An automatic detection method for mixed coal vitrinite

Abstract

The present invention provides a method for automatically identifying and detecting mixed coal vitrinite, comprising: step one, collecting multiple images of various single coals in an actual industrial environment and calibrating the types of coals. Step two, using global median filtering and local mean filtering to eliminate some noise points in the image. Step three, using the clustering-based Otsu algorithm to binarize the image and remove the background image. Step four, using texture features and grayscale histograms to remove inertin and exinin from the image. In step four, the texture features include contrast, correlation, energy and entropy. Step five, using the Random Patchwork algorithm to generate a mixed coal image data set for multiple single coal images. Step six, constructing an image semantic segmentation model. Step seven, inputting the data set into the network for iterative training; and performing detection after the training is completed. The present invention combines traditional feature extraction with an improved deep learning algorithm, and can quickly and accurately detect and identify the types and proportions of mixed coal vitrinite.

Description

Technical Field

[0001] The invention belongs to the technical field of coal rock analysis and detection, relates to mixed coal vitrinite, and particularly relates to an automatic detection method for mixed coal vitrinite. Background Art

[0002] The microscopic composition and vitrinite reflectance of coal determine the physical and chemical properties, process properties and industrial uses of coal. Coal rock parameter indicators have been widely used in commercial coal quality testing and industrial fields such as coal coking, gasification, liquefaction, oil and gas exploration, etc.

[0003] Currently, the determination of coal rock parameter indicators mainly relies on manual identification and measurement, which has the disadvantages of being labor-intensive, time-consuming, and subject to differences in subjective understanding among observers, resulting in poor comparability of identification data between laboratories. Automatic coal rock determination technology has been used both domestically and internationally to automatically determine the vitrinite reflectance of coal, but this suffers from inaccurate vitrinite identification or low vitrinite identification rates. This is especially true for mixed coals with low homogeneous vitrinite content or small particles. Existing identification technologies cannot accurately and efficiently identify it, resulting in data omissions and false detections. Efficient, accurate, and complete identification of vitrinite in mixed coal is the prerequisite and foundation for vitrinite reflectance measurement. Therefore, in the field of coal rock component identification technology, there is an urgent need for an automatic vitrinite identification technology for mixed coal. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for automatically detecting mixed coal vitrinite, so as to solve the technical problems of inaccurate vitrinite identification and low vitrinite identification rate in the existing technology.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for automatically detecting mixed coal vitrinite, the method comprising the following steps:

[0007] Step 1: Collect multiple images of various single coals in actual industrial environments and calibrate the types of coals.

[0008] In step 2, global median filtering and local mean filtering are used to eliminate some noise in the image.

[0009] Step three: Use the clustering-based Otsu algorithm to binarize the image and remove the background image.

[0010] Step 4: Use texture features and grayscale histogram to remove inertin and exintin from the image.

[0011] In step 4, the texture features include contrast, correlation, energy and entropy.

[0012] Step 5: Use the Random Patchwork algorithm to generate a mixed coal image dataset for multiple single coal images, and divide the mixed coal image dataset into a training set, a validation set, and a test set.

[0013] Step 6: Build an image semantic segmentation model.

[0014] The specific process of constructing the image semantic segmentation model includes: building an FCN8s_M network module based on the lightweight VGGNET network architecture, using the FCN8s_M network module as the backbone feature extraction network to extract the local features of each type of coal in the mixed coal; using the ResNet_M network to extract the global features of each type of coal in the single coal image, fusing the local features with the global features, and enhancing the judgment ability of the model's semantic segmentation.

[0015] Step 7: Set the model training parameters. The learning rate is initialized to 0.00001, and the learning rate is reduced by half every 10 epochs. The cross entropy loss function and the mean square error loss function are used as the criteria for model iterative update. The Adam optimizer is selected, and the data set is input into the network for iterative training. The trained model is used to automatically detect mixed coal vitrinite.

[0016] Compared with the prior art, the present invention has the following technical effects:

[0017] (I) The present invention combines traditional feature extraction with an improved deep learning algorithm, which can quickly and accurately detect and identify the types and proportions of mixed coal vitrinite.

[0018] (II) The present invention combines automated detection equipment to solve the shortcomings of manual detection, which is slow, requires high manual experience, and is time-consuming and labor-intensive.

[0019] (III) The present invention can not only identify the types and proportions of vitrinite in mixed coal, but also provide accurate location information, thus providing a basis for industrial screening.

[0020] (IV) The present invention proposes an algorithm for removing inertin and extinct particles from microscopic coal rock images by using texture features such as contrast, correlation, energy and entropy as well as grayscale histogram feature parameters.

[0021] (V) This paper proposes a method based on the lightweight VGGNET network architecture to construct the FCN8s_M network module, which is used as the backbone feature extraction network to extract the local features of each type of coal in the mixed coal. The ResNet_M network is used to extract the global features of each type of coal in the single coal image. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the method of the present invention.

[0023] Figure 2 shows the effect after background removal through binarization.

[0024] Figure 2(a) is the original image before background removal.

[0025] Figure 2(b) is a binary image.

[0026] Figure 2(c) shows the effect after background removal through binarization.

[0027] Figure 3 is the image grayscale histogram.

[0028] FIG4( a ) is the original image before removing the inertin and extinct body.

[0029] Figure 4(b) is a vitrinite image after removing inertinite and extinite.

[0030] Figure 5 This is the FCN8s_M network structure diagram for semantic segmentation of mixed coal vitrinite in the present invention.

[0031] Figure 6 This is the network structure diagram of ResNet_M for semantic segmentation of mixed coal vitrinite in the present invention.

[0032] Figure 7 This is the curve of the change of loss and accuracy of the training set and validation set with epoch during the training of the model of the present invention.

[0033] Figure 8 This is the test result of the present invention on the original image collected in the industrial environment.

[0034] The specific contents of the present invention are further explained in detail below with reference to the embodiments. DETAILED DESCRIPTION

[0035] It should be noted that, unless otherwise specified, all algorithms in the present invention adopt algorithms known in the prior art.

[0036] It should be noted that the mixed coal in the present invention is preferably commercial mixed coal.

[0037] Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent modifications made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0038] Example:

[0039] This embodiment provides a method for automatically detecting mixed coal vitrinite. Figure 1 As shown, the method includes the following steps:

[0040] Step 1: Collect 400 images of each of 15 types of coal in actual industrial environments and calibrate the coal types.

[0041] In step 2, global median filtering and local mean filtering are used to eliminate some noise in the image.

[0042] Step three: Use the clustering-based Otsu algorithm to binarize the image and remove the background image.

[0043] Preferably, the specific steps of step three are:

[0044] Assume that the image grayscale level is L, W0 is the background ratio, U0 is the background mean, W1 is the foreground ratio, U1 is the foreground mean, U is the mean of the entire image, and t is the segmentation threshold.

[0045] Construct the judgment variable g:

[0046] g=W0(t)=(u0(t)-U) 2 +W1(t)×(U1(t)-U) 2 ;

[0047] Traverse each segmentation threshold t and find the value that maximizes g.

[0048] The larger the between-class variance, the greater the difference between the two parts of the segmentation, and the smaller the within-class variance, the smaller the difference between the same class.

[0049] Step 4: Use texture features and grayscale histogram to remove inertin and exintin from the image.

[0050] In step 4, texture features include contrast, correlation, energy and entropy.

[0051] The deeper the texture grooves displayed in the image, the greater the contrast and the clearer the visual effect; conversely, the smaller the contrast, the shallower the grooves and the blurrier the visual effect.

[0052] Correlation measures the degree of similarity between the elements of the spatial grayscale co-occurrence matrix in the row or column direction. Therefore, the magnitude of the correlation value reflects the local grayscale correlation in the image. When the matrix element values ​​are uniformly equal, the correlation value is large; conversely, if the matrix pixel values ​​vary greatly, the correlation value is small.

[0053] Energy reflects the uniformity of image grayscale distribution and texture coarseness characteristics.

[0054] In physics, entropy refers to the regularity of an object: the more ordered, the lower the entropy, while the more disordered, the higher the entropy. Entropy here represents the amount of information in an image. Entropy is high when all elements in the co-occurrence matrix have maximum randomness—that is, when all values ​​in the spatial co-occurrence matrix are nearly equal—and when the elements in the co-occurrence matrix are dispersed. It indicates the degree of texture non-uniformity or complexity within an image.

[0055] In step 4, the contrast is defined as follows:

[0056]

[0057] Where:

[0058] C is contrast;

[0059] δ is the grayscale difference;

[0060] i and j are adjacent pixels;

[0061] δ(i, j) = |ij| is the grayscale difference between adjacent pixels;

[0062] P δ (i, j) is the pixel distribution probability when the grayscale difference between adjacent pixels is δ.

[0063] In step 4, the relevance is defined as follows:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] Where:

[0070] COR is the correlation;

[0071] i is the i-th row of the gray-level co-occurrence matrix of the image;

[0072] j is the jth column of the gray-level co-occurrence matrix of the image;

[0073] k is the grayscale level of the image;

[0074] (ij) represents the multiplication of i and j;

[0075] G(i, j) is the number of times adjacent gray values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the gray-level co-occurrence matrix;

[0076] u i is the mean value of the grayscale relationship in the row direction of the image;

[0077] u j is the mean value of the grayscale relationship in the column direction of the image;

[0078] s i is the standard deviation of the grayscale relationship in the row direction of the image;

[0079] sj is the standard deviation of the grayscale relationship in the column direction of the image.

[0080] In step 4, energy is defined as follows:

[0081]

[0082] Where:

[0083] ASM is energy;

[0084] i is the i-th row of the gray-level co-occurrence matrix of the image;

[0085] j is the jth column of the gray-level co-occurrence matrix of the image;

[0086] k is the grayscale level of the image;

[0087] G(i, j) is the number of times adjacent grayscale values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the grayscale co-occurrence matrix.

[0088] In step 4, entropy is defined as follows:

[0089]

[0090] Where:

[0091] ENT stands for energy;

[0092] i is the i-th row of the gray-level co-occurrence matrix of the image;

[0093] j is the jth column of the gray-level co-occurrence matrix of the image;

[0094] k is the grayscale level of the image;

[0095] G(i, j) is the number of times adjacent grayscale values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the grayscale co-occurrence matrix.

[0096] Step 5: Use the Random Patchwork algorithm to generate a mixed coal image dataset for 15 single coal images, and divide it into a training set, a validation set, and a test set with a ratio of 6:2:2.

[0097] Preferably, in step five, the specific steps of the Random Patchwork algorithm are: randomly select two categories of coal from 15 single coal categories for processing; set the random ratio to m, take the mth part of the i-th image in the first category and the (1-m)th part of the i-th image in the second category to combine to generate a mixed coal image i, and record the types of vitrinite that make up the mixed coal image; continue to randomly select two categories of single coal and repeat the operation until every single coal is traversed.

[0098] Step 6: Build an image semantic segmentation model.

[0099] The specific process of building an image semantic segmentation model includes: building an FCN8s_M network module based on the lightweight VGGNET network architecture, using the FCN8s_M network module as the backbone feature extraction network to extract the local features of each type of coal in the mixed coal; using the ResNet_M network to extract the global features of each type of coal in the single coal image, fusing the local features with the global features, and enhancing the model's semantic segmentation judgment ability.

[0100] Step 7: Set the model training parameters. The learning rate is initialized to 0.00001, and the learning rate is reduced by half every 10 epochs. The cross entropy loss function and the mean square error loss function are used as the criteria for model iterative update. The Adam optimizer is selected, and the data set is input into the network for iterative training. The trained model is used to automatically detect mixed coal vitrinite.

[0101] In step seven, the cross entropy loss function is as follows:

[0102]

[0103] Where:

[0104] H(p,q) is the cross entropy loss function;

[0105] p is the true distribution;

[0106] q represents the predicted distribution;

[0107] i is the serial number of all possible values ​​of the probability distribution, i = 1, ..., n;

[0108] x i is the i-th value among all possible values ​​of the probability distribution.

[0109] In step seven, the mean square error loss function is as follows:

[0110]

[0111] Where:

[0112] J is the mean square error loss function;

[0113] i is the serial number of all possible values ​​of the probability distribution, i = 1, ..., n;

[0114] x i is the i-th value among all possible values ​​of the probability distribution.

[0115] yi is the true value;

[0116] h θ is the predicted value.

[0117] Application examples:

[0118] This application example provides an automatic detection method for mixed coal vitrinite based on the above embodiment. Figure 1 As shown, the method is as follows:

[0119] A dataset of 15 types of single coal vitrinite was constructed, with 400 images of each type of single coal.

[0120] A hybrid approach of global median filtering and local mean filtering is used to address noise in single coal images. Applying median filtering to the entire image eliminates salt-and-pepper noise. Applying mean filtering only to the large area containing vitrinite reduces the difference in grayscale values, making detection easier.

[0121] The clustering-based Otsu algorithm is used to generate binary images, and the background is removed as shown in Figure 2(a), Figure 2(b), and Figure 2(c).

[0122] Based on the pre-processed image, texture features such as contrast, correlation, energy and entropy as well as grayscale histogram are used to remove inertin and extinct particles in the image. Figure 3 The grayscale histogram shown has three peaks. The vitrinite is in the middle, the lower grayscale is the extinctite, and the higher grayscale is the inertinite. By setting the threshold, the vitrinite is well segmented, and the effect is shown in Figure 4(a) and Figure 4(b).

[0123] Based on the single coal vitrinite images, mixed coal images are randomly generated and divided into training set, validation set and test set with a ratio of 6:2:2.

[0124] like Figure 5 and Figure 6 As shown in the figure, the FCN8s_M network and the ResNet_M network are constructed to jointly segment the mixed coal vitrinite image. The FCN8s_M network modifies the convolutional layer parameters based on the original FCN network and alternately uses the Sigmoid activation function and the Tanh activation function to accelerate model convergence and reduce computation time. Practice has shown that this processing method has a good effect on vitrinite semantic segmentation.

[0125] Set the model training parameters, initialize the learning rate to 0.00001, and the loss and accuracy of the training set and validation set during model training as the epoch changes are as follows: Figure 7 As shown, the learning rate is halved every 10 epochs. The cross-entropy loss function and the mean square error loss function are used as the criterion for iterative model updates. The Adam optimizer is selected, and the dataset is input into the network for iterative training. Figure 1 below shows the recognition accuracy of 15 types of coal and rock after using the ResNet_M network to extract global image features.

[0126] Table 1

[0127] Coal rock type Recognition accuracy Coal rock type Recognition accuracy 78-60 0.90263 78-108 0.9872 78-64 0.9124 78-111 0.9271 78-67 0.8647 78-143 0.9974 78-70 0.9872 78-153 0.9927 78-72 0.9892 78-227 0.9278 78-106 0.9087 78-231 0.8872 78-107 0.9612 78-251 0.9214 78-1116 0.9428

[0128] Finally, based on the above algorithm, a software for identifying vitrinite in mixed coal can be developed. By integrating the trained model into the software, the vitrinite in mixed coal can be quickly and accurately segmented and the type and proportion of vitrinite can be displayed. The test results of the present invention on the original image collected in the industrial environment are as follows: Figure 8 It shows 4 examples. The first picture in each row is the original picture, and the second picture is the test result picture.

Claims

1. A method for automatically detecting mixed coal vitrinite, characterized in that: The method comprises the following steps: Step 1: Collect multiple images of various single coals in the actual industrial environment and calibrate the types of coals; Step 2: Use global median filtering and local mean filtering to eliminate some noise in the image; Step 3: Use the clustering-based Otsu algorithm to binarize the image and remove the background image; Step 4: Use texture features and grayscale histogram to remove inertin and exintin from the image; In step 4, the texture features include contrast, correlation, energy and entropy; Step 5: Use the Random Patchwork algorithm to generate a mixed coal image dataset for multiple single coal images, and divide the mixed coal image dataset into a training set, a validation set, and a test set; Step 6: Build an image semantic segmentation model; The specific process of constructing the image semantic segmentation model includes: building an FCN8s_M network module based on the lightweight VGGNET network architecture, using the FCN8s_M network module as the backbone feature extraction network to extract local features of each type of coal in the mixed coal; using the ResNet_M network to extract global features of each type of coal in the single coal image, fusing local features with global features, and enhancing the judgment ability of the model's semantic segmentation; Step 7: Set the model training parameters. The learning rate is initialized to 0.00001, and the learning rate is reduced by half every 10 epochs. The cross entropy loss function and the mean square error loss function are used as the criteria for model iterative update. The Adam optimizer is selected, and the data set is input into the network for iterative training. The trained model is used to automatically detect mixed coal vitrinite.

2. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: The specific steps of step three are: Assume that the image grayscale level is L, W0 is the background ratio, U0 is the background mean, W1 is the foreground ratio, U1 is the foreground mean, U is the mean of the entire image, and t is the segmentation threshold; Construct the judgment variable g: g=W 01 (t)×(U0(t)-U) 2 +W1(t)×(U1(t)-U) 2 ; Traverse each segmentation threshold t and find the value that maximizes g.

3. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 4, the contrast is defined as follows: Where: C is contrast; δ is the grayscale difference; i and j are adjacent pixels; δ(i, j) = |ij| is the grayscale difference between adjacent pixels; P δ (i, j) is the pixel distribution probability when the grayscale difference between adjacent pixels is δ.

4. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 4, the relevance is defined as follows: Where: COR is the correlation; i is the i-th row of the gray-level co-occurrence matrix of the image; j is the jth column of the gray-level co-occurrence matrix of the image; k is the grayscale level of the image; (ij) represents the multiplication of i and j; G(i, j) is the number of times adjacent gray values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the gray-level co-occurrence matrix; u i is the mean value of the grayscale relationship in the row direction of the image; u j is the mean value of the grayscale relationship in the column direction of the image; s i is the standard deviation of the grayscale relationship in the row direction of the image; s j is the standard deviation of the grayscale relationship in the column direction of the image.

5. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 4, the energy is defined as follows: Where: ASM is energy; i is the i-th row of the gray-level co-occurrence matrix of the image; j is the jth column of the gray-level co-occurrence matrix of the image; k is the grayscale level of the image; G(i, j) is the number of times adjacent grayscale values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the grayscale co-occurrence matrix.

6. The method for automatically detecting mixed coal vitrinite according to claim 1, characterized in that: In step 4, the entropy is defined as follows: Where: ENT stands for energy; i is the i-th row of the gray-level co-occurrence matrix of the image; j is the jth column of the gray-level co-occurrence matrix of the image; k is the grayscale level of the image; G(i, j) is the number of times adjacent grayscale values ​​i and j appear in the image, that is, the value of the i-th row and j-th column of the grayscale co-occurrence matrix.

7. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 5, the specific steps of the Random Patchwork algorithm are as follows: randomly select two categories of coal from 15 single coal categories for processing; set the random ratio to m, take the mth part of the i-th image in the first category and the (1-m)th part of the i-th image in the second category to combine to generate a mixed coal image i, and record the types of vitrinite that make up the mixed coal image; continue to randomly select two categories of single coal and repeat the operation until every single coal is traversed.

8. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 7, the cross entropy loss function is as follows: Where: H(p,q) is the cross entropy loss function; p is the true distribution; q represents the predicted distribution; i is the serial number of all possible values ​​of the probability distribution, i = 1, ..., n; x i is the i-th value among all possible values ​​of the probability distribution.

9. The method for automatically detecting mixed coal vitrinite according to claim 1, wherein: In step 7, the mean square error loss function is as follows: Where: J is the mean square error loss function; i is the serial number of all possible values ​​of the probability distribution, i = 1, ..., N; x i is the i-th value among all possible values ​​of the probability distribution; yi is the true value; h θ is the predicted value.

10. The method for automatically detecting mixed coal vitrinite according to claim 1, characterized in that: In step one, there are 400 of each of the multiple single coal images; in steps one and five, there are 15 of the multiple single coal images; in step five, the ratio of the training set, validation set, and test set is 6:2:2.

Images