Rock image classification method based on super-resolution and texture feature fusion

Abstract

The invention provides a rock image classification method based on super-resolution and texture feature fusion. According to the method, the accuracy of mineral classification in the rock slice image is remarkably improved by combining a super-resolution model and a texture feature analysis technology, the enhanced image texture features are extracted by utilizing a Gabor filter, and classification is carried out by adopting a support vector machine. By comparing the performance of various SR models, the optimal model is selected to enhance the image resolution, so that the recognition effect of minerals and pores is improved. In addition, the system further optimizes the classification result through multi-color space feature fusion, finer mineral distribution data is provided for geological analysis, the mineral classification efficiency and precision are improved, and reliable technical support is provided for geological evaluation of oil and gas reservoirs.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of oil and gas exploration and development, and particularly relates to a rock image classification method based on super-resolution and texture feature fusion, which is suitable for application scenarios such as automatic classification of core images, geological lithology analysis and oil and gas reservoir modeling. BACKGROUND

[0002] Identifying minerals and pores in rock thin sections is a core technology for analyzing the geological conditions of oil and gas reservoirs, and its accuracy directly affects key links such as obtaining lithofacies characteristics, analyzing sedimentary diagenesis and modeling pore structure. With the increasing demand for automation and intelligence in geological exploration, high-precision mineral identification technology has become a key to improving the efficiency and reliability of geological evaluation.

[0003] Traditional rock identification methods rely on manual observation, are low in efficiency and dependent on expert experience, and are difficult to adapt to the needs of large-scale core image analysis. In recent years, academia has carried out a lot of exploration in the direction of dense sandstone image recognition in order to improve automation and identification accuracy. For example, Hasan, M, et al. in the document “Rock texture classification using Gabor features and SVM in tight sandstone formations” used Gabor filtering to extract texture features of grayscale images and combined support vector machines (SVM) to classify the lithology of dense sandstone. The method has a certain effect, but is limited to single-channel image processing and does not consider the expression ability of color information. Li, X, et al. in the document “Texture and color fusion for sandstone image classification using Gabor filters and color spaces” introduced the YCbCr color space and fused the Gabor features of part of the color channels to enhance the classification ability, but did not consider multi-directional texture modeling and image resolution problems. Wu, et al. in the document “Multi-channel texture feature fusion for rock thin section classification with deep learning” tried to input the multi-channel texture features into a deep neural network classifier after fusion, which achieved a certain degree of performance improvement, but due to the limitation of image quality, high-frequency details were still missing, which affected the model's ability to distinguish complex structures.

[0004] From these studies, it can be seen that the current rock image recognition technology has the following key problems: low image resolution, which leads to insufficient expression of texture details, especially affecting the discrimination of similar classes such as debris and pores.

[0005] In order to solve the above problems, it is necessary to study the texture enhancement ability of various super-resolution models on rock images, to realize the collaborative optimization of image quality improvement and recognition accuracy by fusing color space, directional texture and machine learning classification method, so as to provide more accurate and efficient technical support for automatic classification of dense sandstone minerals and reservoir modeling. SUMMARY

[0006] In order to overcome the deficiencies of the existing lithology image recognition method in insufficient image resolution, insufficient texture features, low color channel utilization rate and low classification accuracy, the purpose of the present application is to propose a rock image classification method based on super-resolution and texture feature fusion. The method fuses image enhancement and multi-channel texture extraction strategy, and constructs a complete process from data acquisition, image reconstruction, feature extraction to classification discrimination. By introducing a super-resolution model to reconstruct the details of low-resolution images, extracting channel texture features in multiple color spaces, and combining Gabor filter to realize multi-directional response, finally using support vector machine for multi-class lithology identification, so as to significantly improve the image quality and recognition accuracy. The purpose of the present application is to provide a dense sandstone mineral classification method based on super-resolution enhancement and texture feature fusion, to solve the problems of low classification accuracy and insufficient detail preservation of existing image recognition technology in dense sandstone thin section images.

[0007] To solve the above technical problems, the technical scheme adopted by the present application is:

[0008] 1. A rock image classification method based on super-resolution and texture feature fusion, characterized in that it comprises the following steps:

[0009] 1) Collecting dense sandstone thin section images, generating an original high-resolution image set, and down-sampling into a low-resolution image set;

[0010] 2) Inputting the low-resolution image into a super-resolution model for image reconstruction;

[0011] 3) Converting the image to color space;

[0012] 4) Using a two-dimensional Gabor filter to extract multi-directional texture features for each channel image;

[0013] 5) Applying principal component analysis dimensionality reduction to the extracted texture features, inputting the dimensionality reduced feature vectors into a support vector machine classifier based on radial basis function, to realize the identification of mineral types;

[0014] 6) Performance evaluation of the classification results of the support vector machine, outputting the overall classification accuracy and the recognition accuracy of each lithology type.

[0015] 2. The method of claim 1, wherein the step 1) extracts a set of mineral fragment images as a set of original high-resolution images by superpixel segmentation and region merging technology from the scanned image of the collected tight sandstone thin section; and then down-samples the set of original high-resolution images into a set of low-resolution images by using a bicubic interpolation method.

[0016] 3. The method of claim 1, wherein the image reconstruction of the step 2) is performed by constructing a pair of training sets of low-resolution images and corresponding high-resolution images, training a super-resolution model in a supervised learning manner to learn the mapping relationship between the low-resolution images and the high-resolution images, and using the trained super-resolution model to generate a super-resolution image with high resolution and detail enhancement effect according to the input low-resolution image.

[0017] 4. The method of claim 1, wherein the super-resolution model used in the image reconstruction of the step 2) includes SRResNet, EDSR, ESRGAN, and Real-ESRGAN; the model is trained by constructing a supervised pairing relationship between high-resolution images and corresponding low-resolution images, and the loss function used in the training process includes L1 loss, perceptual loss, and adversarial loss; the SRResNet and EDSR models use pixel-level L1 loss.

[0018]

[0019] wherein, represents the i-th original high-resolution image, represents the output image of the super-resolution model, and n represents the number of samples.

[0020] The ESRGAN and Real-ESRGAN models use a total loss function:

[0021] L G =L Percep +λL Ra_G +ηL1

[0022] wherein, L Percep represents the perceptual loss, L Ra_G is the relative adversarial loss, and L1 is the pixel-level L1 loss; λ and η are weighting coefficients.

[0023] 5. The method of claim 1, wherein the two-dimensional function expression of the Gabor filter of the step 4) is as follows:

[0024]

[0025] where x and y are image space coordinates, x' and y' are rotated coordinates, x' = xcosθ + ysinθ, y' = -xsinθ + ycosθ; σ x and σ y are standard deviations of Gaussian function in x and y direction; ω is center frequency; θ is direction angle of filter; j is imaginary unit.

[0026] 6. The method of claim 1, wherein the direction angle θ of the multi-directional texture feature extraction of step 4) is in the range of:

[0027]

[0028] for extracting 12 directional texture features.

[0029] 7. The method of claim 1, wherein the principal component analysis in step 5) is used to reduce the dimension of the multi-dimensional texture features extracted from the multi-directional Gabor filter, and to retain the principal components with cumulative contribution higher than 90% to compress the feature dimension and enhance the stability of feature expression.

[0030] 8. The method of claim 1, wherein the classifier of step 5) is a support vector machine (SVM) using a radial basis function (RBF) kernel, and the kernel function is expressed as:

[0031] k(x i , y i ) = exp(-γ||x i - y j ||) 2

[0032] where x i and x j are feature vectors; γ is an adjustment parameter;

[0033] The decision function of the classifier SVM is as follows:

[0034]

[0035] where α i is a Lagrange multiplier, y i is the label of support vector, k(x i , x) is the value of kernel function, b is a bias term, and the absolute value |f(x)| thereof is used as the confidence evaluation of sample classification.

[0036] ​9. The method of claim 1, wherein the performance evaluation in step 6) comprises calculating the distance of each test sample to the classification hyperplane as a confidence basis using the decision function of the support vector machine, generating precision, recall and F1 value for quantitative analysis of the recognition effect of each type of lithology, and comparing the classification ability of different super-resolution models.

[0037] Compared with the prior art, the present application has the following remarkable technical effects and advantages:

[0038] 1) Significantly improve the image reconstruction quality: by introducing SRResNet, EDSR, ESRGAN and Real-ESRGAN super-resolution models for detail enhancement of low-quality rock images, the high-frequency information of the original image is restored, and the texture blur problem caused by low traditional image resolution is effectively improved;

[0039] 2) Realize multi-channel texture feature fusion: the present application uses H, V, R, Cb and Cr multi-channel images, combines a two-dimensional Gabor filter to extract texture features from multiple color dimensions and direction angles, makes up for the limitations of traditional methods using only grayscale or RGB images, and enhances the feature expression ability;

[0040] 3) Improve the classification performance and generalization ability: combined with principal component analysis (PCA) for dimensionality reduction of high-dimensional features and input of the SVM classifier, the recognition accuracy and calculation efficiency are significantly improved, and overfitting is avoided. Experimental results show that the overall accuracy of the present method in the five-class mineral identification task is more than 85%, which is better than the existing comparative method;

[0041] In summary, the present application not only improves the accuracy and stability of mineral image classification, but also provides a new idea and technical support for intelligent processing of geological information. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Flowchart of rock image classification method based on super-resolution and texture feature fusion

[0043] Figure 2 Typical rock mineral texture diagram

[0044] Figure 3 Image sample construction and annotation diagram

[0045] Figure 4 SRResNet model structure diagram

[0046] Figure 5 EDSR model structure diagram

[0047] Figure 6 ESRGAN generator network structure diagram

[0048] Figure 7 The ESRGAN discriminator network structure diagram is as follows:

[0049] Figure 8 The Real-FSRGAN discriminator structure diagram is as follows:

[0050] Figure 9 The loss curve diagram in the training process of different super-resolution models is as follows:

[0051] Figure 10 The reconstructed image quality comparison diagram of each model is as follows:

[0052] Figure 11 The different color channel texture feature principal component scatter diagram is as follows. DETAILED DESCRIPTION

[0053] Embodiment 1:

[0054] Referring to Figure 1 The rock image classification method based on super-resolution and texture feature fusion provided by the present application comprises the following steps:

[0055] 1) Collecting dense sandstone thin section images, generating an original high-resolution image set, and down-sampling into a low-resolution image set;

[0056] 2) Inputting the low-resolution images into a super-resolution model for image reconstruction;

[0057] 3) Performing color space conversion on the images;

[0058] 4) Using a two-dimensional Gabor filter to extract multi-directional texture features for each channel image;

[0059] 5) Applying principal component analysis dimension reduction to the extracted texture features, inputting the dimension-reduced feature vectors into a support vector machine classifier based on a radial basis function, and realizing the identification of mineral types;

[0060] 6) Evaluating the performance of the support vector machine classification result, and outputting the overall classification accuracy and the identification accuracy of each lithology type.

[0061] 1) Collecting dense sandstone thin section images, generating an original high-resolution image set, and down-sampling into a low-resolution image set:

[0062] 1.1) Collecting representative dense sandstone core samples from the Triassic Yanchang Formation in the Ordos Basin, and imaging the thin sections using a scanning electron microscope. The microscope imaging magnification is set to 1000 times to ensure that the grain boundaries and pore structures can be distinguished at the microscopic scale. Referring to Figure 2As shown, five typical rock mineral texture images are shown, including: (A) Quartz, (B) Matrix, (C) Lithic Fragments, (D) Pore and (E) Kaolinite. The figure serves as the data basis for the training and testing of the classification system of the present application, and intuitively reflects the different characteristics of various minerals in the image.

[0063] 1.2) Use image annotation tools to manually label the typical mineral regions in each image, and construct a data set for supervised learning model. Extract 10 ROI (Region of Interest) from each labeled image, extract 200 labeled samples for each type of mineral, and a total of 1000 high-resolution labeled image samples are constructed. Figure 3 As shown, the image sample construction and labeling schematic diagram shows the process of ROI extraction and mineral label assignment using the labeling tool on the original rock thin section image. Through manual labeling of the image by the tool, each type of mineral sample is extracted, and a high and low resolution image pair is constructed. The total number of images is 100, the original resolution of each image is 512x512 pixels, and the image format is uncompressed PNG to ensure that the image details are not lost.

[0064] 1.3) Using bicubic interpolation, downsample the above high-resolution images (HR) to reduce the image size from 512x512 pixels to 128x128 pixels, and construct a low-resolution image set (LR). This process ensures reasonable compression of the image without introducing obvious artifacts, while preserving the overall texture structure and spatial layout, providing data pairs for subsequent super-resolution reconstruction model training.

[0065] 1.4) Form a data set containing 1000 pairs of HR-LR image pairs for subsequent step 2) training and testing of the super-resolution model.

[0066] 2) Input the low-resolution image into the super-resolution model for image reconstruction:

[0067] 2.1) Input the low-resolution image into the super-resolution model for image reconstruction, and the training data is composed of the image set obtained in step 1), including paired high-resolution images (HR, 512x512 pixels) and low-resolution images (LR, 128x128 pixels) obtained by bicubic interpolation downsampling. All images are stored in PNG format, maintaining lossless compression to ensure reconstruction quality.

[0068] 2.2) Whole dataset construction 1000 pairs of HR-LR images, covering five categories of mineral samples Quartz, Matrix, Lithic Fragments, Pore, Kaolinite, are divided into training set (800 pairs), validation set (100 pairs) and test set (100 pairs) in the ratio of 8:1:1.

[0069] 2.3) Four kinds of super-resolution models are used for image reconstruction, which are SRResNet, EDSR, ESRGAN and Real-ESRGAN.

[0070] Reference Figure 4 As shown in the figure, the structure diagram of the ResNet model is shown. The SRResNet training setting is: batch size is 16, initial learning rate is 0.0001, optimizer uses Adam, total training rounds is 200, and learning rate is decayed once every 50 epochs.

[0071] Reference Figure 5 As shown in the figure, the structure diagram of the EDSR model is shown. The EDSR model training parameter setting is: batch size is 8, learning rate is 0.0001, total epoch number is 300, and the rest of the training settings are the same as SRResNet.

[0072] Reference Figure 6 And Figure 7 As shown in the figure, the network structure diagrams of the ESRGAN generator and discriminator are shown respectively. The training of ESRGAN uses an alternating training method, and the generator is pre-trained only with L1 loss for the first 100 epochs, and then the perceptual loss and adversarial loss are introduced; The training parameter setting is: batch size is 8, learning rate is 0.0001, and total training rounds is 400.

[0073] Reference Figure 8 As shown in the figure, the structure diagram of the Real-ESRGAN discriminator is shown. The training setting of Real-ESRGAN is consistent with that of ESRGAN.

[0074] After the four models are trained on the training set respectively, they are used for inference on the test set, the input is a low-resolution image, and the output is a high-resolution reconstructed image, the resolution is restored to 512x512 pixels.

[0075] 2.4) Reference Figure 9The loss curve changes of SRResNet, EDSR, ESRGAN and Real-ESRGAN in the training process are shown. The convergence speed and stability of each model at different epochs are reflected in the figure, and Real-ESRGAN shows faster convergence speed and lower final loss value, which verifies its superior performance in the real rock image reconstruction task.

[0076] Reference Figure 10 As shown in the figure, the reconstruction results of SRResNet, EDSR, ESRGAN and Real-ESRGAN models on the same low-resolution rock image are compared. Real-ESRGAN performs best in fidelity and edge enhancement, and is particularly suitable for identifying complex mineral particles and pore structures in geological images.

[0077] 3) Color space conversion is performed on the image:

[0078] After completing the super-resolution image reconstruction, in order to enrich the feature expression ability of the image and enhance the texture extraction effect, various color space conversions are performed on the reconstructed high-resolution image, and then multi-channel images are extracted for subsequent texture analysis.

[0079] The input is the super-resolution image output by step 2), with a resolution of 512x512 and a PNG image format, and the image contains complete spatial detail information and mineral structure. In order to more comprehensively express the illumination, color and texture information in the image, the image is converted from the original RGB space to two typical color spaces, HSV and YCbCr, and five channels, H, V, R, Cb and C, are extracted. Among them, the R channel retains the red brightness information in the original image, which is suitable for areas with obvious color structure; the H channel reflects the type change of colors in the image, and is sensitive to the hue difference between quartz and matrix; the V channel enhances the overall brightness and shadow structure of the image, which is beneficial to capture low-brightness areas such as pores; the Cb and Cr channels highlight color difference details, which are particularly effective in identifying the boundaries of complex textures such as kaolinite and debris. Reference Figure 11 In order to evaluate the role of different color channels in distinguishing the ability of texture features, the features extracted by Gabor filtering of each channel image are subjected to principal component analysis (PCA) and visualized as a two-dimensional scatter plot. The results show that the R and H channels have high inter-class discrimination, especially in distinguishing kaolinite and pore categories, which provides a quantitative basis for subsequent channel fusion and feature selection.

[0080] 4) Multi-directional texture feature extraction is performed on each channel image using two-dimensional Gabor filters:

[0081] For each single channel image obtained in step 3), a two-dimensional Gabor filter is applied respectively for multi-direction and multi-frequency texture feature extraction, so as to obtain detailed local structure information, and to provide high-dimensional discriminant features for subsequent dimension reduction and classification.

[0082] The input is the five channel images (H, V, R, Cb, Cr) extracted in step 3), each with a size of 512x512 and a gray value range of 0-255. The two-dimensional Gabor filter function used is defined as follows:

[0083]

[0084] where x' and y' are the rotated coordinates x' = xcosθ + ysinθ, y' = -xsinθ + ycosθ, θ is the direction angle of the filter, and is set to 12 directions for filtering of each channel image to obtain a good direction response, with a value range of:

[0085]

[0086] σ x and σ y The spatial range of the filter is controlled, and is set to ω is the center frequency, and 4 frequencies are selected, i.e. 0.1, 0.2, 0.3, and 0.4; j is the imaginary unit, indicating the sensitivity of the filter to the phase.

[0087] For each color channel image, the Gabor filter is applied in combination with each direction and frequency, and a total of 48 texture response maps are obtained; 5 channels finally obtain 240 texture feature maps. Each response map maintains the same spatial size as the original channel Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 1 Figure 2 image, and then each image is flattened into a one-dimensional vector and spliced into a high-dimensional texture feature vector of a single image.

[0088] In the feature extraction process, each image is flattened into a feature vector with a dimension of 262144 after Gabor filtering. Since a total of 240 Gabor response maps are generated for each image, the total feature dimension is as high as 62914560.

[0089] 5) Principal component analysis is applied to the extracted texture features for dimension reduction, and the dimension-reduced feature vectors are input into a support vector machine classifier based on a radial basis function to realize recognition of the mineral type:

[0090] 5.1) To alleviate the problems of computational complexity, model overfitting, and classification efficiency reduction caused by high-dimensional Gabor texture features, principal component analysis (PCA) is applied to the texture features extracted in step 4) for dimension reduction processing. As known from step 4), the dimension of the texture features of a single image is:

[0091] 3 channels x 32 directions x frequency combination x 262144 pixels = 25165824 dimensions

[0092] The features of 800 images in the training set are spliced to form a training sample matrix:

[0093] X ∈ R 800×25165824

[0094] To avoid memory overflow, the matrix is loaded into memory in blocks, and sparse representation and incremental PCA are used for processing. Before PCA, the training sample matrix X is normalized column by column:

[0095]

[0096] where μ j and σ j are the mean and standard deviation of the j-th feature. After normalization, all features are centered at 0 with unit variance, which helps to correctly extract the principal component direction. Using the PCA method, the first k principal components with cumulative contribution rate greater than 90% are automatically selected. The PCA process includes: first, calculate the covariance matrix of the training set feature matrix, and perform eigenvalue decomposition, select the first k eigenvectors to form the transformation matrix W k ∈ R 25165824×k . Then perform the following linear transformation:

[0097] Z = X·W k

[0098] The original high-dimensional feature X is reduced to low-dimensional feature Z ∈ R 800×k , where k usually fluctuates between 40-60.

[0099] The Gabor features of the validation set and test set (100 images each) divided in step 1) are extracted using the same method, and the transformation matrix W k is used to project the features of the validation set and test set. Projection explanation: PCA maintains the classification performance while reducing the dimensionality by more than 99.999%, significantly reducing the computational cost. In the same low-dimensional space, the feature dimension is kept consistent to ensure uniformity of the classifier input.

[0100] By visualizing the principal component contribution rate curve and the cumulative information amount curve, it is determined that the first 50 principal components cover about 92.7% of the total information amount. The classification result at this threshold is comparable to the original non-reduced data in accuracy (difference <1%), but the calculation speed is significantly improved. The following Table 1 compares the results of the dimensionality reduction experiment.

[0101] Table 1 Comparison of dimensionality reduction experiment results

[0102]

[0103] The high-dimensional texture features are reduced in dimension by principal component analysis, effectively compressing the redundant feature dimension, improving the calculation efficiency and model training stability, and avoiding the convergence difficulty of the classifier in high-dimensional space. This step is a key bridging link in the process of the application, and has an important supporting role for the subsequent classification accuracy and overall system performance.

[0104] 5.2) After the high-dimensional Gabor multi-directional texture features extracted from the image are processed by principal component analysis, the low-dimensional feature vector (about 50 dimensions) corresponding to each image is formed, which constitutes the input data of the training set (800 samples) and the test set (100 samples).

[0105]

[0106] k(x i , x j ) = exp(-γ||x i -x i ||) -5

[0107] where x -3 and x 15 are feature vectors; γ is an adjustment parameter, which controls the complexity of the decision boundary;

[0108] The model learns the classification boundary by optimizing the following objective function:

[0109]

[0110] where a -15 is the Lagrange multiplier, y -13 is the label of the support vector, b is the bias term, and f(x) reflects the distance of the sample from the hyperplane, and its absolute value |f(x)| is used as the confidence evaluation of sample classification.

[0111] In order to obtain the best classification performance, the two key hyperparameters of SVM are optimized by grid search: (1) penalty factor C: control the tolerance of the model to misclassification;

[0112] (2) Kernel width γ: affect the complexity of the model boundary.

[0113] The search space is set as:

[0114] C∈{2 3 , 2 ,..., 2

[0001] } γ∈{2 2 ,..., 2

[0002] } ​​

[0115] The accuracy of each set of parameter combinations was evaluated in the training set (800 samples) using 5-fold cross-validation, and the optimal combination was finally selected for model training.

[0116] 6) The performance of the SVM classification results was evaluated, and the overall classification accuracy and the identification accuracy of each lithology type were output:

[0117] After the model training was completed, the test set samples were input into the classifier to obtain the predicted mineral labels corresponding to each image. Subsequently, the error statistics of the five types of mineral classification results were performed, and the precision (Precision), recall (Recall), F1-score, and sample support (Support) of each class were output, and the overall accuracy (Accuracy) was calculated for performance evaluation.

[0118] The experimental results show that, based on the enhanced images based on Real-ESRGAN and Gabor+PCA features, the SVM classification model has the following performance in the five types of lithology identification task as shown in Table 2.

[0119] Table 2 Classification performance evaluation results of five types of compact sandstone minerals

[0120]

[0121] The overall classification accuracy reached 85.9%, among which the identification effect of Kaolinite and Quartz was the best, showing high structural recognition ability; in the complex texture overlapping and weak contrast areas of debris and pore regions, the F1 value also maintained above 80%, which was significantly better than the SVM classification performance of the unenhanced image (generally below 75%). In addition, the distribution of |f(x)| showed that the texture image after super-resolution enhancement had clearer class boundaries, and the confidence of the classification model was overall improved.

Claims

1. A rock image classification method based on super-resolution and texture feature fusion, characterized in that, Includes the following steps: 1) Acquire images of thin sections of dense sandstone, generate the original high-resolution image set, and downsample them into a low-resolution image set; 2) Input the low-resolution image into the super-resolution model for image reconstruction; 3) Perform color space conversion on the image; 4) Use a two-dimensional Gabor filter to extract multi-directional texture features for each channel image; 5) Principal component analysis is applied to reduce the dimensionality of the extracted texture features. The reduced feature vectors are then input into a support vector machine classifier based on radial basis functions to identify mineral types. 6) Evaluate the performance of the support vector machine classification results and output the overall classification accuracy and the identification accuracy of each lithology type.

2. The method according to claim 1, characterized in that, In step 1), the scanned images of the collected dense sandstone thin sections are used to extract a mineral fragment image set as the original high-resolution image set through superpixel segmentation and region merging techniques; then, the original high-resolution image set is downsampled into a low-resolution image set using bicubic interpolation.

3. The method according to claim 1, characterized in that, The image reconstruction in step 2) involves constructing a pairwise training set of low-resolution images and corresponding high-resolution images, and using supervised learning to train a super-resolution model to learn the mapping relationship between low-resolution and high-resolution images. The trained super-resolution model can generate a super-resolution image with high resolution and enhanced details based on the input low-resolution image.

4. The method according to claim 1, characterized in that, The super-resolution models used for image reconstruction in step 2) include SRResNet, EDSR, ESRGAN, and Real-ESRGAN. These models are trained by constructing supervised pairing relationships between high-resolution images and corresponding low-resolution images. The loss functions used during training include L1 loss, perceptual loss, and adversarial loss. SRResNet and EDSR models use pixel-level L1 loss. in, Represents the i-th original high-resolution image. This represents the output image of the super-resolution model, where n represents the number of samples. The ESRGAN and ReaI-ESRGAN models use a total loss function: L G =L Percep +λL Ra_G +ηL1 Among them, L Percep L represents perceived loss Ra_G It is a relative adversarial loss, L1 is the pixel-level L1 loss; λ and η are weighting coefficients.

5. The method according to claim 1, characterized in that, The two-dimensional function expression of the Gaborr filter in step 4) is as follows: Where x and y are the image space coordinates, x′ and y′ are the rotated coordinates, x′ = xcosθ + ysinθ, y′ = -xsinθ + ycosθ; σ x and σ y ω is the standard deviation of the Gaussian function in the x and y directions; θ is the center frequency; θ is the direction angle of the filter; and j is the imaginary unit.

6. The method according to claim 1, characterized in that, The range of the direction angle θ for multi-directional texture feature extraction in step 4) is: Used to extract texture features in 12 directions.

7. The method according to claim 1, characterized in that, The principal component analysis in step 5) is used to reduce the dimensionality of the multidimensional texture features extracted from the multidirectional GaboT filter, retaining principal components with a cumulative contribution of more than 90%, in order to compress the feature dimension and enhance the stability of feature expression.

8. The method according to claim 1, characterized in that, The classifier in step 5) is a support vector machine, which uses a radial basis function kernel. The kernel function expression is as follows: k(x i ,y i )=exp(-γ||x i -y j || 2 ) Where, x i and x j γ is the feature vector; γ is the adjustment parameter; The decision function of the classifier SVM is as follows: Where, α l It is a Lagrange multiplier, y i These are the labels of the support vectors, k(x) i f(x) is the kernel function value, b is the bias term, and its absolute value |f(x)| is used as the confidence assessment of the sample classification.

9. The method according to claim 1, characterized in that, The performance evaluation in step 6) includes using the decision function of the support vector machine to calculate the distance from each test sample to the classification hyperplane as the confidence basis, generating precision, recall and F1 score, which are used to quantitatively analyze the identification effect of various lithological types and compare the classification capabilities of different super-resolution models.

Images