A rock image classification method based on frequency domain transformation and spatial joint representation

Abstract

The application discloses a rock image classification method based on frequency domain transformation and space joint representation. The method first pre-processes a rock image to obtain a standardized input image; then performs frequency domain transformation on the input image, extracts frequency domain features, and extracts space features representing texture structure, edge contour and local gray level change; then the frequency domain features and the space features are jointly represented and fused to construct a fusion feature vector, which is input into a classification model after dimension reduction mapping, and a rock image class label is output; finally, the performance of the classification result is evaluated in combination with a real label. The method realizes the collaborative representation of frequency domain information and space structure information, and improves the accuracy and stability of classification and recognition in a complex rock image scene.

Description

Technical Field

[0001] This invention belongs to the field of rock image processing and intelligent classification and recognition technology, specifically involving a rock image classification method based on frequency domain transformation and spatial joint representation. Background Technology

[0002] In oil and gas exploration, reservoir evaluation, and lithological analysis, accurately acquiring rock image category information is crucial for rock composition identification, reservoir characteristic analysis, and subsequent geological interpretation. With the development of digital imaging and rock image acquisition technologies, the scale of rock image data is constantly increasing. Traditional classification methods relying on manual observation and experience are inefficient and easily influenced by subjective factors. While existing rock image classification methods can achieve target recognition to some extent, they still suffer from insufficient feature representation and classification stability in situations with complex texture distributions, significant differences in grain structure, and substantial changes in frequency components, making it difficult to meet the intelligent classification requirements of complex rock image scenarios. Therefore, constructing a rock image classification method based on frequency domain transformation and spatial joint representation is of great significance.

[0003] To promote the development of intelligent rock image classification technology, the academic community has conducted various technical explorations on rock image feature extraction and automatic recognition. For example, Su et al. proposed a rock thin section image classification method based on cascaded convolutional neural networks in the paper "Rock classification in petrographic thinsection images based on concatenated convolutional neural networks". By combining polarized light microscopic thin section image information, they achieved automatic rock type identification and achieved good results in thin section image classification. However, their method mainly focuses on deep convolutional feature learning and pays relatively little attention to the joint representation of frequency domain information and spatial features. Aligholi et al. used texture description methods such as local binary patterns in the paper "Mineral Texture Identification Using Local Binary Patterns Equipped with a Classification and Recognition Updating System (CARUS)" to achieve mineral texture identification and improve texture classification efficiency. However, their method mainly relies on local texture features in the spatial domain and still has limitations in the coordinated use of frequency distribution information and spatial structure information in rock images. Therefore, in scenarios with complex texture distribution and significant differences in grain structure, there is still room for further improvement in classification stability and feature representation ability.

[0004] Based on the above technical background, it can be seen that existing rock image classification methods still have the following shortcomings: First, existing methods mostly focus on the extraction and utilization of single spatial domain features or single frequency domain features, making it difficult to comprehensively represent the texture structure, edge contours, and frequency distribution information in rock images; Second, some classification methods mainly rely on deep feature learning or local texture description, which fails to adequately explore the complementary relationship between spatial structure information and frequency component information in rock images, resulting in limited feature expression capabilities in complex scenes; Third, when rock images have complex texture distributions, significant differences in particle structure, and similar features between categories, existing methods are prone to problems such as insufficient feature extraction, insufficient classification stability, and decreased category discrimination ability.

[0005] To address the aforementioned issues, this invention proposes a rock image classification method based on frequency domain transformation and spatial joint representation. By performing frequency domain transformation on rock images to extract frequency distribution and energy change features, and combining this with local texture features in the spatial domain for joint representation and fusion, classification input features are constructed and rock image category determination is completed. This enables the synergistic utilization of frequency information and spatial structure information of rock images, thereby improving the classification accuracy, stability, and intelligent analysis capabilities in complex rock image scenarios. Summary of the Invention

[0006] To overcome the shortcomings of existing rock image classification methods, such as insufficient utilization of single spatial or frequency domain features, difficulty in fully representing the texture structure and frequency distribution information in rock images, and low classification accuracy and stability under conditions of complex texture distribution, significant differences in particle structure, and similar features between categories, this invention proposes a rock image classification method based on joint representation of frequency domain transformation and spatial domain. This method first preprocesses the rock image to obtain a standardized input image for classification; then, it performs a frequency domain transformation on the input image to extract frequency domain features representing the image's frequency distribution and energy changes; simultaneously, it extracts spatial features from the input image to obtain spatial features representing the local texture structure information of the rock; further, it jointly represents and fuses the frequency domain features and spatial features, and constructs classification input features through dimensionality reduction mapping, inputting these features into the classification model to output the corresponding rock category label. By combining the complementary advantages of frequency domain information and spatial structure information, a more complete expression of rock image features is achieved, thereby improving the accuracy, stability, and intelligent analysis capabilities of classification and recognition in complex rock image scenarios. To solve the above technical problems, the technical solution adopted by this invention is:

[0007] 1. A rock image classification method based on frequency domain transformation and spatial joint representation, characterized by comprising the following steps:

[0008] 1) Acquire rock images, preprocess the rock images to obtain standardized input images for classification;

[0009] 2) Perform frequency domain transformation processing on the standardized input image to extract frequency domain features that characterize the frequency distribution and energy changes of the rock image;

[0010] 3) Based on the standardized input image, extract spatial features representing the texture structure, edge contours, and local grayscale changes of the rock image;

[0011] 4) Jointly represent and fuse the frequency domain features and spatial features to construct a fused feature vector for classification and recognition;

[0012] 5) Input the fused feature vector into the classification model and output the corresponding rock image category label;

[0013] 6) Based on the category labels and the real labels, perform a classification performance evaluation to obtain the classification evaluation results of the method.

[0014] 2. The method according to claim 1, wherein the preprocessing in step 1) includes: performing size unification, grayscale normalization and noise reduction on the original rock image to reduce the impact of image acquisition noise and brightness distribution differences on subsequent frequency domain feature extraction and spatial feature extraction.

[0015] 3. The method according to claim 1, characterized in that the frequency domain transformation in step 2) employs a two-dimensional discrete Fourier transform to map the input image from the spatial domain to the frequency domain, obtaining the corresponding spectral representation:

[0016]

[0017] Where I(x, y) represents the input image, F(u, v) represents the frequency domain transformed spectral coefficients, (u, v) represents the frequency domain coordinates, and M and N represent the width and height of the image, respectively; the amplitude spectrum is calculated from the spectral coefficients to obtain:

[0018] A(u, v) = |F(u, v)|

[0019] The amplitude spectrum A(u, v) is used to characterize the energy distribution features of the rock image in different frequency regions; frequency domain feature extraction includes: performing energy statistics on the low-frequency, mid-frequency, and high-frequency regions based on the amplitude spectrum to obtain the frequency domain energy features of the corresponding frequency bands; wherein, the energy of the k-th frequency band region is defined as:

[0020]

[0021] Among them, Ω k E represents the frequency domain region corresponding to the k-th frequency band. kThis represents the energy statistics within the frequency band; the energy statistics of each frequency band are combined to form a frequency domain feature vector:

[0022] F freq = [E1, E2, ..., E n ] T

[0023] The frequency domain feature vector is used to describe the differences in frequency components and structural periodicity of the texture distribution in rock images.

[0024] 4. The method according to claim 1, characterized in that the spatial feature extraction in step 3) includes local texture pattern encoding, used to enhance the characterization ability of differences in rock grain texture structure; the local texture pattern encoding adopts a local binary pattern operator, the expression of which is:

[0025]

[0026] Among them, g c g represents the grayscale value of the center pixel. p Let s(·) represent the grayscale value of the p-th pixel in the neighborhood, where P represents the number of sampling points in the neighborhood. The sign function s(·) is defined as:

[0027]

[0028] The local binary pattern features are used to describe the local texture distribution patterns of rock images.

[0029] 5. The method according to claim 1, characterized in that the joint representation in step 4) adopts a weighted fusion method, uniformly encoding the frequency domain features and spatial features to construct a fused feature vector; the fused feature vector is expressed as:

[0030]

[0031] Among them, F fusion F represents the fused feature vector. freq F represents the frequency domain eigenvector. spa Let represent the spatial feature vector, and α and β represent the weighting coefficients of the frequency domain features and spatial features, respectively. Indicates feature splicing;

[0032] After feature fusion, the fused feature vectors undergo dimensionality reduction mapping to reduce feature redundancy and improve classification ability. The dimensionality reduction mapping employs principal component analysis, and its projection expression is as follows:

[0033] z = W T F fusion

[0034] Among them, F fusionLet represent the fused feature vector, W represent the principal component projection matrix, and z represent the dimensionality-reduced joint representation feature; the dimensionality-reduced joint representation feature is used as the input to the classification model.

[0035] 6. The method according to claim 1, wherein the classification model in step 5) is a support vector machine classification model, and the support vector machine constructs an optimal classification hyperplane based on joint representation features, and its decision function is expressed as:

[0036]

[0037] Where z represents the feature of the sample to be classified, z i α represents the features of the training samples. i Denotes Lagrange multipliers, y i Let K(·,·) represent the training sample label, K(·,·) represent the kernel function, and b represent the bias term; the decision function is used to output the category label of the rock image.

[0038] 7. A rock image classification method based on frequency domain transformation and spatial joint representation according to claim 1, characterized in that the spatial features include local texture features, edge contour features and gray-level statistical features; wherein, the local texture features are extracted using a local binary mode, and the edge contour features are calculated based on an edge detection operator; the gray-level statistical features include gray-level mean, gray-level standard deviation, skewness and kurtosis.

[0039] Compared with existing technologies, the present invention has the following significant technical effects and advantages:

[0040] (1) Improve the comprehensiveness of rock image feature representation: By combining and fusing the frequency distribution and energy change features obtained by frequency domain transformation with the local texture structure features in the spatial domain, the complementary features of rock images at both the frequency information and spatial structure information levels can be utilized to enhance the overall descriptive ability of complex rock images.

[0041] (2) Improve the classification accuracy and stability in complex rock image scenarios: In response to the complex texture distribution, obvious differences in particle structure and similar features between categories in rock images, this invention effectively improves the ability to distinguish different rock categories and the classification stability by using the synergistic expression of frequency domain features and spatial features, combined with dimensionality reduction mapping and classification model discrimination.

[0042] (3) Enhance the applicability and analytical capability of intelligent rock image classification methods: By uniformly modeling frequency domain features, spatial features and their fusion features, this invention can more fully explore the texture variation patterns, structural distribution features and frequency component differences in rock images, thereby improving the reliability of automatic classification results and providing more effective technical support for the intelligent analysis of complex rock images.

[0043] In summary, this invention constructs a rock image classification process consisting of image preprocessing, frequency domain feature extraction, spatial feature extraction, joint characterization and feature fusion, classification and recognition, and performance evaluation. This enables the synergistic utilization of frequency and spatial structure information in rock images, improves the accuracy, stability, and intelligence level of classification and recognition in complex rock image scenarios, and provides a technical solution with practical application value for automatic rock image analysis. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall process of a rock image classification method based on frequency domain transformation and spatial joint representation.

[0045] Figure 2 Schematic diagram of the textures of five typical rock and mineral types;

[0046] Figure 3 This is a schematic diagram of spatial feature extraction and joint characterization of rock images;

[0047] Figure 4 This is a schematic diagram of frequency domain transformation and frequency domain feature extraction of rock images;

[0048] Figure 5 A schematic diagram of image sample construction and annotation. Detailed Implementation

[0049] Example 1:

[0050] See Figure 1 The rock image classification method based on frequency domain transformation and spatial joint representation of the present invention includes the following steps:

[0051] 1) Acquire rock images, preprocess the rock images to obtain standardized input images for classification;

[0052] 2) Perform frequency domain transformation processing on the standardized input image to extract frequency domain features that characterize the frequency distribution and energy changes of the rock image;

[0053] 3) Based on the standardized input image, extract spatial features representing the texture structure, edge contours, and local grayscale changes of the rock image;

[0054] 4) Jointly represent and fuse the frequency domain features and spatial features to construct a fused feature vector for classification and recognition;

[0055] 5) Input the fused feature vector into the classification model and output the corresponding rock image category label;

[0056] 6) Based on the category labels and the real labels, perform a classification performance evaluation to obtain the classification evaluation results of the method.

[0057] 1) Acquire rock images, preprocess the rock images to obtain standardized input images for classification;

[0058] 1.1) Rock Image Acquisition: Acquire rock image samples to be classified, wherein all rock image samples can characterize the texture and structural features of rocks. (Reference) Figure 2 The five typical rock and mineral texture samples shown have certain differences in the clarity of grain boundaries, local texture distribution and grayscale variation characteristics of different categories of images. However, there are also cases where the textures between categories are similar and the local structures are complex. Therefore, the original images need to be preprocessed in a unified manner before being input into the subsequent classification process.

[0059] 1.2) Size Unification and Gray-Level Normalization: Due to differences in resolution, size, and brightness distribution among different original rock images, this embodiment first uniformly adjusts all images to 224×224 pixels to reduce the impact of inconsistent input scales on subsequent feature extraction results. After size unification, gray-level normalization is performed on the images to reduce the influence of image acquisition conditions, lighting intensity, and background brightness variations. The normalized image is denoted as I(x, y), where (x, y) represents the pixel coordinates in the image, and I(x, y) represents the gray-level value at the corresponding position. Its expression is:

[0060]

[0061] Where I0(x, y) represents the grayscale value of the original image, I max and I min These represent the minimum and maximum grayscale values ​​in the current image, respectively.

[0062] 1.3) Image Denoising Processing: After size unification and grayscale normalization, Gaussian filtering is used to smooth the image, so as to better preserve the main texture structure and grain boundary information in the rock image while suppressing noise. After denoising processing, a standardized input image is obtained for subsequent frequency domain feature extraction and spatial feature extraction.

[0063] After the above steps, rock image input samples with consistent size, stable grayscale distribution and controlled noise are obtained, providing a reliable data foundation for subsequent frequency domain transformation, spatial feature extraction and joint characterization.

[0064] 2) Perform frequency domain transformation processing on the standardized input image to extract frequency domain features that characterize the frequency distribution and energy changes of the rock image;

[0065] The standardized rock images obtained after preprocessing in step 1) were selected as input for frequency domain analysis. Each preprocessed image was uniformly 224×224 pixels in size, with grayscale values ​​normalized to the [0, 1] interval. To extract the differences in texture periodicity, grain density, and local detail variations among different rock types, a two-dimensional discrete Fourier transform was performed on each image, and the energy distribution within different frequency bands was statistically analyzed.

[0066] 2.1) Two-dimensional Discrete Fourier Transform Processing: Perform a two-dimensional discrete Fourier transform on the input image I(x, y) to obtain the spectrum F(u, v) coefficients:

[0067]

[0068] Where (x, y) represents the spatial domain pixel coordinates, and (u, v) represents the frequency domain coordinates. After performing frequency domain transformation on the images of the five types of rocks, it was found that samples with clearer grain boundaries and denser local textures have more obvious energy responses in the mid-to-high frequency regions, while samples with more uniform textures and slower changes have more energy concentrated in the low frequency region.

[0069] 2.2) Spectrum Centering and Amplitude Spectrum Calculation: To facilitate subsequent frequency band division and energy statistics, the Fourier transform results are subjected to spectrum centering, moving the low-frequency components to the center of the spectrum. The amplitude spectrum is then calculated.

[0070] A(u, v) = |F(u, v)|

[0071] Because of the large dynamic range of the amplitude spectrum, a logarithmic enhancement method is used for display and analysis in actual processing. That is, the frequency distribution is observed after taking the logarithm of the amplitude spectrum. This step can more clearly distinguish the response differences of different types of rock images in the low-frequency, mid-frequency, and high-frequency regions.

[0072] 2.3) Frequency Band Division: In this embodiment, the spectrum is divided into three concentric frequency band regions based on the distance from the center, using the center of the spectrum as the center: low frequency band, mid frequency band, and high frequency band. The specific division method is as follows:

[0073] Low-frequency region Ω1: radius from the center of the spectrum not greater than 20 pixels;

[0074] Mid-frequency region Ω2: The radius from the center of the spectrum is greater than 20 pixels and not greater than 60 pixels;

[0075] High-frequency region Ω3: The radius from the center of the spectrum is greater than 60 pixels.

[0076] The above division method is determined for a 224×224 image size and can effectively distinguish overall structural information, transitional texture information, and local detail information.

[0077] 2.4) Band Energy Statistics: The energy values ​​within each of the three band regions are statistically analyzed. Band energy is defined as:

[0078]

[0079] Where E1, E2, and E3 represent the energy statistics of the low-frequency, mid-frequency, and high-frequency regions, respectively. In this embodiment, the energy values ​​of the three frequency bands of each image are arranged sequentially to form a frequency domain feature vector:

[0080] F freq =[E1, E2, E3] T

[0081] To avoid the overall brightness levels of different images having an excessive impact on the frequency domain energy statistics, E1, E2, and E3 are further normalized in actual processing, so that the sum of the energies of the three frequency bands is 1. After normalization, the frequency domain features are more suitable for comparison between different categories of images and for subsequent feature fusion.

[0082] 2.5) Frequency Domain Feature Extraction Results: After the above processing, each rock image yields a 3D frequency domain feature vector, which describes the energy distribution ratio in the low-frequency, mid-frequency, and high-frequency regions of the image. This feature reflects the frequency differences in overall structure, local texture, and edge details of different rock types, providing a frequency domain discrimination basis for subsequent classification.

[0083] 3) Based on the standardized input image, extract spatial features representing the texture structure, edge contours, and local grayscale changes of the rock image;

[0084] In this embodiment, to address the issue of insufficient description of local texture patterns and particle boundaries by frequency domain features, spatial domain features are further extracted from the standardized input image obtained in step 1).

[0085] 3.1) Local Texture Feature Extraction: Texture encoding is performed on each standardized input image of size 224×224 using the Local Binary Pattern Operator. In this embodiment, the radius R = 1 and the number of neighborhood sampling points P = 8 are selected for the Local Binary Pattern Operator. That is, grayscale comparison is performed on the 3×3 neighborhood of each pixel to obtain the corresponding local texture pattern value, the expression of which is:

[0086]

[0087] Among them, g c g represents the grayscale value of the center pixel. p This represents the grayscale value of the p-th neighboring pixel. A histogram of the LBP encoding results for the entire image is plotted and divided into 16 intervals, thus forming a 16-dimensional local texture feature vector.

[0088] 3.2) Edge contour feature extraction: The Sobel operator is used to calculate the gradient components of the image in the horizontal and vertical directions, respectively, and a gradient magnitude map is obtained to characterize the changes in grain boundaries and structural transition regions in the rock image. In this embodiment, the mean, standard deviation, and maximum value of the gradient magnitude map are further calculated to form a 3D edge contour feature vector, which is used to describe the overall edge intensity and edge distribution differences of the image.

[0089] 3.3) Gray-level statistical feature extraction and spatial feature vector construction: The standardized input image is further analyzed to obtain a 4-dimensional gray-level statistical feature vector, which includes the gray-level mean, standard deviation, skewness, and kurtosis. Subsequently, the 16-dimensional local texture features, 3-dimensional edge contour features, and 4-dimensional gray-level statistical features are concatenated to construct a 23-dimensional spatial feature vector F. spa In this embodiment, the spatial feature vector is standardized to have a mean of 0 and a standard deviation of 1, in order to reduce the impact of differences in feature dimensions on subsequent fusion and classification.

[0090] 4) Jointly represent and fuse the frequency domain features and spatial features to construct a fused feature vector for classification and recognition.

[0091] In this embodiment, the frequency domain feature vector extracted in step 2) is 3-dimensional, and the spatial feature vector extracted in step 3) is 23-dimensional. To fully utilize the complementarity of the two types of features in terms of frequency distribution information and local texture structure information, it is necessary to jointly represent and fuse the frequency domain features and spatial features to construct a fused feature vector for classification and recognition. The process is as follows: Figure 3 As shown. Specifically, it includes the following steps:

[0092] 4.1) Feature Standardization: Due to the different numerical ranges and dimensions of frequency domain features and spatial features, standardization is performed on both types of features before fusion. In this embodiment, Z-score standardization is used to ensure that each feature dimension has a mean of 0 and a standard deviation of 1, thereby reducing the problem of one type of feature dominating the fusion process due to its larger numerical range. After standardization, the frequency domain feature vector obtained in step 2) is denoted as F. freq The spatial feature vector obtained in step 3) is denoted as F. spa .

[0093] 4.2) Feature-weighted fusion: combining Figure 3 and Figure 4 As shown, this embodiment obtains the frequency domain energy distribution features and spatial texture structure features of the image, respectively. To enhance the joint expressive power of the two types of features, the frequency domain features and spatial features are fused in a weighted manner. The weight of the frequency domain features is set to 0.4, and the weight of the spatial features is set to 0.6. The fusion expression is as follows:

[0094]

[0095] in, This indicates a feature concatenation operation. Since frequency domain features are 3-dimensional and spatial features are 23-dimensional, the initial fused feature vector obtained after fusion has a dimension of 26. In this embodiment, the spatial feature weight is slightly higher than the frequency domain feature weight, mainly because local texture patterns and edge structure information have a more direct discriminative effect on category differentiation during rock image classification.

[0096] 4.3) Construction of Fusion Feature Vector: After weighted fusion, a 26-dimensional fusion feature vector is obtained for each rock image. This fusion feature vector contains energy distribution information in the low-frequency, mid-frequency, and high-frequency regions of the image, as well as local texture patterns, edge contour changes, and gray-level statistical distribution information, which can more comprehensively represent the differences between different types of rock images.

[0097] 5) Input the fused feature vector into the classification model and output the corresponding rock image category label.

[0098] In this embodiment, the 26-dimensional fused feature vector obtained in step 4) is input into the classification model to identify the categories of different rock images and output the corresponding category labels. Specifically, the steps include:

[0099] 5.1) Feature Dimensionality Reduction: To reduce redundant information in the fused features and improve classification efficiency, in this embodiment, principal component analysis (PCA) is first performed on the 26-dimensional fused feature vector for dimensionality reduction. Specifically, the covariance matrix is ​​calculated based on the feature matrix of the training samples, and the principal components with a cumulative contribution rate of 95% are selected as the new feature space. After dimensionality reduction, the original 26-dimensional fused features are compressed into a 12-dimensional joint representation feature vector. This process can reduce feature redundancy and noise interference while preserving as much effective information as possible from the original features.

[0100] 5.2) Classification Model Training: In this embodiment, a Support Vector Machine (SVM) classifier is used as the classification model, and the Radial Basis Function (RBF) is adopted as the kernel function. During training, the... Figure 5The constructed and labeled image sample set shown is divided into a training set and a test set in an 8:2 ratio, where the training set is used for model training and the test set is used for performance validation. The penalty parameter for the Support Vector Machine (SVM) is also shown. The kernel parameter γ is set to 0.1 and the value is set to 10. A five-class classification model is constructed using the 12-dimensional joint representation feature vector and class labels corresponding to the training samples.

[0101] 5.3) Category Label Output: After model training is complete, the 12-dimensional joint representation features corresponding to the test samples are input into the support vector machine classifier, which outputs the category label of each rock image. In this embodiment, the output category labels are set to 5 categories, corresponding to... Figure 2 The five typical rock and mineral texture types shown demonstrate how automatic category determination can be achieved for rock images to be classified.

[0102] 6) Based on the category labels and the real labels, perform a classification performance evaluation to obtain the classification evaluation results of the method.

[0103] In this embodiment, to verify the effectiveness of the rock image classification method based on frequency domain transformation and spatial joint representation, the predicted category label output in step 5) is compared with... Figure 5 The true category labels obtained during the sample labeling stage are compared to calculate the classification performance evaluation index. After inputting the test set samples into the trained support vector machine classification model, the corresponding predicted category labels are obtained, and the various classification evaluation indexes are calculated as shown in Table 1.

[0104] Table 1 Evaluation Indicators for Rock Image Classification Results Based on Frequency Domain Transform and Spatial Joint Representation

[0105]

[0106] Experimental results show that after using the joint representation of frequency domain features and spatial features, the model can achieve relatively stable classification and recognition of five types of rock images, and the precision, recall and F1 score are also well balanced, indicating that the method of the present invention has a good classification effect.

Claims

1. A rock image classification method based on frequency domain transform and spatial joint representation, characterized in that, Includes the following steps: 1) Acquire rock images, preprocess the rock images to obtain standardized input images for classification; 2) Perform frequency domain transformation processing on the standardized input image to extract frequency domain features that characterize the frequency distribution and energy changes of the rock image; 3) Based on the standardized input image, extract spatial features representing the texture structure, edge contours, and local grayscale changes of the rock image; 4) Jointly represent and fuse the frequency domain features and spatial features to construct a fused feature vector for classification and recognition; 5) Input the fused feature vector into the classification model and output the corresponding rock image category label; 6) Based on the category labels and the real labels, perform a classification performance evaluation to obtain the classification evaluation results of the method.

2. The method according to claim 1, characterized in that, The preprocessing in step 1) includes: performing size unification, grayscale normalization, and noise reduction on the original rock image to reduce the impact of image acquisition noise and brightness distribution differences on subsequent frequency domain feature extraction and spatial feature extraction.

3. The method according to claim 1, characterized in that, The frequency domain transformation in step 2) employs a two-dimensional discrete Fourier transform to map the input image from the spatial domain to the frequency domain, obtaining the corresponding spectral representation: Where I(x, y) represents the input image, F(u, v) represents the frequency domain transformed spectral coefficients, (u, v) represents the frequency domain coordinates, and M and N represent the width and height of the image, respectively; the amplitude spectrum is calculated from the spectral coefficients to obtain: A(u, v) = |F(u, v)| The amplitude spectrum A(u, v) is used to characterize the energy distribution features of the rock image in different frequency regions; frequency domain feature extraction includes: performing energy statistics on the low-frequency, mid-frequency, and high-frequency regions based on the amplitude spectrum to obtain the frequency domain energy features of the corresponding frequency bands; wherein, the energy of the k-th frequency band region is defined as: Among them, Ω k E represents the frequency domain region corresponding to the k-th frequency band. k This represents the energy statistics within the frequency band; the energy statistics of each frequency band are combined to form a frequency domain feature vector: F freq [E1, E2, ..., E] n ] T The frequency domain feature vector is used to describe the differences in frequency components and structural periodicity of the texture distribution in rock images.

4. The method according to claim 1, characterized in that, The spatial feature extraction in step 3) includes local texture pattern encoding, which enhances the ability to characterize differences in rock grain texture structure; the local texture pattern encoding uses a local binary pattern operator, the expression of which is: Among them, g c g represents the grayscale value of the center pixel. p Let s(·) represent the grayscale value of the p-th pixel in the neighborhood, where P represents the number of sampling points in the neighborhood. The sign function s(·) is defined as: The local binary pattern features are used to describe the local texture distribution patterns of rock images.

5. The method according to claim 1, characterized in that, The joint representation in step 4) adopts a weighted fusion method, uniformly encoding the frequency domain features and spatial features to construct a fused feature vector; the fused feature vector is expressed as: Among them, F fusion F represents the fused feature vector. freq F represents the frequency domain eigenvector. spa Let represent the spatial feature vector, and α and β represent the weighting coefficients of the frequency domain features and spatial features, respectively. Indicates feature splicing; After feature fusion, the fused feature vectors undergo dimensionality reduction mapping to reduce feature redundancy and improve classification ability. The dimensionality reduction mapping employs principal component analysis, and its projection expression is as follows: with=W T F fusion Among them, F fusion Let represent the fused feature vector, W represent the principal component projection matrix, and z represent the dimensionality-reduced joint representation feature; the dimensionality-reduced joint representation feature is used as the input to the classification model.

6. The method according to claim 1, characterized in that, The classification model in step 5) is a support vector machine (SVM) classification model. The SVM constructs an optimal classification hyperplane based on joint representation features, and its decision function is expressed as: Where z represents the feature of the sample to be classified, z i α represents the features of the training samples. i Denotes Lagrange multipliers, y i Let K(·,·) represent the training sample label, K(·,·) represent the kernel function, and b represent the bias term; the decision function is used to output the category label of the rock image.

7. The rock image classification method based on frequency domain transformation and spatial joint representation according to claim 1, characterized in that, The spatial features include local texture features, edge contour features, and gray-level statistical features; wherein, the local texture features are extracted using a local binary mode, the edge contour features are calculated based on an edge detection operator, and the gray-level statistical features include gray-level mean, gray-level standard deviation, skewness, and kurtosis.

Images