A multi-stage intelligent identification method for rock slice images
By employing a multi-stage intelligent identification method, including data cleaning, enhancement, and multiple classifications, the problem of identifying fine particles in rock thin sections has been solved, enabling precise identification of rock thin sections. This method is applicable to rock and mineral reservoir research and geological exploration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-30
- Publication Date
- 2026-03-17
AI Technical Summary
Existing intelligent identification technologies for rock thin sections are insufficient to accurately identify fine particles and different types of mineral particles in cast sandstone thin sections. In particular, quartz, quartzite fragments and flint are easily confused, and the identification of feldspar subclasses is not precise enough and is greatly affected by subjective factors.
A multi-stage intelligent recognition method for rock thin section images is adopted, including acquiring aligned rock thin section orthogonally polarized light microscopic images, expert annotation, data cleaning, data augmentation, building a classification network model, and improving the soft voting method decision-level fusion method to perform multiple classification recognitions to improve accuracy.
It enables precise identification of rock thin sections, reduces subjective errors in manual identification, and improves the accuracy and reliability of identification, making it suitable for rock and mineral reservoir research and geological exploration.
Smart Images
Figure CN116563661B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of rock thin section image processing technology, and in particular to a multi-stage intelligent recognition method for rock thin section images. Background Technology
[0002] In manual identification of thin sections, the identification of clastic grains in cast sandstone thin sections is the most important and time-consuming task. Identification experts need to continuously switch between rotating grains and polarized light angles under multiple optical conditions and angles, combining the optical characteristics of crystals under single polarized light, crossed polarized light, and conical light to observe the characteristic changes of mineral components and determine the type of grains.
[0003] Due to the diverse types of rock fragments, including granite fragments, acidic extrusive rock fragments, and volcanic breccia fragments, especially when the grains are fine, they are difficult to distinguish manually. Furthermore, because quartz and quartzite fragments exhibit extinction characteristics with flint, they can be easily confused at certain angles. Feldspar subclasses, with their complex twinning types and cleavage characteristics, are significantly influenced by subjective factors, leading to easy confusion during manual identification.
[0004] Currently, existing intelligent identification methods for rock thin sections are only based on the broad categories of quartz and feldspar. They lack methods for identifying specific types of rock fragments, such as quartzite fragments, metamorphic rock fragments, and volcanic rock fragments, as well as more detailed categories within the feldspar subclass, such as plagioclase, orthoclase, stratiform feldspar, and microcline. Furthermore, different types of mineral grains share similar optical properties and exhibit some degree of alteration, making accurate identification difficult with existing intelligent identification technologies. Summary of the Invention
[0005] To address the aforementioned technical problems, this application provides a multi-stage intelligent recognition method for rock thin section images.
[0006] The present invention is achieved by the following technical solution.
[0007] A multi-stage intelligent recognition method for rock thin section images includes the following steps:
[0008] (i) Obtain aligned rock thin sections using orthogonal polarized light microscopy;
[0009] (ii) Perform expert annotation on rock particles and obtain a single-particle image sample library;
[0010] (iii) Perform data cleaning on the single-particle image sample library for each category;
[0011] (iv) Data augmentation was performed on the few categories in the single-particle image sample library, and training and test sets for each category of rock were obtained;
[0012] (v) Construct a classification network model;
[0013] (vi) Train the obtained rock training sets of each category to build a classification network model and obtain the corresponding classification model;
[0014] (vii) A decision-level fusion method based on the improved soft voting method is used to predict the results and obtain the classification results;
[0015] (ⅷ) Perform a second classification result recognition to obtain the final classification result.
[0016] Furthermore, the specific process of obtaining aligned rock thin section orthogonally polarized light micrographs in step (i) is as follows:
[0017] S1. Prepare thin section samples of rock without coverslip;
[0018] S2. Obtain cross-polarized light microscopic images of rock thin sections from different fields of view and multiple angles using an optical microscope;
[0019] S3. Align the cross-polarized light micrographs of the rock thin sections.
[0020] Furthermore, the specific process of step (ii) for expert annotation of rock particles and obtaining a single-particle image sample library is as follows:
[0021] S1. Expert annotation of rock grains in orthogonally polarized light micrographs of rock thin sections;
[0022] S2. Extract the corresponding single particles from the orthogonal polarized light micrographs based on the annotation results;
[0023] S3. Create a single-particle image sample library from the extracted individual particles.
[0024] Data cleaning is necessary because some particles exhibit severe extinction, are too small, and their features are easily confused with those of other particle types.
[0025] Furthermore, the specific process of data cleaning for each category of single-particle image sample library in step (iii) is as follows:
[0026] S1. Remove particles with indistinct characteristics that exhibit severe extinction phenomenon;
[0027] S2. Remove images with a width / height less than 100 pixels;
[0028] S3. Remove incomplete particle images;
[0029] S4. Clear images that are too bright;
[0030] S5. Remove images whose features are not related to the labeled category;
[0031] S6. Remove images with indistinct features;
[0032] S7. Remove images with easily confused features.
[0033] Due to issues such as a small number of training samples and an imbalance in the number of samples between classes in the dataset, image augmentation is necessary.
[0034] Furthermore, step (iv) involves data augmentation of the relatively small number of categories in the single-particle image sample library and the subsequent division into training and testing sets for each category of rock, as follows:
[0035] S1. Enhance images of feldspar and rock grains;
[0036] S2. Divide the single-particle image samples of each category into a training set and a test set according to a 7:3 ratio;
[0037] S3. Repeat steps S1 and S2 to obtain a dataset of N types of rock particles.
[0038] Furthermore, the specific process of building the classification network model in step (v) is as follows:
[0039] S1. The classification network model is used as the multi-channel feature extraction part of a single-particle orthogonal polarized light image, and the EfficientNet network is used as the backbone network.
[0040] S2. The classification network model includes 1 Conv (3×3), 1 MBConv1 (3×3), 2 MBConv6 (3×3), 2 MBConv6 (5×5), 3 MBConv6 (3×3), 3 MBConv6 (5×5), 4 MBConv6 (5×5), 1 MBConv6 (3×3), 1 Conv (1×1), 1 Pooling layer, and 1 FC layer.
[0041] Furthermore, step (vi) involves training the obtained training sets of various rock categories to build a classification network model. The specific process for obtaining the corresponding classification model is as follows:
[0042] S1. Train the dataset of N types of rock particles from step (iv) to build a classification network model and obtain an N-class classification model;
[0043] S2. Train a classification network model using the quartz and quartzite rock fragments dataset to obtain classification models for the two categories of quartz and quartzite rock fragments;
[0044] S3. Use the plagioclase and metamorphic rock fragment datasets to train and build a classification network model to obtain classification models for two categories: plagioclase and metamorphic rock fragments.
[0045] S4. Train a classification network model using the datasets of orthoclase, striped feldspar, and microcline to obtain classification models for the three categories of orthoclase, striped feldspar, and microcline.
[0046] Step (vii) Based on the decision-level fusion method of the improved soft voting method, the results are predicted. In the classification results, most particles in the orthogonal polarized light micrographs will be in one or two images in the interference color or complete extinction state, which makes it impossible to fully display the specific features. In order to comprehensively consider all orthogonal polarized light features, the improved soft voting method is used to predict the classification model results.
[0047] Furthermore, step (vii) involves predicting the results using a decision-level fusion method based on an improved soft voting method. The specific process for obtaining the classification results is as follows:
[0048] S1. A classification model is used to predict a single particle image under orthogonal polarization, and a set of prediction results are obtained, including top1, top2, top3, top4, and top5.
[0049] S2. Five orthogonally polarized single-particle images are used to make predictions using a classification model, resulting in five sets of prediction results;
[0050] S3. Take the top 1 among the five prediction results. If three or more of the five top 1 results belong to the same category, then that category is defined as the classification result.
[0051] S4. If there are no three or more results in the same category among the five top 1 results, then add the classification result of the top 2 results. If the category appears more than four times in the five top 1 results and five top 2 results, then that category is defined as the classification result.
[0052] S5. Except for the two cases described in steps S3 and S4, the classification result is determined by voting.
[0053] Furthermore, step (ⅷ) involves a second classification result recognition to obtain the final classification result. The specific process is as follows:
[0054] S1. Judge the classification result obtained in step (vii) to determine whether a second classification recognition is needed;
[0055] S2. If the classification result obtained in step (vii) does not require a second classification recognition, then this result is the final classification result;
[0056] S3. If the classification result obtained in step (vii) requires a second classification recognition, then the classification model is used to obtain the second classification result, which is the final classification result.
[0057] This application has the following beneficial effects.
[0058] This invention employs data cleaning and data augmentation to perform more accurate and refined processing on graphic samples, while using a secondary recognition method to ensure data reliability. This invention establishes a complete and comprehensive recognition method for rock thin sections, and the recognition method is more refined, overcoming the subjective errors and confusion probability of expert manual recognition, and has broad application prospects in rock and mineral reservoir research and geological exploration. Attached Figure Description
[0059] Figure 1 This is a flowchart of the method of the present invention;
[0060] Figure 2 This is a flowchart of the sample image acquisition and preprocessing process of the present invention;
[0061] Figure 3 This is a schematic diagram of the classification network model in this invention;
[0062] Figure 4 This is a schematic diagram of the improved soft voting method in this invention;
[0063] Figure 5 This is a flowchart illustrating the secondary classification and recognition process in this invention. Detailed Implementation
[0064] The present patent application will be further described below with reference to the embodiments.
[0065] A multi-stage intelligent recognition method for rock thin section images includes the following steps:
[0066] i. Obtaining aligned rock thin sections using orthogonal polarized light microscopy, the specific process is as follows:
[0067] First, prepare thin section samples of rock without coverslips;
[0068] Then, orthogonal polarized light microscopic images of rock thin sections from different fields of view and multiple angles were obtained using an optical microscope;
[0069] Finally, the cross-polarized light micrographs of the rock thin sections were aligned.
[0070] ii. Expert annotation of rock particles was performed to obtain a single-particle image sample library. The specific process is as follows:
[0071] First, the rock grains in the cross-polarized light micrographs of the rock thin sections were annotated by experts.
[0072] Then, based on the annotation results, the corresponding single particles on the orthogonal polarized light micrographs are extracted;
[0073] Finally, the extracted individual particles are used to create a single-particle image sample library.
[0074] iii. Perform data cleaning on the single-particle image sample library for each category, as follows:
[0075] First, remove particles with indistinct features that exhibit severe matting;
[0076] Then, remove images with a width / height less than 100 pixels;
[0077] Next, remove incomplete particle images;
[0078] Next, remove images that are too bright;
[0079] Next, remove images with features that are not related to the labeled category;
[0080] Next, remove images with indistinct features;
[0081] Finally, remove images with easily confused features.
[0082] iv. Data augmentation is performed on the relatively small number of categories in the single-particle image sample library, and training and testing sets are obtained for each category of rock. The specific process is as follows:
[0083] First, the images of feldspar and rock grains were enhanced;
[0084] Then, for each category of single-particle image samples, the training set and test set are obtained by dividing them in a 7:3 ratio;
[0085] Finally, repeat the above two steps to obtain the N-class rock particle dataset.
[0086] v. Building the classification network model, the specific process is as follows:
[0087] First, the classification network model, as the multi-channel feature extraction part of a single-particle orthogonal polarized light image, uses the EfficientNet network as the backbone network.
[0088] Then, the classification network model includes 1 Conv (3×3), 1 MBConv1 (3×3), 2 MBConv6 (3×3), 2 MBConv6 (5×5), 3 MBConv6 (3×3), 3 MBConv6 (5×5), 4 MBConv6 (5×5), 1 MBConv6 (3×3), 1 Conv (1×1), 1 Pooling layer, and 1 FC layer.
[0089] vi. Train the classification network model built using the obtained training sets of each rock category to obtain the corresponding classification model. The specific process is as follows:
[0090] First, the classification network model built by training the dataset of N types of rock particles in step iv is used to obtain an N-class classification model;
[0091] Then, the quartz and quartzite rock fragments datasets were used to train the constructed classification network model to obtain classification models for the two categories of quartz and quartzite rock fragments.
[0092] Next, the classification network model was trained using the plagioclase and metamorphic rock fragment datasets to obtain classification models for the two categories of plagioclase and metamorphic rock fragments.
[0093] Next, the datasets of orthoclase, striped feldspar, and microcline were used to train a classification network model, resulting in classification models for the three categories of orthoclase, striped feldspar, and microcline.
[0094] vii. Based on the improved soft voting method, a decision-level fusion method is used to predict the results and obtain the classification results. The specific process is as follows:
[0095] First, a classification model is used to predict a single particle image of orthogonal polarized light, and a set of prediction results are obtained, including top1, top2, top3, top4, and top5.
[0096] Then, a classification model was used to predict the five orthogonally polarized single-particle images, resulting in five sets of prediction results;
[0097] Next, take the top 1 among the five prediction results. If three or more of the five top 1 results belong to the same category, then that category is defined as the classification result.
[0098] Next, if no three or more results in the top 1 category belong to the same category, the classification result of the top 2 category is added. If the category appears more than four times in the five top 1 and five top 2 categories, then that category is defined as the classification result.
[0099] Finally, except for the two cases mentioned above, the classification result is determined by voting.
[0100] ⅷ. Perform a second classification result recognition to obtain the final classification result. The specific process is as follows:
[0101] First, the classification results obtained in step 1vi are evaluated to determine whether a second classification and recognition is needed.
[0102] Then, if the classification result obtained in step vii does not require a second classification recognition, then this result is the final classification result;
[0103] Finally, if the classification result obtained in step vii requires a second classification recognition, the classification model is used to obtain the second classification result, which is the final classification result.
[0104] Example
[0105] A multi-stage intelligent recognition method for rock thin section images includes the following steps:
[0106] i. Obtaining aligned rock thin sections using orthogonal polarized light microscopy, the specific process is as follows:
[0107] First, prepare thin section samples of rock without coverslips;
[0108] Then, orthogonal polarized light microscopic images of rock thin sections from different fields of view and multiple angles were obtained using an optical microscope;
[0109] Finally, the crossed polarized light micrographs of the rock thin sections were aligned (see...). Figure 2 ).
[0110] ii. Expert annotation of rock particles was performed to obtain a single-particle image sample library. The specific process is as follows:
[0111] First, the rock grains in the cross-polarized light micrographs of the rock thin sections were annotated by experts.
[0112] Then, based on the annotation results, the corresponding single particles on the orthogonal polarized light micrographs are extracted;
[0113] Finally, the extracted individual particles are used to create a single-particle image sample library (see...). Figure 2 ).
[0114] Specifically, experts have categorized rock particles into 15 types: quartz, quartzite fragments, flint, orthoclase, plagioclase, microcline, banded feldspar, metamorphic rock fragments, mica, sedimentary rock fragments, granite rock fragments, andesite rock fragments, basalt rock fragments, acidic extrusive rock fragments, and volcanic clastic rock fragments.
[0115] Read the JSON file containing the expert annotation results, which includes the particle outline coordinates and particle category information.
[0116] Based on the particle outline coordinates in the annotation results, the minimum bounding rectangle of the particle outline is calculated. Then, the corresponding particle images are extracted from the multi-angle orthogonal polarized light sequence images through the four vertices of the rectangle.
[0117] Because there is interference information around the particle image, all information in the extracted single particle image that does not belong to the particle is replaced with a black background to prevent it from affecting the accuracy of neural network recognition.
[0118] The extracted single-particle images are then divided into folders for each category.
[0119] iii. Perform data cleaning on the single-particle image sample library for each category, as follows:
[0120] First, remove particles with indistinct features that exhibit severe matting;
[0121] Then, remove images with a width / height less than 100 pixels;
[0122] Next, remove incomplete particle images;
[0123] Next, remove images that are too bright;
[0124] Next, remove images with features that are not related to the labeled category;
[0125] Next, remove images with indistinct features;
[0126] Finally, remove features that are easily confused (see Figure 2 ).
[0127] iv. Data augmentation is performed on the relatively small number of categories in the single-particle image sample library, and training and testing sets are obtained for each category of rock (here, a dataset of 15 rock particle categories is formed). The specific process is as follows:
[0128] First, the images of feldspar and rock particles are enhanced; specifically, rotation, Gaussian noise and salt-and-pepper noise, symmetry transformation, flipping, and rotation plus flipping are performed on the images to generate six additional enhanced images for each image.
[0129] Then, for each category of single-particle image samples, the training set and test set are obtained by dividing them in a 7:3 ratio;
[0130] Finally, repeat the above two steps to obtain the N-class rock particle dataset.
[0131] v. Constructing a classification network model, the specific process is as follows (see...) Figure 3 ):
[0132] First, the classification network model is used as the multi-channel feature extraction part of the single-particle orthogonal polarized light image, and the EfficientNet-b3 network is used as the backbone network.
[0133] Then, the classification network model includes 1 Conv (3×3), 1 MBConv1 (3×3), 2 MBConv6 (3×3), 2 MBConv6 (5×5), 3 MBConv6 (3×3), 3 MBConv6 (5×5), 4 MBConv6 (5×5), 1 MBConv6 (3×3), 1 Conv (1×1), 1 Pooling layer, and 1 FC layer.
[0134] vi. Train the classification network model using the obtained rock training sets for each category to obtain the corresponding classification model. For the 15 rock particle categories labeled by experts, 15 classification models are formed. Separate classification models are also established for quartz, quartzite fragments, plagioclase, metamorphic rock fragments, orthoclase, banded feldspar, and microcline, as detailed below:
[0135] First, the classification network model built by training the dataset of N types of rock particles in step iv is used to obtain an N-class classification model;
[0136] Then, the quartz and quartzite rock fragments datasets were used to train the constructed classification network model to obtain classification models for the two categories of quartz and quartzite rock fragments.
[0137] Next, the classification network model was trained using the plagioclase and metamorphic rock fragment datasets to obtain classification models for the two categories of plagioclase and metamorphic rock fragments.
[0138] Next, the datasets of orthoclase, striped feldspar, and microcline were used to train a classification network model, resulting in classification models for the three categories of orthoclase, striped feldspar, and microcline.
[0139] vii. Based on the improved soft voting method, a decision-level fusion method is used to predict the results and obtain the classification results. The specific process is as follows (see...). Figure 4 ):
[0140] First, a classification model is used to predict a single particle image of orthogonal polarized light, and a set of prediction results are obtained, including top1, top2, top3, top4, and top5.
[0141] Then, a classification model was used to predict the five orthogonally polarized single-particle images, resulting in five sets of prediction results;
[0142] Next, take the top 1 among the five prediction results. If three or more of the five top 1 results belong to the same category, then that category is defined as the classification result.
[0143] Next, if no three or more results in the top 1 category belong to the same category, the classification result of the top 2 category is added. If the category appears more than four times in the five top 1 and five top 2 categories, then that category is defined as the classification result.
[0144] Finally, except for the two cases mentioned above, the classification result is determined by voting.
[0145] The voting method determines the classification results, and the specific process is as follows:
[0146] The average of the top 1, top 2, top 3, top 4, and top 5 groups is taken to obtain the final top 5 results (top 1, top 2, top 3, top 4, top 5).
[0147] ⅷ. Perform a second classification result recognition to obtain the final classification result. The specific process is as follows:
[0148] First, the classification results obtained in step 1vi are evaluated to determine whether a second classification and recognition is needed.
[0149] Then, if the classification result obtained in step vii does not require a second classification recognition, then this result is the final classification result;
[0150] Finally, if the classification result obtained in step vii requires a second classification recognition, the classification model is used to obtain the second classification result, which is the final classification result.
[0151] Because of the feature confusion phenomenon among different rock grains, i.e., the appearance of features of other grain types on an image of a certain type of grain, this application proposes a "first identification + second identification" method for determination (see...). Figure 5 ).
[0152] First, a 15-class classification model is trained. Then, for categories where the recognition results are easily confused, a smaller classification model is trained. Partial predictions from the 15-class classification model are fed into the smaller classification model for final recognition. The specific process is as follows:
[0153] First, the five orthogonally polarized single-particle images to be predicted are used to perform the first identification using the 15-category classification model trained in step ⅵ, and the first identification result is obtained.
[0154] Then, if the first identification result is quartz or quartzite fragments, the two-category classification model for quartz and quartzite fragments trained in step ⅵ is used for a second classification, and the result of the second classification is the final classification result.
[0155] Then, if the first identification result is plagioclase or metamorphic rock fragments, the two-category classification model for plagioclase and metamorphic rock fragments trained in step ⅵ is used for a second classification, and the result of the second classification is the final classification result.
[0156] Then, if the first identification result is orthoclase, striped feldspar, or microcline, the three-category classification model of orthoclase, striped feldspar, and microcline trained in step ⅵ is used for a second classification, and the result of the second classification is the final classification result.
[0157] Finally, if the initial identification result is another category, then that is the final classification result.
[0158] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A multi-stage intelligent recognition method for rock thin section images, characterized in that: The method comprises the following steps: (i) obtaining aligned rock thin section orthogonal polarized light microscopic images; (ii) expertly labeling rock particles and obtaining a single particle image sample library; (iii) data cleaning is performed on the single particle image sample library of each category respectively; (iv) data augmentation is performed on the categories with less quantity in the single particle image sample library, and training set and test set of each category of rock are divided; (v) a classification network model is built; (vi) the obtained training set of each category of rock is used to train the built classification network model to obtain the corresponding classification model, and the specific process is as follows: S1. The data set of N categories of rock particles in step (iii) is used to train the built classification network model to obtain N category classification models; S2. The quartz and quartz rock debris data set is used to train the built classification network model to obtain two category classification models of quartz and quartz rock debris; S3. The plagioclase and metamorphic debris data set is used to train the built classification network model to obtain two category classification models of plagioclase and metamorphic debris; S4. The orthoclase, striped feldspar and microplagioclase data set is used to train the built classification network model to obtain three category classification models of orthoclase, striped feldspar and microplagioclase; (vii) a decision-level fusion method based on an improved soft voting method is used for result prediction to obtain a classification result, and the specific process is as follows: S1. A single orthogonal polarized light particle image is predicted using the classification model to obtain a group of prediction results, and the prediction results include top1, top2, top3, top4 and top5; S2. Five orthogonal polarized light single particle images are predicted using the classification model to obtain five groups of prediction results; S3. The top1 in the five groups of prediction results is taken, and if three or more results in the five top1 are the same category, the category is determined as the classification result; S4. If three or more results in the five top1 are not the same category, the classification result of the top2 is added, and if the result category appears four or more times in the five top1 and five top2, the category is determined as the classification result; S5. Except for the two cases described in steps S3 and S4, the voting method is used to determine the classification result; (viii) a second classification result identification is performed to obtain a final classification result.
2. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (i) for obtaining aligned rock thin section orthogonal polarized light microscopic images is as follows: S1. Prepare rock thin section samples without cover glass; S2. Obtain orthogonal polarized light microscopic images of rock thin sections at different viewing angles using an optical microscope; S3. Align the orthogonal polarized light microscopic images of the rock thin sections.
3. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (ii) for expertly labeling rock particles and obtaining a single particle image sample library is as follows: S1. Expertly label rock particles in the rock thin section orthogonal polarized light microscopic images; S2. According to the labeling results, the corresponding single particles on the orthogonal polarized light microscopic images are cut out; S3. The cut-out single particles form a single particle image sample library.
4. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (iii) for data cleaning of each category of single particle image sample library is as follows: S1. Remove particles with severe extinction phenomenon and unclear features; S2. Remove images with a width / height less than 100 pixels; S3. Clear incomplete particle images; S4. Clear images with too high brightness; S5. Clear images with no relevant features to the labeled result category; S6. Clear images with no obvious features; S7. Clear images with easily confused features.
5. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (iv) for data augmentation of the small number of categories in the single particle image sample library and division to obtain the training set and test set of each category of rock is as follows: S1. Enhance the feldspar and rock particle images; S2. Divide each category of single particle image sample into training set and test set according to 7:3; S3. Repeat steps S1 and S2 to obtain N categories of rock particle data sets.
6. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (v) for building a classification network model is as follows: S1. The classification network model is used as a multi-channel feature extraction part of the single particle orthogonal polarized light image, and EfficientNet network is used as the backbone network; S2. The classification network model includes 1 Conv(3x3), 1 MBConv1(3x3), 2 MBConv6(3x3), 2 MBConv6(5x5), 3 MBConv6(3x3), 3 MBConv6(5x5), 4 MBConv6(5x5), 1 MBConv6(3x3), 1 Conv(1x1), 1 Pooling layer, 1 FC layer.
7. The multi-stage intelligent identification method for rock slice images according to claim 1, characterized in that: The specific process of step (vii) for the second classification result identification to obtain the final classification result is as follows: S1. Determine whether the classification result obtained in step (vi) needs to be identified for the second time; S2. If the classification result obtained in step (vi) does not need to be identified for the second time, the result is the final classification result; S3. If the classification result obtained in step (vi) needs to be identified for the second time, use the classification model to obtain the second classification result, and the second classification result is the final classification result.
Citation Information
Patent Citations
Image prediction optimization processing method and device
CN114202671A