A method and device for identifying the uniformity of particle size of a coal sample

Abstract

The application discloses a coal sample particle size uniformity identification method and device, and belongs to the technical field of coal quality determination. The identification method comprises the following steps: obtaining baseline features and anchoring features of a coal sample image; based on a preset topological analysis unit, calculating topological parameters between the baseline features and the anchoring features; and calculating the coal sample particle size uniformity in the coal sample image according to the baseline features and the topological parameters. The baseline features and the anchoring features in the coal sample image are simultaneously extracted, and the topological parameters between the baseline features and the anchoring features are calculated, so that the algorithm is more suitable for the appearance characteristics of the coal particles, and the identification precision of the coal particle uniformity is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of coal quality determination, and particularly relates to a coal sample particle size uniformity identification method and device. BACKGROUND

[0002] Currently, the particle uniformity determination methods commonly used in the industry mainly rely on manual visual observation or simple screening methods. These methods are highly subjective and have poor reproducibility, making it difficult to meet the requirements of high-precision detection for sample uniformity. Although there are some high-precision analysis techniques that can be used to evaluate particle size distribution, these methods usually rely on expensive special equipment and have complex operation procedures, making it difficult to apply them in conventional and large-scale coal sample preparation and analysis processes.

[0003] Therefore, there is an urgent need to develop a coal particle uniformity identification technology that is convenient to operate, cost-controllable and reliable in precision, in order to solve the significant shortcomings of existing methods in terms of accuracy, efficiency and cost, and provide a solid sample foundation for high-precision detection of key coal quality and environmental protection indicators. SUMMARY

[0004] Therefore, the present application provides a coal sample particle size uniformity identification method and device, which can improve the identification accuracy of coal sample particle uniformity at a lower cost.

[0005] The technical solution is as follows: a coal sample particle size uniformity identification method, comprising the following steps:

[0006] Obtaining baseline features and anchor features of a coal sample image;

[0007] Based on a preset topological analysis unit, calculating topological parameters between the baseline features and the anchor features; the topological parameters are used to represent the spatial correlation and structure mapping relationship between the baseline features and the anchor features;

[0008] According to the baseline features and the topological parameters, the coal sample particle size uniformity in the coal sample image is calculated.

[0009] In an embodiment of the present application, the baseline features of the coal sample image are obtained by: obtaining a coal sample image, and extracting the baseline features of the coal sample image based on a plurality of baseline feature extraction units connected end to end, wherein the baseline feature extraction unit is internally provided with a convolution layer.

[0010] In an embodiment of the present application, the internal calculation process of the baseline feature extraction unit includes:

[0011] Obtaining the to-be-processed data, performing pre-feature extraction on the to-be-processed data to obtain first basic features;

[0012] dimension reduction is performed on a spatial dimension of the first basic feature to obtain a second basic feature;

[0013] alignment processing is performed on the second basic feature to obtain a third basic feature output by the reference feature extraction unit.

[0014] In an embodiment of the present application, the preposed feature extraction on the to-be-processed data comprises:

[0015] The to-be-processed data is processed by a first convolution layer and activated to obtain a first distribution feature, and a convolution kernel size of the first convolution layer is 1*1;

[0016] The to-be-processed data is processed by a second convolution layer and activated to obtain a second distribution feature, and a convolution kernel size of the second convolution layer is 3*3;

[0017] The first distribution feature and the second distribution feature are multiplied by Hadamard and activated to obtain a third distribution feature;

[0018] The first distribution feature and the to-be-processed data are subtracted and activated to obtain a fourth distribution feature;

[0019] The second distribution feature, the third distribution feature and the fourth distribution feature are spliced to obtain a fifth distribution feature;

[0020] The fifth distribution feature is processed by a third convolution layer and activated to obtain the first basic feature.

[0021] In an embodiment of the present application, the dimension reduction on the spatial dimension of the first basic feature comprises: processing the first basic feature by using a local maximum pooling layer;

[0022] In an embodiment of the present application, the alignment processing on the second basic feature comprises:

[0023] The second basic feature is processed by a deformable convolution layer and activated to obtain a first alignment feature,

[0024] The second basic feature is processed by a fourth convolution layer and activated to obtain a second alignment feature,

[0025] The first alignment feature and the second alignment feature are spliced to obtain a third alignment feature,

[0026] The third alignment feature is processed by a fifth convolution layer and activated to obtain the third basic feature.

[0027] In an embodiment of the present application, the anchor feature of the coal sample image is obtained, comprising: obtaining a coal sample image, after the coal sample image is processed by a sixth convolution layer and activated, a first anchor point feature is obtained; after the first anchor point feature is processed by a seventh convolution layer and activated, a second anchor point feature is obtained; the second anchor point feature is added to the first anchor point feature and activated to obtain the anchor feature.

[0028] In an embodiment of the present application, based on the preset topological analysis unit, the topological parameter between the baseline feature and the anchor feature is calculated, comprising:

[0029] The feature values of each channel position of the baseline feature are integrated respectively to obtain a baseline vector;

[0030] The feature values of each spatial position of the anchor feature are integrated respectively to obtain an anchor matrix;

[0031] The anchor matrix is matched according to the baseline vector to obtain an anchor vector;

[0032] The baseline vector and the anchor vector are integrated to obtain the topological parameter.

[0033] In an embodiment of the present application, the feature values of each channel position of the baseline feature are integrated respectively, comprising: the feature values of each channel position of the baseline feature are respectively globally averaged and pooled and activated;

[0034] In an embodiment of the present application, the feature values of each spatial position of the anchor feature are integrated respectively, comprising: the feature values of each spatial position of the anchor feature are respectively globally maximum pooled and activated;

[0035] In an embodiment of the present application, the anchor matrix is matched according to the baseline vector, comprising: the anchor matrix is input into a first full connection layer after being unfolded, the first full connection layer outputs a connection vector equal to the length of the baseline vector after calculation, and the connection vector is added to the baseline vector and activated to obtain the anchor vector;

[0036] In an embodiment of the present application, the baseline vector and the anchor vector are integrated, comprising: the baseline vector and the anchor vector are Hadamard multiplied and activated.

[0037] In an embodiment of the present application, the uniformity of the coal sample particles in the coal sample image is calculated according to the baseline feature and the topological parameter, comprising:

[0038] The feature values of each channel position of the baseline feature are respectively globally maximum pooled and activated to obtain a first target feature,

[0039] The first target feature is activated after being multiplied with the topological parameter to obtain a second target feature,

[0040] The second target feature is calculated through a second full connection layer to obtain a third target feature,

[0041] The third target feature is calculated through a classifier to obtain the coal sample particle uniformity in the coal sample image.

[0042] The application further provides a coal sample particle size uniformity identification device, comprising a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the method when executing the computer program.

[0043] Compared with the prior art, the application has the following beneficial effects:

[0044] Because the color of the coal particles is very deep, and there is different degree of stacking between different particles, and part of the coal particle edges are not obvious, the existing image-based particle uniformity identification algorithm performs poorly, the application creatively extracts the baseline feature and the anchor feature in the coal sample image, and calculates the topological parameter between them, so that the algorithm is more suitable for the appearance characteristics of the coal particles, and the identification accuracy of the coal particle size uniformity is improved. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a flowchart of a coal sample particle size uniformity identification method of the application. DETAILED DESCRIPTION

[0046] The application will be further described in detail below in combination with the drawings.

[0047] The embodiment provides a coal sample particle size uniformity identification method, as shown in the figure, comprising the following steps: Figure 1

[0048] Step 1: Obtain a coal sample image and extract a baseline feature of the coal sample image. When the coal sample image is obtained by shooting, the background is preferably a pure color, and preferably a light color (for example, white), so that the contrast between the coal and the background can be improved, and the interference of the environment can be reduced.

[0049] ​To extract the baseline feature from the coal sample image, usually, multiple reference feature extraction units connected in series (i.e., in tandem) are needed to complete the task. In this embodiment, five reference feature extraction units are used to extract the baseline feature of the coal sample image. For example, the first reference feature extraction unit directly receives the coal sample image, extracts features from it, and outputs the results. The second unit no longer processes the coal sample image, but takes the features output by the first reference feature extraction unit as input, continues to extract features, and outputs them. The third and fourth reference feature extraction units also follow the same logic, taking the output features of the previous unit as input for processing. Finally, the fifth reference feature extraction unit receives the output features of the fourth unit, and after calculation, the results obtained are the baseline features extracted from the coal sample image. The reference feature extraction unit is internally provided with a convolution layer, and the reference feature extraction unit can be implemented using existing convolutional neural network-based feature extraction units, for example, in some embodiments, the reference feature extraction unit can be a feature extraction module used in GoogLeNet, MobileNet, or SENet.

[0050] In other embodiments, the internal calculation process of the reference feature extraction unit includes:

[0051] The data to be processed (i.e., the data input into the reference feature extraction unit from its head) is obtained, pre-feature extraction is performed on the data to be processed to obtain first basic features, the spatial dimensions of the first basic features are reduced to obtain second basic features, and the second basic features are aligned to obtain third basic features output at the end of the reference feature extraction unit. The third basic features output at the end of the fifth reference feature extraction unit are the baseline features.

[0052] In some embodiments, pre-feature extraction is performed on the data to be processed, including:

[0053] The first distribution feature is obtained after the to-be-processed data is processed by a first convolution layer and activated, and the convolution kernel size of the first convolution layer is 1*1; the second distribution feature is obtained after the to-be-processed data is processed by a second convolution layer and activated, and the convolution kernel size of the second convolution layer is 3*3; the third distribution feature is obtained after the first distribution feature and the second distribution feature are multiplied by Hadamard and activated; the fourth distribution feature is obtained after the first distribution feature and the to-be-processed data are subtracted (for example, the first distribution feature can be subtracted from the to-be-processed data) and activated; the fifth distribution feature is obtained after the second distribution feature, the third distribution feature and the fourth distribution feature are spliced; and the first basic feature is obtained after the fifth distribution feature is processed by a third convolution layer (the convolution kernel size can be 3*3) and activated. In some embodiments, the height and width sizes of the first basic feature correspond to and are equal to the height and width of the to-be-processed data input into the reference feature extraction unit, and the number of channels of the first basic feature is twice the number of channels of the to-be-processed data input into the reference feature extraction unit.

[0054] In some embodiments, the spatial dimension of the first basic feature is reduced, including: the first basic feature is processed by a local maximum pooling layer to obtain a second basic feature. As an example, the pooling window size of the local maximum pooling layer can be 2*2, and when the pooling operation is performed, the pooling window slides along the spatial direction (i.e., the width and height direction), and the sliding step can be 2.

[0055] In some embodiments, the second basic feature is processed for alignment, including: the second basic feature is processed by a (existing) deformable convolution layer and activated to obtain a first alignment feature, the second basic feature is processed by a fourth convolution layer and activated to obtain a second alignment feature, the first alignment feature and the second alignment feature are spliced to obtain a third alignment feature, and the third alignment feature is processed by a fifth convolution layer and activated to obtain a third basic feature.

[0056] After the above calculation process is used in the reference feature extraction unit, more refined feature extraction and robust alignment processing are realized, the feature expression ability and feature deformation adaptability are improved, the information is efficiently utilized in stages, and the effect of subsequent recognition and classification tasks can be effectively strengthened.

[0057] Step 2: Obtain a coal sample image, and the first anchor point feature is obtained after the coal sample image is processed by a sixth convolution layer (for example, the convolution kernel size can be 3*3) and activated; the second anchor point feature is obtained after the first anchor point feature is processed by a seventh convolution layer (for example, the convolution kernel size can be 3*3) and activated; and the anchor feature is obtained after the corresponding elements of the second anchor point feature and the first anchor point feature are added and activated.

[0058] Step 3: Both the baseline feature and the anchor feature are three-dimensional (height, width, and channel) feature matrix data. Based on the preset topological analysis unit, the topological parameters between the baseline feature and the anchor feature are calculated; the topological parameters are used to represent the spatial correlation and structural mapping relationship between the baseline feature and the anchor feature.

[0059] In some embodiments, based on the preset topological analysis unit, the topological parameters between the baseline feature and the anchor feature are calculated, including: respectively integrating the feature values of each channel position of the baseline feature to obtain a baseline vector; respectively integrating the feature values of each spatial position of the anchor feature to obtain an anchor matrix; performing matching processing on the anchor matrix according to the baseline vector to obtain an anchor vector; and integrating the baseline vector and the anchor vector to obtain the topological parameters.

[0060] In some embodiments, respectively integrating the feature values of each channel position of the baseline feature includes: respectively performing global average pooling and post-activation on the feature values of each channel position of the baseline feature to obtain the baseline vector. For example, for a baseline feature with a height of A1, a width of A2, and a channel number of A3, each channel position of the baseline feature contains (A1*A2) feature values. Global average pooling is performed on the feature values of each channel position, that is, the average value of (A1*A2) feature values in each channel is calculated, and then a baseline vector with a length of A3 is obtained.

[0061] In some embodiments, respectively integrating the feature values of each spatial position of the anchor feature includes: respectively performing global maximum pooling and post-activation on the feature values of each spatial position of the anchor feature to obtain the anchor matrix. For example, for an anchor feature with a height of B1, a width of B2, and a channel number of B3, the anchor feature contains B1*B2 different spatial positions, and each spatial position contains B3 feature values. Global maximum pooling is performed on the feature values of each spatial position, that is, the maximum value of B3 feature values in each spatial position is calculated, and then an anchor matrix with a size of (B1*B2) is obtained.

[0062] In some embodiments, according to the baseline vector, the matching processing is performed on the anchor matrix, including: inputting the anchor matrix after expansion into a first fully connected layer, the first fully connected layer calculating and outputting a connection vector with a length equal to that of the baseline vector, and the connection vector and the baseline vector corresponding elements being added and activated to obtain the anchor vector.

[0063] In some embodiments, integrating the baseline vector and the anchor vector includes: performing Hadamard product on the baseline vector and the anchor vector and then activating to obtain the topological parameters.

[0064] The topology analysis unit is designed in a progressive manner of "differential integration-precise matching-deep fusion", so that the topology parameters can not only reflect the overall characteristic benchmark of the coal sample, but also capture the key local characteristics, while the correlation between the two is strengthened, thereby greatly making up for the deficiency of the baseline characteristics.

[0065] Step 4: Calculate the coal sample particle uniformity in the coal sample image according to the baseline characteristics and the topology parameters. In some embodiments, the coal sample particle uniformity in the coal sample image is calculated according to the baseline characteristics and the topology parameters, including: respectively performing global maximum pooling on the feature values of each channel position of the baseline characteristics to obtain a first target feature, performing Hadamard product on the first target feature and the topology parameters to obtain a second target feature, and performing calculation on the second target feature through a second fully connected layer to obtain a third target feature, and performing calculation on the third target feature through a classifier (such as a softmax classifier) to obtain the coal sample particle size uniformity in the coal sample image.

[0066] The third target feature described above is one-dimensional vector data, and its length can be set to different values according to actual needs, so as to adjust the number of classification categories of the particle size uniformity. For example, the length of the third target feature can be set to 20, and the result output by the classifier contains 20 categories, which can be represented by numbers 1-20 respectively. The larger the number is, the better the particle size uniformity is. A suitable threshold value can also be set to determine whether the coal sample particle size uniformity is qualified (for example, the threshold value can be 10). When it is identified that the particle uniformity is lower than the preset threshold value, it means that the coal sample particle uniformity is unqualified, and the sample needs to be processed again or replaced with another sample. Otherwise, it means that the coal sample particle size uniformity meets the requirements and has little effect on the subsequent coal measurement results.

[0067] As known to those skilled in the art, the above-mentioned activation refers to inputting relevant data into a nonlinear activation function for operation to obtain activated data. As an example but not limitation, the above-mentioned activation can be activated by selecting ReLU, Sigmoid, Swish or ELU functions. In some embodiments, the steps of extracting the baseline characteristics and the anchor characteristics of the coal sample image, calculating the topology parameters between the baseline characteristics and the anchor characteristics based on the preset topology analysis unit, and calculating the coal sample particle size uniformity according to the baseline characteristics and the topology parameters can be realized by building an artificial neural network through a computer program. These steps can be completed by operating a computer, a tablet computer or a mobile phone terminal and other electronic devices.

[0068] Finally, it should be noted that the above description is only for the preferred embodiments of the present application, and those of ordinary skill in the art can make various similar expressions under the inspiration of the present application without departing from the purpose and scope of the present application, and such changes fall within the protection scope of the present application.

Claims

1. A method for identifying the particle size uniformity of coal samples, characterized in that, Includes the following steps: Obtain baseline and anchoring features from coal sample images; Based on a preset topology analysis unit, the topology parameters between the baseline feature and the anchoring feature are calculated; the topology parameters are used to characterize the spatial association and structural mapping relationship between the baseline feature and the anchoring feature. The uniformity of coal sample particle size in the coal sample image is calculated based on the baseline features and the topology parameters. Based on a preset topology analysis unit, the topology parameters between the baseline feature and the anchoring feature are calculated, including: The feature values ​​of each channel position of the baseline feature are integrated to obtain the baseline vector; The feature values ​​of each spatial location of the anchoring feature are integrated to obtain the anchoring matrix; The anchoring matrix is ​​matched based on the baseline vector to obtain the anchoring vector; The baseline vector and the anchoring vector are integrated to obtain the topology parameters; Integrating the feature values ​​of each channel position of the baseline feature includes: performing global average pooling on the feature values ​​of each channel position of the baseline feature and then activating them; Or / and, respectively integrate the feature values ​​of each spatial location of the anchor feature, including: respectively perform global max pooling on the feature values ​​of each spatial location of the anchor feature and then activate; Or / and, matching the anchoring matrix according to the baseline vector includes: expanding the anchoring matrix and inputting it into a first fully connected layer; the first fully connected layer calculates and outputs a connection vector with the same length as the baseline vector; the connection vector is added to the baseline vector and then activated to obtain the anchoring vector; Or / and, integrating the baseline vector with the anchor vector includes: activating the baseline vector after performing a Hadamard product with the anchor vector; The uniformity of coal sample particle size in the coal sample image is calculated based on the baseline features and the topological parameters, including: After performing global max pooling on the feature values ​​at each channel position of the baseline feature, the first target feature is obtained. The first target feature is activated by performing a Hadamard product with the topological parameters to obtain the second target feature. The second target feature is processed by the second fully connected layer to obtain the third target feature. The third target feature, after being calculated by the classifier, yields the uniformity of coal sample particle size in the coal sample image.

2. The method for identifying the particle size uniformity of coal samples according to claim 1, characterized in that, Obtaining baseline features of a coal sample image includes: acquiring a coal sample image, and extracting baseline features of the coal sample image based on multiple baseline feature extraction units connected end-to-end, wherein each baseline feature extraction unit has a convolutional layer.

3. The method for identifying the particle size uniformity of coal samples according to claim 2, characterized in that, The internal calculation process of the benchmark feature extraction unit includes: Obtain the data to be processed, and perform preliminary feature extraction on the data to be processed to obtain the first basic feature; The spatial dimension of the first basic feature is reduced to obtain the second basic feature; The second basic feature is aligned to obtain the third basic feature output by the reference feature extraction unit.

4. The method for identifying the particle size uniformity of coal samples according to claim 3, characterized in that, Preliminary feature extraction is performed on the data to be processed, including: The data to be processed is processed and activated by the first convolutional layer to obtain the first distribution feature. The kernel size of the first convolutional layer is 1*1. The data to be processed is processed and activated by the second convolutional layer to obtain the second distribution feature. The kernel size of the second convolutional layer is 3*3. The first distribution feature and the second distribution feature are activated by performing a Hadamard product to obtain the third distribution feature; The first distribution feature is activated by subtracting the data to be processed, resulting in a fourth distribution feature. The second distribution feature, the third distribution feature, and the fourth distribution feature are concatenated to obtain the fifth distribution feature; The fifth distribution feature is processed and activated by the third convolutional layer to obtain the first basic feature.

5. The method for identifying the particle size uniformity of coal samples according to claim 3, characterized in that, Dimensionality reduction of the first basic feature includes: processing the first basic feature using a local max pooling layer; Or / and, aligning the second basic feature, including: The second basic feature, after being processed and activated by a deformable convolutional layer, yields the first alignment feature. The second basic feature, after being processed and activated by the fourth convolutional layer, yields the second alignment feature. The first alignment feature is concatenated with the second alignment feature to obtain the third alignment feature. The third alignment feature is processed and activated by the fifth convolutional layer to obtain the third basic feature.

6. The method for identifying the particle size uniformity of coal samples according to claim 1, characterized in that, The method for obtaining anchoring features from a coal sample image includes: acquiring a coal sample image; processing and activating the coal sample image through a sixth convolutional layer to obtain a first anchoring feature; processing and activating the first anchoring feature through a seventh convolutional layer to obtain a second anchoring feature; and adding and activating the second anchoring feature to the first anchoring feature to obtain the anchoring feature.

7. A device for identifying the particle size uniformity of coal samples, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method as described in any one of claims 1 to 6.

Images