Rock slice image intelligent analysis system and method based on deep learning

Abstract

The invention relates to the technical field of rock slice image intelligent analysis systems, in particular to a rock slice image intelligent analysis system based on deep learning and a method thereof. The mineral and pore structure identification module is used for extracting local and multi-scale information of the image by using a deep neural network model, and processing the preprocessed rock slice image through feature mapping to obtain a high-confidence segmentation mask image of minerals and pores; the result visualization module is used for carrying out spatial feature fusion on the segmentation mask image and the original rock slice image to realize fusion display of the segmentation result and the original image, and the multiple receptive field image pooling module uses convolution kernels of different sizes and can capture mineral particle features of different sizes at the same time to realize fusion display of the segmentation result and the original image. The method is crucial for accurately identifying complex mineral aggregates. The multi-scale feature fusion module greatly improves the segmentation precision by organically combining global and local features.

Description

Technical Field

[0001] The present invention relates to the technical field of rock thin section image intelligent analysis systems, and in particular to a rock thin section image intelligent analysis system and method based on deep learning. Background Art

[0002] With the deepening of geological research, rock thin section analysis plays an increasingly important role in fields such as mineralogy, sedimentology and petroleum geology. Traditional rock thin section analysis mainly relies on professional geologists to observe through microscopes. Although this method is reliable, it is time-consuming and labor-intensive, and has subjective and consistency problems. In recent years, with the development of computer vision and image processing technology, automated rock thin section analysis methods have gradually emerged.

[0003] Early automated methods were mainly based on traditional image processing techniques, such as threshold segmentation and edge detection. These methods performed well when dealing with simple mineral compositions, but when faced with complex rock thin sections, it was often difficult to accurately identify multiple minerals and complex pore structures. Subsequently, machine learning methods, especially support vector machines (SVMs) and random forests, were introduced into rock thin section analysis, improving the accuracy of the analysis. However, these methods still require manual feature design and are difficult to fully utilize all the information in the image.

[0004] In recent years, the rise of deep learning technology has brought new opportunities for rock thin section analysis. Some researchers have tried to use convolutional neural networks (CNN) for mineral identification and pore structure segmentation. These methods have significantly improved over traditional methods, but there are still some key problems. First, most of the existing deep learning methods use standard network structures, such as U-Net or FCN. Although these structures perform well in fields such as medical images, they are not optimized for the particularity of rock thin sections. The size of mineral particles in rock thin sections varies significantly, ranging from microns to millimeters, and standard network structures are difficult to capture these different scale features at the same time.

[0005] Secondly, existing methods often face the problem of low computational efficiency when processing high-resolution rock thin section images. In order to ensure recognition accuracy, processing speed has to be sacrificed, which greatly limits the promotion of the method in practical applications. In addition, existing methods perform poorly when processing complex mineral aggregates, and it is difficult to accurately distinguish different mineral particles that are closely connected. Finally, most existing methods lack effective visualization methods, making it difficult to intuitively display the analysis results, which creates obstacles for geologists to conduct further research and interpretation.

[0006] In view of the above problems, there is an urgent need for an intelligent analysis system that can efficiently and accurately analyze high-resolution rock thin section images and provide intuitive visualization results. Summary of the invention

[0007] The present invention is proposed to meet this demand. Through innovative network structure design, the present invention can simultaneously process mineral features of different scales, greatly improving the recognition accuracy. At the same time, the present invention adopts an efficient feature fusion strategy, which significantly reduces the computational complexity while improving the accuracy, and realizes efficient processing. In addition, the present invention also provides a flexible visualization module, making the analysis results more intuitive and easy to understand.

[0008] The present invention proposes a rock thin section image intelligent analysis system based on deep learning, which includes: a rock thin section image preprocessing module, which is used to perform grayscale, gamma correction, filtering and histogram equalization operations on the collected high-resolution rock thin section images to remove uneven illumination and noise interference and improve image quality;

[0009] The mineral and pore structure recognition module uses a deep neural network model to extract local and multi-scale information of the image, processes the pre-processed rock thin section image through feature mapping, and obtains a high-confidence segmentation mask image of minerals and pores;

[0010] The result visualization module performs spatial feature fusion on the segmentation mask image and the original rock slice image to realize the fusion display of the segmentation result and the original image.

[0011] Preferably, the rock slice image preprocessing module comprises:

[0012] The image grayscale conversion unit reduces the dimension of the data by reducing the dimension of the rock slice image channel; the image gamma correction unit adjusts the contrast of the image, where the gamma correction formula is: Among them I out is the output image, I in is the input image, A is a constant, and γ is the gamma value;

[0013] Image filtering unit, including low-pass filtering, median filtering and Gaussian filtering;

[0014] The image histogram equalization unit changes the grayscale histogram of the original image from a relatively concentrated grayscale interval to a uniform distribution in the entire grayscale range.

[0015] Preferably, the mineral and pore structure identification module includes:

[0016] The training image segmentation unit labels the rock thin section images and segmentation mask images, binds the rock thin section image pixels and image categories, and then combines them together to form a training data set; the rock thin section image deep neural network model, which fuses feature maps of the rock thin section images at both global and local levels, and then generates a high-confidence segmentation mask image.

[0017] Preferably, the rock thin section image deep neural network model comprises:

[0018] The global feature extraction submodule extracts global features from the original rock slice image through convolutional layers and pooling layers, and then processes it through ReLu activation function and batch normalization.

[0019] The multi-receptive field image pooling module uses three convolution kernels of different sizes as filters in the same block, extracts features of different receptive fields through depthwise separable convolution operations, and aggregates global and local feature information through feature splicing;

[0020] The multi-scale feature fusion module combines the multi-scale features of the global feature extraction submodule and the features of the multi-receptive field image pooling module in the channel direction to form a multi-scale fusion feature.

[0021] The multi-category image segmentation module classifies the multi-scale fusion features at the pixel level, predicts the probability of belonging to each category at each pixel on the feature map, and obtains a high-confidence segmentation mask image through the softmax activation function, where the softmax function is defined as: where z i is the prediction score of the i-th category, σ(z) i is the predicted probability of the ith category.

[0022] Preferably, the multiple receptive field image pooling module includes:

[0023] The original rock slice image input unit performs a depthwise convolution operation with multiple filters with different convolution kernel sizes but the same stride;

[0024] A global feature extraction unit processes the original rock slice image to obtain global features;

[0025] The deep feature extraction unit includes filters of different scales, a feature splicing unit and a feature compressor, wherein filters of different scales respectively extract features of different receptive fields through deep convolution operations, the feature splicing unit splices the features extracted by filters of all different scales in the same channel, and the feature compressor compresses the number of channels of local features extracted by filters of all different scales to the same number of channels as the global features.

[0026] Preferably, the result visualization module adopts an image fusion method to merge the mask image with the original rock thin section image to intuitively observe the microstructure of the rock thin section, including the splicing of the rock thin section image and the mask image and the overlapping display of the rock thin section image and the mask image, wherein the image fusion adopts the weighted average method, and the fusion formula is: F(x, y)=w1·I1(x, y)+w2·I2(x, y), wherein F(x, y) is the fused image, I1(x, y) and I2(x, y) are the original image and the mask image respectively, w1 and w2 are weight coefficients, and w1+w2=1.

[0027] According to the system's deep learning-based rock slice image intelligent analysis method, the method includes the following steps:

[0028] Obtain full-color images of rock thin sections;

[0029] Preprocess the acquired original image, including grayscale conversion, gamma correction, filtering and histogram equalization operations;

[0030] Collect mineral and pore image data and corresponding image mask image data;

[0031] Label the preprocessed images, bind the rock slice image pixels and image categories, and then combine them together to form a training data set;

[0032] The training data set and label data are cross-validated to obtain the optimal parameters and model training of the classification model, and the mineral and pore structure recognition model is trained;

[0033] Using the mineral and pore structure recognition model to segment the rock thin section image to obtain a high-confidence segmentation mask image, and performing morphological processing to improve the quality and edge smoothness of the mask image;

[0034] Performing spatial feature fusion on the mask image and the original rock slice image to form a merged image;

[0035] The rock slice image, the mask image and the merged image are visualized.

[0036] Preferably, the training process of the mineral and pore structure recognition model includes:

[0037] The preprocessed image is input into the global feature extraction submodule, and global feature extraction is performed through the convolution layer and pooling layer, and then processed by the ReLu activation function and batch normalization;

[0038] The features output by the global feature extraction module are fused with the multi-scale channel splicing features through the multi-scale feature fusion module;

[0039] The above fusion results are input into the multi-receptive field image pooling module, and then pass through the multi-scale feature fusion module;

[0040] Finally, enter the multi-category image segmentation module to perform multi-category image segmentation to obtain mineral and pore mask images.

[0041] Preferably, the morphological processing comprises:

[0042] The morphological method is used to open and smooth the mask image, fill the connected pore structure image, eliminate small area holes in the image, and remove the interference objects connected to the contour of the target area;

[0043] The mask image is supplemented by closing operation to make the image outline clear and remove unnecessary fragments;

[0044] The mask image is processed by corrosion, expansion and skeleton operations to make the edge transition of the mask image smooth; the mask image data is classified and processed by grayscale threshold segmentation method to remove image noise, where the grayscale threshold T is selected by Otsu algorithm, that is, the maximum inter-class variance method, and the calculation formula is: T = argmax{w1(T)·w2(T)·[μ1(T)-μ2(T)] 2}, where w1(T) and w2(T) are the probabilities of the two classes under the threshold T, μ1(T) and μ2(T) are the average gray values ​​of the two classes;

[0045] The mask image is segmented by morphological operation. The mask image data is binarized through opening and closing operations to retain the large area and main boundaries of minerals and pores and remove small area connected structures. The pore mask image data is segmented by pore area division.

[0046] Preferably, the visualization step comprises:

[0047] The mask image is fused with the original rock slice image and the merged image is displayed by setting different grayscale ranges;

[0048] When the grayscale range of the mineral mask image is 0-255 and the grayscale range of the pore mask image is 0-100, the grayscale range of the merged image is 0-255, among which the mineral mask image has the largest grayscale and is displayed as red in the merged image, and the pore mask image has the smallest grayscale and is displayed as blue in the merged image;

[0049] When the grayscale range of the original rock thin section image is also 0 to 255, the mineral mask image in the merged image is displayed in red, the original rock thin section image is displayed in black, and the pore mask image is displayed in blue;

[0050] The mineral mask image is directly fused and displayed with the original rock slice image in the pixel dimension;

[0051] The mineral mask image, the pore mask image and the original rock thin section image are respectively channel-joined to obtain a multi-channel mineral mask image, a multi-channel rock thin section image and a multi-channel pore mask image;

[0052] The mineral area of ​​the multi-channel mineral mask image is inverted, and then multiple channels are remapped to balance the image color and maintain the contrast between image data. The image remapping adopts a histogram matching method to adjust the histogram of the source image to the shape of the target histogram. The adjustment formula is: g(x)=T(f(x)), where f(x) is the source image, g(x) is the target image, and T is the transformation function.

[0053] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0054] The system of the present invention effectively solves the problems existing in the prior art by cleverly combining the multi-receptive field image pooling module and the multi-scale feature fusion module. The multi-receptive field image pooling module uses convolution kernels of different sizes to simultaneously capture the characteristics of mineral particles of different sizes, which is crucial for accurately identifying complex mineral aggregates. The multi-scale feature fusion module greatly improves the accuracy of segmentation by organically combining global and local features. The synergy of these two modules not only improves the recognition accuracy of the system, but also significantly improves the processing efficiency.

[0055] It is worth mentioning that the present invention not only improves performance but also reduces the complexity of the model. This seemingly contradictory result actually reflects the ingenuity of the present invention in algorithm design. Through efficient feature extraction and fusion strategies, the system can achieve better performance with fewer parameters, which not only reduces the risk of overfitting but also improves the generalization ability of the model.

[0056] In addition, the visualization module of the present invention adopts an innovative image fusion method, which can seamlessly combine the segmentation results with the original image, providing geologists with intuitive and easy-to-understand analysis results. This visualization method not only facilitates experts to conduct further research and interpretation, but also makes it possible for non-professionals to understand the analysis results.

[0057] In summary, the deep learning-based rock thin section image intelligent analysis system of the present invention successfully solves the problems of insufficient accuracy, low efficiency and visualization difficulties existing in the prior art through innovative network design and efficient processing strategies. It not only greatly improves the efficiency and accuracy of rock thin section analysis, but also provides new perspectives and tools for geological research, and is expected to promote the rapid development of related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall logic block diagram of the system of the present invention.

[0059] Figure 2 It is a logic block diagram of the rock slice image preprocessing module of the present invention.

[0060] Figure 3 It is a logic block diagram of the mineral and pore structure identification module of the present invention.

[0061] Figure 4 It is a logic block diagram of the multi-receptive field image pooling module of the present invention.

[0062] Figure 5 It is a logic block diagram of the result visualization module of the present invention.

[0063] Figure 6 It is a logic block diagram of the morphological processing steps of the present invention. DETAILED DESCRIPTION

[0064] In order to further explain the technical means and effects of the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments thereof. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0065] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0066] Example 1

[0067] See also Figure 1-6 The present invention proposes a rock thin section image intelligent analysis system based on deep learning. The system is mainly used to perform intelligent analysis on rock thin section images, identify the mineral composition and pore structure therein, and provide important support for geological research.

[0068] First, let us describe the system structure of the present invention in detail. The system of the present invention includes a rock thin section image preprocessing module 1, a mineral and pore structure recognition module 2 and a result visualization module 3. The rock thin section image preprocessing module 1 is used to grayscale, gamma correct, filter and histogram equalization operations on the collected high-resolution rock thin section images to remove uneven illumination and noise interference and improve image quality. Then, the mineral and pore structure recognition module 2 uses a deep neural network model to extract local and multi-scale information of the image, and processes the preprocessed rock thin section image through feature mapping to obtain a high-confidence segmentation mask image of minerals and pores. Finally, the result visualization module 3 performs spatial feature fusion of the segmentation mask image with the original rock thin section image to achieve a fusion display of the segmentation result and the original image.

[0069] Next, we further explain the specific structure of the rock thin section image preprocessing module 1. The module includes an image grayscale conversion unit 11, an image gamma correction unit 12, an image filtering unit 13 and an image histogram equalization unit 14. The image grayscale conversion unit 11 reduces the dimension of the data by reducing the dimension of the rock thin section image channel. This step can not only reduce the amount of calculation for subsequent processing, but also highlight the structural features of the image. Preferably, in one embodiment of the present invention, we use a weighted average method for grayscale conversion, and the weight ratio is R:G:B=0.299:0.587:0.114. This weight setting takes into account the sensitivity of the human eye to different colors.

[0070] The image gamma correction unit 12 adjusts the contrast of the image, wherein the gamma correction formula is:

[0071] Among them I out is the output image, I in is the input image, A is a constant, and γ is the gamma value. By adjusting the γ value, we can change the overall brightness and contrast of the image. In one embodiment of the present invention, we found that when A=1 and γ=0.8, good results can be obtained for most rock slice images. This is because rock slice images usually have low contrast. Using γ<1 can increase the contrast of the image and make the boundaries of mineral particles clearer.

[0072] The image filtering unit 13 includes low-pass filtering, median filtering and Gaussian filtering. The main purpose of these filtering operations is to reduce noise and enhance image edges. Preferably, we use a 3x3 Gaussian filter kernel with σ=0.5. This parameter setting can effectively remove high-frequency noise in the image while retaining edge information. For some rock slice images with more noise, we can first perform median filtering and then Gaussian filtering. This combination can better remove salt and pepper noise and Gaussian noise.

[0073] The image histogram equalization unit 14 changes the grayscale histogram of the original image from a relatively concentrated grayscale interval to a uniform distribution within the entire grayscale range. This step is very important for enhancing image contrast and highlighting image details. In practical applications, we have found that for some rock slice images with particularly low contrast, local histogram equalization can be performed first, and then global histogram equalization can be performed. This method can better retain local details.

[0074] Next, we describe the mineral and pore structure recognition module 2 in detail. The module includes a training image segmentation unit 21 and a rock thin section image deep neural network model 22. The training image segmentation unit 21 labels the rock thin section image and the segmentation mask image, binds the rock thin section image pixels and image categories, and then combines them together to form a training data set. In one embodiment of the present invention, we used 1,000 labeled rock thin section images as a training set, which included various common mineral types (such as quartz, feldspar, mica, etc.) and pore structures. These images come from different rock types, including sedimentary rocks, igneous rocks, and metamorphic rocks, to ensure the generalization ability of the model.

[0075] The rock thin section image deep neural network model 22 is the core of the present invention. As described in claim 4, the model includes a global feature extraction submodule 221, a multi-receptive field image pooling module 222, a multi-scale feature fusion module 223 and a multi-category image segmentation module 224. The global feature extraction submodule 221 uses 5 convolution layers and 3 pooling layers, and each convolution layer uses a ReLu activation function and batch normalization. This design can effectively extract the global features of the image and lay the foundation for subsequent fine segmentation.

[0076] The multiple receptive field image pooling module 222 is an innovative point of the present invention. As described in claim 5, the module includes an original rock slice image input unit 2221, a global feature extraction unit 2222 and a deep feature extraction unit 2223. The deep feature extraction unit 2223 uses three sizes of convolution kernels: 3x3, 5x5 and 7x7. This design can capture features of different scales at the same time, and is particularly suitable for processing mineral particles of different sizes in rock slices. For example, a 3x3 convolution kernel is suitable for capturing fine mineral particles or pore structures, while a 7x7 convolution kernel can capture larger mineral particles or overall structures.

[0077] The multi-scale feature fusion module 223 splices the global features and the local features together in the channel direction to form a multi-scale fusion feature. This fusion method can retain more spatial and semantic information, which helps to improve the accuracy of segmentation. In practical applications, we found that this feature fusion method is particularly effective for identifying complex mineral aggregates.

[0078] Finally, the multi-class image segmentation module 224 uses the softmax activation function to obtain a high confidence segmentation mask image. The softmax function is defined as: where z i is the prediction score of the i-th category, σ(z) i is the predicted probability of the i-th category. In one embodiment of the present invention, we classify minerals into 10 categories, including common minerals such as quartz, feldspar, mica, and pore structures. This multi-category segmentation method can identify multiple minerals at the same time, greatly improving the analysis efficiency.

[0079] The result visualization module 3 uses the weighted average method to perform image fusion, and the fusion formula is:

[0080] F(x, y) = w1·I1(x, y) + w2·I2(x, y), where F(x, y) is the fused image, I1(x, y) and I2(x, y) are the original image and mask image respectively, w1 and w2 are weight coefficients, and w1+w2=1. In practical applications, we found that when w1=0.7 and w2=0.3, the segmentation results can be clearly displayed while retaining the original image information. This visualization method allows geologists to intuitively observe the mineral distribution and pore structure without losing the details of the original image.

[0081] Next, we describe the method of the present invention in detail. The method includes the steps of obtaining a full-color image of a rock slice, preprocessing the obtained original image, collecting mineral and pore image data and corresponding image mask image data, labeling the preprocessed image, training a mineral and pore structure recognition model, using the model to perform image segmentation, and performing feature fusion and visualization. The sequential arrangement of these steps ensures the fluency and accuracy of the analysis process.

[0082] During the model training process, we adopted a feature extraction strategy from global to local. First, the preprocessed image is input into the global feature extraction submodule, and global feature extraction is performed through the convolution layer and pooling layer, and then processed by the ReLu activation function and batch normalization. Next, the features output by the global feature extraction module are fused with the multi-scale channel splicing features through the multi-scale feature fusion module. Then, the fusion result is input into the multi-receptive field image pooling module and passed through the multi-scale feature fusion module again. Finally, the multi-category image segmentation module is entered to perform image segmentation of multiple categories to obtain mineral and pore mask images. This strategy can capture the overall structure and local details of rock slices at the same time, and is particularly suitable for processing complex rock slice images.

[0083] Morphological processing is a key step to improve the quality of segmentation results. We use a series of morphological operations, including opening, closing, erosion, dilation, and skeleton operations. First, the morphological method is used to open and smooth the mask image, fill the connected pore structure image, eliminate small holes in the image, and remove interference connected to the contour of the target area. Then, the mask image is supplemented by closing operations to make the image contour clear and remove unnecessary fragments. Next, the mask image is processed by erosion, dilation, and skeleton operations to make the edge transition of the mask image smooth.

[0084] In one embodiment of the present invention, we found that using a 3x3 structure element for opening and closing operations can effectively remove small area noise while retaining the main structure. For some rock slices containing a large number of tiny mineral particles, we can use a 2x2 structure element to better retain these tiny structures.

[0085] In addition, we also use the grayscale threshold segmentation method to classify the mask image data and remove image noise. The grayscale threshold T is selected using the Otsu algorithm, that is, the maximum inter-class variance method, and the calculation formula is: T = argmax{w1(T)·w2(T)·[μ1(T)-μ2(T)] 2}, where w1(T) and w2(T) are the probabilities of the two classes under the threshold T, and μ1(T) and μ2(T) are the average grayscale values ​​of the two classes. Through a large number of experiments, we found that the best segmentation effect can be obtained when T is between 100-150. This range is suitable for most rock slice images, but for some special rock types (such as dark igneous rocks), the threshold may need to be adjusted to between 80-120.

[0086] Finally, the visualization step includes multiple display modes. We fuse the mask image with the original rock thin section image and display the merged image by setting different grayscale ranges. When the grayscale range of the mineral mask image is 0-255 and the grayscale range of the pore mask image is 0-100, the grayscale range of the merged image is 0-255. Under this setting, the mineral mask image has the largest grayscale and is displayed in red in the merged image, and the pore mask image has the smallest grayscale and is displayed in blue in the merged image. When the grayscale range of the original rock thin section image is also 0-255, the mineral mask image in the merged image is displayed in red, the original rock thin section image is displayed in black, and the pore mask image is displayed in blue. This color coding scheme makes the analysis results clear at a glance and easy for geologists to interpret.

[0087] In addition, we also used the histogram matching method for image remapping. The adjustment formula is: g(x) = T(f(x)), where f(x) is the source image, g(x) is the target image, and T is the transformation function. In practical applications, we found that the best visual effect can be obtained when the target histogram adopts a uniform distribution. This method is particularly suitable for processing rock thin section images with low contrast and can effectively enhance the detail information of the image.

[0088] In general, the rock thin section image intelligent analysis system and method based on deep learning proposed in the present invention can analyze rock thin section images efficiently and accurately through innovative network structure design and carefully designed image processing flow. The system can not only identify common mineral types and pore structures, but also process complex mineral aggregates, providing a powerful tool support for geological research. In addition, the visualization function of the system makes the analysis results more intuitive, which helps geologists conduct in-depth research and interpretation.

[0089] In order to verify the superiority of the rock thin section image intelligent analysis system based on deep learning of the present invention, we conducted a series of experiments. The following is a detailed description and analysis of the embodiments and comparative examples:

[0090] Example 1: In this example, we used the complete system of the present invention, including a rock thin section image preprocessing module, a multi-receptive field image pooling module, and a multi-scale feature fusion module. We used a dataset of 5,000 high-resolution rock thin section images, which cover a variety of rock types such as sedimentary rocks, igneous rocks, and metamorphic rocks. The image resolution is 2048x2048 pixels, and contains annotations of 10 common minerals (such as quartz, feldspar, mica, etc.) and pore structures.

[0091] We first preprocess the image, including grayscale, gamma correction (γ = 0.8), and Gaussian filtering (3x3 kernel, σ = 0.5). Then, we extract features using a multi-receptive field image pooling module, which contains three convolution kernels: 3x3, 5x5, and 7x7. Finally, we combine global and local features through a multi-scale feature fusion module.

[0092] Comparative Example 1: In the comparative example, we used the traditional U-Net network structure, without using the multi-receptive field image pooling module and the multi-scale feature fusion module. The preprocessing steps are the same as those in Example 1, but the network structure only includes the standard encoder-decoder architecture.

[0093] Test indicators and methods:

[0094] 1. Mineral identification accuracy: Use pixel-level accuracy to evaluate the model’s ability to identify different minerals.

[0095] 2. Pore structure segmentation accuracy: IoU (intersection over union) is used to evaluate the accuracy of pore structure segmentation.

[0096] 3. Processing time: records the average time required to process a 2048x2048 resolution image.

[0097] 4. Model parameter quantity: The total number of parameters of the statistical model, reflecting the complexity of the model.

[0098] We used 1,000 test set images for evaluation, which came from different rock samples than the training set to ensure fairness. The test results are shown in the following table:

[0099] index Example 1 Comparative Example 1 Mineral identification accuracy 94.5％ 89.2％ Pore ​​structure segmentation accuracy (IoU) 0.91 0.85 Processing time (seconds / picture) 0.8 1.2 Model parameters (millions) 15.3 23.7

[0100] It can be clearly seen from the test results that the system of the present invention outperforms the traditional U-Net structure in all indicators. First, in terms of mineral recognition accuracy, the system of the present invention reached 94.5%, 5.3 percentage points higher than the control ratio. This significant improvement is mainly due to the application of the multi-receptive field image pooling module, which can capture features of different scales at the same time, and is particularly suitable for processing mineral particles of different sizes in rock slices.

[0101] Secondly, in terms of pore structure segmentation accuracy, the IoU value of the system of the present invention reaches 0.91, which is significantly higher than the 0.85 of the control group. This shows that our system can not only accurately identify minerals, but also accurately segment the pore structure. This is crucial for studying key properties such as rock porosity and permeability.

[0102] In terms of processing time, although our system structure is more complex, due to the use of efficient multi-scale feature fusion strategy, the actual processing speed is 33% faster than the traditional U-Net. This means that more rock thin section images can be analyzed in the same time using our system, greatly improving work efficiency.

[0103] Finally, it is worth noting that although our system has better performance, the number of model parameters is 35% less than that of the traditional U-Net. This not only reduces the storage requirements of the model, but also reduces the risk of overfitting and improves the generalization ability of the model.

[0104] These results fully demonstrate the superiority of the system of the present invention. The innovative design of the multi-receptive field image pooling module and the multi-scale feature fusion module enables the system to better adapt to the complexity and diversity of rock thin section images. The high-precision mineral identification and pore structure segmentation capabilities, combined with the fast processing speed, make this system a powerful tool for geological research. It can not only help geologists analyze rock samples more quickly and accurately, but also reveal subtle structural features that are difficult to discover with traditional methods, providing new possibilities for in-depth understanding of the rock formation process and prediction of its physical properties.

[0105] In summary, the deep learning-based rock thin section image intelligent analysis system of the present invention shows obvious advantages in accuracy, efficiency and applicability, providing a powerful new tool for geological research.

[0106] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A rock thin section image intelligent analysis system based on deep learning, characterized by: The system includes: The rock thin section image preprocessing module is used to grayscale, gamma correct, filter and histogram equalize the collected high-resolution rock thin section images to remove uneven illumination and noise interference and improve image quality; The mineral and pore structure recognition module uses a deep neural network model to extract local and multi-scale information of the image, processes the pre-processed rock thin section image through feature mapping, and obtains a high-confidence segmentation mask image of minerals and pores; The result visualization module performs spatial feature fusion on the segmentation mask image and the original rock slice image to realize the fusion display of the segmentation result and the original image.

2. The system according to claim 1, characterized in that The rock slice image preprocessing module comprises: The image grayscale conversion unit reduces the dimension of the data by reducing the dimension of the rock slice image channel; the image gamma correction unit adjusts the contrast of the image, where the gamma correction formula is: Among them I out is the output image, I in is the input image, A is a constant, and γ is the gamma value; Image filtering unit, including low-pass filtering, median filtering and Gaussian filtering; The image histogram equalization unit changes the grayscale histogram of the original image from a relatively concentrated grayscale interval to a uniform distribution in the entire grayscale range.

3. The system according to claim 1, characterized in that The mineral and pore structure identification module includes: The training image segmentation unit labels the rock slice image and the segmentation mask image, binds the rock slice image pixels and image categories, and then combines them together to form a training data set; A deep neural network model for rock thin section images, which fuses feature maps of rock thin section images at both global and local levels, and then generates a high-confidence segmentation mask image.

4. The system according to claim 3, characterized in that The rock thin section image deep neural network model includes: The global feature extraction submodule extracts global features from the original rock slice image through convolutional layers and pooling layers, and then processes it through ReLu activation function and batch normalization. The multi-receptive field image pooling module uses three convolution kernels of different sizes as filters in the same block, extracts features of different receptive fields through depthwise separable convolution operations, and aggregates global and local feature information through feature splicing; The multi-scale feature fusion module combines the multi-scale features of the global feature extraction submodule and the features of the multi-receptive field image pooling module in the channel direction to form a multi-scale fusion feature. The multi-category image segmentation module classifies the multi-scale fusion features at the pixel level, predicts the probability of belonging to each category at each pixel on the feature map, and obtains a high-confidence segmentation mask image through the softmax activation function, where the softmax function is defined as: where z i is the prediction score of the i-th category, σ(z) i is the predicted probability of the ith category.

5. The system according to claim 4, characterized in that The multiple receptive field image pooling module includes: The original rock slice image input unit performs a depthwise convolution operation with multiple filters with different convolution kernel sizes but the same stride; A global feature extraction unit processes the original rock thin section image to obtain global features; The deep feature extraction unit includes filters of different scales, a feature splicing unit and a feature compressor, wherein filters of different scales respectively extract features of different receptive fields through deep convolution operations, the feature splicing unit splices the features extracted by filters of all different scales in the same channel, and the feature compressor compresses the number of channels of local features extracted by filters of all different scales to the same number of channels as the global features.

6. The system according to claim 1, characterized in that The result visualization module adopts an image fusion method to merge the mask image with the original rock thin section image to intuitively observe the microstructure of the rock thin section, including the splicing of the rock thin section image and the mask image and the overlapping display of the rock thin section image and the mask image. The image fusion adopts a weighted average method, and the fusion formula is: F(x, y)=w1·I1(x, y)+w2·I2(x, y), where F(x, y) is the fused image, I1(x, y) and I2(x, y) are the original image and the mask image respectively, w1 and w2 are weight coefficients, and w1+w2=1.

7. The rock thin section image intelligent analysis method based on deep learning according to any one of claims 1 to 6, characterized in that: The method comprises the following steps: Obtain full-color images of rock thin sections; Preprocess the acquired original image, including grayscale conversion, gamma correction, filtering and histogram equalization operations; Collect mineral and pore image data and corresponding image mask image data; Label the preprocessed images, bind the rock slice image pixels and image categories, and then combine them together to form a training data set; The training data set and label data are cross-validated to obtain the optimal parameters and model training of the classification model, and the mineral and pore structure recognition model is trained; Using the mineral and pore structure recognition model to segment the rock thin section image to obtain a high-confidence segmentation mask image, and performing morphological processing to improve the quality and edge smoothness of the mask image; The mask image and the original rock slice image are spatially fused to form a merged image; and the rock slice image, the mask image and the merged image are visualized.

8. The method according to claim 7, characterized in that The training process of the mineral and pore structure recognition model includes: The preprocessed image is input into the global feature extraction submodule, and global feature extraction is performed through the convolution layer and pooling layer, and then processed by the ReLu activation function and batch normalization; The features output by the global feature extraction module are fused with the multi-scale channel splicing features through the multi-scale feature fusion module; The above fusion results are input into the multi-receptive field image pooling module, and then pass through the multi-scale feature fusion module; Finally, enter the multi-category image segmentation module to perform multi-category image segmentation to obtain mineral and pore mask images.

9. The method according to claim 7, characterized in that: The morphological processing includes: The morphological method is used to open and smooth the mask image, fill the connected pore structure image, eliminate small area holes in the image, and remove the interference objects connected to the contour of the target area; The mask image is supplemented by closing operation to make the image contour clear and remove unnecessary fragments; the mask image is processed by corrosion, expansion and skeleton operations to make the edge transition of the mask image smooth; the mask image data is classified by gray threshold segmentation and gray threshold segmentation method to remove image noise, where the gray threshold T is selected by Otsu algorithm, that is, the maximum inter-class variance method, and the calculation formula is: T = argmax{w1(T)·w2(T)·[μ1(T)-μ2(T)] 2 }, where w1(T) and w2(T) are the probabilities of the two classes under the threshold T, μ1(T) and μ2(T) are the average gray values ​​of the two classes; The mask image is segmented by morphological operation. The mask image data is binarized through opening and closing operations to retain the large area and main boundaries of minerals and pores and remove small area connected structures. The pore mask image data is segmented by pore area division.

10. The method according to claim 7, characterized in that The visualization step includes: The mask image is fused with the original rock slice image and the merged image is displayed by setting different grayscale ranges; When the grayscale range of the mineral mask image is 0-255 and the grayscale range of the pore mask image is 0-100, the grayscale range of the merged image is 0-255, among which the mineral mask image has the largest grayscale and is displayed in red in the merged image, and the pore mask image has the smallest grayscale and is displayed in blue in the merged image; when the grayscale range of the original rock thin section image is also 0-255, the mineral mask image in the merged image is displayed in red, the original rock thin section image is displayed in black, and the pore mask image is displayed in blue; The mineral mask image is directly fused and displayed with the original rock slice image in the pixel dimension; The mineral mask image, the pore mask image and the original rock thin section image are respectively channel-joined to obtain a multi-channel mineral mask image, a multi-channel rock thin section image and a multi-channel pore mask image; The mineral area of ​​the multi-channel mineral mask image is inverted, and then multiple channels are remapped to balance the image color and maintain the contrast between image data. The image remapping adopts the histogram matching method to adjust the histogram of the source image to the shape of the target histogram. The adjustment formula is: g(x)=T(f(x)), where f(x) is the source image, g(x) is the target image, and T is the transformation function.

Images