Mine dust concentration image recognition method based on prior features and multi-kernel residual attention network

Abstract

The invention discloses a mine dust concentration image recognition method based on prior features and a multi-kernel residual attention network, and the method comprises the steps: collecting mine dust images under different concentration conditions, extracting multi-dimensional image features, and constructing a prior feature library; linear and nonlinear feature subsets significantly related to the mine dust concentration are screened out; respectively mapping the linear and nonlinear feature subsets to a high-dimensional space by adopting a multi-kernel learning framework and carrying out weighted fusion to obtain fused high-order features; a multi-kernel residual attention network is constructed and trained, the gradient disappearance problem is relieved through a residual block structure, the weight of each feature channel is learned through a channel attention mechanism, and dynamic enhancement of key mine dust image features is carried out; and inputting the fused high-order features into a multi-kernel residual attention network for regression calculation, and outputting a mine dust concentration predicted value. According to the method, high-precision, non-contact and high-interpretability image recognition of mine dust concentration can be realized, and the method is suitable for real-time mine safety monitoring in a complex environment.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and mine safety monitoring technology, and in particular to a method for recognizing mine dust concentration images based on prior features and multi-kernel residual attention networks. Background Technology

[0002] Mine dust is one of the main hazardous substances generated during coal mining, seriously threatening the health of workers and posing significant safety risks such as dust explosions and fires. Therefore, real-time and accurate monitoring of mine dust concentration is crucial for ensuring safe coal mine production.

[0003] Existing methods for detecting mine dust concentration can be mainly divided into two categories: offline detection and online detection. Offline detection methods, represented by the filter membrane weighing method, offer high measurement accuracy, but their operation procedures are cumbersome, the detection cycle is long, and the results are severely delayed, failing to meet the urgent need for real-time monitoring and early warning in underground mines. Online detection methods, such as light scattering, beta-ray, and oscillating balance methods, can achieve continuous monitoring, but they generally suffer from problems such as complex equipment structures, high purchase and maintenance costs, poor adaptability to harsh underground environments, and limited detection range.

[0004] With the development of computer vision technology, image-based methods for detecting mineral dust concentration have become a research hotspot due to their advantages such as non-contact operation, low cost, and visualization. However, existing image-based detection methods, especially end-to-end deep learning models, have the following significant drawbacks: First, the model acts as a "black box," with its internal decision-making logic being opaque and severely lacking interpretability, resulting in low reliability of the detection results in practical engineering. Second, such models rely on massive amounts of data and complex network structures for extracting image features, ignoring the guiding role of prior physical features of mineral dust images. They are not adaptable enough to complex scenarios such as variable lighting and uneven dust distribution underground, resulting in weak generalization ability and low recognition accuracy. Summary of the Invention

[0005] The problem to be solved by this invention is to provide a method for image recognition of coal dust concentration based on prior features and multi-kernel residual attention network, which can be used for real-time, non-contact detection of coal dust concentration in underground mines. While ensuring high accuracy, it enhances the interpretability of the model, reduces the dependence on massive amounts of data, and enhances the adaptability to complex underground environments.

[0006] This invention adopts the following technical solution: a method for recognizing mineral dust concentration images based on prior features and a multi-kernel residual attention network, comprising the following steps:

[0007] S1. Collect images of mineral dust under different concentration conditions, and extract multi-dimensional image features from the mineral dust images to construct a priori feature library of mineral dust images;

[0008] S2. Based on the Pearson correlation coefficient and mutual information coefficient, select linear feature subsets and nonlinear feature subsets that are significantly related to the concentration of mineral dust from the prior feature library;

[0009] S3. Based on the principle of implicit high-dimensional mapping of kernel function and complementary fusion of multiple features, a multi-kernel learning framework is constructed. The linear feature subset and the nonlinear feature subset are mapped to a high-dimensional space respectively, and weighted fusion is performed to obtain the fused high-order features.

[0010] S4. Construct a multi-kernel residual attention network and train the model, including: a multi-kernel feature fusion module, a residual block structure, and a channel attention mechanism; the residual block structure consists of a fully connected layer and a ReLU activation function to alleviate the gradient vanishing problem; the channel attention mechanism learns the weights of each feature channel through global average pooling and a fully connected layer to dynamically enhance key mineral dust image features.

[0011] S5. Input the fused high-order features into the trained multi-core residual attention network for regression calculation, and output the predicted value of mineral dust concentration through the fully connected layer.

[0012] Preferably, in step S1, the multi-dimensional image features include: color features, texture features, geometric features, image quality features, grayscale statistical features, and transmittance features.

[0013] The color characteristics include: mode, uniformity, skewness, kurtosis, and diversity;

[0014] The texture features include: contrast, dissimilarity, energy, homogeneity, correlation, and entropy;

[0015] The geometric features include: total energy moment, axial length moment, corner feature moment, second corner feature moment, eccentricity moment, compactness moment, and asymmetry moment;

[0016] The image quality features include: peak signal-to-noise ratio, visual information fidelity, and modulation index;

[0017] The gray-level statistical features include: gray-level mean, gray-level standard deviation, gray-level skewness, and gray-level kurtosis;

[0018] The transmittance characteristics include: mean transmittance and median transmittance.

[0019] Preferably, in step S2, the Pearson correlation coefficient and mutual information coefficient between each type of feature and the dust concentration are calculated respectively. Based on the set threshold of the Pearson correlation coefficient and the significance of the mutual information coefficient, several linear and nonlinear features are selected to form a subset of linear and nonlinear features.

[0020] The linear feature subset includes at least: contrast, homogeneity, compactness moment, asymmetry moment, and modulation index;

[0021] The nonlinear feature subset includes at least: mode, uniformity, diversity, axial length moment, gray mean, and transmission median.

[0022] Preferably, in step S3, the multi-kernel learning framework is constructed based on the principle of implicit high-dimensional mapping of kernel functions and complementary fusion of multiple features. The dust image features are divided according to linear separability, and linear and nonlinear features are modeled separately, avoiding the information interference problem caused by the unified modeling of all features in traditional methods. A multi-kernel fusion mechanism is introduced, and multi-kernel learning is performed through linear weighted fusion. At the same time, the high-dimensional mapping of linear and nonlinear features is considered, so that the model takes into account both the overall grayscale trend and the differences in local particle structure.

[0023] Specifically, for the input linear and nonlinear feature subsets, a radial basis kernel function is used to perform high-dimensional mapping on the linear and nonlinear feature subsets respectively, resulting in the corresponding kernel matrix. and Multi-core learning is performed using a linear weighted fusion method, as follows:

[0024] ；

[0025] in, The fusion weight coefficients are adjustable, and .

[0026] Preferably, in step S4, the multi-core residual attention network is composed of multiple residual block structures stacked together. Each residual block structure contains two fully connected layers and a ReLU activation function, and introduces skip connections.

[0027] The output of the residual block structure is represented as follows:

[0028] ;

[0029] in, and These are the input and output of the residual block, respectively. Let {Wi} represent the residual function consisting of a fully connected layer and a ReLU activation function, where {Wi} are learnable parameters.

[0030] The jump connection is introduced within the residual block, specifically by: converting the input of the residual block... The residual function directly interacts with the fully connected layer and the ReLU activation function. The processed results are then element-wise summed, meaning the input features of the residual block are added together. Without going through the core computational process of fully connected layers, it is directly passed to the position in the residual block after the fully connected layer and ReLU activation function calculation. Add them together to get the output of the residual block. This helps to alleviate the gradient vanishing problem during deep network training and enhances the effectiveness of feature propagation.

[0031] Preferably, in step S4, the channel attention mechanism is specifically processed as follows:

[0032] Global average pooling is performed on the input features to obtain channel-level statistical information;

[0033] The statistical information is passed sequentially through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function to generate weights for each feature channel;

[0034] The generated weights are multiplied channel-by-channel with the original input features to dynamically enhance key mineral dust image features. Let the original input features be... The weights of each generated feature channel are W=[w1,w2,.....,w c ]∈R C Then output features Represented as:

[0035]

[0036] in, For feature map height, The width of the feature map. The number of feature map channels. , , , This represents the numerical values ​​of the original input features at height h, width w, and channel c. This indicates the numerical value of the output feature at the corresponding position. This represents the channel weight.

[0037] Preferably, in step S4, the multi-core feature fusion module performs feature fusion on the output features of the channel attention mechanism to obtain the fused mineral dust concentration image features, as follows:

[0038] First, establish associations between the features output by the channel attention mechanism and the kernel matrix features obtained by mapping linear features through the radial basis kernel function, and the kernel matrix features obtained by mapping nonlinear features through the radial basis kernel function.

[0039] Then, the fusion is carried out using a linear weighted method, and the fusion formula is:

[0040] ;

[0041] in, It is a feature resulting from the combination of the output features of the channel attention mechanism and the features of the linear feature kernel matrix. It is a feature resulting from the combination of the output features of the channel attention mechanism and the features of the nonlinear feature kernel matrix. The fusion weight coefficients are adjustable, and Ultimately, the fused multi-core features are obtained. .

[0042] Preferably, in step S4, the residual attention network is trained using the Adam optimizer, and the learning rate is set to... The mean squared error is used as the loss function, and an early stopping strategy is adopted to prevent overfitting.

[0043] Preferably, in step S5, the fused high-order features are input into the trained multi-kernel residual attention network for regression calculation, as follows:

[0044] After the fused high-order features are converted into tensor form, they are input into a multi-kernel residual attention network composed of multiple stacked residual blocks;

[0045] The network first performs a linear transformation and ReLU activation on the input features. Then, it extracts features step by step through a residual block structure (each residual block contains two fully connected layers and a ReLU activation function, and introduces skip connections so that the output is the sum of the input and the result of the residual function). At the same time, it uses a multi-kernel feature fusion module to learn the output feature weights of the channel attention mechanism and performs weighted fusion on the features to further optimize the feature representation.

[0046] Finally, the features are mapped to a single-value output through a linear layer. This output is the predicted value of the dust concentration, thus realizing the regression calculation from the fusion of high-order features to the dust concentration.

[0047] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0048] 1. High interpretability: The method of this invention introduces prior image features with clear physical meaning and filters them based on statistical correlation, so that the input and decision-making process of the model have clear physical basis, thus overcoming the defects of traditional "black box" models.

[0049] 2. High recognition accuracy: The method of this invention adopts a multi-core learning framework to effectively integrate the representation capabilities of linear and nonlinear features, and combines a residual attention network to enhance key features, which significantly improves the recognition accuracy and robustness of mineral dust concentration.

[0050] 3. Good generalization ability: The residual structure of this invention alleviates the gradient problem of deep networks, and the attention mechanism enables the model to adapt to complex and ever-changing downhole environments, enhancing the stability and adaptability of the model in practical applications. Attached Figure Description

[0051] Figure 1 The overall flowchart of the mineral dust concentration image recognition method of the present invention is shown below;

[0052] Figure 2 This is a schematic diagram of the categories of the prior feature library of mineral dust images constructed in this embodiment of the invention;

[0053] Figure 3 This is a LOSS diagram drawn in an embodiment of the present invention;

[0054] Figure 4 This is a diagram illustrating the recognition effect in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0056] In a preferred embodiment of the present invention, a method for recognizing mineral dust concentration images based on prior features and a multi-kernel residual attention network is provided, the overall process of which is as follows: Figure 1 As shown, it includes the following steps:

[0057] First, in a mine dust scene simulation experimental system built in the laboratory, images of mine dust at different concentrations were collected, and the actual concentration values ​​measured by the online dust detector were recorded simultaneously, constructing an image-concentration paired dataset containing 143 sets of samples.

[0058] 1. Construction and screening of prior feature library

[0059] The prior feature library constructed in this embodiment (as shown in Table 1) comprehensively covers six categories: color features (such as mode and uniformity), texture features (contrast and homogeneity extracted based on the gray-level co-occurrence matrix), geometric features (such as compactness moments and asymmetry moments of the Hu invariant moment series), image quality features (such as peak signal-to-noise ratio and modulation index), gray-level statistical features (such as gray-level mean and skewness), and transmittance features (such as transmittance mean and transmittance median).

[0060] Table 1 Image Feature Library of Mine Dust Concentration

[0061]

[0062] Subsequently, the Pearson correlation coefficient (PCC) and mutual information (MI) between each feature and the dust concentration were calculated.

[0063] Based on a set threshold (e.g., |PCC|>0.7) and the significance of the MI value, five linear features (contrast, homogeneity, compactness moment, asymmetry moment, modulation index) and six nonlinear features (mode, uniformity, diversity, axial length moment, gray mean, transmission median) are selected to form a subset of linear and nonlinear features for subsequent model input, such as... Figure 2 As shown.

[0064] 2. Multi-core feature fusion

[0065] The two feature subsets, linear and nonlinear, are input into the multi-kernel learning framework respectively.

[0066] In this embodiment, the input linear feature subset (dimension 143×5) and nonlinear feature subset (dimension 143×11) are mapped to high dimensions using radial basis kernel functions, respectively. Different kernel bandwidth parameters are set to accommodate different scales of linear and nonlinear relationships. The optimal fusion weight α is determined through grid search, and weighted fusion is performed according to the following formula to obtain a fusion kernel matrix K that has both interpretability and strong representational capabilities:

[0067]

[0068] 3. Residual Attention Network Regression Recognition

[0069] First, the fused high-order features are input into a residual attention network, which consists of multiple stacked residual blocks. Each residual block contains two fully connected layers and a ReLU activation function, and skip connections are introduced to effectively alleviate the vanishing gradient problem.

[0070] Then, the fused high-order features of dimension n×n are input into the residual attention network, where n is the number of samples. Here, the number of samples is 143, that is, the dimension is 143×143.

[0071] This residual attention network consists of multiple stacked residual blocks. For the input high-order features, the first fully connected layer (nn.Linear(n, 64)) transforms the feature dimension from 143×143 to 143×64, followed by a non-linear transformation using the ReLU activation function (nn.ReLU()). Next, the residual block structure is introduced. Each residual block contains two fully connected layers (both nn.Linear(64, 64)) and a ReLU activation function, while also introducing skip connections: the residual block first performs two fully connected and ReLU activation operations on the input 143×64 features to obtain the processed features. Then, the processed features are element-wise added to the original input features of the residual block, and then ReLU activation is applied again to obtain the processed 143×64 features, effectively mitigating the gradient vanishing problem.

[0072] Subsequently, the embedded channel attention mechanism processes the 143×64 features output by the residual block. This mechanism learns 143×64 channel weights through a network consisting of two fully connected layers (first nn.Linear(64, 32), then nn.Linear(32, 64)), a ReLU activation function, and a Sigmoid activation function. These weights are then multiplied element-wise with the features output by the residual block to dynamically enhance the key features of the mineral dust image, resulting in 143×64 features after channel attention processing.

[0073] Afterwards, it will go through a residual block, a fully connected layer and other operations to gradually extract and optimize features, which will eventually be used for regression identification of mineral dust concentration.

[0074] 4. Model Training and Testing

[0075] As a preferred option, the Adam optimizer is used, with a learning rate set to [value missing]. The network is trained using mean squared error as the loss function. An early stopping strategy is employed during training: training is terminated when the validation set loss no longer decreases for several consecutive epochs (training rounds or cycles) to prevent overfitting.

[0076] Specifically, based on the image-concentration paired dataset of 143 samples in this embodiment, the training set and validation set are divided in an 8:2 ratio. Model training is performed on the training set using the Adam optimizer, and the difference between predicted and true values ​​is calculated using the mean squared error loss function. Simultaneously, the model reliability is verified using model data output on the validation set. When the validation set loss no longer decreases for several consecutive epochs, an early stopping strategy is triggered, terminating training. A loss diagram is plotted, as shown below. Figure 3As shown, when both training loss and validation loss decrease rapidly with increasing iterations and then stabilize and become similar in value, it indicates that the model has not overfitted and has good generalization ability.

[0077] On the test set of this embodiment, the model achieved a mean squared error of 0.7802 mg²·m⁻. 6 With an average absolute error of 0.6187 mg·m⁻³ and a high coefficient of determination R² of 0.9363, it outperforms traditional machine learning models such as support vector machine regression and random forest.

[0078] In summary, this embodiment fully verifies the effectiveness of the method proposed in this invention. Through the synergistic effect of prior feature selection, multi-kernel fusion, residual connections, and attention mechanisms, high-precision and highly interpretable image recognition of mineral dust concentration is successfully achieved, such as... Figure 4 As shown, this provides a reliable solution for mine safety monitoring.

[0079] Furthermore, embodiments of the present invention also provide an electronic device, including: one or more processors; a storage device storing one or more programs thereon; when the one or more programs are executed by the one or more processors, the one or more processors implement the mineral dust concentration image recognition method based on prior features and multi-core residual attention network of the present invention.

[0080] In this embodiment of the invention, a computer-readable storage medium is also provided, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the mineral dust concentration image recognition method based on prior features and multi-core residual attention network of the present invention.

[0081] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for recognizing mineral dust concentration images based on prior features and a multi-kernel residual attention network, characterized in that, Includes the following steps: S1. Collect images of mineral dust under different concentration conditions, and extract multi-dimensional image features from the mineral dust images to construct a priori feature library of mineral dust images; S2. Based on the Pearson correlation coefficient and mutual information coefficient, select linear feature subsets and nonlinear feature subsets that are significantly related to the concentration of mineral dust from the prior feature library; S3. Based on the principle of implicit high-dimensional mapping of kernel function and complementary fusion of multiple features, a multi-kernel learning framework is constructed. The linear feature subset and the nonlinear feature subset are mapped to a high-dimensional space respectively, and weighted fusion is performed to obtain the fused high-order features. S4. Construct a multi-kernel residual attention network and train the model, including: a multi-kernel feature fusion module, a residual block structure, and a channel attention mechanism; the residual block structure consists of a fully connected layer and a ReLU activation function to alleviate the gradient vanishing problem; The channel attention mechanism learns the weights of each feature channel through global average pooling and fully connected layers to dynamically enhance key mineral dust image features. S5. Input the fused high-order features into the trained multi-core residual attention network for regression calculation, and output the predicted value of mineral dust concentration through the fully connected layer.

2. The method for image recognition of mineral dust concentration according to claim 1, characterized in that, The multi-dimensional image features include: color features, texture features, geometric features, image quality features, grayscale statistical features, and transmittance features; The Pearson correlation coefficient and mutual information coefficient between each feature and the dust concentration are calculated separately. Based on the set threshold of the Pearson correlation coefficient and the significance of the mutual information coefficient, several linear and nonlinear features are selected to form a subset of linear and nonlinear features.

3. The method for image recognition of mineral dust concentration according to claim 2, characterized in that, The color characteristics include: mode, uniformity, skewness, kurtosis, and diversity; The texture features include: contrast, dissimilarity, energy, homogeneity, correlation, and entropy; The geometric features include: total energy moment, axial length moment, corner feature moment, second corner feature moment, eccentricity moment, compactness moment, and asymmetry moment; The image quality features include: peak signal-to-noise ratio, visual information fidelity, and modulation index; The gray-level statistical features include: gray-level mean, gray-level standard deviation, gray-level skewness, and gray-level kurtosis; The transmittance characteristics include: mean transmittance and median transmittance.

4. The method for image recognition of mineral dust concentration according to claim 3, characterized in that, In step S2, the linear feature subset includes at least five features: contrast, homogeneity, compactness moment, asymmetry moment, and modulation index. The nonlinear feature subset includes at least six features: mode, uniformity, diversity, axial length moment, gray mean, and median transmission.

5. The method for image recognition of mineral dust concentration according to claim 2, characterized in that, In step S3, the multi-kernel learning framework divides the dust image features according to linear separability, models linear and nonlinear features respectively, and introduces a multi-kernel fusion mechanism. By considering the high-dimensional mapping of linear and nonlinear features at the same time, the model can take into account both the overall grayscale trend and the local particle structure differences. For the input linear and nonlinear feature subsets, a radial basis kernel function is used to perform high-dimensional mapping on the linear and nonlinear feature subsets respectively, resulting in the corresponding kernel matrix. and Multi-core learning is performed using a linear weighted fusion method, as follows: ; in, The fusion weight coefficients are adjustable, and .

6. The method for image recognition of mineral dust concentration according to claim 1, characterized in that, In step S4, the multi-core residual attention network is composed of multiple residual block structures stacked together. Each residual block structure contains two fully connected layers and a ReLU activation function, and introduces skip connections. The output of the residual block structure is represented as follows: ； in, and These are the input and output of the residual block, respectively. Let {Wi} represent the residual function consisting of a fully connected layer and a ReLU activation function, where {Wi} are learnable parameters. The skip connection is introduced inside the residual block, taking the input features of the residual block into account. The result is directly passed to the position in the residual block after the fully connected layer and ReLU activation function calculation, and... Add them together to get the output of the residual block. This is used to alleviate gradient vanishing during deep network training.

7. The method for recognizing mineral dust concentration images according to claim 6, characterized in that, In step S4, the channel attention mechanism is specifically processed as follows: Global average pooling is performed on the input features to obtain channel-level statistical information; The statistical information is passed sequentially through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function to generate weights for each feature channel; The generated weights are multiplied channel-by-channel with the original input features to dynamically enhance key mineral dust image features. Let the original input features be... The weights of each generated feature channel are W=[w1,w2,.....,w c ]∈R C Then output features Represented as: ； in, For feature map height, The width of the feature map. The number of feature map channels. , , , This represents the numerical values ​​of the original input features at height h, width w, and channel c. This indicates the numerical value of the output feature at the corresponding position. This represents the channel weight.

8. The method for image recognition of mineral dust concentration according to claim 7, characterized in that, In step S4, the multi-kernel feature fusion module performs feature fusion on the output features of the channel attention mechanism to obtain the fused mineral dust concentration image features, as follows: The features output by the channel attention mechanism are associated with the kernel matrix features obtained by mapping linear features through the radial basis kernel function and the kernel matrix features obtained by mapping nonlinear features through the radial basis kernel function. The fusion is carried out using a linear weighted method, and the fusion formula is as follows: ； in, It is a feature resulting from the combination of the output features of the channel attention mechanism and the features of the linear feature kernel matrix. It is a feature resulting from the combination of the output features of the channel attention mechanism and the features of the nonlinear feature kernel matrix. The fusion weight coefficients are adjustable, and Ultimately, the fused multi-core features are obtained. .

9. The method for image recognition of mineral dust concentration according to claim 1, characterized in that, In step S4, the residual attention network is trained using the Adam optimizer, with a learning rate set to... The mean squared error is used as the loss function, and an early stopping strategy is adopted to prevent overfitting.

10. The method for image recognition of mineral dust concentration according to claim 1, characterized in that, In step S5, the fused high-order features are input into the trained multi-kernel residual attention network for regression calculation, as follows: The fused high-order features are converted into tensor form and input into a multi-kernel residual attention network composed of multiple stacked residual blocks. After performing linear transformation and ReLU activation on the input features, the multi-kernel residual attention network extracts features step by step through the residual block structure. Each residual block contains two fully connected layers and a ReLU activation function, and a skip connection is introduced to make the output the sum of the input and the result of the residual function. The output feature weights of the channel attention mechanism are learned by using a multi-kernel feature fusion module, and the features are weighted and fused to optimize feature representation. By mapping features to single-value outputs through a linear layer, the predicted value of dust concentration is obtained, realizing the regression calculation from fusing high-order features to dust concentration.

Images