A Coal Rock Image Recognition and Classification Method Based on Fusion Improved CLBP and Receptive Field Theory

Abstract

This invention discloses a coal and rock image recognition and classification method based on the fusion of improved CLBP and receptive field theory, belonging to the field of image processing technology. First, the coal and rock image is converted to grayscale, and the pixel values ​​of each point in the grayscale image are extracted to form a pixel value matrix. Then, median processing is used to denoise the pixel matrix to enhance the algorithm's noise resistance. Based on the new pixel matrix obtained after median processing, an improved Complete Local Binary Pattern (CLBP) image feature descriptor is introduced for coal and rock feature extraction. Finally, a network model is used for recognition and classification. Experimental results show that the recognition accuracy of coal and rock in the coal and rock database reaches 94%, and the recognition speed is also better than the original algorithm. This invention can effectively solve the problem of low recognition efficiency of the original algorithm and can efficiently solve practical problems in coal and rock recognition methods, possessing certain practical value.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a coal and rock image recognition and classification method based on the fusion of improved CLBP and receptive field theory. Background Technology

[0002] Texture features are crucial in image recognition and classification, playing a wide range of roles across many fields. In some classification tasks, such as accurately distinguishing between coal and rock, inadequate prior knowledge can negatively impact classification results. Therefore, employing more targeted feature extraction methods is key to addressing these challenges.

[0003] Current coal and rock identification methods mainly fall into two categories: contact and non-contact. Contact methods rely on frequent interaction between the sensing device and the coal and rock mass, leading to rapid device wear and tear. Non-contact methods, on the other hand, face several other challenges. Therefore, image-based coal and rock identification methods hold great promise. Effective extraction of texture features from images is therefore crucial in coal and rock identification.

[0004] The CLBP texture descriptor is an extension of the LBP descriptor. It effectively describes the missing information of the LBP type, thereby achieving better texture classification performance. Analysis of coal and rock images reveals that the essential difference lies in texture features (including primitive structure, statistical regularities, gray-level distribution, etc.). Image texture information is largely contained in the mid-to-high frequency domain information of the image. Existing feature extraction methods only study the overall gray-level pixel distribution of the image without focusing on mid-to-high frequency information, necessitating improvements to current feature extraction methods.

[0005] Given the current limitations of complete local binary patterns in terms of recognition efficiency, the redundant feature information generated by feature extraction in coal and rock texture images, and the low classification efficiency of traditional machine learning algorithms, improvements are needed in reducing feature dimensionality and increasing computational speed. Summary of the Invention

[0006] To address the problems of low classification efficiency, redundant feature extraction in coal and rock texture image recognition, and the lack of interpretability of deep learning network models in feature extraction, this invention provides a coal and rock image recognition and classification method that integrates improved CLBP and receptive field theory.

[0007] This invention is achieved through the following technical solution: a coal and rock image recognition and classification method based on the fusion of improved CLBP and receptive field theory, comprising the following steps:

[0008] Step 1: Collect image data of coal and rocks in the mine;

[0009] Step 2: When preparing experimental data, the coal and rock images are augmented and their dimensions are standardized to obtain the coal and rock texture image experimental dataset;

[0010] Step 3: Perform grayscale processing on the experimental dataset obtained in Step 2 to obtain a coal and rock image database;

[0011] Step 4: Extract texture features from the coal and rock image database obtained in Step 3 using median processing and an improved complete local binary mode method. These texture features include sign components, amplitude components, and center pixel components.

[0012] Step 4.1: Use the median processing method to perform noise reduction on the grayscale image obtained in Step 3. Specifically, use a filter kernel of a certain size to traverse the entire pixel matrix from left to right or from top to bottom, set the filter kernel step size, and use the extracted medians to form a new pixel value matrix.

[0013] Step 4.2: Perform feature extraction on the obtained new image pixel matrix, and obtain the sign component, amplitude component and center pixel component of the coal and rock image through improved local difference operation;

[0014] Step 4.2.1: Use a 5*5 sampling neighborhood for feature extraction, which is divided into the center pixel, the neighboring pixels and the outermost sampling pixels. Pixels are collected at the horizontal, vertical and diagonal positions respectively. Excluding the center pixel, the number of pixels sampled in each layer is 8, and a new local sampling topology structure is obtained.

[0015] Step 4.2.2: Extract features from the local sampling region, and obtain the sign component and amplitude component through second-order difference calculation, as shown in formula (1). This is to improve the pixel difference obtained by second-order difference for the complete local binary pattern:

[0016] d i =g c +g 1i -2×g 2i i = 1, 2, ..., 8 (1)

[0017] Where, d i g represents the local pixel difference value. c The pixel value at the center point of the local sampled pixel, where i is the number of sampling points, and g is the pixel value at the center point of the local sampled pixel. 1i g 2i These are the sampled pixel values ​​in the horizontal, vertical, and diagonal directions, with the subscripts representing the positions of the sampled pixels in the horizontal, vertical, and diagonal directions.

[0018] d obtained from formula (1) i Substitute di =s i ×m i Then, the symbol component and amplitude component of the local sampling region are obtained respectively, and the calculation method is as follows:

[0019] The symbolic component after binary encoding is:

[0020]

[0021] Where: s i S represents the sign value, and S is the discriminant function;

[0022] The amplitude component after binary encoding is:

[0023]

[0024] Where: m i The absolute value of the local difference, C is the threshold, and is set to the average value of the current local amplitude;

[0025] Step 4.3: Set the pixel value g of the center point of the locally sampled pixel. c The feature value of the center pixel is obtained by comparing it with the global threshold I. The feature value is calculated using the following formula:

[0026] Center pixel feature value = t(g c ,I) (4)

[0027] In the formula, I is the threshold, which is set as the average value of the pixel values ​​of the entire image;

[0028] Step 5: Finally, the feature values ​​obtained from each sampling region are used to form the final feature vector. At the same time, the feature information of the symbol component and amplitude component obtained in Step 4.2 are combined and used in a series and parallel joint structure to combine the CLBP_S, CLBP_M and CLBP_C descriptors and cascade them to form the final complete texture feature. The symbol information CLBP_S and amplitude information CLBP_M are combined to form CLBP_SM, and finally CLBP_C is connected to form the histogram of CLBP_SM\C.

[0029] Step 6: Because the extracted feature matrix has too high a dimension and contains a lot of redundant information, it not only interferes with the improvement of recognition accuracy, but also causes a sharp increase in computer time and space overhead, limiting the recognition speed and hindering the practicality of the algorithm. Therefore, the histogram obtained in Step 5 is used as a feature matrix and input into the network model. After several receptive field modules, the final classification result is output through softmax. The network model is set as follows: First, two convolutional layers are set, each activated by the hyperbolic tangent function; the next layer is a max pooling layer with a filter kernel of 3 and a stride of 2, which performs further dimensionality reduction through max pooling; another convolutional layer is used to extract abstract features, and the output is fed into the next max pooling layer for the final dimensionality reduction; the obtained feature information is output through the Flatten operation; finally, softmax is used to determine whether the feature information is coal or rock.

[0030] Preferably, in step 2, the data amplification method is to flip, translate, rotate, and transform the contrast of the coal and rock image, etc.

[0031] Compared with the prior art, the present invention has the following beneficial effects: The coal and rock image recognition and classification method provided by the present invention integrates improved CLBP and receptive field theory. It addresses the problem that the original algorithm excessively ignores the high-order information of the image and extracts redundant information in the features, resulting in low efficiency in recognizing and classifying coal and rock images. By using max pooling and improved fully local binary mode, the present invention effectively solves the problem of low recognition efficiency and ensures high accuracy. Attached Figure Description

[0032] Figure 1 This is a flowchart of the coal and rock image recognition and classification method of the present invention, which integrates the improved CLBP and receptive field theory.

[0033] Figure 2 It is a local sampling topology map of a coal and rock image recognition and classification method that integrates improved CLBP and receptive field theory.

[0034] Figure 3 This is a schematic diagram of a local difference sign-amplitude transformation example of a coal and rock image recognition and classification method that integrates improved CLBP and receptive field theory.

[0035] Figure 4 This is the network model structure diagram from step 6. Detailed Implementation

[0036] The present invention will be further described below with reference to specific embodiments.

[0037] A coal and rock image recognition and classification method that integrates improved CLBP and receptive field theory, such as Figure 1 As shown, it includes the following steps:

[0038] Step 1: Collect image data of coal and rocks in the mine;

[0039] Step 2: When creating experimental data, the coal and rock images are amplified and their sizes are standardized by performing operations such as flipping, translating, rotating, and contrast transformation to obtain the coal and rock texture image experimental dataset.

[0040] Step 3: Perform grayscale processing on the experimental dataset obtained in Step 2 to obtain a coal and rock image database;

[0041] Step 4: Extract texture features from the coal and rock image database obtained in Step 3 using median processing and an improved complete local binary mode method. These texture features include sign components, amplitude components, and center pixel components.

[0042] Step 4.1: Use the median processing method to perform noise reduction on the grayscale image obtained in Step 3. Specifically, use a filter kernel of a certain size to traverse the entire pixel matrix from left to right or from top to bottom, set the filter kernel step size, and use the extracted medians to form a new pixel value matrix.

[0043] Step 4.2: Perform feature extraction on the obtained new image pixel matrix, and obtain the sign component, amplitude component and center pixel component of the coal and rock image through improved local difference operation;

[0044] Step 4.2.1: Use a 5*5 sampling neighborhood for feature extraction, which is divided into the center pixel, the neighboring pixels and the outermost sampling pixels. Pixels are collected at the horizontal, vertical and diagonal positions respectively. Excluding the center pixel, the number of pixels sampled in each layer is 8, and a new local sampling topology structure is obtained.

[0045] Step 4.2.2: Extract features from the local sampling region, and obtain the sign component and amplitude component through second-order difference calculation, as shown in formula (1). This is to improve the pixel difference obtained by second-order difference for the complete local binary pattern:

[0046] d i =g c +g 1i -2×g 2i i = 1, 2, ..., 8 (1)

[0047] Where, d i g represents the local pixel difference value. c The pixel value at the center point of the local sampled pixel, where i is the number of sampling points, and g is the pixel value at the center point of the local sampled pixel. 1i g 2i These are the sampled pixel values ​​in the horizontal, vertical, and diagonal directions, with the subscripts representing the positions of the sampled pixels in the horizontal, vertical, and diagonal directions.

[0048] d obtained from formula (1) i Substitute d i =s i ×m i Then, the symbol component and amplitude component of the local sampling region are obtained respectively, and the calculation method is as follows:

[0049] The symbolic component after binary encoding is:

[0050]

[0051] Where: s i S represents the sign value, and S is the discriminant function;

[0052] The amplitude component after binary encoding is:

[0053]

[0054] Where: m i The absolute value of the local difference, C is the threshold, and is set to the average value of the current local amplitude;

[0055] Step 4.3: Set the pixel value g of the center point of the locally sampled pixel. c The feature value of the center pixel is obtained by comparing it with the global threshold I. The feature value is calculated using the following formula:

[0056] Center pixel feature value = t(g c ,I) (4)

[0057] In the formula, I is the threshold, which is set as the average value of the pixel values ​​of the entire image;

[0058] Step 5: Finally, the feature values ​​obtained from each sampling region are used to form the final feature vector. At the same time, the feature information of the symbol component and amplitude component obtained in Step 4.2 are combined and used in a series and parallel joint structure to combine the CLBP_S, CLBP_M and CLBP_C descriptors and cascade them to form the final complete texture feature. The symbol information CLBP_S and amplitude information CLBP_M are combined to form CLBP_SM, and finally CLBP_C is connected to form the histogram of CLBP_SM\C.

[0059] Step 6: The histogram obtained in Step 5 is used as a feature matrix and input into the network model. After several receptive field modules, the final classification result is output through softmax. The network model is set as follows: First, two convolutional layers are set, each activated by the hyperbolic tangent function; the next layer is a max pooling layer with a filter kernel of 3 and a stride of 2, which performs further dimensionality reduction; another convolutional layer is used to extract abstract features, and the output is fed into the next max pooling layer for the final dimensionality reduction; the obtained feature information is output through the Flatten operation; finally, softmax is used to determine whether the feature information is coal or rock.

[0060] The experimental results of this embodiment show that the accuracy of coal and rock identification and classification of the present invention is 99.47%, and the running time is reduced by about 90% compared with the original algorithm. The present invention effectively solves the problem that the original algorithm is inefficient and cannot meet the real-time requirements, and greatly improves the running efficiency of the algorithm while ensuring accuracy.

[0061] The scope of protection claimed by this invention is not limited to the specific embodiments described above. Moreover, for those skilled in the art, this invention can have various modifications and alterations. Any modifications, improvements, and equivalent substitutions made within the concept and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A coal rock image recognition and classification method based on fusion of improved CLBP and receptive field theory, characterized in that: Comprising the following steps: Step 1: Collecting picture data of coal and rock taken in the mine; Step 2: When making experimental data, the coal and rock images are data augmented and unified in size to obtain a coal and rock texture image experimental data set; Step 3: The experimental data set obtained in step 2 is subjected to grayscale processing to obtain a coal and rock image database; Step 4: The coal and rock image database of the grayscale image obtained in step 3 is subjected to median processing and improved complete local binary pattern method to extract texture features of the coal and rock image, the texture features including sign component, amplitude component and center pixel point component: Step 4.1: The grayscale image obtained in step 3 is subjected to noise reduction operation by median processing method, specifically: a filter kernel of a certain size is used to traverse the entire pixel matrix from left to right or from top to bottom, the filter kernel step is set, and the extracted median is composed into a new pixel value matrix; Step 4.2: The new image pixel matrix obtained is subjected to feature extraction, and the sign component, amplitude component and center pixel point component of the coal and rock image are obtained by improved local difference operation; Step 4.2.1: A 5*5 sampling neighborhood is used for feature extraction, which is divided into a center pixel point, a neighborhood pixel point and an outermost circle sampling pixel point, and the pixel points are collected at horizontal, vertical and diagonal positions, and the number of pixel points in each layer is 8, and a new local sampling topology is obtained; Step 4.2.2: The features of the local sampling region are extracted, and the sign component and amplitude component are obtained by calculating the second-order difference, as shown in formula (1), which is the pixel difference obtained by improving the complete local binary pattern by second-order difference: （1） wherein, is a local pixel difference value, is a local sampling pixel center point pixel value, i is the number of sampling points, , is a horizontal, vertical and diagonal direction sampling pixel point value, the subscript represents the position of the sampling pixel point in the horizontal, vertical and diagonal direction; d = 2π / 2N i Substitute Then, the sign component and the amplitude component of the local sampling region are obtained respectively, and the calculation method is as follows: The sign component after binary coding is: ， （2） wherein: represents a symbol value, S is a discriminant function; The amplitude component after binary coding is: ， （3） wherein: is the absolute value of the local difference, and C is a threshold value, set to the average of the current local amplitudes; Step 4.3: Compare the local sampling pixel center point pixel value with the global threshold I to obtain a center pixel feature value, and the feature value calculation formula is as follows: The center pixel feature value is compared with the global threshold I, and the feature value calculation formula is as follows: center pixel feature value (4) In the formula, I is a threshold value, which is set as the average value of the whole image pixel value; Step 5: Finally, the feature values obtained from each sampling region form the final feature vector, and the feature information of the sign component and amplitude component obtained in step 4.2 is combined, and the CLBP_S, CLBP_M and CLBP_C descriptors are combined and used in series and parallel joint structure, and are cascaded and fused into the final complete texture feature, the sign information CLBP_S and the amplitude information CLBP_M are combined in series to become CLBP_SM, and finally CLBP_C is connected in series to form the histogram of CLBP_SM\C; Step 6: The histogram obtained in step 5 is a feature matrix, which is input into the network model, and after several receptive field module operations, the final classification result is output through softmax; the network model is set as follows: first, two convolutional layers are set, each convolutional layer is activated by hyperbolic tangent function; followed by a maximum pooling layer, the filter kernel is set to 3 and the step is set to 2, and further dimension reduction operation is performed by maximum pooling; again, a convolutional layer is used to continue to extract abstract features, and the output enters the next maximum pooling layer for the last time; the obtained feature information is output by Flatten operation; finally, the softmax is used to determine whether the feature information is coal or rock.

2. The coal rock image recognition and classification method of fusing improved CLBP and receptive field theory according to claim 1, characterized in that: In step 2, the data augmentation is performed by flipping, translating, rotating and contrast transforming the coal rock image.

3. The coal rock image recognition and classification method of fusing improved CLBP and receptive field theory according to claim 1, characterized in that: In step 4.1, the size of the filter kernel is 2*2, and the filter kernel step is 2.

Images