SENet-ConvNeXt-FL network-based rock slice image classification method, system and device, and medium

Abstract

The invention discloses a rock slice image classification method, system, equipment and medium based on an SENet-ConvNeXt-FL network, and the method comprises the steps: firstly, building a rock slice image data set; secondly, an improved ConvNeXt network model is established; a SENet module is introduced to optimize the feature extraction capability of ConvNeXt, a loss function in the training process is changed into a Focal Loss function, and an improved model is named as SENet-ConvNeXt-FL; and finally, taking a PyCharm platform as a deep learning framework, training the improved model, carrying out rock slice image classification, and carrying out multi-dimensional evaluation on a rock slice image classification result by adopting a confusion matrix and a model performance evaluation index. The invention further comprises a system, equipment and a medium based on the method. According to the method, the SENet attention mechanism is fused into the ConvNeXt network, the Focal Loss function is introduced, the feature representation capability of the model is enhanced, the problem of class imbalance is effectively solved, and rock slice image recognition and classification are more accurate and reliable when the method is used for performing rock slice image recognition and classification.

Description

Technical Field

[0001] The present invention relates to the field of petroleum geology technology, and in particular to a rock slice image classification method, system, equipment and medium based on a SENet-ConvNeXt-FL network. Background Art

[0002] Rock thin section identification can obtain information such as rock mineral composition, particle size, cementation mode, pore morphology and structure, and is an important basic work for understanding reservoir characteristics and obtaining relevant parameters of oil and gas reservoirs during oil and gas field exploration and development. Traditional rock thin section identification is mainly based on naked eye observation and content estimation. However, the multi-source and multi-mode characteristics of mineralization information will lead to blurred boundaries between components of rock thin sections under the microscope and difficulty in identification; at the same time, human subjective factors (naked eye observation errors, expert personal experience, etc.) will have a certain impact on the results of rock thin section identification.

[0003] With the development of deep learning technology, intelligent recognition of rock thin section images has become a trend. In recent years, scholars have applied a variety of artificial intelligence methods such as convolutional neural networks, GBDT, VGG, SVM, MobileNetV2, Inception v3, SqueezeNet lightweight neural networks, lightweight MobileViT and ResNet101 to rock thin section image classification, and achieved good results. For example, the patent application with application number 202410592576.5 discloses a rock thin section image lithology recognition method based on deep learning, which is to help interpret data by imitating the mechanism of the human brain to achieve automatic recognition of rock thin section images, thereby reducing the influence of subjective factors and improving model recognition accuracy. However, the above-mentioned artificial intelligence methods are heavily dependent on the quality of digital features extracted by researchers or by imitating the mechanism of the human brain to help interpret data and achieve automatic recognition of rock thin section images. However, in practical applications, there are still some inherent defects and limitations, and it is impossible to completely eliminate the negative impact of the imbalance of original data categories and the complexity of image features on the prediction results. These problems limit the generalization ability and prediction accuracy of the model to a certain extent. Especially when dealing with tasks with strong heterogeneity and unbalanced data volume, the model may find it difficult to effectively capture global features, which in turn affects the final decision performance. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a rock thin section image classification method, system, device and medium based on the SENet-ConvNeXt-FL network, integrate the SENet attention mechanism into the ConvNeXt network, and introduce the Focal Loss function to enhance the feature representation ability of the model, effectively solve the category imbalance problem, so that rock thin section image recognition and classification can be more accurate and reliable.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] The rock thin section image classification method based on SENet-ConvNeXt-FL network includes the following steps:

[0007] Step 1: Create a rock thin section image dataset

[0008] The original rock slice images are fused into the same data set, and the data enhancement methods such as Gaussian filtering and random flipping, magnification, displacement, cropping, and scaling are used to generate an image-enhanced data set, which is then divided into a training set and a test set.

[0009] Step 2: Establish SENet-ConvNeXt-FL network model

[0010] The SENet module is introduced to optimize the feature extraction capability of the ConvNeXt network model, and the loss function in the training process is changed to the Focal Loss function, thus obtaining the SENet-ConvNeXt-FL network model;

[0011] Step 3: Train the SENet-ConvNeXt-FL network model and classify rock slice images

[0012] Using PyCharm platform as the deep learning framework, the SENet-ConvNeXt-FL network model is used to classify rock thin section images;

[0013] Step 4: Model stability and convergence evaluation

[0014] Confusion matrix and model performance evaluation indicators are used to conduct multi-dimensional evaluation of rock thin section image classification results.

[0015] The specific implementation steps of step 1 include:

[0016] (1.1) For the original rock thin section image data set, the Gaussian filtering method is used to remove the noise in the image. The basic principle is to use the Gaussian function to generate a weight matrix, and then perform a convolution operation on the image based on the weight matrix. The specific expression is as follows;

[0017]

[0018] Where σ is the standard deviation, which controls the width of the filter, G(x, y) refers to the value of the Gaussian function at the point (x, y), which represents the weight of the two-dimensional Gaussian function at the coordinate (x, y);

[0019] (1.2) Based on the data enhancement module Augmentor in Python, the rock thin section images after denoising in step (1.1) are subjected to data enhancement processing such as random flipping, magnification, displacement, cropping, and scaling, and the original data set is expanded to n times of the original data set to construct a new rock thin section image enhancement data set; the random flipping refers to randomly flipping the image horizontally or vertically; the magnification, displacement, cropping, and scaling transformations are to change the size and position of the image to simulate the image changes under different shooting conditions;

[0020] (1.3) The rock thin section image enhancement dataset is randomly divided into p% as a training set and q% as a test set to construct a rock thin section image classification dataset, and the p%+q%=100%.

[0021] The specific implementation steps of step 2 include:

[0022] (2.1) The SENet module is introduced to optimize the ConvNeXt network model. The SENet module contains three operations: Squeeze operation, Extitation operation and Reweight operation:

[0023] ①Squeeze operation For each input feature map, global average pooling is used to convert it into a scalar value to capture the global information of each channel. The specific expression is as follows:

[0024]

[0025] In the formula, z c is the compressed output; u c is the eigenvalue on the feature map in channel c;

[0026] ②The Excitation operation maps the global information obtained in the previous step to a channel-related weight vector through a set of fully connected layers and activation functions. The specific expression is as follows:

[0027] s c =δ{σ2[ReLU(σ1z c )]}

[0028] In the formula, σ1 and z c Multiplication means output z c The first layer of fully connected layer operation, followed by ReLU activation function and multiplication with σ2 represents the second layer of fully connected operation, δ represents the Sigmoid function, and the output result s after multiplication c ;

[0029] ③The Reweight operation applies the weight vector calculated in the Excitation operation to the input feature map, reweighting the feature response of each channel in the feature map, thereby enhancing the influence of important feature channels. The specific expression is as follows:

[0030] o=u c s c

[0031] In the formula, s c is the weight of channel c, and o is the reweighted output of channel c;

[0032] (2.2) The loss function in the training process is changed to the Focal Loss function. The Focal Loss function reduces the impact of class imbalance by reducing the weight of samples with a large number and increasing the weight of samples with a small number. The specific expression is as follows:

[0033] F L (p t )=-α t (1-p t ) γ lbp t

[0034] In the formula, p t Represents the prediction confidence score of the model, α t represents the importance index of balancing positive and negative samples, γ is the focusing function, and F L Represents the Focal Loss function.

[0035] The step three is specifically as follows: training the SENet-ConvNeXt-FL network model, including re-segmenting the data at the beginning of each iteration, and training the model using the training set data; during the training process, the weights of the model are updated according to the training data; after the training, the updated model weights are saved for subsequent use or further analysis, and the training of the SENet-ConvNeXt-FL network model includes setting the training round (Epoch), batch size (Bacth Size), learning rate and regularization coefficient.

[0036] The step 4 is specifically as follows: the confusion matrix is ​​used to evaluate the performance of the classification model. The confusion matrix compares the real label and the label predicted by the model, and displays the result in a table. The table consists of four elements, namely true positive examples, false positive examples, true negative examples and false negative examples; the model performance evaluation indicators include accuracy (ACC), precision (Pre), recall (Re) and F1 value; the SENet-ConvNeXt-FL network model is evaluated in multiple dimensions, and the specific expressions of the evaluation indicators are as follows:

[0037]

[0038]

[0039]

[0040]

[0041] In the formula, TP refers to the number of positive samples that are correctly classified as positive; TN refers to the number of samples that are actually negative and correctly classified as negative; FP refers to the number of samples that are actually negative and incorrectly classified as positive; FN refers to the number of samples that are actually positive and incorrectly classified as negative.

[0042] The rock thin section image classification system based on the SENet-ConvNeXt-FL network includes a processing instruction module for executing each step in the rock thin section image classification method based on the SENet-ConvNeXt-FL network.

[0043] Rock thin section image classification equipment based on SENet-ConvNeXt-FL network, including:

[0044] Memory: a computer-readable device for storing the computer program of the rock thin section image classification method based on the SENet-ConvNeXt-FL network;

[0045] Processor: used to implement the rock thin section image classification method based on SENet-ConvNeXt-FL network when executing the computer program.

[0046] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the rock thin section image classification method based on the SENet-ConvNeXt-FL network can be implemented.

[0047] Compared with the prior art, the beneficial technical effects of the present invention are:

[0048] (1) In step 1 of the present invention, Gaussian filtering and random flipping, magnification, displacement, cropping, and scaling are used as data enhancement methods to effectively enhance the quality and quantity of rock thin section images and the generalization ability of the model.

[0049] (2) In step 2 of the present invention, the SENet module is introduced to optimize the ConvNeXt network model. The SENet module dynamically adjusts the importance of channels so that the network can better focus on channels that are useful for specific tasks, and can improve the feature extraction capability and task performance of the model while adding fewer parameters to the original model.

[0050] (3) In step 2 of the present invention, the loss function in the training process is changed to a Focal Loss function. The Focal Loss function reduces the impact of uneven sample distribution by increasing the weight of samples with a smaller number, so that the model pays more attention to samples that are difficult to classify, thereby improving the accuracy of rock thin section image classification.

[0051] In summary, the present invention integrates the SENet attention mechanism into the ConvNeXt network, so that the network can better focus on channels that are useful for specific tasks, improve the feature extraction ability and task performance of the model; introduces the Focal Loss function to effectively solve the problem of category imbalance; and utilizes the rock thin section image classification method, system, device and medium based on the SENet-ConvNeXt-FL network proposed in the present invention to effectively improve the recognition and classification accuracy of rock thin section images, making them more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the rock thin section image classification method and system based on the SENet-ConvNeXt-FL network model of the present invention.

[0053] Figure 2 Schematic diagram of the structure of the SENet-ConvNeXt-FL network model of the present invention.

[0054] Figure 3 This is a schematic diagram comparing the confusion matrix of the present invention with different models; wherein, Figure 3 (a) is the AlexNet model, (b) is the VGG model; (c) is the ConvNeXt model; (d) is the SENet-ConvNeXt-F network model of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] like Figure 1 As shown, the present invention provides a rock thin section image classification method and system based on SENet-ConvNeXt-FL network, which specifically includes the following steps:

[0057] Step 1: Establish a rock thin section image dataset;

[0058] The original rock thin section images are merged into the same data set, and the data enhancement methods such as Gaussian filtering and random flipping, magnification, displacement, cropping, and scaling are used to improve the quality and quantity of rock thin section images, generate new data sets, and divide them into training sets and test sets according to a certain ratio;

[0059] The specific implementation steps of step 1 include:

[0060] (1.1) Based on the rock thin section images required for the experiment obtained from the China Science Data Network, five rock types, including dolomite, siliceous rock, gneiss, schist and sandstone, were selected to construct a rock thin section image dataset. The Gaussian filtering method was used to remove noise from the original rock thin section image dataset. The basic principle is to use the Gaussian function to generate a weight matrix, and then perform a convolution operation on the image based on the weight matrix to achieve the purpose of smoothing the image while retaining the main features in the image. The specific expression is as follows;

[0061]

[0062] Where σ is the standard deviation, which controls the width of the filter, and G(x, y) refers to the value of the Gaussian function at the point (x, y), which represents the weight of the two-dimensional Gaussian function at the coordinates (x, y).

[0063] (1.2) Based on the data enhancement module Augmentor in Python, the rock thin section images after denoising in step (1.1) were subjected to data enhancement processing such as random flipping, magnification, displacement, cropping, and scaling, and the original data set was expanded to five times of the original data set. A total of 4,400 rock thin section images were generated to construct a new rock thin section image enhanced data set. The random flipping refers to randomly flipping the image horizontally or vertically, which can increase the diversity of the data and enable the model to learn features in different directions. The transformations such as magnification, displacement, cropping, and scaling can change the size and position of the image, simulating image changes under different shooting conditions, thereby improving the generalization ability of the model.

[0064] (1.3) The rock thin section image enhancement dataset is randomly divided into 90% as the training set and 10% as the test set to construct a rock thin section image classification dataset.

[0065] Step 2: Establish the SENet-ConvNeXt-F network model;

[0066] The SENet module is introduced to optimize the feature extraction capability of ConvNeXt, and the loss function in the training process is changed to the Focal Loss function, thus obtaining the SENet-ConvNeXt-FL network model. Figure 2 ;

[0067] The specific implementation steps of step 2 include:

[0068] (2.1) The SENet module is introduced to optimize the ConvNeXt network model. The SENet module dynamically adjusts the importance of channels so that the network can better focus on channels that are useful for specific tasks, and can improve the feature extraction capability of the model while adding fewer parameters to the original model. The SENet module contains three operations, namely, the Squeeze operation, the Excitation operation, and the Reweight operation:

[0069] ①Squeeze operation For each input feature map, global average pooling is used to convert it into a scalar value to capture the global information of each channel. The specific expression is as follows:

[0070]

[0071] In the formula, z c is the compressed output; u c is the eigenvalue on the feature map in channel c.

[0072] ②The Excitation operation maps the global information obtained in the previous step to a channel-related weight vector through a set of fully connected layers and activation functions. The specific expression is as follows:

[0073] s c =δ{σ2[ReLU(σ1z c )]}

[0074] In the formula, σ1 and z c Multiplication means output z c The first layer of fully connected layer operation is then multiplied by σ2 after the ReLU activation function to represent the second layer of fully connected operation. δ represents the Sigmoid function, and the output result s is output after multiplication with it.

[0075] ③The Reweight operation applies the weight vector calculated in the Excitation operation to the input feature map, reweighting the feature response of each channel in the feature map, thereby enhancing the influence of important feature channels. The specific expression is as follows:

[0076] o=u c s c

[0077] In the formula, s c is the weight of channel c, and o is the re-weighted output of channel c.

[0078] (2.2) The loss function in the training process is changed to the Focal Loss function to solve the extreme imbalance between categories. The Focal Loss function reduces the impact of category imbalance by reducing the weight of samples with a large number and increasing the weight of samples with a small number. The specific expression is as follows:

[0079] F L (p t )=-α t (1-p t ) γ lbp t

[0080] In the formula, p t Represents the prediction confidence score of the model, α t represents the importance index of balancing positive and negative samples, γ is the focusing function, and F L Represents the Focal Loss function.

[0081] Step 3: Train the SENet-ConvNeXt-FL network model and classify rock thin section images. Use the SENet-ConvNeXt-FL network model to classify rock thin section images using the PyCharm platform as a deep learning framework.

[0082] The specific implementation steps of step three include:

[0083] (1) This experiment uses the PyCharm platform as the deep learning framework on a computer with an InterCorei5-12400f processor, a graphics card model Nvidia GeForce RTX 4060Ti, and 32GB of memory. The software configuration uses Python 3.8 combined with PyTorch 2.0, and the corresponding CUDA version is 11.8.

[0084] (2) During the training process, the training epoch was set to 300, the batch size was 32, the optimizer was SGD, the loss function was focal loss, the learning rate was set to 0.0006, and the regularization coefficient was 0.06. After optimization, the SENet-ConvNeXt-FL model for rock thin section image recognition was finally obtained. Subsequently, the model was verified using the test set to evaluate its performance in the classification task.

[0085] Step 4: Model stability and convergence evaluation;

[0086] Confusion matrix and model performance evaluation indicators are used to conduct multi-dimensional evaluation of rock thin section image classification results.

[0087] The specific implementation steps of step 4 include:

[0088] (1) In order to verify the superiority of SENet-ConvNeXt-FL model over the classic classification models (AlexNet, VGG, ConvNeXt), the confusion matrix is ​​introduced to evaluate the rock thin section image classification results. Figure 3 , comparative analysis Figure 3 It can be seen that compared with the other three classic classification models, the SENet-ConvNeXt-FL model has improved the recognition accuracy of dolomite, siliceous rock, gneiss, schist and sandstone, and the recognition accuracy can reach more than 84.0%, which are 99.0%, 98.0%, 85.0%, 84.4% and 95.0% respectively. The recognition accuracy of the SENet-ConvNeXt-FL model for schist is significantly lower than that of other lithologies. Analysis of the microscope images of schist and gneiss shows that schist and gneiss are both metamorphic rocks and have flaky structural characteristics, which increases the difficulty of the model to extract microscopic image features, and ultimately leads to low classification accuracy of schist.

[0089] (2) By comparing the results of training rock thin section images on AlexNet, VGG, ConvNeXt and SENet-ConvNeXt-FL models (Table 1), it can be found that the AlexNet model was the first to use GPU for training, which greatly accelerated the training speed. However, compared with the VGG and ConvNeXt models, the architecture of the AlexNet model is relatively simple and has fewer parameters, which limits its expressive power. The ConvNeXt model combines the characteristics of convolutional neural networks and Transformer, and has high accuracy and generalization ability. The SENet-ConvNeXt-FL model improves the model's attention to different channel features and enhances the feature representation ability by introducing SENet, and introduces the Focal Loss loss function to effectively solve the category imbalance problem. By comparing the rock thin section classification results of different models, it can be found that the reference Figure 3 From Table 1, we can see that the SENet-ConvNeXt-FL model has the best stability and the highest accuracy.

[0090] Table 1 Experimental results of different models

[0091] Model Accuracy Accuracy Recall F1 value AlexNet 83.0％ 84.6％ 83.2％ 84.0％ VGG 85.0％ 85.8％ 84.0％ 85.0％ ConvNeXt 88.6％ 88.4％ 90.0％ 89.2％ SENet-ConvNeXt-FL 92.5％ 93.0％ 92.3％ 92.5％

[0092] The present invention also includes a rock thin section image classification system based on the SENet-ConvNeXt-FL network, including a processing instruction module for executing each step in the above-mentioned rock thin section image classification method based on the SENet-ConvNeXt-FL network.

[0093] Rock thin section image classification equipment based on SENet-ConvNeXt-FL network, including:

[0094] Memory: a computer-readable device for storing the computer program of the rock thin section image classification method based on the SENet-ConvNeXt-FL network;

[0095] Processor: used to implement the rock thin section image classification method based on SENet-ConvNeXt-FL network when executing the computer program.

[0096] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the rock thin section image classification method based on the SENet-ConvNeXt-FL network can be implemented.

[0097] In summary, the principles of the present invention are:

[0098] Firstly, a rock thin section image dataset was established. Gaussian filtering and data enhancement methods such as random flipping, magnification, displacement, cropping, and scaling were used to improve the quality and quantity of rock thin section images, generate new datasets, and divide the training set and test set into training set and test set according to a certain ratio. Secondly, an improved ConvNeXt network model was established. The SENet module was introduced to optimize the feature extraction capability of ConvNeXt, and the loss function in the training process was changed to the Focal Loss function. The improved model was named SENet-ConvNeXt-FL. Finally, the PyCharm platform was used as the deep learning framework to train the improved model and classify rock thin section images. The confusion matrix and model performance evaluation indicators were used to evaluate the classification results of rock thin section images in multiple dimensions.

[0099] Any person skilled in the art of the present invention may make any modification and change in the form and details of the implementation without departing from the spirit and scope disclosed by the present invention. However, the scope of patent protection of the present invention shall still be subject to the scope defined by the attached claims.

Claims

1. A rock thin section image classification method based on SENet-ConvNeXt-FL network, characterized in that: The following steps are involved: Step 1: Establish a rock thin section image dataset; The original rock slice images are fused into the same data set, and the data enhancement methods such as Gaussian filtering and random flipping, magnification, displacement, cropping, and scaling are used to generate an image-enhanced data set, which is then divided into a training set and a test set. Step 2: Establish the SENet-ConvNeXt-FL network model; The SENet module is introduced to optimize the feature extraction capability of the ConvNeXt network model, and the loss function in the training process is changed to the Focal Loss function, thus obtaining the SENet-ConvNeXt-FL network model; Step 3: Train the SENet-ConvNeXt-FL network model and perform rock thin section image classification; Using PyCharm platform as the deep learning framework, the SENet-ConvNeXt-FL network model is used to classify rock thin section images; Step 4: Model stability and convergence evaluation; Confusion matrix and model performance evaluation indicators are used to conduct multi-dimensional evaluation of rock thin section image classification results.

2. The rock thin section image classification method based on SENet-ConvNeXt-FL network according to claim 1 is characterized in that: The specific implementation steps of step 1 include: (1.1) For the original rock thin section image data set, the Gaussian filtering method is used to remove the noise in the image. The basic principle is to use the Gaussian function to generate a weight matrix, and then perform a convolution operation on the image based on the weight matrix. The specific expression is as follows; In the formula, σ is the standard deviation, which controls the width of the filter, G(x,y) refers to the value of the Gaussian function at the point (x,y), which represents the weight of the two-dimensional Gaussian function at the coordinate (x,y); (1.2) Based on the data enhancement module Augmentor in Python, the rock thin section images after denoising in step (1.1) are subjected to data enhancement processing such as random flipping, magnification, displacement, cropping, and scaling, and the original data set is expanded to n times of the original data set to construct a new rock thin section image enhancement data set; the random flipping refers to randomly flipping the image horizontally or vertically; the magnification, displacement, cropping, and scaling transformations are to change the size and position of the image to simulate the image changes under different shooting conditions; (1.3) The rock thin section image enhancement dataset is randomly divided into p% as a training set and q% as a test set to construct a rock thin section image classification dataset, and the p%+q%=100%.

3. The rock thin section image classification method based on SENet-ConvNeXt-FL network according to claim 1 is characterized in that: The specific implementation steps of step 2 include: (2.1) The SENet module is introduced to optimize the ConvNeXt network model. The SENet module contains three operations: Squeeze operation, Excitation operation and Reweight operation: ①Squeeze operation For each input feature map, global average pooling is used to convert it into a scalar value to capture the global information of each channel. The specific expression is as follows: In the formula, z c is the compressed output; u c is the eigenvalue on the feature map in channel c; ②The Excitation operation maps the global information obtained in the previous step to a channel-related weight vector through a set of fully connected layers and activation functions. The specific expression is as follows: s c =δ{σ2[ReLU(σ1z c )]} In the formula, σ1 and z c Multiplication means output z c The first layer of fully connected layer operation, followed by ReLU activation function and multiplication with σ2 represents the second layer of fully connected operation, δ represents the Sigmoid function, and the output result s after multiplication c ; ③The Reweight operation applies the weight vector calculated in the Excitation operation to the input feature map, reweighting the feature response of each channel in the feature map, thereby enhancing the influence of important feature channels. The specific expression is as follows: o=u c s c In the formula, s c is the weight of channel c, and o is the reweighted output of channel c; (2.2) The loss function in the training process is changed to the Focal Loss function. The Focal Loss network reduces the impact of class imbalance by reducing the weight of samples with a large number and increasing the weight of samples with a small number. The specific expression is as follows: F L (p t )=-α t (1-p t ) γ lbp t In the formula, p t Represents the prediction confidence score of the model, α t represents the importance index of balancing positive and negative samples, γ is the focusing function, and F L Represents the Focal Loss function.

4. The rock thin section image classification method based on SENet-ConvNeXt-FL network according to claim 1 is characterized in that: The step three specifically includes: training the SENet-ConvNeXt-FL network model, including re-segmenting the data at the beginning of each iteration, and training the model using the training set data; During training, the model’s weights are updated based on the training data; After the training is completed, the updated model weights are saved for subsequent use or further analysis. The training improved SENet-ConvNeXt-FL network model includes setting the training round (Epoch), batch size (Bacth Size), learning rate and regularization coefficient.

5. The rock thin section image classification method based on SENet-ConvNeXt-FL network according to claim 1 is characterized in that: The step 4 is specifically as follows: the confusion matrix is ​​used to evaluate the performance of the classification model, the confusion matrix compares the real label and the label predicted by the model, and displays the result in a table, which consists of four elements, namely true positive examples, false positive examples, true negative examples and false negative examples; The model performance evaluation indicators include accuracy (ACC), precision (Pre), recall (Re) and F1 value, and the SENet-ConvNeXt-FL network model is evaluated in multiple dimensions. The specific expressions of the evaluation indicators are as follows: In the formula, TP refers to the number of positive samples that are correctly classified as positive; TN refers to the number of samples that are actually negative and correctly classified as negative; FP refers to the number of samples that are actually negative and incorrectly classified as positive; FN refers to the number of samples that are actually positive and incorrectly classified as negative.

6. Rock thin section image classification system based on SENet-ConvNeXt-FL network, characterized by: It includes a processing instruction module for executing each step in the rock thin section image classification method based on the SENet-ConvNeXt-FL network as described in any one of claims 1 to 5.

7. Rock thin section image classification device based on SENet-ConvNeXt-FL network, characterized in that: include: Memory: a computer-readable device storing the computer program of the rock thin section image classification method based on the SENet-ConvNeXt-FL network as described in any one of claims 1 to 5; Processor: used to implement the rock thin section image classification method based on SENet-ConvNeXt-FL network as described in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the rock thin section image classification method based on the SENet-ConvNeXt-FL network described in any one of claims 1 to 5.

Images