Coal rock maceral classification method and classification model based on computer vision

Abstract

The invention discloses a coal rock maceral classification method and classification model based on computer vision. The method comprises the steps that a first sample set is used for training a generative adversarial network to obtain an image generation model, a generator is a multi-scale convolutional neural network, and a discriminator classifies images through an adaptive sparse convolution matrix; a second sample set is acquired, image expansion is performed by using an image generation model, a third sample set is established, and the first sample set and the second sample set both comprise a coal rock vitrinite microscopic image and an inertinite microscopic image; training a fractional order neural network based on the third sample set to obtain a classification model; and determining the maceral type of the coal rock microscopic image based on the classification model. According to the method, training data of an adversarial network expansion classification model is introduced, and the problems of insufficient data and poor model generalization ability are solved; meanwhile, in combination with a fractional order neural network classification model, the processing capability and classification accuracy of the model on complex geological data are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a method and a classification model for classifying microscopic components of coal and rock based on computer vision. Background Art

[0002] The accurate identification of vitrinite and inertinite is of great significance for the research of deep coalbed methane. Vitrinite and inertinite are the main microscopic components in coal and rock, and their content and characteristics directly affect the physical properties, reservoir performance and gas storage capacity of coal and rock. For example, vitrinite has a relatively high reflectivity and luster, and is often associated with high gas production capacity and reservoir density, while inertinite reflects the structural complexity and fracture development degree of coal and rock, and is closely related to the gas migration ability of the reservoir. In addition, the microscopic characteristics of these two components can also indicate the origin and evolution process of coal and rock, providing key information for the geological modeling and resource evaluation of the reservoir. By accurately identifying the content of vitrinite and inertinite, the reservoir classification and exploration strategy can be optimized, and the development efficiency and economic value of coalbed methane can be improved. There are significant differences in the microscopic structures of these two microscopic components. Therefore, accurately classifying them through microscope images based on computer vision is of great significance for the classification and identification of minerals and related research. Using the microscopic image classification method based on computer vision to identify the microscopic components of coal and rock can not only significantly improve the classification efficiency, but also reduce the errors in manual analysis, providing an advanced technical means for the research and application of deep coalbed methane. Summary of the Invention

[0003] The inventors found in their work that the microscopic image classification task of coal and rock components based on computer vision faces multiple technical challenges. First, microscopic images usually have very high resolution and complex microstructures; second, the amount of sample data required for microscopic image classification is usually very large, and the existing data collection and annotation work is time-consuming and labor-intensive, and the number of existing samples is often insufficient to train a high-precision classification model; in addition, there are many types of mineral samples and they have high similarity, especially between vitrinite and inertinite, and their microscopic morphology and optical properties have small differences, resulting in high classification difficulty.

[0004] In the prior art, although there are many research methods based on image processing and machine learning that attempt to solve the microscopic image classification problem, there are still many deficiencies. For example, traditional image classification methods rely on manually extracted features or shallow neural networks, and these methods usually have difficulty capturing complex details and small changes in images, resulting in low classification accuracy. In addition, existing image augmentation techniques mostly rely on simple image transformations such as rotation and scaling. Although these methods can increase the sample size, they cannot effectively improve the sample quality and are prone to causing model overfitting.

[0005] The Chinese invention patent with application number 202110509387.3 proposes a fungus microscopic image classification method and system based on a multi-scale attention mechanism, including: obtaining a training sample, the training sample includes a plurality of fungus microscopic images, and each fungus microscopic image has a corresponding fungus category label; constructing a deep learning image classification model including an attention module, using the training sample to train the deep learning image classification model, and using the trained deep learning image classification model as a fungus image classification model; inputting the fungus microscopic image to be classified into the fungus image classification model to obtain its fungus category. The present invention adds an attention module to the network so that the network focuses on the area where the fungi are located and ignores the influence of the background area as much as possible. Thereby, fungi of the same genus and different species with small morphological differences can be more accurately identified.

[0006] The Chinese invention patent with application number CN201910714652.4 proposes an intelligent identification method for coal rock microscopic components based on image segmentation and classification, including: first, based on the double-layer K-means image segmentation algorithm, the microscopic image is segmented into four categories: background resin, exinite, vitrinite and inertinite; then the segmented image is extracted using the multimodal feature extraction method designed by the present invention, the geometric features, texture features and grayscale features of the coal rock components are integrated together to form a total of 172 dimensional features, and the random forest method is used to classify the above multimodal features, further subdivided into background resin, vitrinite, spores, cutinoids, silk bodies, semi-silk bodies, clastic inertinoids and fragments. The present invention combines image segmentation and image classification algorithms to complete the analysis of coal rock microscopic components, which has the advantages of high accuracy, strong robustness, and fast and simple.

[0007] The Chinese invention patent with application number CN201810240366.4 proposes a fungus microscopic image recognition method based on a fully convolutional neural network, which includes the following steps: fungus image acquisition; image preprocessing; construction of a fully convolutional neural network; training of a fungus recognition neural network; verification of recognition effect and parameter adjustment. The present invention can extract sufficient feature information for recognition from a large number of fungus images, thereby being applied to the recognition of multiple types of fungus images. In addition, the use of a fully convolutional neural network improves the recognition efficiency and accuracy of the image, while making it easier to visualize the learning features. The present invention achieves efficient and accurate fungus image recognition, and therefore has high practical value.

[0008] The above prior art has many defects, including:

[0009] 1. In the microscopic image classification task, data collection, annotation, and preprocessing are usually time-consuming and cumbersome, and the number of natural mineral specimens and laboratory synthetic samples is often limited. Traditional image classification methods rely on a large amount of annotated data, and there are few methods in the existing technology that can effectively expand the dataset. Existing data augmentation techniques usually rely on simple operations such as image rotation and scaling. These methods cannot generate high-quality and realistic samples, which easily lead to overfitting or low generalization ability during the training process of the model.

[0010] 2. In traditional neural network training, due to the selection of activation functions or unreasonable network structures, it is easy to cause gradient vanishing, gradient explosion, or getting stuck in local optimal solutions. Especially in the training of high-dimensional data (such as microscopic images), common gradient problems will lead to unstable training and performance degradation. In the existing technology, many neural network models cannot effectively solve these problems, thus affecting the training efficiency and model performance.

[0011] 3. Traditional feature extraction methods often use fixed hand-designed features (such as texture features, edge features, etc.). These methods are often difficult to capture the complex details in microscopic images. Even when using deep learning methods, the existing technology mostly relies on traditional convolutional neural networks. These methods may not be able to fully learn the microscopic structure in microscopic images. Especially when classifying vitrinite and inertinite, the accuracy of the model is often affected. Most of the existing classifier models do not have an effective mechanism to dynamically adjust feature weights, which results in the importance of features not being adjusted in time, thus affecting the classification accuracy.

[0012] In order to solve the above problems at least partially, the embodiments of the present invention provide a method and a classification model for classifying coal petrographic macerals based on computer vision. By introducing a generative adversarial network to expand the training data of the classification model, the problems of insufficient data and poor model generalization ability are solved; at the same time, combined with a fractional-order neural network classification model, the processing ability and classification accuracy of the model for complex geological data are further improved.

[0013] In the first aspect, the embodiments of the present invention provide a method for classifying coal petrographic macerals based on computer vision, including:

[0014] Determining the maceral type of the coal petrographic microscopic image based on a classification model, and the classification model is established through the following steps:

[0015] Training the generator and discriminator of the generative adversarial network using a first sample set to obtain an image generation model, where the generator is a multi-scale convolutional neural network, and the discriminator classifies images through an adaptive sparse convolutional matrix;

[0016] Obtain a second sample set, perform image augmentation using the image generation model, and establish a third sample set. Both the first sample set and the second sample set contain vitrinite micrographs and inertinite micrographs of coal rock.

[0017] Train a fractional-order neural network based on the third sample set to obtain a classification model.

[0018] Optionally, training the generator and discriminator of the generative adversarial network using the first sample set includes:

[0019] Initialize the weight parameters of the generator and discriminator of the generative adversarial network respectively.

[0020] Input a set of random noise vectors into the generator, input the image set generated by the generator through multi-layer convolution processing from low frequency to high frequency and the first sample set into the discriminator, and obtain the discrimination result of the authenticity of the input image through adaptive sparse convolution matrix processing.

[0021] Determine the generator loss function by maximizing the misclassification probability of the discriminator, and determine the discriminator loss function by maximizing the sum of the correct classification of real images and the misclassification of generated images.

[0022] Update the weight parameters of the generator based on the gradient of the generator loss function, and update the weight parameters of the discriminator based on the gradient of the discriminator loss function.

[0023] If the iteration termination condition is not satisfied, return to execute the step of inputting the set of random noise vectors into the generator.

[0024] Optionally, the fractional-order neural network includes an input layer, a feature extraction layer, multiple dynamic recalibration layers, and an output layer. The dynamic recalibration layer dynamically adjusts the importance weight of each feature according to the current state and context of the input features.

[0025] Optionally, training the fractional-order neural network based on the third sample set includes:

[0026] Input the third sample set into the fractional-order neural network, and initialize the weight parameters and biases of the fractional-order neural network.

[0027] The dynamic recalibration layer of the fractional-order neural network adjusts the feature importance weight matrix of the input data through an activation function according to the current bias, and combines the current network weight parameters and biases to obtain the output data through the activation function.

[0028] Determine the probability distribution of the microscopic component types of each input image according to the output result, and combine the microscopic component type labels and the current weight matrix to determine the loss function of the fractional-order neural network.

[0029] Updating the network weight parameters of the fractional-order neural network based on the gradient of the loss function of the fractional-order neural network with respect to the network weight parameters, and updating the bias of the fractional-order neural network based on the gradient of the loss function of the fractional-order neural network with respect to the bias;

[0030] If the iteration termination condition is not satisfied, return to execute the dynamic recalibration layer of the fractional-order neural network, and adjust the feature importance weight matrix of the input data through the activation function according to the current bias.

[0031] Optionally, determining the probability distribution of the maceral types of each input image according to the output result includes:

[0032] Determining the probability distribution of the maceral types of each input image through the Softmax function according to the output result.

[0033] Optionally, after establishing the third sample set, it further includes:

[0034] Using the third sample set to train a fully connected neural network with the dynamic adaptive oscillation algorithm to obtain a feature extraction model; correspondingly,

[0035] Training the fractional-order neural network based on the third sample set includes:

[0036] Obtaining a fourth sample set from the output data of the feature extraction model and training the fractional-order neural network.

[0037] Optionally, using the third sample set to train a fully connected neural network with the dynamic adaptive oscillation algorithm includes:

[0038] Inputting the third sample set into the fully connected neural network, and initializing the parameters to be optimized of the fully connected neural network and the initial phase, amplitude, and frequency of the parameter oscillation, where the parameters to be optimized include weight parameters and biases;

[0039] Determining the loss function based on the output result, and updating the frequency, phase, and amplitude of the parameter oscillation according to the loss function, the current weight parameters, and the historical weight parameters;

[0040] Updating each parameter to be optimized according to the frequency, phase, and amplitude of the parameter oscillation;

[0041] If the iteration termination condition is not satisfied, return to execute determining the loss function based on the output result.

[0042] Optionally, updating each parameter to be optimized according to the frequency, phase, and amplitude of the parameter oscillation includes:

[0043] Summing the frequency and phase of the parameter oscillation and then taking the cosine value, and determining the product of the cosine value and the amplitude of the oscillation as the update amount of the weight parameter to update the weight parameter;

[0044] Sum the frequency and phase of the parametric oscillation, then calculate the sine value, determine the product of the sine value and the oscillation amplitude as the update amount of the bias, and update the bias.

[0045] Optionally, the coal maceral vitrinite microscopic image includes the microscopic images of structured vitrinite, unstructured vitrinite, and vitrinite debris. The unstructured vitrinite microscopic image includes the microscopic images of homogeneous vitrinite, matrix vitrinite, and massive vitrinite.

[0046] The coal maceral inertinite microscopic image includes the microscopic images of fusinite, semifusinite, funginite, secretinite, macrinite, micrinite, and inertinite debris.

[0047] In a second aspect, an embodiment of the present invention provides a method for establishing a coal maceral classification model based on computer vision, including:

[0048] Determine the maceral type of the coal maceral microscopic image based on the classification model, and the classification model is established through the following steps:

[0049] Use the first sample set to train the generator and discriminator of the generative adversarial network to obtain an image generation model. The generator is a multi-scale convolutional neural network, and the discriminator classifies the image through an adaptive sparse convolutional matrix.

[0050] Obtain a second sample set, use the image generation model for image augmentation, and establish a third sample set. Both the first sample set and the second sample set include coal maceral vitrinite microscopic images and inertinite microscopic images.

[0051] Train a fractional-order neural network based on the third sample set to obtain a classification model.

[0052] In a third aspect, an embodiment of the present invention provides a coal maceral classification model based on computer vision, which is used to determine the maceral type based on the input coal maceral microscopic image. The classification model is established through the following steps:

[0053] Use the first sample set to train the generator and discriminator of the generative adversarial network to obtain an image generation model. The generator is a multi-scale convolutional neural network, and the discriminator classifies the image through an adaptive sparse convolutional matrix.

[0054] Obtain a second sample set, use the image generation model for image augmentation, and establish a third sample set. Both the first sample set and the second sample set include coal maceral vitrinite microscopic images and inertinite microscopic images.

[0055] Train a fractional-order neural network based on the third sample set to obtain a classification model.

[0056] Fourthly, an embodiment of the present invention provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, any of the above methods is implemented.

[0057] Fifthly, an embodiment of the present disclosure provides a server, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, any of the above methods is implemented.

[0058] The beneficial effects of the above technical solutions provided by the embodiments of the present invention at least include:

[0059] (1) The computer vision-based coal and rock maceral classification method provided by the embodiment of the present invention, through an automated classification method, greatly reduces the time and errors required for manual analysis, improves the work efficiency of deep coalbed methane reservoir exploration, and provides a fast and accurate technical means for reservoir classification and identification.

[0060] (2) Different from the traditional single-scale image generation method, the computer vision-based coal and rock maceral classification method provided by the embodiment of the present invention uses a generative adversarial network based on multi-scale coupling for image data augmentation. When generating images through a multi-scale convolutional neural network, it learns features of different scales layer by layer, and gradually generates finer and finer image features from low-level to high-level to adapt to image details of different scales; in addition, the sparse convolutional matrix of the discriminator effectively reduces redundant calculations when processing large-scale microscopic images, has higher computational efficiency, and at the same time avoids interference from too much irrelevant information, enhancing the network's sensitivity to changes in image details.

[0061] (3) The computer vision-based coal and rock maceral classification method provided by the embodiment of the present invention uses a fractional-order neural network and a dynamic recalibration layer for classification. The dynamic recalibration layer adjusts the importance weights of features according to the current state and context of the data, thereby improving the classification accuracy, being able to better handle the non-linear relationship of the data, and enhancing the fitting ability of the model. And the dynamic recalibration layer can adjust its weight in real time according to the state of each feature, further optimizing the model's attention to different features, thereby improving the classification accuracy.

[0062] (4) The computer vision-based coal and rock maceral classification method provided by the embodiment of the present invention uses a dynamic adaptive oscillation algorithm to train a fully connected neural network for feature extraction, introduces a dynamic adaptive oscillation mechanism, simulates the non-linear oscillation behavior in physical phenomena, and improves the gradient problems (such as gradient disappearance, explosion, and local optimal solution problems) in traditional neural network training, significantly improving the performance of the model in complex data.

[0063] (5)The coal maceral classification method based on computer vision provided by the embodiments of the present invention realizes the automatic recognition of 12 subclasses in vitrinite and inertinite of coal maceral microscopic images for the first time.

[0064] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the written specification, claims, and drawings.

[0065] The technical solutions of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0066] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0067] Figure 1 is the flowchart of the coal maceral classification method based on computer vision in Embodiment 1 of the present invention;

[0068] Figure 2 is Figure 1 the specific implementation flowchart of step S11 in

[0069] Figure 3 is Figure 1 the specific implementation flowchart of step S13 in

[0070] Figure 4 is the training flowchart of the fully connected neural network in Embodiment 1 of the present invention;

[0071] Figures 5A - 5L is the image feature map of typical vitrinite and inertinite under plane-polarized light in Embodiment 2 of the present invention. Detailed Embodiments

[0072] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0073] It should be understood that the terms used in the present invention are only for describing specific embodiments and are not intended to limit the present invention. In addition, for the numerical ranges in the present invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Each intermediate value within any stated value or stated range, as well as each smaller range between any other stated value or intermediate value within the stated range, is also included in the present invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0074] Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention pertains. Although the present invention only describes preferred methods and materials, any methods and materials similar or equivalent to those described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods and / or materials related to the documents. In case of conflict with any incorporated document, the content of this specification shall prevail.

[0075] Example 1

[0076] Example 1 of the present invention provides a method for establishing a coal-rock microscopic component classification model based on computer vision, and its process is as follows Figure 1 shown, including the following steps:

[0077] Step S11: Use the first sample set to train the generator and discriminator of the generative adversarial network to obtain an image generation model.

[0078] Among them, the generator is a multi-scale convolutional neural network, and the discriminator classifies images through an adaptive sparse convolutional matrix. Higher-quality training data is generated through the confrontation between the generator and the discriminator, thereby effectively expanding the training sample set.

[0079] Assume that the pixels of the input coal-rock microscopic image are 8192×8192×3. The goal of the generator is to generate a high-resolution image (8192×8192×3) from a low-dimensional random noise. The network structure of the generator is:

[0080] Input layer: Input a random noise vector, and assume the dimension of the noise is 100 dimensions;

[0081] First layer: Transposed convolutional layer, with an output image size of 256×256×512;

[0082] Second layer: Transposed convolutional layer, with an output image size of 512×512×256;

[0083] Third layer: Transposed convolutional layer, with an output image size of 1024×1024×128;

[0084] The fourth layer: Transposed convolutional layer, with the output image size of 2048×2048×64;

[0085] The fifth layer: Transposed convolutional layer, with the output image size of 4096×4096×32;

[0086] The sixth layer: Transposed convolutional layer, with the output image size of 8192×8192×3.

[0087] The task of the discriminator is to distinguish whether the input image comes from real samples. The input image is a high-resolution image of 8192×8192×3. The network structure of the discriminator is as follows:

[0088] The first layer: Convolutional layer, with the input image size of 8192×8192×3, using a 4×4 convolutional kernel, a stride of 2, and the output of 4096×4096×64;

[0089] The second layer: Convolutional layer, with the output size of 2048×2048×128;

[0090] The third layer: Convolutional layer, with the output size of 1024×1024×256;

[0091] The fourth layer: Convolutional layer, with the output size of 512×512×512;

[0092] The fifth layer: Convolutional layer, with the output size of 256×256×1024;

[0093] The sixth layer: Convolutional layer, with the output size of 128×128×2048;

[0094] The seventh layer: Average pooling layer, with the output of 1×1×2048;

[0095] The output layer: The final output of the discriminator is a scalar value, which is the authenticity score of the output image (0 represents the generated image, and 1 represents the real image).

[0096] Specifically, see Figure 2 As shown, the training process of the generative adversarial network algorithm based on multi-scale coupling is as follows:

[0097] Step S111: Initialize the weight parameters of the generator and discriminator of the generative adversarial network respectively.

[0098] The way to initialize the weight parameters of the generator and discriminator is expressed as:

[0099]

[0100]

[0101] In the formula, represents the mean of 0 and the variance of of the normal distribution, indicates being subject to a specific distribution; represents the weights of the generator; represents the weights of the discriminator; represents the variance for initializing the parameters of the generative adversarial network. Preferably, is set to 0.01.

[0102] Step S112: Input a set of random noise vectors into the generator, and input the set of images generated by the generator through multi-layer convolutional processing from low frequency to high frequency and the first sample set into the discriminator, and obtain the authenticity discrimination result of the input image through adaptive sparse convolutional matrix processing.

[0103] The starting input for generator training is a set of random noise vectors generated by a uniform distribution or a Gaussian distribution. The input noise vectors are processed through the multi-scale convolutional layer in the generator, and images are gradually generated from low frequency to high frequency. The resolution and details of the generated images are gradually improved as the network depth increases, so as to be able to generate high-quality microscopic images. The generation process is expressed as:

[0104]

[0105] In the formula, is the input noise vector, and assume the dimension of the input noise vector is ; is the generator function, which maps the noise to the image space based on the weights ; is the generated image.

[0106] Specifically, in the process of generating images, the generator uses multi-scale coupled convolutional layers to gradually extract feature information of different scales. Each convolutional layer will gradually generate more refined image features from low frequency to high frequency, ensuring that the generated images can capture important local and global features at different scales. The convolution operation can be expressed as:

[0107]

[0108] In the formula, is the feature map of the th layer; is the feature map of the th layer; is the convolution operation, is the convolution kernel matrix of the th layer of the generator.

[0109] Furthermore, the convolution operation in the generator is actually a weighted fusion of multiple scales of the image. The convolution process of each layer is expressed as:

[0110]

[0111] In the formula, represents a convolution operation; is the feature map of the th layer; is the feature map of the th layer; is the weighting coefficient for each scale of the th layer, representing the weighting of image features at different scales, obtained through adaptive learning; is the number of layers of the convolutional layer.

[0112] The image generated by the generator and the real image in the first sample set are input into the discriminator together. The input image is processed by the adaptive sparse convolution matrix of the discriminator. The sparse convolution matrix can effectively improve the discriminator's perception ability of the details of microscopic images, while reducing redundant calculations and enhancing the discriminator's computational efficiency and accuracy. The output of the discriminator is expressed as:

[0113]

[0114] In the formula, is the discriminator function; is the output of the discriminator, representing the authenticity score of the discriminator for the generated image ; is the Sigmoid activation function, compressing the output to the interval; is the sparse convolution operation of the th layer; is the adaptive convolution kernel weight of the discriminator; is the input image of the discriminator.

[0115] Furthermore, the sparse convolution operation of the discriminator reduces redundant calculations and improves the accuracy of feature learning by adaptively optimizing the parameters of each convolutional layer. The sparse convolution calculates by only focusing on the important regions in the image, enhancing the perception of details and structural information. Especially when processing microscopic images, the learning of the sparse convolution matrix can effectively reduce the computational amount and improve the accuracy of feature extraction. The specific implementation method of the sparse convolution is expressed as:

[0116]

[0117] In the formula, is the th convolutional kernel weight matrix of the th layer of the discriminator; is the pixel value of the input image at the th position; is the adaptive sparse coefficient.

[0118] Furthermore, the adaptive sparse coefficient controls the sparsity of the convolutional matrix, dynamically adjusts the sparsity of the convolutional matrix according to the features of the input image, thereby extracting important local features in the discriminator and avoiding redundant calculations. The calculation method is expressed as:

[0119]

[0120] In the formula, is the sparsity adjustment parameter, which controls the degree of sparsity (a constant, for example, set to 2); is the bias term at this position, which is used to fine-tune each coefficient of the sparse convolutional matrix.

[0121] Step S113: Determine the generator loss function by maximizing the misclassification probability of the discriminator, and determine the discriminator loss function by maximizing the sum of the correct classification of the real image and the misclassification of the generated image.

[0122] The goal of the generator is to maximize the misclassification probability of the discriminator, while the goal of the discriminator is to maximize the correct classification probability of the real image. The calculation method of the generator loss function is expressed as:

[0123]

[0124] In the formula, is the generator loss function; represents the expectation; represents being subject to a specific distribution; is the probability distribution of the input noise; is the output of the discriminator for the generated image, representing the authenticity of the generated image; is the coefficient used to control the gradient penalty term, which is used to avoid the problem of gradient disappearance; is the number of convolutional layers, is the gradient penalty term, which ensures that the features of the generated image conform to physical constraints; the gradient of the generated data; represents the L2 norm.

[0125] Moreover, the discriminator loss function is calculated by maximizing the sum of the correct classification of the real image and the misclassification of the generated image, and the calculation method is expressed as:

[0126]

[0127] In the formula, is the discriminator loss function; is the distribution of the real image; is the output of the discriminator for the real image; is the coefficient used to control the sparse convolutional penalty term; Denote the norm of the weight matrix of the discriminator at the layer, which is used to control the sparsity of the convolutional kernel of the discriminator; is the L1 norm.

[0128] Step S114: Update the weight parameters of the generator based on the gradient of the generator loss function, and update the weight parameters of the discriminator based on the gradient of the discriminator loss function.

[0129] Based on the gradients calculated by backpropagation, the weights of the generator and the discriminator are optimized respectively, so that the network weights gradually converge after multiple rounds of training. The generator can generate more realistic images, and the discriminator can more accurately distinguish between real images and generated images. The parameter update method of the generator is:

[0130]

[0131] Moreover, the parameter update method of the discriminator is:

[0132]

[0133] In the formula, is the learning rate of the generative adversarial network; is the parameter update operation; and are the gradients of the generator and discriminator loss functions respectively; represents the gradient of the generator weight parameter; represents the gradient of the discriminator weight parameter; and are the regularization coefficients of the generator and discriminator respectively.

[0134] Step S115: Determine whether the iteration termination condition is satisfied.

[0135] The preset iteration termination condition can be reaching the preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000 times.

[0136] If the judgment in step S115 is no, return to step S112; if the judgment in step S115 is yes, end the training of the generative adversarial network, and use the trained generator as the image generation model.

[0137] Step S12: Obtain the second sample set, use the image generation model for image augmentation, and establish the third sample set.

[0138] Both the above-mentioned first sample set and the second sample set contain vitrinite micrographs and inertinite micrographs of coal and rock, which will be introduced in detail later.

[0139] After the image generation model is trained, the trained image generation model is used to expand the number of samples.

[0140] In one embodiment, assume that there are 800 samples in the originally collected second sample set. After image expansion using the image generation model, 200 samples are obtained. Then, the expanded third data set contains 1000 samples.

[0141] Step S13: Train a fractional-order neural network based on the third sample set to obtain a classification model for determining the microscopic component type based on the input coal-rock microscopic image.

[0142] The fractional-order neural network includes an input layer, a feature extraction layer, multiple dynamic recalibration layers, and an output layer. The dynamic recalibration layer dynamically adjusts the importance weight of each feature according to the current state and context of the input features, thereby improving the classification accuracy.

[0143] In one embodiment, it can also be that the fractional-order neural network does not include a feature extraction layer, but a feature extraction model specifically trained for image feature extraction. The image features extracted by the feature extraction model are input into the fractional-order neural network for training of the classification model. The training of the feature extraction model will be introduced in detail later.

[0144] See Figure 3 As shown, the training process of the fractional-order neural network is as follows:

[0145] Step S131: Input the third sample set into the fractional-order neural network and initialize the weight parameters and biases of the fractional-order neural network.

[0146] Let be the initial weight of the fractional-order neural network, be the initial bias of the fractional-order neural network, and the initialization method is random initialization.

[0147] Step S132: The dynamic recalibration layer of the fractional-order neural network adjusts the feature importance weight matrix of the input data according to the current bias through the activation function, and combines the current network weight parameters and biases to obtain the output data through the activation function.

[0148] The data input into the fractional-order neural network is subjected to feature extraction by the feature extraction layer. The extracted features are processed through multiple dynamic recalibration layers. Each dynamic recalibration layer performs a specific transformation on the input data and uses a non-linear activation function to enhance the expression ability of the model. The forward propagation process is expressed as:

[0149]

[0150] In the formula, is the output after activation by the fractional-order neural network layer; is the ReLU activation function; is the weight of the fractional-order neural network; is the bias of the fractional-order neural network; is the data input to the fractional-order neural network; denotes element-wise multiplication, that is, applying the re-scaled weights to the original features to adjust the contribution of each feature; is the re-scaled feature.

[0151] Furthermore, to enhance the model's adaptability to the characteristics of dynamic data, a dynamic feature re-scaling network layer is utilized, aiming to adjust the importance weights of each feature in real time in response to changes in the data stream. The calculation method of feature re-scaling is expressed as:

[0152]

[0153] In the formula, is the learnable re-scaling weight matrix; is the bias vector; is S activation function, which is used to adjust the weights to the interval (0, 1) to represent the importance of each feature.

[0154] Step S133: Determine the probability distribution of the microscopic component types of each input image according to the output result, and combine the microscopic component type labels and the current weight matrix to determine the loss function of the fractional-order neural network.

[0155] The classification probability predicted by the model is calculated by the Softmax function, which is used to convert the linear output into probabilities in multi-class classification problems. The calculation method is expressed as:

[0156]

[0157] In the formula, is the probability distribution predicted by the model, is the element corresponding to the th output of the network output layer, is the total number of classes; Softmax ensures that the sum of all outputs is 1 and each output is non-negative, which is suitable for interpretation as probabilities.

[0158] The difference between the model output of the fractional-order neural network and the true label is quantified by the composite loss function, and the regularization term is used to control the model complexity to prevent overfitting. The calculation method of the loss function is expressed as:

[0159]

[0160] In the formula, is the loss function of the fractional-order neural network, is the one-hot encoded form of the true label; is the regularization coefficient, used to control the influence of L1 regularization and thus affect the sparsity of the weights; is the number of elements in the weight matrix in the fractional-order neural network; represents the th element in the weight matrix. L1 regularization promotes the sparsity of the model weights and helps improve the generalization ability of the model.

[0161] Step S134: Update the network weight parameters of the fractional-order neural network based on the gradient of the fractional-order neural network loss function with respect to the network weight parameters, and update the bias of the fractional-order neural network based on the gradient of the fractional-order neural network loss function with respect to the bias.

[0162] According to the gradient calculated from the loss function, adjust the parameters of the model through the backpropagation algorithm until the value of the loss function converges or reaches the predetermined number of iterations. The parameter update method of the fractional-order neural network is expressed as:

[0163]

[0164]

[0165] In the formula, is the learning rate of the fractional-order neural network, which controls the learning step size; and are the gradients of the loss function with respect to the weights and biases; is the weight of the fractional-order neural network at the th iteration; is the bias of the fractional-order neural network at the th iteration; is the weight of the fractional-order neural network at the th iteration; is the bias of the fractional-order neural network at the th iteration.

[0166] Step S135: Determine whether the iteration termination condition is satisfied.

[0167] The preset iteration termination condition can be reaching the preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000 times.

[0168] If the judgment in step S135 is no, return to step S132; if the judgment in step S135 is yes, end the training of the fractional-order neural network and obtain a classification model for determining the microscopic component type based on the input coal-rock microscopic image.

[0169] The coal-rock microscopic component classification method based on computer vision provided in the first embodiment of the present invention, through an automated classification method, greatly reduces the time and error required for manual analysis, improves the work efficiency of deep coal-rock gas reservoir exploration, and provides a fast and accurate technical means for reservoir classification and identification.

[0170] Different from the traditional single-scale image generation method, the coal-rock microscopic component classification method based on computer vision provided in the first embodiment of the present invention uses a generative adversarial network based on multi-scale coupling for image data augmentation. When generating images through a multi-scale convolutional neural network, it gradually learns features of different scales layer by layer, generating increasingly fine image features from low-level to high-level to adapt to image details of different scales. In addition, the sparse convolutional matrix of the discriminator effectively reduces redundant calculations when processing large-scale microscopic images, has higher computational efficiency, and at the same time avoids interference from too much irrelevant information, enhancing the network's sensitivity to image detail changes.

[0171] The coal-rock microscopic component classification method based on computer vision provided in the first embodiment of the present invention uses a fractional-order neural network and a dynamic recalibration layer for classification. The dynamic recalibration layer adjusts the importance weights of features according to the current state and context of the data, thereby improving the classification accuracy, being able to better handle the non-linear relationship of the data, and enhancing the fitting ability of the model. And the dynamic recalibration layer can adjust its weights in real time according to the state of each feature, further optimizing the model's attention to different features, thereby improving the classification accuracy.

[0172] In one embodiment, the fractional-order neural network does not include a feature extraction layer, but uses the third sample set obtained after image augmentation and trains a fully connected neural network using a dynamic adaptive oscillation algorithm to obtain a feature extraction model; the image features extracted by the feature extraction model are used to obtain a fourth sample set, which is input into the fractional-order neural network for training of the classification model.

[0173] Specifically, the number of nodes in each layer of the fully connected layer of the fully connected neural network used in this embodiment is 128, that is, the s layer contains 128 neurons.

[0174] See Figure 4 As shown, the training process of the fully connected neural network is as follows:

[0175] Step S41: Input the third sample set into the fully connected neural network, and initialize the parameters to be optimized of the fully connected neural network and the initial phase, amplitude, and frequency of parameter oscillation.

[0176] The parameters to be optimized include weight parameters and biases.

[0177] In one embodiment, the initialization method is expressed as:

[0178]

[0179]

[0180]

[0181]

[0182] wherein, represents being subject to a specific distribution; represents a normal distribution with a mean of 0 and a variance of ; represents a normal distribution; represents the initialization variance of the neural network; is the initial weight of the neural network, is the initial bias of the neural network; is the initial phase of the oscillation; is the initial amplitude of the oscillation. Preferably, is set to 0.01.

[0183] Step S42: Determine the loss function based on the output result, and update the frequency, phase, and amplitude of the parameter oscillation according to the loss function, the current weight parameter, and the historical weight parameter.

[0184] Calculate the oscillation frequency of each parameter based on the current loss function. The adjustment of the oscillation frequency depends on the local curvature estimation of the loss surface and is adjusted using historical gradient information to make the parameter update smoother and enable more effective search in the parameter space. The adjustment method is expressed as:

[0185]

[0186] wherein, is the historical gradient weight factor, which controls the influence of the historical gradient in the current frequency adjustment; is the oscillation frequency, is the th iteration's oscillation frequency; is the th iteration's neural network weight; is the th iteration's neural network weight; is the neural network's loss function; is the partial derivative symbol; is the base oscillation frequency; is the hyperparameter for adjusting the oscillation responsiveness. Preferably, is set to 5, is set to 0.3; the neural network's loss function uses cross-entropy loss and is calculated from the output of the last layer of the neural network through a preset Softmax.

[0187] The update of each parameter is affected by the phase difference at the same time. The phase difference is determined by the success rate of the previous update and the interaction strength between parameters. The update of the phase depends on the effect of the previous parameter update, which is used to simulate the delay effect of the causal relationship. The update method is expressed as:

[0188]

[0189] In the formula, is the phase of the oscillation, is the phase of the oscillation at the th iteration; is the phase of the oscillation at the th iteration; is the phase adjustment factor; is the hyperbolic tangent function, which can limit the adjustment range of the phase and avoid the instability of the algorithm caused by excessive adjustment; is the hyperparameter for adjusting the phase sensitivity. Preferably, is set to 0.1, is set to 0.3.

[0190] According to the effect of parameter updates in the past few iterations, automatically adjust the amplitude size. If the update of a certain parameter continuously leads to a reduction in loss, increase its amplitude; otherwise, decrease the amplitude. The adaptive adjustment method of the amplitude is expressed as:

[0191]

[0192] In the formula, is the amplitude of the oscillation, is the amplitude of the oscillation at the th iteration; is the amplitude of the oscillation at the th iteration; is the amplitude adjustment coefficient. Preferably, is set to 0.95.

[0193] Step S43: Update each parameter to be optimized according to the frequency, phase, and amplitude of the parameter oscillation.

[0194] Combined with the oscillation behavior for parameter update, use the sine function and cosine function to simulate the oscillation behavior, allowing the parameter to perform periodic exploration in the gradient direction of the loss function. The update method is expressed as:

[0195]

[0196]

[0197] In the formula, is the The weights of the neural network for the -th iteration; The biases of the neural network for the -th iteration; The weights of the neural network for the -th iteration; The biases of the neural network for the -th iteration;

[0198] Step S44: Determine whether the iteration termination condition is satisfied.

[0199] The preset iteration termination condition may be reaching the preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000 times.

[0200] If the determination in step S44 is no, return to step S42; if the determination in step S44 is yes, end the training of the fully connected neural network to obtain a feature extraction model for image feature extraction.

[0201] In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient explosion, or getting stuck in local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The method for classifying coal and rock microscopic components based on computer vision provided by the embodiments of the present invention uses a dynamic adaptive oscillation algorithm to train a fully connected neural network for feature extraction, introduces a dynamic adaptive oscillation mechanism, simulates the non-linear oscillation behavior in physical phenomena. The algorithm can effectively explore and utilize local extrema in the high-dimensional parameter space, improves the gradient problems (such as gradient vanishing, explosion, and local optimal solution problems) in traditional neural network training, and significantly improves the performance of the model in complex data.

[0202] The first sample set and the second sample set are introduced above, both of which include coal and rock vitrinite microscopic images and inertinite microscopic images. Further, the coal and rock vitrinite microscopic images include structured vitrinite, unstructured vitrinite, and vitrinite debris microscopic images, and the unstructured vitrinite microscopic images include homogeneous vitrinite, matrix vitrinite, and massive vitrinite microscopic images; the coal and rock inertinite microscopic images include fusinite, semifusinite, funginite, secretoinite, macrinite, micrinite, and inertinite debris microscopic images.

[0203] The method for classifying coal and rock microscopic components based on computer vision provided by the embodiments of the present invention realizes the automatic recognition of 12 subclasses in the vitrinite and inertinite of coal and rock microscopic images for the first time.

[0204] The following introduces the data acquisition and label annotation process:

[0205] The data collection sources in the embodiments of the present invention mainly include image data captured by microscopes. The data are from a variety of natural mineral specimens and laboratory-synthesized samples, and are captured by professional microscope equipment, including different types and morphological characteristics of vitrinite and inertinite. The specific collection methods include:

[0206] Microscope image collection: High-resolution optical microscopes and high-performance cameras are used for microscope image collection. The microscopic details of vitrinite and inertinite, such as boundary morphology, fracture state, and texture features, are captured through imaging in the plane-polarized light mode. During the sample preparation process, the surface is finely polished to ensure imaging quality, and representative areas are selected under the low-power microscope and then imaged at high magnification. In addition, by combining multiple imaging modes (plane-polarized light, reflected light, fluorescence) at the same position, a multi-modal contrast data set is formed to provide rich information dimensions for classification.

[0207] Image data format: The captured image data are stored in the lossless compressed TIFF or high-quality PNG format to ensure information integrity. At the same time, the resolution is set to 3000x3000 pixels or higher, with a magnification of 400 times or 1000 times to capture microscopic details. Single-channel grayscale images or multi-channel RGB images are selected according to needs to highlight optical reflection or fluorescence characteristics, and metadata (such as sample number, shooting conditions, and microscope parameters) are attached to each image for subsequent tracking and data organization;

[0208] Furthermore, the collected microscopic component images are labeled. The microscopic components are subdivided into the following categories (Table 1): By combining multiple imaging modes (plane-polarized light, reflected light, fluorescence) at the same position, a multi-modal contrast data set is formed.

[0209] Table 1. Summary of the differences in microscopic characteristics of vitrinite and inertinite

[0210]

[0211] Vitrinite includes minerals with a vitreous structure and no crystals. The typical characteristics of vitrinite samples are relatively high transparency, strong luster, and neat edges. It mainly includes the following subcategories:

[0212] (1) Telinite: Characterized by obvious plant tissue structures, usually in strip or tubular forms, retaining the original forms such as plant fibers and ducts. Under the microscope, the structure is clear and arranged in an orderly manner, showing high luster and high reflectivity in reflected light, which is a key indicator for identifying the plant origin (Figure 5A);

[0213] (2) Homogeneous vitrinite: Homogeneous vitrinite has no obvious plant tissue residues and appears as a dense block or evenly distributed form. It is optically isotropic under a microscope, has a uniform luster and low reflectivity under reflected light, and is a common type of vitrinite in coal ( Figure 5B ).

[0214] (3) Matrix vitrinite: Matrix vitrinite is characterized by irregular diffuse distribution and mixed existence with other vitrinites. Under a microscope, the boundary is fuzzy, the optical anisotropy is weak, and it shows medium luster and medium reflectivity under reflected light, and microcracks are relatively developed (Figure 5C).

[0215] (4) Agglomerated vitrinite: Agglomerated vitrinite usually presents an irregular block shape, with possible cracks or inclusions inside and unclear boundaries. Under a microscope, the luster is average under reflected light, the reflectivity is lower than that of homogeneous vitrinite, and it has relatively low optical heterogeneity (Figure 5D).

[0216] (5) Mirror crumbs: Mirror crumbs are characterized by tiny debris-like shapes of varying sizes and clear boundaries. Under a microscope, they have strong optical anisotropy, with obvious changes in light and dark as they rotate, and they exhibit high gloss and reflectivity under reflected light ( Figure 5E ).

[0217] The inertinite group includes minerals with a distinct crystal structure, which are typically characterized by the presence of fine crystals and structures on the surface, and a strong reflective luster. It mainly includes the following categories:

[0218] (1) Silky bodies: Silky bodies are fibrous or strip-like, with consistent arrangement and compact structure. They have significant optical anisotropy under a microscope, strong luster under reflected light, and high reflectivity. They are typical components of high-metamorphic coal ( Figure 5F ).

[0219] (2) Semi-silk bodies: Semi-silk bodies are strip-shaped or fibrous, retaining some plant structural characteristics, with a slightly rough surface. The optical anisotropy is weak under the microscope, the luster is weak under reflected light, and the fluorescence characteristics are not obvious, but there is occasional dark yellow light (Figure 5G).

[0220] (3) Fungal bodies: Fungal bodies are characterized by spherical, elliptical or branched shapes, with a relatively loose structure, and are commonly found in low-grade metamorphic coal. Under a microscope, the luster is weak under reflected light, and it appears bright yellow or orange-yellow under fluorescence excitation, which is an important indicator of biological origin ( Figure 5D ).

[0221] (4) Secretory bodies: Secretory bodies are irregular blocks with clear boundaries and are mostly colloid or transparent. They are optically isotropic under a microscope, have a weak luster under reflected light, and usually appear bright yellow or orange-yellow under fluorescence excitation, making them highly recognizable (Figure 5H).

[0222] (5) Macrograins: Macrograins are characterized by relatively large granular or massive forms, with rough surfaces and blurred boundaries. Under the microscope, they have weak optical anisotropy, average luster under reflected light, and mostly exhibit weak yellowish-green light under fluorescence excitation ( Figure 5I -J).

[0223] (6) Microparticles: Microparticles appear as extremely small granular or fine massive forms, with uniform particle distribution. Under the microscope, they have poor light transmittance, weak luster and low reflectivity under reflected light, and usually have weak or no response to fluorescence characteristics ( Figure 5K ).

[0224] (7) Inert macerals: Inert macerals are in irregular clastic shapes, with variable forms and blurred boundaries. Under the microscope, they have strong optical anisotropy, obvious luster and relatively high reflectivity under reflected light, and weak fluorescence characteristics, and are common inertinite components ( Figure 5L ).

[0225] Example Two

[0226] Example Two of the present invention provides a specific application of a coal petrographic maceral classification method based on computer vision.

[0227] Taking the deep coalbed methane sample X of Well A1 as the research object, the sample is from the deep coalbed methane reservoir and contains various subclass characteristics of vitrinite and inertinite. The research goal is to automatically identify 13 subclasses of vitrinite and inertinite based on computer vision technology, including 5 subgroups of vitrinite (tectovitrinite, telovitrinite, desmocollinite, corpocollinite, and vitrinite clastic) and 7 subgroups of inertinite (fusinite, semifusinite, funginite, secretoinite, macrinite, micrinite, and inertinite clastic). High-resolution optical microscopy is used to collect sample images, including multi-modal data of plane-polarized light, reflected light, and fluorescence. The resolution of each image is 3000×3000 pixels and is stored in the lossless TIFF format. The images are manually annotated to mark the boundary morphology, fracture state, and texture features of vitrinite and inertinite, and the annotation accuracy is ensured through expert review. There are no less than 50 images for each subclass sample to ensure that the data volume meets the training requirements. A multi-scale coupled generative adversarial network (GAN) is used for sample augmentation, and high-quality samples are gradually generated through multi-scale convolutional layers. The final generated training set includes 530 vitrinite images and 725 inertinite images, significantly improving the problem of sample imbalance. Microscopic image features of the sample are extracted based on a neural network with dynamic adaptive oscillation, including geometric features, texture features, and gray-scale features. The model training adopts a hierarchical feature extraction method, and the network performance is optimized by dynamically adjusting the oscillation parameters, effectively solving the gradient problem in high-dimensional data training. A fractional-order neural network is used to classify the extracted features, and the feature weights are adjusted in real time by combining a dynamic recalibration layer to improve the classification accuracy. The output layer of the model uses the Softmax function to classify the images into 12 subclasses of vitrinite and inertinite. The trained model is applied to the sample images of Well A1 to identify the subclasses of vitrinite and inertinite in each image. By comparing with the manually annotated results and calculating the difference between the automatically identified component content and the manually identified component content, the classification accuracy of the model is verified, and the overall error is less than 0.6%.

[0228] Table 2 Comparison of automatically identified content and manually identified content of microscopic components of vitrinite and inertinite in coal rock sample X of Well A1

[0229]

[0230] Based on the plane-polarized light, reflected light, and fluorescence characteristics of microscopic components under an optical microscope, the embodiments of the present invention comprehensively utilize various technical means such as data acquisition and annotation, data augmentation, feature extraction, and deep learning classifiers. By introducing a generative adversarial network to augment the training data, the problems of insufficient data and poor model generalization ability are solved; an optimization method for neural network parameters based on ecosystem optimization is adopted to improve the training efficiency and accuracy of the model; at the same time, combined with a fractional-order neural network classifier, the processing ability and classification accuracy of the model for complex geological data are further improved.

[0231] Based on the inventive concept of the present invention, an embodiment of the present invention further provides a method for establishing a coal petrographic maceral classification model based on computer vision, including:

[0232] Determining the maceral type of the coal petrographic microscopic image based on the classification model, and the classification model is established through the following steps:

[0233] Training the generator and discriminator of the generative adversarial network using the first sample set to obtain an image generation model, where the generator is a multi-scale convolutional neural network, and the discriminator classifies the image through an adaptive sparse convolutional matrix;

[0234] Obtaining a second sample set, using the image generation model for image augmentation, and establishing a third sample set. Both the first sample set and the second sample set include coal vitrinite microscopic images and inertinite microscopic images;

[0235] Training a fractional-order neural network based on the third sample set to obtain a classification model.

[0236] Based on the inventive concept of the present invention, an embodiment of the present invention further provides a coal petrographic maceral classification model based on computer vision. The classification model is used to determine the maceral type based on the input coal petrographic microscopic image, and the classification model is established through the following steps:

[0237] Training the generator and discriminator of the generative adversarial network using the first sample set to obtain an image generation model, where the generator is a multi-scale convolutional neural network, and the discriminator classifies the image through an adaptive sparse convolutional matrix;

[0238] Obtaining a second sample set, using the image generation model for image augmentation, and establishing a third sample set. Both the first sample set and the second sample set include coal vitrinite microscopic images and inertinite microscopic images;

[0239] Training a fractional-order neural network based on the third sample set to obtain a classification model.

[0240] Regarding the establishment method of the above classification model, it has been described in detail in Embodiment 1, and will not be elaborated here.

[0241] Based on the inventive concept of the present invention, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, the above method is implemented.

[0242] Based on the inventive concept of the present invention, an embodiment of the present invention further provides a server, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above method is implemented.

[0243] Unless otherwise specifically stated, terms such as processing, computing, calculating, determining, displaying, etc. can refer to the actions and / or processes of one or more processing or computing systems, or similar devices, which operate on and transform data represented as a physical (such as electronic) quantity within the registers or memories of the processing system into other data similarly represented as a physical quantity within the memories, registers, or other such information storage, transmission, or display devices of the processing system. Information and signals can be represented using any of a variety of different technologies and methods. For example, the data, instructions, commands, information, signals, bits, symbols, and chips referred to throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0244] It should be understood that the specific order or hierarchy of steps in the disclosed processes are examples of exemplary methods. Based on design preferences, it should be understood that the specific order or hierarchy of steps in a process can be rearranged without departing from the scope of the present disclosure. The appended method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy recited.

[0245] In the above detailed description, various features are combined in a single embodiment to simplify the present disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are stated in each claim. On the contrary, as reflected by the appended claims, the invention lies in less than all of the features of the single disclosed embodiment. Accordingly, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0246] Those skilled in the art should also understand that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments herein can be implemented as electronic hardware, computer software, or a combination thereof. To clearly illustrate the interchangeability of hardware and software, the above description of the various illustrative components, blocks, modules, circuits, and steps has been generally described in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and the design constraints imposed on the overall system. Skilled artisans may implement the described functionality in a flexible manner for each particular application, but such implementation decisions should not be interpreted as departing from the scope of the present disclosure.

[0247] The steps of the methods or algorithms described in conjunction with the embodiments of this specification may be directly embodied as hardware, software modules executed by a processor, or a combination thereof. The software modules may be located in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Of course, the processor and the storage medium may also exist as discrete components in the user terminal.

[0248] For a software implementation, the techniques described in this application may be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes may be stored in a memory unit and executed by a processor. The memory unit may be implemented within the processor or outside the processor. In the latter case, it is communicatively coupled to the processor via various means, which are well known in the art.

[0249] The foregoing description includes examples of one or more embodiments. Of course, it is not possible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but those of ordinary skill in the art should recognize that the various embodiments can be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, this term is inclusive in a manner similar to the term "including" as interpreted when used as a transitional word in a claim. In addition, any use of the term "or" in the claims or the specification is to mean "non-exclusive or". The terms "first", "second", etc. are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

Claims

1. A classification method for macerals of coal based on computer vision, characterized in that, Including: Determining the maceral type of the coal-rock microscopic image based on a classification model, and the classification model is established through the following steps: Training the generator and discriminator of the generative adversarial network using a first sample set to obtain an image generation model, where the generator is a multi-scale convolutional neural network, and the discriminator classifies images through an adaptive sparse convolutional matrix; Obtaining a second sample set, using the image generation model for image augmentation to establish a third sample set, and both the first sample set and the second sample set include coal-rock vitrinite microscopic images and inertinite microscopic images; Training a fractional-order neural network based on the third sample set to obtain a classification model.

2. The method according to claim 1, wherein The training of the generator and discriminator of the generative adversarial network using the first sample set includes: Initializing the weight parameters of the generator and discriminator of the generative adversarial network respectively; Inputting a random noise vector set into the generator, inputting the image set generated by the generator through multi-layer convolutional processing from low frequency to high frequency and the first sample set into the discriminator, and obtaining the authenticity discrimination result of the input image through the processing of the adaptive sparse convolutional matrix; Determining the generator loss function by maximizing the misclassification probability of the discriminator, and determining the discriminator loss function by maximizing the sum of the correct classification of real images and the misclassification of generated images; Updating the weight parameters of the generator based on the gradient of the generator loss function, and updating the weight parameters of the discriminator based on the gradient of the discriminator loss function; If the iteration termination condition is not satisfied, return to execute the step of inputting the random noise vector set into the generator.

3. The method according to claim 1, characterized in that The fractional-order neural network includes an input layer, a feature extraction layer, multiple dynamic recalibration layers and an output layer, and the dynamic recalibration layer dynamically adjusts the importance weight of each feature according to the current state and context of the input features.

4. The method according to claim 3, characterized in that, The training of the fractional-order neural network based on the third sample set includes: Inputting the third sample set into the fractional-order neural network, and initializing the weight parameters and biases of the fractional-order neural network; The dynamic recalibration layer of the fractional-order neural network adjusts the feature importance weight matrix of the input data through an activation function according to the current bias, and combines the current network weight parameters and biases to obtain output data through the activation function; Determining the probability distribution of the maceral type of each input image according to the output result, and combining the maceral type label and the current weight matrix to determine the loss function of the fractional-order neural network; Updating the network weight parameters of the fractional-order neural network based on the gradient of the fractional-order neural network loss function with respect to the network weight parameters, and updating the bias of the fractional-order neural network based on the gradient of the fractional-order neural network loss function with respect to the bias; If the iteration termination condition is not satisfied, return to execute the step of the dynamic recalibration layer of the fractional-order neural network adjusting the feature importance weight matrix of the input data through an activation function according to the current bias.

5. The method according to claim 4, wherein The determining the probability distribution of the maceral type of each input image according to the output result includes: Determining the probability distribution of the maceral type of each input image through the Softmax function according to the output result.

6. The method according to claim 1, characterized in that, After establishing the third sample set, it further includes: Using the third sample set, train a fully connected neural network with a dynamic adaptive oscillation algorithm to obtain a feature extraction model; correspondingly, Training the fractional-order neural network based on the third sample set includes: Obtain a fourth sample set from the output data of the feature extraction model and train the fractional-order neural network.

7. The method according to claim 6, wherein The using the third sample set to train a fully connected neural network with a dynamic adaptive oscillation algorithm includes: Input the third sample set into the fully connected neural network, and initialize the parameters to be optimized of the fully connected neural network and the initial phase, amplitude, and frequency of parameter oscillation. The parameters to be optimized include weight parameters and biases; Determine the loss function based on the output result, and update the frequency, phase, and amplitude of parameter oscillation according to the loss function, the current weight parameters, and the historical weight parameters; Update each parameter to be optimized according to the frequency, phase, and amplitude of parameter oscillation; If the iteration termination condition is not satisfied, return to execute the step of determining the loss function based on the output result.

8. The method according to claim 7, wherein The updating each parameter to be optimized according to the frequency, phase, and amplitude of parameter oscillation includes: Sum the frequency and phase of parameter oscillation and then calculate the cosine value, and determine the product of the cosine value and the amplitude of oscillation as the update amount of the weight parameter to update the weight parameter; Sum the frequency and phase of parameter oscillation and then calculate the sine value, and determine the product of the sine value and the amplitude of oscillation as the update amount of the bias to update the bias.

9. The method according to any one of claims 1 to 8, characterized in that, Coal rock vitrinite microscopic images include structural vitrinite, non-structural vitrinite, and vitrinite debris microscopic images. The non-structural vitrinite microscopic images include homogeneous vitrinite, matrix vitrinite, and massive vitrinite microscopic images; Coal rock inertinite microscopic images include fusinite, semifusinite, fungal body, secretory body, macrinite, micrinite, and inertinite debris microscopic images.

10. A method for establishing a coal maceral classification model based on computer vision, characterized in that, Includes: Use the first sample set to train the generator and discriminator of the generative adversarial network to obtain an image generation model. The generator is a multi-scale convolutional neural network, and the discriminator classifies images through an adaptive sparse convolutional matrix; Obtain a second sample set, use the image generation model for image augmentation, and establish a third sample set. The first sample set and the second sample set both contain coal rock vitrinite microscopic images and inertinite microscopic images; Train a fractional-order neural network based on the third sample set to obtain a classification model for determining the microscopic component type based on the input coal rock microscopic image.

11. A coal maceral classification model based on computer vision, characterized in that, The classification model is used to determine the microscopic component type based on the input coal rock microscopic image. The classification model is established through the following steps: Use the first sample set to train the generator and discriminator of the generative adversarial network to obtain an image generation model. The generator is a multi-scale convolutional neural network, and the discriminator classifies images through an adaptive sparse convolutional matrix; Obtain a second sample set, use the image generation model for image augmentation, and establish a third sample set. The first sample set and the second sample set both contain coal rock vitrinite microscopic images and inertinite microscopic images; Train a fractional-order neural network based on the third sample set to obtain a classification model.

12. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

13. A server, characterized in that, Including: A memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the method according to any one of claims 1 to 10 is implemented.

Images