Lithology intelligent identification method based on median-enhanced attention mechanism

Abstract

The invention discloses a lithology intelligent identification method based on a median enhanced attention mechanism, and the method comprises the steps: carrying out the image data enhancement of rock slice image data, and obtaining an enhanced rock slice image data set; the method comprises the following steps: constructing a convolutional neural network EfficentNetV2 model, improving the network structure of the convolutional neural network EfficentNetV2 model, combining channel attention and space attention mechanisms, improving the effect and robustness of model feature extraction, training and verifying the improved EfficentNetV2 model, obtaining a trained rock slice image classification model, and predicting the rock category. According to the method, the influence of subjective factors can be avoided, and efficient intelligent lithology identification is realized.

Description

Technical Field

[0001] The present invention belongs to the intersection of geology and deep learning technology, and specifically designs a lithology intelligent identification method based on median enhanced attention mechanism. Technical Background

[0002] Lithology is a summary of the physical characteristics of rocks. Observing and analyzing the lithology of thin sections made from rock samples using a polarizing microscope is an important basic work in geological research.

[0003] Since the various components in rocks have certain differences in morphology, composition structure, color, etc., they can be analyzed and identified manually. Traditional manual identification methods are often limited by the work efficiency and experience accumulation of identification personnel, and have high time and labor costs. Therefore, the intelligent identification method of rock samples has extremely important application prospects and practical value.

[0004] In recent years, the computing power of computers has been continuously improved, and deep learning methods have also shown excellent performance in the field of image recognition and classification. Using deep learning methods to intelligently identify lithology can free geological researchers from heavy work. Compared with other types of images, rock thin section images have richer feature information, more noise and interference information in the images, and higher requirements for the structural design and performance of the classification model. In the field of computer vision, the performance of the model is highly dependent on the model's ability to extract feature information. Although the traditional convolutional neural network model can capture local features in the image, there is still room for optimization in the extraction of global information and multi-scale feature information. Therefore, while ensuring the model's feature extraction capability and efficiency, it is very important to enhance the model's robustness to noise information for intelligent lithology identification. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, the present invention provides a lithology intelligent identification method based on the median enhanced attention mechanism, which improves the anti-noise ability of the model through median pooling, combines spatial and channel attention mechanisms to capture feature information at different scales, and improves the accuracy of lithology intelligent identification and classification.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is a lithology intelligent identification method based on the median enhanced attention mechanism, and the lithology intelligent identification method specifically includes the following steps:

[0007] Step S1, performing image data enhancement on rock thin section image data to obtain an enhanced rock thin section image data set;

[0008] Step S2, dividing the enhanced rock thin section image data set into a training data set and a test data set according to proportion;

[0009] Step S3, constructing a convolutional neural network EfficientNetV2 model, and modifying the network structure of the convolutional neural network EfficientNetV2 model to obtain an improved EfficientNetV2 model;

[0010] Step S4, inputting the training data set into the improved EfficientNetV2 model for training, and verifying the training effect of the improved EfficientNetV2 model through the test data set until the improved EfficientNetV2 model meets the test accuracy, thereby obtaining a trained rock thin section image classification model;

[0011] Step S5, using the trained rock thin section image classification model to identify and classify the rock thin section image, and obtaining the name and confidence of the rock type identified by the model;

[0012] Preferably, the rock thin section image dataset includes 1882 rock thin section images of 76 types of rocks such as andalusite, andesite, hornblende, gneiss, basalt, porphyry pyroxenite, granodiorite, and perlite and 753 rock thin sections;

[0013] Preferably, the data enhancement technology includes: sequentially performing center cropping, random photometric enhancement, random flipping, boundary expansion and random rotation on each of the rock slice images.

[0014] Preferably, the center cropping includes: taking the geometric center of each rock slice image as a reference, cropping at a ratio of 224 pixels×224 pixels to obtain center cropped image data;

[0015] Preferably, the random photometric enhancement comprises: filling the pixel value of any one of the RGB three color channels of each of the central cropped images with a random value between 0 and 255 with a probability of 10% to obtain optically enhanced image data;

[0016] Preferably, the random flipping comprises: horizontally or vertically flipping each of the optically enhanced images in sequence with a probability of 50% to obtain randomly flipped image data;

[0017] Preferably, the border expansion includes: using edge pixel values ​​as border colors, adding a 100-pixel border to each edge of the randomly flipped image for expansion, and obtaining edge-expanded image data;

[0018] Preferably, the random rotation comprises: randomly selecting any degree in the range of negative 15 degrees to 15 degrees, and rotating each edge-expanded image data in turn to obtain the enhanced rock thin section image data set;

[0019] Preferably, S3 is specifically:

[0020] Construct a convolutional neural network EfficientNetV2 model, set the number of iterations of the EfficientNetV2 model to 500 times, the learning rate to 0.01, add a median enhanced space and channel attention module after the head convolution of the EfficientNetV2 model, and change the last fully connected layer FC1000 of the EfficientNetV2 model to FC76 according to the number of categories of the rock slice image being 76, to obtain an improved EfficientNetV2 model;

[0021] Preferably, S4 is specifically:

[0022] The training data set is input into the improved EfficientNetV2 model for training. During the training process, the model is continuously optimized by continuously adjusting the hyperparameters of the improved EfficientNetV2 model until the model performance meets the accuracy requirements, thereby obtaining an accurate and efficient rock thin section image classification model;

[0023] Preferably, the deep learning hyperparameters include: training batch size, number of iterations, initial learning rate, EfficientNetV2 model parameters, and optimizer threshold.

[0024] Preferably, S5 specifically comprises: identifying and classifying the rock slice image by using the improved EfficientNetV2 model, and taking the category with the highest confidence as the identification result of the improved EfficientNetV2 model according to the calculation result of the fully connected layer;

[0025] Preferably

[0026] The advantages of the present invention are:

[0027] 1) The proposed model can automatically extract feature information from rock thin section images without manual intervention. Compared with traditional technologies, it has higher recognition efficiency for rock thin section images.

[0028] 2) Compared with the prior art, the present invention further improves the recognition rate of rock thin section images by combining the median-enhanced spatial and channel attention methods, and can effectively identify rock thin section images with complex structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0030] Figure 1 A flowchart of a lithology intelligent identification method based on a median enhanced attention mechanism provided in an embodiment of the present invention;

[0031] Figure 2 A structure diagram of a median-enhanced space and channel attention module of a lithology intelligent identification method based on a median-enhanced attention mechanism provided in an embodiment of the present invention;

[0032] Figure 3 An improved EfficientNetV2 model structure diagram of a lithology intelligent identification method based on a median enhanced attention mechanism provided in an embodiment of the present invention;

[0033] Figure 4 An MB convolution module structure diagram of a lithology intelligent identification method based on a median enhanced attention mechanism provided by an embodiment of the present invention;

[0034] Figure 5 A Fused-MB convolution module structure diagram of a lithology intelligent identification method based on a median enhanced attention mechanism provided in an embodiment of the present invention; DETAILED DESCRIPTION

[0035] The present invention is described clearly and completely below in conjunction with the accompanying drawings. A person skilled in the art will be able to implement the present invention based on these descriptions. Before describing the present invention in conjunction with the accompanying drawings, it should be particularly noted that:

[0036] The technical solutions and technical features provided in each part of the present invention, including the following description, may be combined with each other if there is no conflict.

[0037] In addition, the embodiments of the present invention involved in the following description are generally only a part of the embodiments of the present invention, rather than all the embodiments. Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0038] About the terms and units in the present invention: The term "comprises" and any variations thereof in the description and claims of the present invention and related parts are intended to cover non-exclusive inclusions.

[0039] Example 1

[0040] Reference Figure 1 As shown, the present invention proposes a lithology intelligent identification method based on median enhanced attention mechanism, which specifically includes the following steps:

[0041] Step S1, performing image data enhancement on rock thin section image data to obtain an enhanced rock thin section image data set;

[0042] This step is used to expand the rock thin section image dataset to prevent the model from failing to converge due to insufficient number of rock thin section images in the dataset, and to further enhance the performance of the model.

[0043] In this embodiment, the data enhancement technology includes: sequentially performing center cropping, random photometric enhancement, random flipping, boundary expansion and random rotation on each of the rock slice images;

[0044] In this embodiment, the center cropping includes: taking the geometric center of each rock slice image as a reference, cropping at a ratio of 224 pixels×224 pixels to obtain center cropped image data;

[0045] In this embodiment, the random photometric enhancement includes: filling the pixel value of any one of the RGB three color channels of each of the central cropped images with a random value between 0 and 255 with a probability of 10% to obtain optically enhanced image data;

[0046] In this embodiment, the random flipping includes: horizontally or vertically flipping each of the optically enhanced images in sequence with a probability of 50% to obtain randomly flipped image data;

[0047] In this embodiment, the border expansion includes: using the edge pixel value as the border color, adding a 100-pixel border to each edge of the randomly flipped image for expansion, and obtaining edge-expanded image data;

[0048] In this embodiment, the random rotation includes: randomly selecting any degree in the range of negative 15 degrees to 15 degrees, rotating each edge expansion image data in turn, and obtaining the enhanced rock thin section image data set. The enhanced rock thin section image data set includes 76 kinds of rocks and a total of 19376 rock thin section images.

[0049] Step S2: Divide the enhanced rock thin section image data set into a training data set and a test data set according to a certain ratio. The training data set includes 15,501 rock thin section images, and the test data set includes 3,875 rock thin section images.

[0050] Step S3, constructing a convolutional neural network EfficientNetV2 model, and modifying the network structure of the convolutional neural network EfficientNetV2 model to obtain an improved EfficientNetV2 model;

[0051] The median-enhanced spatial and channel attention modules are added after the head convolution of the EfficientNetV2 model. The structure of the median-enhanced spatial and channel attention modules is as follows: Figure 2 As shown in the figure, the input feature map is first processed by three different pooling methods: global average pooling (AvgPool), global maximum pooling (MaxPool) and global median pooling (MedianPool). Each pooling result is processed by a shared multi-layer perceptron (Shared MLP), which contains two 1x1 convolutional layers and a ReLU activation function. Finally, the output value is compressed to the range of [0, 1] using the Sigmoid activation function. The attention maps of the three pooling results are added to obtain the final channel attention map, and the final channel attention map is multiplied with the original input feature map to obtain the weighted feature map. The input feature map in the spatial attention module first extracts basic feature information through a 5×5 deep convolutional layer, and further advances multi-scale feature information through multiple deep convolutional layers of different sizes. The outputs of all deep convolutional layers are generated by element-by-element addition to generate a fused feature map, and the final spatial attention map is generated through a 1x1 convolutional layer. The final spatial attention map is multiplied with the channel weighted feature map to obtain the final output feature. Combining the channel and spatial attention mechanisms, the model can capture and fuse feature information at different scales. Median pooling is introduced into the channel attention mechanism, combined with global average pooling and global maximum pooling to enhance the robustness of the model to noise information. According to the number of categories of the rock slice image being 76, the last fully connected layer FC1000 of the EfficientNetV2 model is changed to FC76 to obtain the improved EfficientNetV2 model.

[0052] The structure of the improved EfficientNetV2 model is as follows Figure 3 As shown in the figure, Conv is a convolution layer for convolution processing, Conv3x3 represents a convolution operation with a convolution kernel size of 3x3, Conv1x1 represents a convolution operation with a convolution kernel size of 1x1, Average Pooling represents global average pooling, and the MB convolution module (MBConv) combines depthwise separable convolution, inverted residual connection, and SE module to further optimize the computational efficiency and model performance. The structure of the MB convolution module is shown in Figure 4As shown in the figure, MBConv4 means that the number of input channels will be expanded to 4 times the original number, and MBConv6 means that the number of input channels will be expanded to 6 times the original number. The Fused-MBConv module is an improvement on the MB convolution module, which combines deep convolution and standard convolution to provide flexible computing and feature extraction capabilities for different scenarios. The structure of the Fused-MBConv module is as follows: Figure 5 As shown in the figure, Fused-MBConv1 does not expand the number of channels, and Fused-MBConv4 expands the number of channels to 4 times the number of input channels.

[0053] Step S4, inputting the training data set into the improved EfficientNetV2 model for training, and verifying the training effect of the improved EfficientNetV2 model through the test data set until the improved EfficientNetV2 model meets the test accuracy, thereby obtaining a trained rock thin section image classification model;

[0054] The deep learning hyperparameters include: training batch size, number of iterations, initial learning rate, EfficientNetV2 model parameters, and optimizer threshold.

[0055] Step S5, using the trained rock thin section image classification model to identify and classify the rock thin section image, and obtaining the name and confidence of the rock type identified by the model;

[0056] The improved EfficientNetV2 model identifies and classifies rock thin section images, identifies the specific category name of the rock, and records the confidence of the rock category, and selects the name with the highest confidence as the final rock category name.

[0057] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

[0058] It should be noted that the drawings are in a very simplified form and are not in precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purposes, features and advantages of the present invention more obvious and easy to understand, please refer to the drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope of the technical content disclosed in the present invention.

Claims

1. A lithology intelligent identification method based on median enhanced attention mechanism, characterized in that: The specific steps include: Step S1, performing image data enhancement on rock thin section image data to obtain an enhanced rock thin section image data set; Step S2, dividing the enhanced rock thin section image data set into a training data set and a test data set according to proportion; Step S3, constructing a convolutional neural network EfficientNetV2 model, and modifying the network structure of the convolutional neural network EfficientNetV2 model to obtain an improved EfficientNetV2 model; Step S4, inputting the training data set into the improved EfficientNetV2 model for training, and verifying the training effect of the improved EfficientNetV2 model through the test data set until the improved EfficientNetV2 model meets the test accuracy, thereby obtaining a trained rock thin section image classification model; Step S5: Using the trained rock thin section image classification model to identify and classify the rock thin section image, and obtain the name and confidence of the rock type identified by the model.

2. The lithology intelligent identification method based on median enhanced attention mechanism according to claim 1 is characterized in that: The rock thin section image dataset includes 1882 rock thin section images of 76 types of rocks, including andalusite, andesite, hornblende, gneiss, basalt, porphyry pyroxenite, granodiorite, and perlite, and 753 rock thin sections.

3. The lithology intelligent identification method based on median enhanced attention mechanism according to claim 1 is characterized in that: The data enhancement operation includes: sequentially performing center cropping, random photometric enhancement, random flipping, boundary expansion and random rotation on each of the rock slice images.

4. The lithology intelligent identification method based on median enhanced attention mechanism according to claim 1 is characterized in that: The S3 is specifically: A convolutional neural network EfficientNetV2 model was constructed, the number of iterations of the EfficientNetV2 model was set to 500 times, the learning rate was set to 0.01, and the median enhanced spatial and channel attention modules were added after the head convolution of the EfficientNetV2 model. According to the number of categories of the rock thin section images being 76, the last fully connected layer FC1000 of the EfficientNetV2 model was changed to FC76 to obtain an improved EfficientNetV2 model.

5. The lithology intelligent identification method based on median enhanced attention mechanism according to claim 1 is characterized in that: The S4 is specifically: The training data set is input into the improved EfficientNetV2 model for training. During the training process, the model is continuously optimized by continuously adjusting the hyperparameters of the improved EfficientNetV2 model until the model performance meets the accuracy requirements, thereby obtaining an accurate and efficient rock thin section image classification model.

6. The lithology intelligent identification method based on median enhanced attention mechanism according to claim 1 is characterized in that: The S5 is specifically: The rock thin section images are identified and classified by the improved EfficientNetV2 model, and according to the calculation results of the fully connected layer, the category with the highest confidence is used as the recognition result of the improved EfficientNetV2 model.

Images