features; region merging module, based on similarity for superpixel merging; Semi-supervised learning module, containing VGG16 classifier and ResNet18 discriminator to realize self-training framework; result output module, used for displaying segmentation results, recognition results and performance evaluation indicators.

Images