Carbonate rock image labeling system based on deep learning technology

Abstract

The invention relates to the technical field of data processing, in particular to a carbonate rock image labeling system based on a deep learning technology. The system comprises a multi-source feature extraction unit, an atlas classification unit and a type identification unit. According to the method, technologies such as deep learning, multi-source data fusion and map intelligent identification are introduced into carbonate rock slice identification, and map identification marks of typical carbonate rock slice types are established by constructing map labels, establishing an identification method system and summarizing the correlation between the typical carbonate rock slice types and rock material components, structures and physical property parameters. According to the method, comprehensive analysis is carried out from the perspective of graph, spectrum and multi-source physical property parameter information, the technical route and the technology are feasible, improvement of the type identification precision is facilitated, traditional slice naked eye identification work is optimized and improved, and technical support can be provided for intelligent identification, nondestructive testing, in-situ parameter analysis and other work of carbonate rock slices.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a carbonate rock image labeling system based on deep learning technology. Background Art

[0002] Geology is a course that involves the composition, structure and tectonics of earth materials, the principles of geological action and the history of earth evolution. Mineral rocks (ores) are important record carriers of geological evolution history and are the "natural books" of geology. Observing rocks and conducting rock thin section identification are important contents and important skills in geology work, and are also an important basis for further carrying out geological action and geological process analysis, resource and environmental special application and other work. At present, most rock thin section identification mainly relies on a large number of geologists to use microscopes for visual identification. Rock and mineral identification personnel need a lot of training and experience accumulation to be competent for the job, and they face a lot of heavy repetitive work in their work. There are problems such as strong subjectivity of identification results, long judgment cycle, high judgment cost, and the inability to balance the energy and efficiency of identification personnel. At the same time, due to the differences in geological characteristics and development mechanisms of rocks in different regions, the identification levels of petrologists involved in the identification are often uneven, which will lead to inconsistent analysis, thus affecting subsequent work such as cross-source identification and comparative analysis; With the development of graphics and image processing technology, information extraction using rock thin section images has become an important research direction. Early thin section image information extraction was mainly based on image texture, color, grayscale and other information. It combined color information with texture description by using image enhancement processing, unsupervised classification, image density segmentation and other methods to assist expert knowledge decision-making and extract more obvious trace feature information in rock thin sections. A human-computer combined rock microstructure image system analysis method was proposed. However, since there are often certain differences in the material composition and particle size in thin sections, thin section graphics of different types and scales require different image processing methods for information extraction. The image processing calculation is complex, and the information extraction accuracy is not high and the efficiency is low, and it is impossible to obtain relatively stable calculation results. In recent years, the application of image intelligent processing technology in various fields has continued to deepen, and the classification and information extraction of rock images have made great progress. Neural network models and various numerical analysis methods have been gradually introduced into the research work of image processing and intelligent recognition of rock thin sections, which has achieved the automatic output of analysis results to a certain extent. However, the accuracy of the analysis results is greatly affected by the sample screening and image parameter extraction effects. Especially in the identification of carbonate rock thin sections, a single graphic parameter limits the recognition accuracy. In view of this, we propose a carbonate rock image labeling system based on deep learning technology. Summary of the invention

[0003] The present invention aims to solve the problems in the prior art that image processing calculations are complex and that single graphic parameters limit the accuracy of recognition. While extracting information through rock thin section microscopic images of different spatial scales, the present invention comprehensively utilizes cathode luminescence images and element spectrum surface scanning images to extract shape structure and material composition information, constructs a three-dimensional spatial grid of various types of rock thin sections, realizes analysis of multi-source and multi-level data in the context of the three-dimensional spatial grid, performs correlation analysis and type classification, realizes rock thin section type identification, and avoids the accuracy and mobility problems faced by a single data source.

[0004] To achieve the above-mentioned purpose, the present invention provides a carbonate rock image labeling system based on deep learning technology, including a multi-source feature extraction unit, a graph classification unit and a type recognition unit; The multi-source feature extraction unit is used to establish a carbonate rock particle image data set, use remote sensing image processing algorithms to extract spatial trace information from rock thin section microscopic images of different spatial scales, and comprehensively use cathode luminescence images and element spectrum surface scanning images to extract trace structure and material composition information, model the structural characteristics of carbonate rocks in three-dimensional space, and construct a three-dimensional spatial grid of various types of rock thin sections through transfer learning; The atlas classification unit is used to screen the structural features and material composition characteristic atlas information of carbonate rocks, preliminarily construct an atlas classifier, reflect the correlation between material composition and structural features through the Pearson correlation coefficient, summarize the content and spatial distribution of its material components, and explore its influencing factors; The type recognition unit is used to fuse the morphological structure and material composition characteristic map information in a unified three-dimensional space framework, establish the map fingerprint of the carbonate rock thin section, use the BP neural network model to take the map fingerprint as input, and comprehensively consider the coupling law of the carbonate rock thin section map information and the results of transfer learning to identify the rock thin section type, and the output result is the predicted carbonate rock thin section type.

[0005] As a further improvement of the technical solution, the multi-source feature extraction unit includes the following steps when extracting spatial trace information: Scan the two-dimensional digital images of various types of rock slices, classify and establish a graphic library to form a carbonate rock particle image data set; The structural framework is described graphically in the two-dimensional grid plane by using the morphology, texture, color, brightness and roughness characteristic parameters in the two-dimensional digital image, including the position and orientation changes, to obtain the average morphology of the image and extract the spatial trace information.

[0006] As a further improvement of the technical solution, the multi-source feature extraction unit comprehensively utilizes the cathode luminescence image and the element spectrum surface scanning image to extract the shape structure and material composition information, including the following steps: The electron beam parameters of the electron microscope are adjusted so that the electron beam is focused on the surface of the rock slice to excite the luminescent center in the sample. The energy spectrometer is installed on the electron microscope to record the X-ray signal emitted by the sample under the excitation of the electron beam, and the recorded X-ray signal is converted into a digital image. The grayscale value is used to represent the luminescence intensity. The thin-section microscopic image is first subjected to image denoising and image enhancement preprocessing to keep good edges and textures in the area where the carbonate rock particles are located, and to enhance the grayscale difference between the particle area and the background area in the image to form a cathode luminescence image and an element spectrum surface scanning image. Align the cathode luminescence image and the element spectrum scan image to fuse the information of the two images to generate a comprehensive image; Clustering algorithms are used to classify features, identify different mineral phases, and visualize the generation of trace composition maps and material composition maps.

[0007] As a further improvement of the technical solution, the information fusion of the two images to generate a comprehensive image adopts a multi-scale fusion algorithm, including the following steps: Read the cathode luminescence image and element spectrum scanning image, and normalize the image data to the same dynamic range; The feature points of the image are extracted using the feature point detection method, the feature points of the two images are matched, the transformation matrix is ​​calculated, one image is aligned to the other image, the aligned image is decomposed into sub-bands of different scales using wavelet transform, and the sub-bands of different scales are fused to form a comprehensive image.

[0008] As a further improvement of the present technical solution, the structural features include spatial trace information, trace structure and material composition information.

[0009] As a further improvement of the technical solution, the three-dimensional spatial grid of various types of rock slices is constructed through transfer learning, including the following steps: Read the composite image, convert the image data into a format suitable for 3D modeling, create a 3D mesh, and interpolate the composite image data into the 3D mesh; Transfer learning based on traces: According to the trace features, a trace model is constructed for transfer learning to adapt to the classification task of trace features; Material-based transfer learning: Based on the material composition information, a transfer model is constructed for transfer learning to adapt to the material composition classification task; Coupling-based transfer learning: Combine shape features and material composition information to extract coupling features, build a coupling model for transfer learning, and adapt to the classification task of coupling features; Model-based transfer learning: Combining shape features, material composition information and coupling features, a multimodal model is constructed for transfer learning to adapt to the classification task of multimodal features; The classification results are applied to a three-dimensional spatial grid to generate a three-dimensional model of the carbonate rock thin section.

[0010] As a further improvement of the technical solution, the graph classifier is preliminarily constructed in the graph classification unit, including the following steps: In a unified spatial scale, the structural features and material composition characteristics of the main rock slices are screened from the three-dimensional model according to the different optical characteristics of the internal biological framework, grains, and cement materials to form multiple single vector layers; The content and spatial distribution of material components of a single vector layer in the three-dimensional model are analyzed and corresponded to the marked carbonate rock thin section types.

[0011] As a further improvement of the technical solution, the output result of the BP neural network model in the type recognition unit is the predicted carbonate rock thin section type, including the following steps: Use Keras to build a BP neural network model, compile the model, specify the loss function, optimizer and evaluation indicators, use the atlas fingerprints of multiple carbonate rock sections as the training set data to train the model, use the test set data to evaluate the performance of the model, use the trained model to predict new carbonate rock sections, input the atlas fingerprint, and output the rock section type.

[0012] Compared with the prior art, the present invention has the following beneficial effects: In this carbonate rock image labeling system based on deep learning technology, deep learning, multi-source data fusion, and intelligent atlas recognition technologies are introduced into carbonate rock thin section identification. By constructing atlas labels, an identification method system is established, the correlation between typical carbonate rock thin section types and rock material composition, structure, and physical property parameters is summarized, and atlas recognition marks for typical carbonate rock thin section types are established. A comprehensive analysis is conducted from the perspectives of images, spectra, and multi-source physical property parameter information. The technical route and technology are feasible and conducive to improving the accuracy of type recognition. The "atlas fingerprint" of carbonate thin sections is input into the BP neural network model of the type recognition unit, and the predicted carbonate thin section type is output, which optimizes and improves the work of traditional thin section naked eye identification, and can provide technical support for the intelligent identification, non-destructive testing and in-situ parameter analysis of carbonate thin sections. Moreover, in modern geological teaching and work, the virtual simulation platform formed by the carbonate rock image labeling system based on deep learning technology plays a vital role. The system can more intuitively understand geological phenomena and provide technical support for petrology teaching and virtual simulation platform construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is the overall structural principle block diagram of the present invention; Figure 2 It is a demonstration schematic diagram of the present invention.

[0014] The meaning of each number in the figure is: 100. Multi-source feature extraction unit; 200. Graph classification unit; 300. Type recognition unit. DETAILED DESCRIPTION

[0015] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0016] See also Figure 1-Figure 2 As shown, this embodiment provides a carbonate rock image labeling system based on deep learning technology, including a multi-source feature extraction unit 100, a graph classification unit 200 and a type recognition unit 300; The multi-source feature extraction unit 100 is used to establish a carbonate rock particle image data set, extract spatial trace information from rock thin-section microscopic images of different spatial scales using remote sensing image processing algorithms, and extract trace structure and material composition information by comprehensively utilizing cathode luminescence images and element spectrum surface scanning images, model the structural characteristics of carbonate rocks (structural characteristics include spatial trace information, trace structure and material composition information) in three-dimensional space, and construct a three-dimensional spatial grid of various types of rock thin sections through transfer learning, so as to realize the analysis of multi-source and multi-level data in the context of three-dimensional spatial grid, and lay a spatial grid foundation for the subsequent fusion of carbonate rock structural data, spectral database and auxiliary attribute data; The atlas classification unit 200 is used to screen the structural features and material composition characteristic atlas information of carbonate rocks, preliminarily construct an atlas classifier, reflect the correlation between material composition and structural features through the Pearson correlation coefficient, summarize the content and spatial distribution of its material components, and explore its influencing factors; The type recognition unit 300 is used to fuse the morphological structure and material composition characteristic map information in a unified three-dimensional spatial framework, establish the map fingerprint of the carbonate rock thin section, use the BP neural network model to take the map fingerprint as input, and comprehensively consider the coupling law of the carbonate rock thin section map information and the results of transfer learning to identify the rock thin section type. The output result is the predicted carbonate rock thin section type, avoiding the accuracy and mobility problems faced by a single data source.

[0017] On the basis of the above, specifically: When extracting spatial trace information, the multi-source feature extraction unit 100 includes the following steps: Scan the two-dimensional digital images of various types of rock slices, classify and establish a graphic library to form a carbonate rock particle image data set. The source of rock slices can be collected from teaching, scientific research, and predecessors' accumulated data. The collected images are pre-processed, including denoising, contrast enhancement, cropping and other operations to ensure image quality. According to the standard rock type of the acquired image data and the annotation content, a graphic library is established to realize the classification of graphic data; Using the morphology, texture, color, brightness, and roughness characteristic parameters in the two-dimensional digital image, the structural framework is described graphically in the two-dimensional grid plane, including the position and orientation changes, and the average morphology of the image is obtained by superimposing, and the spatial trace information is extracted; Among them, when acquiring the characteristic parameters: extracting the morphological features of the rock slices such as shape, size, and boundary, extracting the texture features of the rock slices, such as the gray-level co-occurrence matrix GLCM, extracting the color features of the rock slices, such as the RGB value or the HSV value, extracting the brightness features of the rock slices, such as the gray value, and extracting the surface roughness features of the rock slices; When describing in a graphical way: determine the position of each feature in the two-dimensional grid, describe the orientation change of the feature, overlay the graphical descriptions of multiple features to form an average shape, and enrich the image parameters such as mean, variance, correlation, etc. through statistical methods.

[0018] Furthermore, the multi-source feature extraction unit 100 comprehensively utilizes the cathode luminescence image and the element spectrum scan image to extract the shape structure and material composition information, including the following steps: The electron beam parameters of the electron microscope are adjusted so that the electron beam is focused on the surface of the rock slice to excite the luminescent center in the sample. The energy spectrometer is installed on the electron microscope to record the X-ray signal emitted by the sample under the excitation of the electron beam, and the recorded X-ray signal is converted into a digital image. The grayscale value is used to represent the luminescence intensity. The thin-section microscopic image is first subjected to image denoising and image enhancement preprocessing to keep good edges and textures in the area where the carbonate rock particles are located, and to enhance the grayscale difference between the particle area and the background area in the image to form a cathode luminescence image and an element spectrum surface scanning image. Align the cathode luminescence image and the element spectrum scan image to fuse the information of the two images to generate a comprehensive image; Use clustering algorithms to classify features, identify different mineral phases, and visualize the generation of trace composition maps and material composition maps. During visualization, generate pseudo-color images based on clustering results and draw distribution contour maps of different elements or minerals; Trace structure: interpret the morphological features in the image, such as cracks, bedding, deformation zones, etc., and identify the cracks and bedding in the rock through morphological and texture feature analysis; Material composition: Interpret the material composition information in the image, such as the distribution of different elements, mineral phases, etc., and identify different mineral phases and their distribution through spectral and luminescence feature analysis.

[0019] It is worth noting that the information fusion of the two images to generate a comprehensive image adopts a multi-scale fusion algorithm, which includes the following steps: Read the cathode luminescence image and element spectrum scan image, normalize the image data to the same dynamic range, such as between 0 and 1, and perform denoising on the image to reduce the impact of noise on subsequent processing; Feature point detection methods such as SIFT and ORB are used to extract feature points of images, match feature points of two images, calculate transformation matrices, align one image to another, decompose the aligned image into sub-bands of different scales using wavelet transform, and fuse sub-bands of different scales to form a comprehensive image; this method changes the limitations of traditional visual identification and single graphic recognition such as the "blind men touching an elephant", conducts a comprehensive analysis from the perspectives of graphs, spectra and multi-source physical property parameter information, and the technical route and technology are feasible. The research results are conducive to improving the accuracy of type recognition.

[0020] Through transfer learning, a three-dimensional spatial framework of various types of rock slices is constructed, including the following steps: Read the comprehensive image, convert the image data into a format suitable for 3D modeling, create a 3D mesh using software such as GOCAD, and interpolate the comprehensive image data into the 3D mesh; Transfer learning based on traces: According to the trace features, a trace model is constructed for transfer learning to adapt to the classification task of trace features; Material-based transfer learning: Based on the material composition information, a transfer model is constructed for transfer learning to adapt to the material composition classification task; Coupling-based transfer learning: Combine shape features and material composition information to extract coupling features, build a coupling model for transfer learning, and adapt to the classification task of coupling features; Model-based transfer learning: Combining shape features, material composition information and coupling features, a multimodal model is constructed for transfer learning to adapt to the classification task of multimodal features; The classification results are applied to a three-dimensional spatial grid to generate a three-dimensional model of the carbonate rock thin section.

[0021] The graph classification unit 200 preliminarily constructs a graph classifier, including the following steps: In a unified spatial scale, the structural features and material composition characteristics of the main rock slices are screened from the three-dimensional model according to the different optical characteristics of the internal biological framework, grains, and cement materials to form multiple single vector layers; Analyze the content and spatial distribution of material components in a single vector layer in a three-dimensional model and correspond them to the marked carbonate thin section types; Among them, the random forest model RF is an integrated learning method that combines multiple decision trees. It can be effectively applied to the linear regression analysis of earth science big data and is conducive to the quantitative analysis of material composition. The artificial neural network ANN is an operation model composed of many nodes. It can have multiple connection methods, has adaptive information processing capabilities, and has a wide range of applications. It has unique advantages in knowledge mining of nonlinear data in geoscience. The support vector machine SVM can reduce the complexity of training data based on limited sample information and obtain the optimal number of support vectors of the model. It is suitable for the nonlinear classification of material composition and structural traces of atypical carbonate rocks. The appropriate model is selected according to the situation. The Pearson correlation coefficient is a statistic used to reflect the degree of linear correlation between two variables. The correlation coefficient is represented by r, where n is the sample size, and is the observed value and mean of the two variables, respectively. r describes the strength of the linear correlation between the two variables. The larger the absolute value of r, the stronger the correlation. The Pearson correlation coefficient can reflect the correlation between different variables. The Pearson correlation coefficient is selected to analyze the correlation between material composition and structural traces, summarize the content and spatial distribution of its material composition, and explore its influencing factors.

[0022] The output result of the BP neural network model in the type identification unit 300 is the predicted carbonate rock thin section type, which includes the following steps: Use Keras to build a BP neural network model, compile the model, specify the loss function, optimizer and evaluation index, use the atlas fingerprints of multiple carbonate rock slices as the training set data to train the model, use the test set data to evaluate the performance of the model, use the trained model to predict new carbonate rock slices, input the atlas fingerprint, and output the rock slice type; BP neural network is a multi-layer feedforward neural network, similar to the signal transmission process of human neurons. The overall structure is 3 layers: input layer - hidden layer - output layer. This model has great processing ability for nonlinear data. To build a BP neural network model, first determine the number of input layer neuron nodes. , the number of hidden layer neuron nodes , the number of neuron nodes in the output layer . Secondly, determine the hidden layer threshold , output layer threshold , learning rate, number of trainings, minimum error of training target, activation function, etc. Among them, the connection weight between the input layer and the hidden layer is , the connection weight between the hidden layer and the output layer is , BP neural network models with different structures are constructed in Matlab, the input layer neuron nodes are the "atlas fingerprint" of the thin section established, and the output result is the predicted carbonate thin section type. The coupling law of carbonate thin section atlas information and the results of transfer learning are combined to identify the type of rock thin section, avoiding the accuracy and transferability problems faced by a single data source.

[0023] In summary, taking granular limestone as an example, granular limestone has three types of material components. The multi-source feature extraction unit 100 obtains the shape trace composition map a and the material composition map b, realizes the visualization of the spatial shape trace information, shape trace structure and material composition information, and enables the atlas classification unit 200 to screen out the single vector map layers of the three material components respectively, which is conducive to analyzing the correlation between the material components and the structural shape traces, summarizing the content and spatial distribution law of its material components, and exploring its influencing factors. When determining the type of carbonate rock, the "atlas fingerprint" of the carbonate rock thin section is obtained again through the multi-source feature extraction unit 100 and the atlas classification unit 200 and input into the BP neural network model of the type recognition unit 300, and the predicted carbonate rock thin section type is output, which optimizes and improves the work of traditional thin section naked eye identification, and can provide technical support for the intelligent identification, non-destructive testing and in-situ parameter analysis of carbonate rock thin sections. Moreover, in modern geological teaching and work, the virtual simulation platform formed by the carbonate rock image labeling system based on deep learning technology plays a vital role. The system can more intuitively understand geological phenomena and provide technical support for petrology teaching and virtual simulation platform construction.

[0024] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Carbonate rock image labeling system based on deep learning technology, characterized by: It includes a multi-source feature extraction unit (100), a graph classification unit (200) and a type recognition unit (300); The multi-source feature extraction unit (100) is used to establish a carbonate rock particle image data set, use a remote sensing image processing algorithm to extract spatial trace information from rock thin section microscopic images of different spatial scales, and comprehensively use cathode luminescence images and element spectrum surface scanning images to extract trace structure and material composition information, perform three-dimensional spatial modeling on the structural characteristics of carbonate rocks, and construct a three-dimensional spatial framework of various types of rock thin sections through transfer learning; The atlas classification unit (200) is used to screen the structural features and material composition characteristic atlas information of carbonate rocks, preliminarily construct an atlas classifier, reflect the correlation between material composition and structural features through the Pearson correlation coefficient, summarize the content and spatial distribution of its material components, and explore its influencing factors; The type recognition unit (300) is used to fuse the morphological structure and material composition characteristic spectrum information in a unified three-dimensional space framework to establish the spectrum fingerprint of the carbonate rock thin section, use the BP neural network model to take the spectrum fingerprint as input, and comprehensively analyze the coupling law of the carbonate rock thin section spectrum information and the results of transfer learning to perform rock thin section type identification, and output the predicted carbonate rock thin section type.

2. The carbonate rock image labeling system based on deep learning technology according to claim 1 is characterized by: When extracting spatial trace information, the multi-source feature extraction unit (100) comprises the following steps: Scan the two-dimensional digital images of various types of rock slices, classify and establish a graphic library to form a carbonate rock particle image data set; The structural framework is described graphically in the two-dimensional grid plane by using the morphology, texture, color, brightness and roughness characteristic parameters in the two-dimensional digital image, including the position and orientation changes, to obtain the average morphology of the image and extract the spatial trace information.

3. The carbonate rock image labeling system based on deep learning technology according to claim 2 is characterized by: The multi-source feature extraction unit (100) comprehensively utilizes cathode luminescence images and element spectrum surface scanning images to extract information on shape structure and material composition, and comprises the following steps: The electron beam parameters of the electron microscope are adjusted so that the electron beam is focused on the surface of the rock slice to excite the luminescent center in the sample. The energy spectrometer is installed on the electron microscope to record the X-ray signal emitted by the sample under the excitation of the electron beam, and the recorded X-ray signal is converted into a digital image. The grayscale value is used to represent the luminescence intensity. The thin-section microscopic image is first subjected to image denoising and image enhancement preprocessing to keep good edges and textures in the area where the carbonate rock particles are located, and to enhance the grayscale difference between the particle area and the background area in the image to form a cathode luminescence image and an element spectrum surface scanning image. Align the cathode luminescence image and the element spectrum scan image to fuse the information of the two images to generate a comprehensive image; Clustering algorithms are used to classify features, identify different mineral phases, and visualize the generation of trace composition maps and material composition maps.

4. The carbonate rock image labeling system based on deep learning technology according to claim 3 is characterized by: The information fusion of the two images to generate a comprehensive image adopts a multi-scale fusion algorithm, including the following steps: Read the cathode luminescence image and element spectrum scanning image, and normalize the image data to the same dynamic range; The feature points of the image are extracted using the feature point detection method, the feature points of the two images are matched, the transformation matrix is ​​calculated, one image is aligned to the other image, the aligned image is decomposed into sub-bands of different scales using wavelet transform, and the sub-bands of different scales are fused to form a comprehensive image.

5. The carbonate rock image labeling system based on deep learning technology according to claim 4 is characterized in that: The structural features include spatial trace information, trace structure and material composition information.

6. The carbonate rock image labeling system based on deep learning technology according to claim 5 is characterized by: The three-dimensional spatial framework of various types of rock slices is constructed through transfer learning, including the following steps: Read the composite image, convert the image data into a format suitable for 3D modeling, create a 3D mesh, and interpolate the composite image data into the 3D mesh; Transfer learning based on traces: According to the trace features, a trace model is constructed for transfer learning to adapt to the classification task of trace features; Material-based transfer learning: Based on the material composition information, a transfer model is constructed for transfer learning to adapt to the material composition classification task; Coupling-based transfer learning: Combine shape features and material composition information to extract coupling features, build a coupling model for transfer learning, and adapt to the classification task of coupling features; Model-based transfer learning: Combining shape features, material composition information and coupling features, a multimodal model is constructed for transfer learning to adapt to the classification task of multimodal features; The classification results are applied to a three-dimensional spatial grid to generate a three-dimensional model of the carbonate rock thin section.

7. The carbonate rock image labeling system based on deep learning technology according to claim 1, characterized in that: The graph classification unit (200) preliminarily constructs a graph classifier, including the following steps: In a unified spatial scale, the structural features and material composition characteristics of the main rock slices are screened from the three-dimensional model according to the different optical characteristics of the internal biological framework, grains, and cement materials to form multiple single vector layers; The content and spatial distribution of material components of a single vector layer in the three-dimensional model are analyzed and corresponded to the marked carbonate rock thin section types.

8. The carbonate rock image labeling system based on deep learning technology according to claim 7, characterized in that: The output result of the BP neural network model in the type identification unit (300) is the predicted carbonate rock thin section type, comprising the following steps: Use Keras to build a BP neural network model, compile the model, specify the loss function, optimizer and evaluation indicators, use the atlas fingerprints of multiple carbonate rock sections as the training set data to train the model, use the test set data to evaluate the performance of the model, use the trained model to predict new carbonate rock sections, input the atlas fingerprint, and output the rock section type.

Images