Particle material ct image segmentation method and device based on random forest algorithm

Abstract

The application discloses a kind of granular material CT image segmentation method and device based on random forest algorithm.Method as follows:S1: input granular material CT image, set the image feature to be used;S2: the importance of the image feature to be used is evaluated by Gini coefficient, and the image feature with high importance is selected;S3: pixel classification model of granular material CT image is trained based on the selected image feature using random forest machine learning method;S4: the trained pixel classification model is applied to all CT images, and three-dimensional segmentation of granular material is completed.The importance of each image feature is evaluated by Gini coefficient in the application, and a small number of features that play a key role in model training are selected, which effectively improves the training efficiency.In the model training process, the granules and the pores between the granules are manually marked, which improves the robustness and classification effect of the model.The method can accurately and efficiently segment the CT images of irregular-shaped and complex-component granules.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a granular material CT image segmentation method and device based on a random forest algorithm. BACKGROUND

[0002] As a kind of material widely existing in nature and industry, the physical properties of granular material have been widely concerned by researchers. Random internal particles lead to structural and dynamic heterogeneity of granular material, making it difficult to study using conventional mechanics and mathematical analysis methods. Numerical simulation can explore the macro and micro mechanical properties of granular material to some extent, but since the calculation formula is mostly a theoretical formula, there is a difference between the actual situation, so physical experiments need to be used to verify and further supplement the research results.

[0003] With the progress of science and technology, X-ray computed tomography (X-ray CT) has gradually matured and been used to study the internal structure of granular material. At present, it is convenient and fast to obtain high-resolution CT images of granular material, but the image processing techniques used are still traditional methods. Most of these methods are based on pixel value image processing and can only segment uniform material granular materials. When processing irregularly shaped particles, over-segmentation and under-segmentation often occur, the accuracy of separating single particle morphology is low, and the types of granular materials that can be studied are severely limited. Therefore, developing a granular material CT image segmentation method and device based on a random forest algorithm to accurately segment irregularly shaped and complex component particles has become a technical problem that needs to be solved in the industry. SUMMARY

[0004] To solve the above problems in the prior art, the present application provides a granular material CT image segmentation method and device based on a random forest algorithm for the segmentation of irregularly shaped and complex component particles.

[0005] To achieve the above purpose, the technical solutions provided by the present application are as follows:

[0006] In a first aspect, the present application provides a granular material CT image segmentation method based on a random forest algorithm, the steps are as follows:

[0007] S1: input the CT image of the granular material, and set the image features to be used;

[0008] S2: evaluate the importance of the image features to be used by Gini coefficient, and select the image features with high importance;

[0009] S3: Based on the image features selected in S2, a random forest machine learning method is used to train a pixel classification model for CT images of particulate materials, so that it has the ability to classify particles and the pores between particles.

[0010] S4: Apply the trained pixel classification model to all CT images to ensure that the particle morphology is correct and that the particles are independent of each other, thus completing the three-dimensional segmentation of the granular material.

[0011] Furthermore, the particulate material is an aggregate of discrete solid substances that exist in nature or are artificially manufactured.

[0012] Furthermore, the CT images of the particulate material are a series of two-dimensional images obtained by scanning the particulate material using X-ray computed tomography (CT) technology.

[0013] Furthermore, in step S1, the types of image features to be used include Gaussian smoothing, Laplacian of Gaussian operator, difference of Gaussian, Hessian matrix eigenvalues, and tensor eigenvalues.

[0014] Furthermore, in step S1, the specific image features are as follows:

[0015] Gaussian smoothing: Performs a two-dimensional convolution operation to blur images and eliminate noise. The convolution kernel formula is:

[0016]

[0017] Where (x,y) represents the position of a pixel in the image, and σ is the standard deviation parameter that needs to be set;

[0018] Laplacian of Gaussian operator: The Laplacian of Gaussian operator is used to filter images. The formula for the convolution kernel is:

[0019]

[0020] Where (x,y) represents the position of a pixel in the image, and σ is the standard deviation parameter that needs to be set;

[0021] Difference of Gaussians (DG): The difference of Gaussians image is obtained by subtracting two Gaussian-filtered images. The formula is:

[0022]

[0023] Where (x,y) represents the position of a pixel in the image, and σ1 and σ2 are two different standard deviation parameters;

[0024] Hessian matrix eigenvalues: After filtering the image using the second derivative of Gaussian, a Hessian matrix is ​​constructed for each pixel. The eigenvalues ​​are then solved to form the feature image.

[0025] Hessian matrix calculation formula:

[0026]

[0027] where I(x, y) is the pixel value of the image at position (x, y), and * is the convolution symbol;

[0028] Tensor eigenvalue: distinguish the parts with strong local information and the parts with weak local information in the image, and the calculation formula is:

[0029]

[0030] Further, different image features are distinguished by the category of the image feature and different σ values, such as Gaussian smoothing with σ = 0.3 and Gaussian smoothing with σ = 1 are two image features.

[0031] Further, in the step S2, the evaluation of the Gini coefficient is based on the structure of the random forest machine learning method, which includes a plurality of decision trees, each of which includes a plurality of nodes, and the importance of the image feature is evaluated by calculating the degree of decrease of the Gini coefficient before and after each node. The specific steps are:

[0032] S2.1: Randomly select a small number of images from the input CT image, and randomly manually mark a small number of particle and pore pixels in them;

[0033] S2.2: Calculate the degree of decrease of the Gini coefficient: for an image feature, first calculate the Gini coefficient of the feature on a node, then calculate the difference in Gini coefficient between the node and the child node, and finally aggregate the Gini coefficient differences in all decision trees, and the aggregation result represents the importance of the image feature;

[0034] The calculation formula is:

[0035] Gini coefficient:

[0036] In the formula, x is an image feature, is the proportion of pixel class j in the qth node of the dth tree in the random forest machine learning method; J is the class of all pixels, since only two classes of particle and pore pixels are divided, then J = 2;

[0037] Difference in Gini coefficient between parent node and child node:

[0038] In the formula, and are the Gini coefficients of the xth image feature in the rth node and the sth node of the dth tree in the random forest machine learning method, respectively, and r and s are the two child nodes of q;

[0039] Importance of image features:

[0040] In the formula, Q represents all nodes in the d-th tree that contain image feature x, D represents the number of decision trees, and X represents the total number of image features;

[0041] S2.3: Sort all image features from highest to lowest importance and select a small number of image features with high importance.

[0042] Furthermore, in step S3, the specific steps for training the pixel classification model using the random forest machine learning method are as follows:

[0043] S3.1: Randomly select a small number of images from the input CT images, and manually label a small number of particle and pore pixels in them. The labeled pixels will be used to train the pixel classifier.

[0044] S3.2: Based on the image features selected in S2, generate corresponding feature images for the labeled images, form a pixel matrix, and input it into the random forest algorithm;

[0045] S3.3: The Random Forest algorithm randomly divides the pixel matrix into multiple sub-matrices equal to the number of decision trees. Each decision tree uses the CART method to divide a sub-matrix into granular pixels and porous pixels. After the division is completed, the classification results of all decision trees are summarized, and the final category of the pixel is determined by majority vote.

[0046] S3.4: Evaluate the out-of-bag error rate of the model and complete model training when it is less than the set threshold.

[0047] Out-of-bag error rate: Decision trees do not use all pixels in the submatrix. The unused pixels are called out-of-bag data. After the decision tree model is completed, the out-of-bag data is classified. The number of incorrect classifications out of the total number of out-of-bag data is called the out-of-bag error rate.

[0048] Furthermore, in step S3.3, the CART method proceeds as follows: The image feature with the smallest Gini coefficient and its corresponding threshold are selected from all image features and all possible thresholds as the optimal feature and optimal threshold; the pixel data is classified using the optimal feature and optimal threshold as nodes, and this process is repeated until multiple classification nodes are created using all image features, thus completing the construction of the decision tree; the threshold refers to an integer from 1 to 254, and the threshold is selected by traversing from 1 to 254 to calculate the Gini coefficient.

[0049] In a second aspect, the application provides an application device of the method in the first aspect, comprising: an input module for importing all two-dimensional images obtained by CT scanning, preselecting image features, manually marking particle and pore pixels; a core processing module for determining final used image features and completing training of a pixel classifier; and an application module for applying the trained pixel classifier to all two-dimensional images and composing a three-dimensional image.

[0050] The application has the following beneficial effects:

[0051] (1) The particle material CT image segmentation method and device based on the random forest algorithm provided by the application make full use of pixel information in the image, the process of training the pixel classifier is efficient, and the device has the ability to correct error segmentation areas, thereby overcoming the defects of traditional particle image segmentation methods, such as limitation and high error rate, and can accurately segment particles with irregular shapes and complex compositions, and has the characteristics of high efficiency, high precision and strong applicability.

[0052] (2) The application manually marks particles and pores between particles in the model training process, identifies particles from two aspects of pixel value size and particle contour, and improves the robustness and classification effect of the model. The method can accurately and efficiently segment CT images of particles with irregular shapes and complex compositions. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0054] Figure 1 A flowchart of the particle material CT image segmentation method based on the random forest algorithm provided by the embodiment of the application is shown in the figure.

[0055] Figure 2 An inputted CT image of river sand particle material provided by the embodiment of the application is shown in the figure.

[0056] Figure 3 A schematic diagram of manually marking particles and pores provided by the embodiment of the application is shown in the figure.

[0057] Figure 4 A schematic diagram of a feature image provided by the embodiment of the application is shown in the figure.

[0058] Figure 5 A schematic diagram of a pixel matrix inputted into the random forest algorithm provided by the embodiment of the application is shown in the figure.

[0059] Figure 6 A two-dimensional segmentation result of a CT image of a river sand particle material provided by an embodiment of the present application;

[0060] Figure 7 A three-dimensional segmentation result of a CT image of a river sand particle material provided by an embodiment of the present application. DETAILED DESCRIPTION

[0061] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form a feasible technical solution, and such combination is not restricted by the order of steps and / or structure mode, but should be based on the fact that it can be realized by those of ordinary skill in the art. When the combination of technical solutions appears to be contradictory or unfeasible, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection required by the present application.

[0062] The present application provides a particle material CT image segmentation method based on a random forest algorithm, which comprises the following steps: Figure 1 S1: inputting a particle material CT image and setting image features to be used;

[0063] S2: evaluating the importance of the image features to be used by a Gini coefficient and selecting image features with high importance;

[0064] S3: training a pixel classification model of a particle image based on the selected image features by using a random forest machine learning method, so as to have the ability to classify particles and inter-particle pores;

[0065] S4: applying the trained pixel classification model to all images to ensure that the particle morphology is correct and the particles are independent of each other, and completing three-dimensional segmentation of the particle material.

[0066]

[0067] Embodiment 1

[0068] The CT image of river sand is segmented by using the particle material CT image segmentation method based on the random forest algorithm provided by the present application, and the specific steps are as follows:

[0069] S1: inputting a particle material CT image and setting image features to be used. The image feature types include Gaussian smoothing, Gaussian Laplacian operator, Gaussian difference, Hessian matrix eigenvalue and tensor eigenvalue.​

[0070] The image features are as follows:

[0071] Gaussian smoothing: a two-dimensional convolution operation is performed to blur the image to eliminate noise, and the formula of the convolution kernel is:

[0072]

[0073] where (x, y) is the position of the pixel in the image, and σ is the standard deviation parameter to be set;

[0074] Gaussian Laplace operator: the image is filtered using a Gaussian Laplace operator, and the formula of the convolution kernel is:

[0075]

[0076] where (x, y) is the position of the pixel in the image, and σ is the standard deviation parameter to be set;

[0077] Gaussian difference: the Gaussian difference image is obtained by subtracting two Gaussian filtered images, and the formula is:

[0078]

[0079] where (x, y) is the position of the pixel in the image, and σ1 and σ2 are two different standard deviation parameters;

[0080] Hessian matrix eigenvalue: after filtering the image by the Gaussian second derivative, a Hessian matrix is constructed for each pixel point, and the eigenvalue is solved to form a feature image;

[0081] Hessian matrix calculation formula:

[0082]

[0083] where I(x, y) is the pixel value at position (x, y) of the image, and * is the convolution symbol;

[0084] Tensor eigenvalue: the parts with prominent local information and the parts with weak local information in the image are distinguished, and the calculation formula is:

[0085]

[0086] Specifically, 1000 CT images of river sand are selected as input images, and the size is unified to 899x1003, Figure 2 One of the images used in this embodiment is shown; the image features to be used are shown in Table 1, a total of 5 image features are used, and 6 σs are set for each, i.e. a total of 30 image features are used.

[0087] Table 1

[0088] Image feature class Parameter σ Gaussian smoothing 0.7,1.0,1.5,3.5,5.0,10.0 Gaussian Laplacian 0.7,1.0,1.5,3.5,5.0,10.0 Gaussian difference [σ1: 0.7, 1.0, 1.5, 3.5, 5.0, 10.0; σ2: 0.3] Hessian matrix eigenvalue 0.7,1.0,1.5,3.5,5.0,10.0 Tensor eigenvalue 0.7,1.0,1.5,3.5,5.0,10.0

[0089] S2: Evaluate the importance of the image features to be used by Gini coefficient, and select the image features with high importance, the specific steps are:

[0090] S2.1: Randomly select a small number of images from the input CT images, and randomly manually label a small number of particle and pore pixels in them; specifically, randomly select 2 images to label particle and pore pixels from 1000 images, and the labeling is as shown in Figure 3

[0091] S2.2: Calculate the degree of decrease of Gini coefficient.

[0092] The calculation of Gini coefficient is based on the structure of random forest algorithm, which includes multiple decision trees, and each decision tree includes multiple nodes. For an image feature, first calculate the Gini coefficient of the feature on a node, then calculate the Gini coefficient difference between the node and the child node, and finally aggregate the Gini coefficient differences in all decision trees, which represents the importance of the image feature. The calculation formula is:

[0093] Gini coefficient:

[0094] In the formula, x is an image feature, is the proportion of pixel class j in the qth node of the dth tree in the random forest machine learning method, and J is the class of all pixels. Since only two classes of particle and pore pixels are divided, J = 2;

[0095] Difference in Gini coefficient between parent node and child node:

[0096] In the formula, and are the Gini coefficients of the xth image feature in the rth node and the st node in the dth tree, and r and s are the two child nodes of q.

[0097] Importance of image feature:

[0098] In the formula, Q is all nodes in the dth tree containing image feature x, D is the number of decision trees, and X is the total number of image features.

[0099] S2.3: Sort the importance of all image features from high to low, and select a small number of image features with high importance.

[0100] ​S2.4: Repeat steps S2.1 to S2.3 multiple times, and select the common high importance image features.

[0101] Specifically, 2 times of S2.1 to S2.3 are completed, and the importance ranking of image features is shown in Table 2 (only the top 10 most important ones are shown), and the selected common high importance image features can be Gaussian smoothing σ = 1.5, Gaussian smoothing σ = 1.0, Gaussian smoothing σ = 0.7, structure tensor σ = 0.7, and Gaussian difference σ1 = 3.5, a total of 5.

[0102] Table 2

[0103]

[0104]

[0105] S3: Based on the selected image features, a pixel classification model of the particle image is trained using a random forest method, so as to have the ability to classify particles and inter-particle pores, and the specific steps are as follows:

[0106] S3.1: A small number of images are randomly selected from the input CT image, and a small number of particle and pore pixels are randomly manually labeled in the images, and the labeled pixels will be used to train the pixel classifier;

[0107] S3.2: According to the image features selected in S2, the corresponding feature images of the labeled images are generated to form a pixel matrix input into the random forest algorithm;

[0108] Specifically, the schematic diagram of the feature image is as shown in Figure 4 , and the schematic diagram of the pixel matrix is as shown in Figure 5 .

[0109] S3.3: The random forest algorithm randomly divides the pixel matrix into multiple sub-matrices with the same number as the number of decision trees, and each decision tree uses the CART method to divide a sub-matrix into particle pixels and pore pixels, and after the division is completed, the classification results of all decision trees are summarized to determine the final class of the pixel by majority voting;

[0110] CART method: Among all image features and all possible threshold values, the image feature and its corresponding threshold value with the smallest Gini coefficient are selected as the optimal feature and the optimal threshold value. The pixel data is classified using the optimal feature and the optimal threshold value as the node, and the same is repeated until multiple classification nodes are created using all image features, and the construction of the decision tree is completed. The threshold value refers to an integer in 1 to 254, and the threshold value is selected by traversing 1 to 254 to calculate the Gini coefficient.

[0111] S3.4: Evaluate the out-of-bag error rate of the model, and complete the model training when it is less than a set threshold value.

[0112] Out-of-bag error rate: the decision tree does not use all the pixels within the sub-matrix, the pixels not used are called out-of-bag data, when the decision tree model is completed, the out-of-bag data is classified again, and the number of division errors accounts for the total number of out-of-bag data, which is called out-of-bag error rate.

[0113] Specifically, the out-of-bag error rate of the model in the implementation process is 0.16%, indicating that the accuracy of the model is very high and meets the requirements.

[0114] S4: apply the trained pixel classification model to all two-dimensional images to complete the three-dimensional segmentation of the granular material, which needs to ensure that the particle morphology is correct and the particles are independent of each other, if the segmentation result is problematic, manually mark and return to S3.2 to further train the model to meet the requirements.

[0115] Specifically, Figure 6 The segmentation result of a two-dimensional CT image of river sand is shown, Figure 7 The partial three-dimensional segmentation result of river sand is shown.

[0116] Example 2

[0117] An application device of a granular material CT image segmentation method based on a random forest algorithm, comprising an input module for importing all two-dimensional images obtained by CT scanning, preselecting image features, manually marking particle and pore pixels; a core processing module for determining the final used image features and completing the training of the pixel classifier; an application module for applying the trained pixel classifier to all two-dimensional images and forming a three-dimensional image.

[0118] The granular material CT image segmentation method based on the random forest algorithm provided by the embodiment of the application introduces accurate information into the training process of machine learning by manually marking particle and pore pixels, and uses the Gini coefficient to evaluate the importance of image features, which can ensure the training efficiency and training accuracy. The method fully utilizes the pixel value information and particle contour information of the image, and does not cause over-segmentation of the particles, and the contact between the particles can be solved by further manual marking. Compared with the traditional image segmentation method which only relies on pixel value for particle segmentation, the method can accurately and efficiently segment the granular material with irregular morphology and complex composition.

[0119] Those skilled in the art can clearly understand the implementation of the various embodiments by means of software and necessary universal hardware platforms through the above description of the embodiments, and the various embodiments can also be implemented by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0120] In this patent, the terms "comprise", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the elements defined by the statement "comprise" do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0121] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for segmentation of a CT image of a granular material based on a random forest algorithm, characterized by the steps of The method comprises the following steps: S1: inputting a particle material CT image and setting image features to be used; S2: evaluating the importance of the image features to be used by a Gini coefficient, and selecting image features with high importance; S3: training a pixel classification model of the particle material CT image by using a random forest machine learning method based on the image features selected in S2, so that the pixel classification model has the ability to classify particles and inter-particle pores; In the step S3, the specific steps of training the pixel classification model by using the random forest machine learning method are as follows: S3.1: a small amount of images are randomly selected from the input CT image, and a small amount of particle and pore pixels are randomly manually labeled in the images, and the labeled pixels are used to train the pixel classifier; S3.2: generating corresponding feature images of the labeled images according to the image features selected in S2, and inputting the pixel matrix into the random forest algorithm; S3.3: the random forest algorithm randomly divides the pixel matrix into a plurality of sub-matrices with the same number as the number of decision trees, each decision tree uses the CART method to divide a sub-matrix into particle pixels and pore pixels, and after the division is completed, the classification results of all decision trees are summarized to determine the final category of the pixel by majority voting; the steps of the CART method are as follows: selecting the image feature with the smallest Gini coefficient and the corresponding threshold value as the optimal feature and the optimal threshold value; using the optimal feature and the optimal threshold value as the node to classify the pixel data, and iteratively classifying until all image features are used to create a plurality of classification nodes, and the construction of the decision tree is completed; the threshold value refers to an integer in 1 to 254, and the threshold value is selected by traversing 1 to 254 to calculate the Gini coefficient; S3.4: evaluating the out-of-bag error rate of the model, and completing the model training when the error rate is less than a set threshold value; S4: applying the trained pixel classification model to all CT images to ensure that the particle morphology is correct and the particles are independent of each other, and completing the three-dimensional segmentation of the particle material.

2. The method of claim 1, wherein: The particle material is a collection of discrete solid substances existing in nature or artificially manufactured.

3. The method of claim 1, wherein: The particle material CT image is a series of two-dimensional images obtained by scanning the particle material by X-ray electronic computed tomography technology.

4. The method of claim 1, wherein: In the step S1, the types of image features to be used include Gaussian smoothing, Gaussian Laplace operator, Gaussian difference, Hessian matrix eigenvalue, and tensor eigenvalue.

5. The method of claim 4, wherein, In the step S1, each image feature is as follows: Gaussian smoothing: performing a two-dimensional convolution operation to blur the image to eliminate noise, and the convolution kernel formula is: wherein, is the position of a pixel in the image, is a standard deviation parameter to be set; Gaussian Laplace operator: filtering the image using the Gaussian Laplace operator, and the formula of the convolution kernel is: wherein, is the position of a pixel in the image, is a standard deviation parameter to be set; Gaussian difference: subtracting two Gaussian filtered images to obtain a Gaussian difference image, and the formula is: wherein is the position of a pixel in the image, and are two different standard deviation parameters; Hessian matrix eigenvalue: constructing a Hessian matrix for each pixel point after filtering the image by a Gaussian second derivative, and forming a feature image after solving the eigenvalues; Hessian matrix calculation formula: wherein is an image of the pixel value at the position, is a convolution symbol; Tensor eigenvalue: distinguishing between parts with prominent local information and parts with weak local information in the image, and the calculation formula is:

6. The method of claim 5, wherein: Different image features are distinguished by the category of the image feature and different values.

7. The method of claim 1, wherein: The step S2, the evaluation of Gini coefficient is based on the structure of random forest machine learning method, which includes multiple decision trees, each decision tree includes multiple nodes, the importance of image features is evaluated by calculating the decline degree of Gini coefficient before and after each node, the specific steps are: S2.1: randomly select a small amount of images from the input CT image, and randomly manually mark a small amount of particle and pore pixels in them; S2.2: calculate the decline degree of Gini coefficient: for an image feature, first calculate the Gini coefficient of the feature on a node, then calculate the Gini coefficient difference between the node and the sub node, and finally summarize the Gini coefficient differences in all decision trees, the summary result represents the importance of the image feature; The calculation formula is: Gini coefficient: wherein, is a feature of an image, is the proportion of pixels of a class at a node in a tree of a random forest machine learning method; of a class; is the class of all pixels, since only two classes of pixels, particle and pore, are distinguished, ; Difference between Gini coefficients between parent and child nodes: wherein and are the Gini coefficients of the th image feature at the th node and the th node in the th tree, respectively, and are the two child nodes of . Importance of image features: wherein is the number of decision trees, is the number of decision trees, is the number of nodes in the tree containing the image features, is the number of decision trees, is the total number of image features; S2.3: sort all image features from high to low in importance, and select a small amount of image features with high importance.

8. An application device of a method for segmenting a granular material CT image according to any one of claims 1 to 7, characterized in that, Including: The input module is used for importing all two-dimensional images obtained by CT scanning, preselecting image features, and manually marking particle and pore pixels; The core processing module is used for determining the final used image features and completing the training of the pixel classifier; the application module applies the trained pixel classifier to all two-dimensional images and forms a three-dimensional image.

Images