Face feature extraction method in underground operation space dust environment

Abstract

The present application provides a kind of underground operating space dust environment under face feature extraction method, belong to underground operating space dust environment under face feature extraction technical field, the present application is by polarized imaging sensor collection face image and calculates five channel polarization feature map, combine particle counter measured dust concentration, utilize the physical self-adapting dust removal algorithm based on Mie scattering theory and dark color priori dual constraint to restore clear face image, again by anisotropic diffusion and fractal dimension regularization restoration algorithm output edge-preserving denoising face image, five channel polarization feature map and edge-preserving denoising face image are input into stokes vector perception map artificial intelligence model, output identity feature embedding vector, and pass through face key point robust matching and thin plate spline interpolation virtual key point completion constructs complete key point set, solves the technical problem that face identity feature extraction accuracy rate is seriously reduced under the condition of high concentration dust dynamic change.

Description

Technical Field

[0001] This invention belongs to the field of facial feature extraction technology in dusty environments of underground work spaces. Specifically, it relates to a method for facial feature extraction in dusty environments of underground work spaces. Background Technology

[0002] Personnel identification in underground work spaces is an important part of mine safety management. Traditional facial feature extraction methods use visible light image sensors to collect facial images and combine them with convolutional neural networks or graph neural networks to extract identity feature embedding vectors. These methods have achieved high recognition accuracy under standard lighting and clean environments and have been widely used in scenarios such as ground access control and video surveillance.

[0003] However, in high-intensity dust-generating operations such as mine excavation and blasting, the concentration of suspended coal dust and rock dust in the air can reach hundreds to thousands. Dust particles exert a strong Mie scattering effect on light, causing images to exhibit overall hazy and low-contrast degradation characteristics. Traditional dehazing algorithms rely on prior estimation of transmittance using dark color, but this prior assumption fails in highly absorbent dust environments, leading to a significant increase in transmittance estimation errors and resulting in color shifts and loss of detail in the restored image. Simultaneously, traditional convolutional neural networks lack physical perception of image degradation and cannot distinguish between pseudo-textures caused by dust scattering and genuine skin textures, resulting in severely dust-affected feature vectors.

[0004] In existing technologies, due to the superposition of dust-scattered light and facial reflection light in the intensity domain, and the dynamic variation of scattering degree with dust concentration, current methods lack a mechanism to combine measured dust concentration information with image physical degradation models. They also fail to utilize polarization information to physically decouple the scattered background and facial reflection signals. This results in the robustness and accuracy of identity feature extraction failing to meet practical requirements under time-varying dust concentration conditions in mines. In other words, existing technologies suffer from the technical problem of failing to effectively decouple dust-scattered signals from facial reflection signals under dynamic conditions of high dust concentrations, leading to a significant decrease in the accuracy of facial identity feature extraction. Summary of the Invention

[0005] In view of this, the present invention provides a method for facial feature extraction in a dusty environment in an underground work space, which can solve the problem

[0006] In existing technologies, facial images are severely degraded due to dust scattering in high-concentration dust environments such as underground working spaces like mines. Existing facial feature extraction methods cannot effectively distinguish between dust scattering background and facial reflection signals, resulting in a sharp decline in the accuracy of identity feature extraction.

[0007] This invention is implemented as follows: This invention provides a method for extracting facial features in a dusty environment in an underground work space, including the following steps:

[0008] A polarization imaging sensor is used to acquire face images, a five-channel polarization feature map is calculated, and the five-channel polarization feature map is output to subsequent steps;

[0009] The dust concentration measured by the particle counter is read, and combined with the dark primary color prior, a clear face image is recovered through a physical adaptive dust removal algorithm.

[0010] An anisotropic diffusion and fractal dimension regularization restoration algorithm is applied to a clear face image to output an edge-preserving and denoised face image.

[0011] The five-channel polarization feature map and the edge-preserving and denoised face image are input into the Stokes vector perception model. The node weight allocation is adjusted according to the dust concentration exponential function, and a 512-dimensional identity feature embedding vector is output.

[0012] Based on the dust concentration distribution map, the transmittance of each key point area is calculated, and a complete set of key points is constructed using a robust facial key point matching algorithm.

[0013] Calculate the mean transmittance of the current frame. If the mean transmittance is lower than the occlusion threshold, enable the thin plate spline interpolation virtual key point completion processing. Otherwise, directly output the complete key point set and the 512-dimensional identity feature embedding vector.

[0014] The polarization imaging sensor refers to an image sensor that integrates polarization filters in four directions (0°, 45°, 90°, and 135°) on the same focal plane, and outputs four polarization intensity images simultaneously in a single exposure.

[0015] The five-channel polarization feature map is calculated from the four-channel polarization intensity images, with the five channels representing the total intensity. Linear polarization difference Diagonal polarization difference Linear polarization degree With polarization angle ,in , .

[0016] Specifically, the physical adaptive dust removal algorithm is as follows: the volume scattering coefficient is calculated from the dust concentration using the empirical formula of Mie scattering theory; a variational optimization functional is constructed with the prior estimate of transmittance as the data fidelity term and the theoretical transmittance as the physical constraint term; the transmittance map is solved using the Split Bregman iterative method; and after refinement by guided filtering, a clear face image is recovered using the atmospheric scattering inverse model.

[0017] Specifically, the anisotropic diffusion and fractal dimension regularization restoration algorithm is as follows: based on the Perona-Malik anisotropic diffusion equation, the fractal box dimension is calculated for the local region of the image. A regularization penalty term is constructed based on the deviation between the fractal box dimension and the prior fractal box dimension interval of the skin. The solution is accelerated by Gauss-Seidel iteration combined with multigrid. Finally, wavelet multi-scale adaptive thresholding is performed on the output image to remove residual high-frequency noise pulses.

[0018] Specifically, the Stokes vector-sensing graph model involves an input layer that receives a concatenated five-channel polarization feature map and an edge-preserving, denoised face image, totaling eight channels. A graph structure is constructed using superpixel nodes, and the graph convolutional backbone employs an improved graph attention network, using linear polarization degrees... The attention coefficients are adjusted as physical gating factors and then aggregated step by step through differentiable pooling layers to output a 512-dimensional identity feature embedding vector.

[0019] Wherein, the dust concentration exponential function In the middle, when When the concentration is above the high concentration threshold, the default attention weight allocation is maintained. When in the middle range, the degree of linear polarization The weighting coefficients of the physical gating factor are amplified to a low magnification value when The degree of linear polarization will be lower than the low concentration threshold. The weighting coefficients of the physical gating factor are amplified to a high magnification value and the bypass gain of the polarization consistency check jump connection is activated.

[0020] Specifically, the robust matching algorithm for facial key points is as follows: using the set of 68 key points from a clear reference face template as the left vertex set of the bipartite graph, and the set of candidate key points detected in the edge-preserving and denoising face image as the right vertex set. The edge weights are weighted and fused by appearance similarity, geometric consistency and visibility probability. The Hungarian algorithm is used to solve the maximum weight bipartite graph matching. For unmatched key points, thin plate spline interpolation is used to infer their positions to form a complete set of key points.

[0021] The visibility probability term is derived from the mean of the corresponding region in the transmittance map, and the geometric consistency term adopts the affine invariance constraint of triangle area ratio.

[0022] The thin-plate spline interpolation refers to establishing a thin-plate spline deformation field from the reference clear face template space to the edge-preserving and denoising face image space, using the matched key point pairs as control point pairs. The coordinates of unmatched key points in the reference clear face template are mapped to the edge-preserving and denoising face image space using the thin-plate spline deformation field, which serves as the estimated position of the virtual key points.

[0023] The training of the Stokes vector perception map model employs a joint optimization of InfoNCE contrastive learning loss and polarization reconstruction auxiliary loss. The optimizer uses AdamW with a learning rate warm-up and cosine decay strategy, with the lowest InfoNCE contrastive learning loss on the validation set serving as the early stopping criterion.

[0024] The training dataset for the Stokes vector sensing model consists of polarized four-channel images and synchronous particle counter data collected from no fewer than 200 workers at three dust concentration gradients (low, medium, and high). The training set is then augmented with random horizontal flipping, random brightness perturbation, and random dust concentration interpolation.

[0025] The occlusion threshold is obtained by collecting multiple sets of face images with different dust concentrations, fitting a monotonic mapping relationship between the average transmittance and the visible proportion of key points, selecting the average transmittance corresponding to a key point visible proportion of less than 50% as the initial value of the occlusion threshold, and finally determining it after verification by no less than 5 rounds of on-site experiments.

[0026] Among them, the edge sensitivity parameter The search is performed stepwise within the range of 0.05 to 0.30 with a step size of 0.01. The typical range of the skin prior fractal box dimension is 2.2 to 2.6, and the fractal box dimension of the dust pseudo-texture is higher than 2.7 or lower than 2.0.

[0027] Among them, the sensitivity coefficient in the dust concentration exponential function The concentration was determined experimentally within the range of 0.5 to 2.0; the high concentration threshold value corresponds to... The value is 0.7, and the low concentration threshold corresponds to... The magnification is 0.4; the low-magnification value is 1.5 to 2.0 times, and the high-magnification value is 2.0 to 3.0 times; the polarization angle threshold is... By statistically analyzing the polarization angles of the scattering nodes and the skin nodes The optimal quantile for separating differential distributions and distributions was determined through at least three rounds of experimental verification.

[0028] This invention combines polarization imaging physical mechanisms, dust concentration adaptive removal algorithms, fractal dimension regularization restoration algorithms, and Stokes vector perception graph models to construct a face feature extraction method that collaboratively suppresses dust interference at three levels: physical layer, image layer, and feature layer. This solves the technical problem of severely reduced accuracy in face identity feature extraction under dynamic conditions of high-concentration dust.

[0029] This invention utilizes a polarization imaging sensor to achieve initial decoupling between the scattering background and facial reflection signals at the physical layer of image formation. Then, by embedding the measured dust concentration into a transmittance-optimized functional, the image restoration process achieves physical fidelity, overcoming the fundamental defect of traditional dark color priors failing under highly absorbent dust. Furthermore, fractal dimension regularization enables the diffusion process to distinguish between dust artifacts and real skin textures, providing a more geometrically reliable image foundation for feature extraction. The Stokes vector perception graph model uses linear polarization as a physical gating factor to systematically suppress information propagation from scattering background nodes at the feature extraction layer, and dynamically adjusts weight allocation based on dust concentration, ensuring that the identity feature embedding vector maintains discriminative power even under time-varying dust concentration conditions.

[0030] In summary, this invention solves the technical problem mentioned in the background art of the inability to effectively decouple dust scattering signals and facial reflection signals under dynamic changes in high-concentration dust, which leads to a serious decrease in the accuracy of facial identity feature extraction. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention.

[0032] Figure 2 For different dust concentrations Comparison of channel spatial distribution and face region.

[0033] Figure 3 A trend chart showing the changing ratio of intra-class distance to inter-class distance for 512-dimensional identity feature embedding vectors under different dust concentration gradients.

[0034] Figure 4 This is a monotonic mapping relationship between the visible proportion of facial key points and the mean transmittance under three dust concentration gradients.

[0035] Figure 5 A comparison chart of the mean fractal box dimension of the output images for each processing step under different dust concentration gradients. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0037] like Figure 1 The diagram shown is a flowchart of a method for extracting facial features in a dusty environment in an underground work space, provided by the present invention. This method includes the following steps:

[0038] S01. Acquire face images using a polarization imaging sensor, calculate a five-channel polarization feature map, and output the five-channel polarization feature map to subsequent steps;

[0039] S02. Read the dust concentration measured by the particle counter, combine it with the dark primary color prior, and restore the clear face image through the physical adaptive dust removal algorithm;

[0040] S03. Perform anisotropic diffusion and fractal dimension regularization restoration algorithm on the clear face image to output an edge-preserving and denoised face image.

[0041] S04. Input the five-channel polarization feature map and the edge-preserving and denoising face image into the Stokes vector perception model, adjust the node weight allocation according to the dust concentration exponential function, and output a 512-dimensional identity feature embedding vector.

[0042] S05. Calculate the transmittance of each key point area based on the dust concentration distribution map, and construct a complete set of key points using a robust face key point matching algorithm.

[0043] S06. Calculate the average transmittance of the current frame. When the average transmittance is lower than the occlusion threshold, enable the thin plate spline interpolation virtual key point completion processing. Otherwise, directly output the complete key point set and the 512-dimensional identity feature embedding vector.

[0044] Among them, the polarization imaging sensor refers to an image sensor that integrates polarization filters in four directions (0°, 45°, 90°, and 135°) on the same focal plane, and outputs four polarization intensity images simultaneously in a single exposure, making it suitable for real-time acquisition in dusty mine environments.

[0045] The five-channel polarization feature map is calculated from the four-channel polarization intensity images, with the five channels representing the total intensity. Linear polarization difference Diagonal polarization difference Linear polarization degree With polarization angle The formulas for each channel are expressed as follows: ;in , , , These are polarization intensity images corresponding to the polarization direction. The linear polarization degree of dust-scattered light is low, while the linear polarization degree of face-mirror-reflected light is high. This difference is used to decouple the scattered background and face-reflected signals in the physical layer of image formation, providing a physically meaningful input for subsequent feature extraction.

[0046] The specific steps of the physical adaptive dust removal algorithm are as follows: read the dust concentration measured by the particle counter. (unit: The volume scattering coefficient is calculated using empirical formulas based on Mie scattering theory. (unit: ); through the theoretical transmittance formula Calculate the theoretical transmittance, where To estimate the depth of field (unit: m). Normalized reference depth of field (unit: m); transmittance estimated using dark primary color prior. For data fidelity, theoretical transmittance is used. To address the physical constraints, a total variational smoothing term is added to construct a variational optimization functional; the Split Bregman iterative method is then used to solve for the transmittance map. ; Using the original image as a guide to analyze the transmittance map The algorithm refines the guided filtering; recovers clear face images using an inverse atmospheric scattering model; performs contrast-limited adaptive histogram equalization post-processing on the clear face images; and incorporates the laws of photophysics propagation into the optimization framework using a physical adaptive dust removal algorithm. This allows the transmittance estimation to be constrained by both data-driven priors and measured dust concentrations, overcoming the problem of failure of relying solely on dark color priors in high-absorption dust environments. This ensures that clear face images retain physical fidelity even under time-varying dust concentration conditions.

[0047] Wherein, volume scattering coefficient The empirical formula coefficients were obtained as follows: Multiple groups of coal dust, rock dust, and cement dust of known concentrations were set up in a laboratory dust simulation chamber. The extinction coefficients at different concentrations were measured using an integrating sphere photometer, and the volume scattering coefficients were determined. With dust concentration The measurement data were fitted using least squares; no fewer than 5 concentration gradients were set, and each group was measured at least 3 times. The final coefficient values ​​were determined based on the minimum error variance.

[0048] Among them, guided filtering refers to using the original noisy image as a guide image to filter the transmittance map. Perform structure-preserving filtering to improve the transmittance map. The edge structure is aligned with the original image to eliminate block artifacts generated by the Split Bregman iteration.

[0049] Among them, contrast-limited adaptive histogram equalization refers to performing histogram equalization within local image blocks and limiting excessive contrast amplification by cropping threshold, thereby improving local facial contrast while suppressing noise amplification; the cropping threshold is determined through stepwise search experiments on low-light facial image samples collected at the mine site, with the goal of maximizing the mean contrast gradient of manually annotated facial key points.

[0050] The specific steps of the anisotropic diffusion and fractal dimension regularization restoration algorithm are as follows: Based on the Perona-Malik anisotropic diffusion equation, the diffusion coefficient formula is expressed as follows: ,in This is the edge sensitivity parameter (dimensionless). To normalize the reference gradient magnitude; calculate the fractal box dimension for local regions of the image. ; based on the measured fractal box dimension A regularization penalty term is constructed based on the deviation from the skin's prior fractal box dimension interval and incorporated into the energy functional. Gauss-Seidel iteration combined with multigrid is used to accelerate the solution, converging after 20-50 iterations, outputting an edge-preserving denoised face image. Wavelet multi-scale adaptive thresholding is applied to the edge-preserving denoised face image to remove residual high-frequency noise pulses. Anisotropic diffusion and fractal dimension regularization restoration algorithms encode the self-similar geometric prior of skin texture into constraints of the diffusion process, enabling diffusion to preferentially eliminate pseudo-textures in fractal box dimension anomalous regions (i.e., dust regions) while protecting the fractal self-similar structure of normal skin texture, providing a more geometrically reliable image basis for subsequent keypoint matching.

[0051] Among them, edge sensitivity parameters The prior fractal box dimension interval for the skin was obtained as follows: Facial image samples of different workers and different dust concentrations were collected in a mining environment. The skin region and dust-occluded region were manually labeled for each image. The fractal box dimension was then calculated in the skin region. The distribution is used, with the 5%–95th percentile as the range for the skin prior fractal box dimension; for The search was conducted stepwise within the range of 0.05 to 0.30 with a step size of 0.01. The dual objectives were to achieve the highest signal-to-noise ratio and the highest edge preservation rate in the denoised skin area. The optimal result was determined through experimental data analysis. The typical range of the skin prior fractal box dimension is 2.2 to 2.6, while the fractal box dimension of the dust pseudo-texture is higher than 2.7 or lower than 2.0. The typical range was obtained through the above experiments.

[0052] The specific structure of the Stokes vector sensing model is as follows: the input layer receives five-channel polarization feature maps ( , , , , The graph convolutional backbone is stitched together with the edge-preserving and denoising face image, totaling 8 channels. The superpixel segmentation layer uses the SLIC algorithm to divide the face region into 80-150 superpixels, with each superpixel serving as a node in the graph. The node feature vector is constructed by concatenating the mean and variance of the superpixel in the 8-channel input, with a dimension of 16. The edge weights between nodes are linearly weighted by the spatial Euclidean distance and the cosine similarity of the features in each channel. The ratio of spatial distance weight to feature similarity weight is determined experimentally within the range of 1:1 to 1:3. The graph convolutional backbone uses an improved graph attention network with 4 layers, each with a hidden dimension of 128. Linear polarization is introduced in the calculation of the attention coefficients. As a physical gating factor, linear polarization degree Higher nodes are assigned higher attention weights to suppress linear polarization. Information propagation from nodes with low scattering background; setting a polarization consistency check skip connection between the graph convolutions of layers 2 and 4, if the polarization angle across two skip nodes... The difference exceeds the polarization angle threshold (Unit: rad) Then, activating the bypass connection injects the original node features into the 4th layer graph convolution; polarization angle threshold The polarization angles of scattering nodes and skin nodes were statistically analyzed on polarized image samples with different dust concentrations. The differential distribution was determined by selecting the optimal quantile point through at least three rounds of experiments. Four layers of graph convolution were followed by differentiable pooling layers, progressively aggregating the number of nodes from 80-150 to 40, 10, and 1 in three stages, outputting a 512-dimensional identity feature embedding vector. In CUDA stream allocation, superpixel segmentation and five-channel polarization feature map calculation were allocated to independent CUDA streams for parallel execution. The four-layer forward propagation of the graph convolution was allocated to the main CUDA stream in layer order. The three-stage aggregation of the differentiable pooling was allocated to three sub-CUDA streams and merged at the synchronization point of the main CUDA stream. In memory allocation, the 8-channel input feature map and the graph node feature matrix shared the memory pool, while the graph convolution weight matrix remained resident in memory. Intermediate activation values ​​were released and recalculated as needed using a gradient checkpoint strategy. The dust concentration exponential function was used. middle, Normalized reference dust concentration (unit: Sensitivity coefficient (Dimensionless) Determined experimentally within the range of 0.5–2.0; Dust concentration exponential function As a node weight adjustment factor, when dust concentration When it rises The model automatically increases the degree of linear polarization as the polarization is reduced. The weight coefficients of the physical gating factor concentrate more computational resources on feature propagation of nodes with high linear polarization. The Stokes vector-sensing graph model directly constructs polarization physical quantities as graph topological weights and gating, giving the graph convolution information propagation process physical prior constraints. The influence of scattering background nodes is systematically suppressed by low linear polarization gating, and the extraction of identity discrimination features is protected by both data-driven and physical constraints.

[0053] The specific steps for establishing the training dataset for the Stokes vector sensing image model include: collecting polarization four-channel images and synchronous particle counter data from no fewer than 200 workers at low, medium, and high dust concentration gradients in a mine site or laboratory simulation chamber; collecting images of each worker under no fewer than 5 lighting directions; and establishing no fewer than 3000 sets of polarization images and identity tag pairs; and synchronously recording the dust concentration for each set of data. The image was processed sequentially with five-channel polarization feature map calculation, physical adaptive dust removal algorithm, and anisotropic diffusion and fractal dimension regularization restoration algorithm to generate 8-channel input samples. The training set, validation set, and test set were divided according to identity labels in a ratio of 7:1.5:1.5. The training set was augmented with random horizontal flipping, random brightness perturbation, and random dust concentration interpolation, and the augmented sample size was no less than 15,000 groups.

[0054] The training steps for the Stokes vector perception map model specifically include: jointly optimizing the model using InfoNCE contrastive learning loss and polarization reconstruction auxiliary loss, with the polarization reconstruction auxiliary loss supervising the model in terms of total intensity. Reconstruction quality in components; the optimizer uses AdamW, with an initial learning rate of... ~ During this period, a learning rate warm-up and cosine decay strategy is adopted; the batch size is set to 8-32 based on the GPU memory capacity, and the gradient checkpoint strategy is automatically activated when GPU memory is insufficient; the training epochs are 80-150, and the lowest learning loss compared to the validation set InfoNCE is used as the early stopping criterion; after each training epoch, the ratio of intra-class distance to inter-class distance of the 512-dimensional identity feature embedding vectors on different dust concentration groups is used. If the ratio of a certain dust concentration group is lower than the separation threshold, the sensitivity coefficient is increased. The separation threshold was determined through validation set ablation experiments.

[0055] Among them, the dust concentration exponential function refers to the dust concentration measured by the particle counter. According to the formula Calculate the node weight adjustment factor; when When, maintain the default attention weight allocation; when belong At that time, the degree of linear polarization The weighting coefficient of the physical gating factor is amplified by 1.5 to 2.0 times; when At that time, the degree of linear polarization The weighting coefficients of the physical gating factor are amplified by 2.0 to 3.0 times, while activating the bypass gain of the polarization consistency check skip connection. The amplification factor is determined through grid search experiments on a validation set with three dust concentration gradients, aiming to minimize the ratio of intra-class distance to inter-class distance of the 512-dimensional identity feature embedding vector. The two cutoff values ​​of 0.7 and 0.4 are determined through stepwise search on the validation set. The interval boundaries are determined through at least three rounds of experiments, based on the criterion of minimizing the variance of recognition accuracy within each interval.

[0056] The specific steps of the robust facial landmark matching algorithm are as follows: The set of 68 landmarks from a clear reference face template is used as the left vertex set of the bipartite graph, and the set of candidate landmarks detected in the edge-preserving and denoising face image is used as the right vertex set; the edge weights are calculated by weighted fusion of three terms: appearance similarity (cosine distance of local feature descriptors), geometric consistency (affine invariance constraint of triangle area ratio), and visibility probability; the visibility probability is calculated from the transmittance map. The value is derived from the mean of the corresponding region; the Hungarian algorithm is used to solve the maximum weight bipartite graph matching problem; the edge weights are below the edge weight threshold. The matching edge augmentation path is re-searched to replace the match; for unmatched reference keypoints, thin-plate spline interpolation is used to infer their most likely positions, forming a complete set of keypoints; edge weight thresholding. The algorithm was determined through a stepwise search on a test set labeled with dust occlusion levels, aiming to achieve the highest keypoint matching accuracy, and was repeated for at least three rounds. The robust face keypoint matching algorithm transforms the dust occlusion problem into a combinatorial optimization problem with theoretical optimality guarantees, using transmittance maps... The visibility probability term of the embedded edge weights automatically reduces the matching confidence of key points in high dust occlusion areas during the matching process, and completes the invisible key points under geometric constraints through thin plate spline interpolation, so that the complete set of key points maintains its integrity under any degree of occlusion.

[0057] Thin plate spline interpolation refers to establishing a thin plate spline deformation field from the reference clear face template space to the edge-preserving and denoising face image space using the matched key point pairs as control point pairs. The coordinates of unmatched key points in the reference clear face template are mapped to the edge-preserving and denoising face image space using the thin plate spline deformation field as the estimated position of the virtual key points.

[0058] The occlusion threshold was obtained by collecting multiple sets of face images with different dust concentrations at the mine site, calculating the mean transmittance of each frame and the visible proportion of manually annotated key points, fitting the monotonic mapping relationship between the mean transmittance and the visible proportion of key points, selecting the mean transmittance corresponding to a key point visible proportion of less than 50% as the initial value of the occlusion threshold, and finally determining it after verification by no less than 5 rounds of on-site experiments.

[0059] Optionally, the present invention also provides a method for implementing a facial feature extraction system in a dusty environment of an underground work space using a computer. The computer is equipped with a readable storage medium that stores program instructions, which are used to execute the above-mentioned method when the computer is run.

[0060] The specific implementation of step S01 is as follows: The polarization imaging sensor integrates polarization filters in four directions (0°, 45°, 90°, and 135°) on the same focal plane, allowing four polarization intensity images to be acquired simultaneously in a single exposure. , , , This avoids motion artifacts caused by multiple exposures. Based on the four-channel polarization intensity images, the polarization feature maps of the five channels are calculated sequentially: total intensity Linear polarization difference diagonal polarization difference linear polarization degree polarization angle After multiple scatterings, the polarization state of dust-scattered light tends to become random. The light level is too low, while the mirror reflection of the skin on the face is too high. The high polarization feature map of the five channels quantizes and distinguishes the two types of signals at the physical layer, providing a physically meaningful input for all subsequent processing steps.

[0061] The specific implementation of step S02 is as follows: read the dust concentration measured in real time by the particle counter. (unit The volume scattering coefficient is calculated using empirical formulas based on Mie scattering theory. Then, according to the theoretical transmittance formula The theoretical transmittance is obtained, where To estimate depth of field, To normalize the reference depth of field, transmittance is estimated using the dark primary color prior. For data fidelity items, To address the physical constraints, a total variational smoothing term is added, and a variational optimization functional is constructed. The Split Bregman iterative method is then used to solve for the transmittance map. Then, the original noisy image was used as the guide image to... Guided filtering is performed to align the edge structure of the transmittance map with the original image, eliminating block artifacts generated during iteration. The refined transmittance map is then substituted into the inverse atmospheric scattering model to recover a clear face image. Finally, contrast-limited adaptive histogram equalization post-processing is performed. Histogram equalization is performed within local image blocks, and a cropping threshold is used to limit excessive contrast amplification. The cropping threshold reference value is determined on samples from the mine site with the goal of maximizing the mean contrast gradient of manually annotated facial key points.

[0062] The specific implementation of step S03 is as follows: based on the Perona-Malik anisotropic diffusion equation, the diffusion coefficient... In the middle, edge sensitivity parameters The reference value range is 0.05 to 0.30. The actual value is determined through a stepwise search using two targets: the skin region with the highest signal-to-noise ratio and the highest edge preservation rate, on face image samples from the mine site. The fractal box dimension is calculated for each local region of the image. The typical range of fractal box dimension for skin regions is 2.2–2.6, while the fractal box dimension for dust pseudo-textures is higher than 2.7 or lower than 2.0. (Based on actual measurements...) The deviation from the skin prior region is used to construct a regularization penalty term, which is then added to the energy functional to preferentially eliminate the diffusion process. The pseudo-texture in abnormal regions is used while preserving the self-similar structure of normal skin texture. A Gauss-Seidel iterative method combined with multigrid acceleration is employed, converging after 20–50 iterations. Finally, wavelet multi-scale adaptive thresholding is performed on the output image to remove residual high-frequency noise impulses, resulting in an edge-preserving and denoised face image.

[0063] The specific implementation of step S04 is as follows: The five-channel polarization feature map ( , , , , The image is concatenated with the edge-preserving and denoised face image as an 8-channel input, and then fed into the Stokes vector perception graph model. The superpixel segmentation layer uses a simple linear iterative clustering algorithm to divide the face region into 80–150 superpixels. Each superpixel serves as a node in the graph, and the node feature vector is constructed by concatenating the mean and variance of 8 channels, with a dimension of 16. The edge weights between nodes are calculated by linearly weighting the spatial Euclidean distance and the cosine similarity of the features in each channel. The ratio of the spatial distance weight to the feature similarity weight is determined within the range of 1:1 to 1:3. The graph convolutional backbone uses an improved graph attention network with 4 layers, each with a hidden dimension of 128. As a physical gating factor, nodes with high linear polarization degrees are assigned higher attention weights. A polarization consistency check skip connection is set between layers 2 and 4; when crossing two skip nodes... The difference exceeds the polarization angle threshold The bypass connection is activated at that time. Dust concentration exponential function. In the middle, sensitivity coefficient The reference value range is 0.5 to 2.0: when Maintain the default weight when; belong When The gating weight is amplified by 1.5 to 2.0 times; when The time-amplified layer is 2.0 to 3.0 times and the bypass gain is activated. After 4 layers of graph convolution, a differentiable pooling layer is applied, and the number of nodes is gradually aggregated from 80 to 150 to 40, 10 and 1 in 3 levels, outputting a 512-dimensional identity feature embedding vector.

[0064] The specific implementation of step S05 is as follows: The transmittance of each key point region is calculated based on the dust concentration distribution map, and the average transmittance is converted into a visibility probability term. The set of 68 key points from a reference clear face template is used as the left vertex set of the bipartite graph, and the set of candidate key points detected in the edge-preserving and denoising face image is used as the right vertex set. The edge weight is calculated by weighted fusion of three terms: appearance similarity (cosine distance of local feature descriptors), geometric consistency (affine invariance constraint of triangle area ratio), and visibility probability. The Hungarian algorithm is used to solve the maximum weight bipartite graph matching problem, and edge weights below the edge weight threshold are considered. The matching edge augmentation path is re-searched for alternative matches, and the edge weight threshold is used. The key points were determined through a stepwise search on a test set labeled with dust occlusion levels and repeated experiments for at least three rounds. For unmatched reference key points, thin-plate spline interpolation was used to infer their most likely locations. A deformation field was established using matched key point pairs as control point pairs, and the coordinates of unmatched key points in the reference template were mapped to the edge-preserving and denoising face image space to form a complete set of key points.

[0065] The specific implementation of step S06 is as follows: Calculate the average transmittance of each pixel in the current frame and determine whether it is lower than the occlusion threshold. The reference method for determining the occlusion threshold is: collect multiple sets of face images under different dust concentrations, fit the monotonic mapping relationship between the average transmittance and the visible proportion of key points, select the average transmittance corresponding to the key point visible proportion being lower than 50% as the initial value, and finally determine it after verification by no less than 5 rounds of on-site experiments. When the average transmittance is lower than the occlusion threshold, enable the thin plate spline interpolation virtual key point completion processing, and output the virtual key points inferred in step S05 after incorporating them into the complete key point set; otherwise, directly output the complete key point set constructed in step S05 and the 512-dimensional identity feature embedding vector output in step S04.

[0066] It should be noted that the key technologies of this invention include: Physically adaptive dust removal technology embeds the measured dust concentration into the transmittance optimization functional, making transmittance estimation subject to both data-driven priors and physical laws, overcoming the fundamental defect of traditional dark color priors leading to transmittance estimation bias due to prior assumption failure under highly absorbent dust; Fractal dimension regularization restoration technology encodes the self-similar geometric prior of skin texture as a constraint on the diffusion process, enabling denoising diffusion to distinguish between dust pseudo-textures with aberrant fractal box dimensions and normal skin texture, providing a geometrically reliable image basis for feature extraction; and the Stokes vector perception graph model... A physical gating factor, directly constructed as a graph topology, systematically suppresses the influence of scattering background nodes at the feature propagation layer, and dynamically adjusts the gating strength based on dust concentration. These three key technologies form a synergistic dust suppression chain at the image restoration, image reconstruction, and feature extraction levels, respectively, ensuring targeted suppression of dust interference at each processing stage. The overall suppression effect of the three technologies working together is far superior to that of any single technology applied individually.

[0067] It should be noted that shortly after a mine face excavation or blasting operation, the concentration of dust in the air surges to extremely high levels, with a wide particle size distribution. Dust particles of different sizes exhibit significantly different light-scattering characteristics. At this time, the degradation degree in different spatial regions of a facial image is extremely uneven; some areas have extremely low transmittance, making key points almost invisible, while other areas still retain some image information. Under these conditions, the spatial integrity of facial key points is severely damaged, posing significant challenges to subsequent identity recognition and behavior analysis. The above technical problems arise because: the sudden increase in dust concentration causes the transmittance of local areas to approach zero, resulting in extremely scarce image information. Key point detection algorithms cannot reliably locate key points due to a lack of sufficient local texture gradients; simultaneously, the spatial unevenness of degradation degrees in different regions makes it difficult for global image enhancement methods to simultaneously address the processing effects of high-transmittance and low-transmittance areas. Common solutions to these technical problems include directly marking failed key points as missing, using interpolation for uniform completion, or using a fixed template for forced alignment. However, directly labeling missing features leads to inconsistent feature vector dimensions, making subsequent matching impossible; uniform interpolation ignores the non-linear characteristics of facial geometric deformation, resulting in large positional errors; fixed template alignment does not consider the actual facial pose and deformation in the current frame, leading to particularly significant errors when the pitch or side angle is large. This invention effectively solves this technical problem by using transmittance maps... By embedding the edge-weighted visibility probability term of the robust facial landmark matching algorithm, the matching process automatically reduces the matching confidence of landmarks in high-dust-occlusion areas, prioritizing the use of reliable landmarks in low-occlusion areas to establish matching relationships. Based on this, a thin-plate spline deformation field is established using matched landmark pairs as control points. Utilizing the minimum bending energy principle of the thin-plate spline deformation field, and under the condition of satisfying facial geometric continuity constraints, the coordinates of unmatched landmarks in the reference template are nonlinearly mapped to the current frame space to obtain the estimated positions of virtual landmarks. Compared to linear interpolation, thin-plate spline interpolation can better capture the nonlinear geometric deformation caused by facial pose changes. Compared to fixed template alignment, it fully utilizes the actual pose information carried by matched landmarks in the current frame, making the completed position closer to the true position under geometric constraints. When the mean transmittance is lower than the occlusion threshold, the system automatically enables virtual landmark completion processing, ensuring that the complete landmark set can be output under any dust occlusion level, thereby supporting the normal operation of subsequent identity recognition processes.

[0068] Specifically, the principle of this invention is:

[0069] The reason why this invention can solve the above-mentioned core technical problems is that it starts from the laws of photophysics propagation and forms a progressive chain of physical constraints in the four stages of image acquisition, image restoration, image reconstruction and feature extraction.

[0070] First, during the image acquisition phase, the polarization imaging sensor utilizes the fundamental difference in linear polarization between dust-scattered light and specular reflection from the human face to achieve physical layer decoupling. Dust scattering is a multiple scattering process, resulting in a more randomized polarization state and a lower degree of linear polarization; while specular reflection from the human face follows Fresnel's law, leading to a higher degree of linear polarization. Through calculation... and The five-channel polarization feature map quantizes and encodes the difference between the scattered background and the reflected signal during the signal formation stage, providing a physically meaningful input for all subsequent processing steps.

[0071] Secondly, in the image restoration stage, the physically adaptive dust removal algorithm converts the measured dust concentration from the particle counter into the volume scattering coefficient based on Mie scattering theory, and then calculates the theoretical transmittance. This transmittance is then used as a physical constraint term in the optimization functional, jointly constraining the transmittance estimation with the dark color prior fidelity term. This mechanism subjectes transmittance estimation to both data-driven and physical constraints. When highly absorbent dust causes the dark color prior to fail, the physical constraint term can correct the prior bias, ensuring the physical fidelity of the restored image.

[0072] Furthermore, in the image restoration stage, the anisotropic diffusion and fractal dimension regularization restoration algorithm encodes the self-similar geometric prior of skin texture as a constraint on the diffusion process. Skin texture has a certain range of fractal box dimensions, while the fractal box dimension of pseudo-textures caused by dust adhesion deviates from the normal skin range. By constructing the fractal box dimension deviation as a penalty term, the diffusion process is guided to prioritize the elimination of noise in dust pseudo-texture regions, while protecting the self-similar structure of normal skin texture, thus enabling subsequent keypoint matching to obtain a geometrically more reliable image basis.

[0073] Finally, in the feature extraction stage, the Stokes vector-sensing graph model inputs the five-channel polarization feature map and the restored image into the graph neural network to... As an attention gating factor, it systematically suppresses the influence of scattering background nodes during graph convolution information propagation, and dynamically adjusts the gating weights according to the dust concentration exponential function, automatically strengthening the physical gating constraint when the dust concentration increases. Differentiable pooling layers aggregate node features step by step into a 512-dimensional identity feature embedding vector, ensuring the discriminativeness and robustness of the features.

[0074] Because the above four stages form a synergistic dust suppression mechanism at the physical layer, image layer and feature layer, this invention can effectively decouple scattering signals and facial reflection signals in a mining environment with dynamically changing dust concentration, ensuring the accuracy and robustness of facial identity feature extraction, thus solving the core technical problems faced by existing methods from the perspective of technical principles.

[0075] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0076] The specific implementation of step S01 is as follows: The polarization imaging sensor integrates polarization filters of four directions (0°, 45°, 90°, and 135°) on the same focal plane, and outputs four polarization intensity images simultaneously in a single exposure. , , , The calculation formulas for each channel of the five-channel polarization feature map are as follows:

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] In the formula, The total intensity (dimensionless, and...) , , , (same measurement range) It represents the linear polarization difference (dimensionless). This is the diagonal polarization difference (dimensionless). It represents the degree of linear polarization (dimensionless). Polarization angle (unit: ), , , , Polarization intensity images at 0°, 45°, 90°, and 135° (dimensionless, normalized to). Dust-scattered light Low, mirror reflection of the face The value is slightly higher. This difference is used to decouple the scattering background and face reflection signals at the physical layer, providing a physically meaningful input for subsequent steps.

[0083] The specific implementation of step S02 is as follows: read the dust concentration measured by the particle counter. (unit: The volume scattering coefficient is calculated using empirical formulas based on Mie scattering theory. (unit: The formula is as follows:

[0084] ;

[0085] In the formula, Normalized reference scattering coefficient (unit: (Experience value is 1) Normalized reference concentration (unit: (Experience value is 1) , These are the least squares fitting coefficients (dimensionless). This represents the fitting error term (dimensionless). The steps for obtaining the fitting coefficients are as follows: Step 1: Set up no fewer than 5 groups of coal dust, rock dust, and cement dust of known concentrations in a laboratory dust simulation chamber, and repeat the measurement for each group no fewer than 3 times; Step 2: Measure the extinction coefficients at different concentrations using an integrating sphere photometer. and The measurement data are fitted using least squares, and the minimum error variance is used as the criterion for determining... , The theoretical transmittance formula is as follows:

[0086] ;

[0087] In the formula, The theoretical transmittance (dimensionless). To estimate depth of field (unit: ), Normalized reference depth of field (unit: (Experience value is 1) and All are dimensionless ratios, and the exponent is dimensionless overall. The variational optimization functional estimates transmittance using a dark primary color prior. For data authenticity items, with To address the physical constraints, a total variational smoothing term is added, as shown in the following formula:

[0088] ;

[0089] In the formula, The transmittance diagram to be determined (dimensionless). Estimate the transmittance (dimensionless) a priori for the dark primary color. , , These are the weighting coefficients for each item (dimensionless). To prevent small quantities (dimensionless, empirically valued) from being divided by zero ), The gradient of the transmittance map (dimensionless). This is the normalized reference gradient (dimensionless, empirically the mean of the transmittance gradients of each training sample). for The square of the norm, for Norm, Let be the functional energy value (dimensionless). The split Bregman iteration method is used to solve for . Then, using the original noisy image as a guide, Guided filtering is applied to refine the image, aligning the edge structure of the transmittance map with the original image and eliminating blocky artifacts. A clear face image is then recovered using the inverse atmospheric scattering model, as shown in the following formula:

[0090] ;

[0091] In the formula, To recover a clear human face image (dimensionless). For the original noisy input image (dimensionless, and...) (same measurement range) Atmospheric light estimate (dimensionless, compared with) (same measurement range) This is the normalized reference transmittance (dimensionless, empirically 1). To prevent the lower limit of transmittance from being divided by zero (dimensionless, empirical value is 0.1). The ratio is dimensionless, with the numerator and denominator having the same dimensions. Contrast-limited adaptive histogram equalization post-processing is performed on clear face images. The cropping threshold is determined through stepwise search experiments, aiming to maximize the mean contrast gradient of manually annotated key points on low-light face image samples from the mine site.

[0092] The specific implementation of step S03 is as follows: based on the Perona-Malik anisotropic diffusion equation, the diffusion coefficient formula is as follows:

[0093] ;

[0094] In the formula, This represents the current image gradient magnitude (dimensionless). This is the normalized reference gradient magnitude (dimensionless, taken as the average gradient magnitude of the training samples). This is the edge sensitivity parameter (dimensionless). The diffusion coefficient (dimensionless, with a value range of...) ), and Both are dimensionless, and the exponent as a whole is dimensionless. Calculate the fractal box dimension for a local region of the image. (Dimensionless), based on actual measurement Center value of the interval of the prior fractal box dimension of the skin The bias constructs a regularization penalty term, and the energy functional is as follows:

[0095] ;

[0096] In the formula, Define the domain for the image. is the regularization weight coefficient (dimensionless). The center value of the skin prior fractal box dimension interval (dimensionless). and All are dimensionless ratios. This represents the functional energy value (dimensionless). The typical range of the skin prior fractal box dimension is 2.2–2.6, and the dust pseudo-texture... For values ​​above 2.7 or below 2.0, the range was determined by manually labeling skin areas and dust-occluded areas in facial images taken in a mine environment. The distribution is determined by taking the 5% to 95th percentiles. Take the mean of this interval. The search was performed incrementally within the range of 0.05–0.30 with a step size of 0.01. The dual objectives, determined experimentally, were the highest signal-to-noise ratio and the highest edge preservation rate in the denoised skin region. A Gaussian-Seidel iteration combined with multigrid acceleration was employed to solve the problem. The discrete update equations are as follows:

[0097] ;

[0098] In the formula, For the first Pixel at the next iteration The grayscale value (dimensionless). The iteration step size is dimensionless, with an empirical value of 0.1 to 0.25. For pixels The set of neighboring pixels, The number of neighboring pixels (dimensionless). For pixels With neighboring pixels The gradient magnitude (dimensionless) between them. Spatial stride between adjacent pixels (unit: ), Normalized reference space step size (unit: (Experience value is 1) It is a dimensionless ratio. For pixels The fractal box dimension (dimensionless) of the local region. The normalized reference iteration step size (dimensionless, empirically 1) is used, and all terms in the entire update equation are dimensionless. The process converges after 20–50 iterations, outputting an edge-preserving, denoised face image, and then performing wavelet multi-scale adaptive thresholding to remove residual high-frequency noise pulses.

[0099] The specific implementation of step S04 is as follows: The input layer receives the concatenation of a five-channel polarization feature map and a face image with edge preservation and denoising, resulting in a total of 8 channels, denoted as... ,in , Image height and width (unit: The superpixel segmentation layer uses a simple linear iterative clustering algorithm to divide the face region into segments. Superpixels ( (Taking a range of 80-150), each superpixel is used as a node in the graph. Node feature vectors By the The superpixel is constructed by concatenating the mean and variance of the 8-channel input:

[0100] ;

[0101] In the formula, For the first The superpixel in the... The mean of the channel (dimensionless). For the first The superpixel in the... The variance of the channel (dimensionless). Channel index. Edge weights between nodes. It is composed of a linear weighted average of spatial Euclidean distance and cosine similarity of features in each channel:

[0102] ;

[0103] In the formula, For nodes With nodes Spatial Euclidean distance (unit: ), Normalized reference spatial distance (unit: The empirical value is the length of the image diagonal. The mean cosine similarity of features for each channel (dimensionless). , These are the spatial distance weight and the feature similarity weight (dimensionless), respectively. It is a dimensionless ratio. For dimensionless boundary weights, and The ratio was determined experimentally within the range of 1:1 to 1:3. The graph convolutional backbone uses an improved graph attention network with 4 layers, each with a hidden dimension of 128. Layer nodes For neighboring nodes Attention coefficient introduction As a physical gating factor:

[0104] ;

[0105] In the formula, For the first Layer-learnable attention vectors (dimensionless). For the first Layer learnable weight matrix (dimensionless). For the first Layer nodes eigenvectors (dimensionless). For nodes The set of neighboring nodes, This represents a vector concatenation operation. For nodes within the area The mean (dimensionless). The attention coefficient is dimensionless, with the numerator and denominator having the same dimensions. Leaking linear rectified function. For input The definition of is: when hour ,when hour ,in The slope coefficient is negative (dimensionless, empirically estimated at 0.2). The feature update formula for each layer of nodes is as follows:

[0106] ;

[0107] In the formula, It is a non-linear activation function (the default is a linear rectified function). For the first Layer nodes The feature vector (dimensionless). A polarization consistency check skip connection is set between the graph convolutions of layer 2 and layer 4. If it spans two skip nodes... The difference exceeds the polarization angle threshold (unit: Then, activating the bypass connection injects the original node features into the 4th layer graph convolution. By statistically analyzing scattering nodes and skin nodes on polarized image samples with different dust concentrations... The differential distribution was determined by selecting the optimal quantile point for separation and validating it through at least three rounds of experiments. The dust concentration exponential function is:

[0108] ;

[0109] In the formula, Normalized reference dust concentration (unit: ), The sensitivity coefficient (dimensionless) is determined experimentally within the range of 0.5 to 2.0. It is a dimensionless ratio. This is a dimensionless node weight adjustment factor. When... Maintain the default attention weight allocation; when When The weighting coefficient of the physical gating factor is amplified by 1.5 to 2.0 times; when When The physical gating factor weight coefficient is amplified by 2.0 to 3.0 times, while the bypass gain of the polarization consistency check skip connection is activated. The amplification factor is determined through grid search experiments. Four layers of graph convolution are followed by differentiable pooling layers, which aggregate the number of nodes in three stages from 80 to 150, then to 40, 10, and finally 1, outputting a 512-dimensional identity feature embedding vector. (Dimensionless). Training employs a joint optimization of information-noise contrastive estimation contrastive learning loss and polarization reconstruction auxiliary loss. The optimizer uses an adaptive weight decay optimization algorithm, with an initial learning rate of... ~ The training sessions consist of 80 to 150 rounds.

[0110] The specific implementation of step S05 is as follows: using the set of 68 key points of the reference clear face template as the left vertex set of the bipartite graph. The set of candidate keypoints detected in the edge-preserving and denoising face image is used as the right vertex set. , border rights The calculation is based on a weighted fusion of three terms: appearance similarity, geometric consistency, and visibility probability.

[0111] ;

[0112] In the formula, Cosine similarity (dimensionless) is used to describe local features. The geometric consistency score for the affine invariance of the triangle area ratio (dimensionless). For the first Mean transmittance of the region corresponding to each reference key point (dimensionless). , , These are the weighting coefficients for each item (dimensionless). The normalized reference edge weight (dimensionless, empirically valued at 1). The edge weights are dimensionless. The formula for geometric consistency score is as follows:

[0113] ;

[0114] In the formula, For reference templates, key points Area of ​​the local triangle formed by vertices (unit: ), Candidate keypoints in the current image Area of ​​the corresponding triangle formed by the vertices (unit: ), The area of ​​the normalized reference triangle (unit: (Take the average area of ​​all triangles in the reference template). and All are dimensionless ratios. The score is a dimensionless geometric consistency score, with a range of values ​​of [value range missing]. A higher value indicates higher geometric consistency. The Hungarian algorithm is used to solve the maximum weight bipartite graph matching problem, where edge weights are below a threshold. (Dimensionless) matching edge augmenting path re-search for alternative matches, The keypoints were determined by progressively searching on a test set labeled with dust occlusion levels, aiming for the highest keypoint matching accuracy, and through at least three rounds of repeated experiments. For unmatched reference keypoints, thin-plate spline interpolation was used to infer their positions, forming a complete keypoint set.

[0115] The specific implementation of step S06 is: calculate the average transmittance of the current frame. (dimensionless), when Below the occlusion threshold When the image is dimensionless, thin-plate spline interpolation is used to complete the virtual keypoints. A thin-plate spline deformation field is established from the reference clear face template space to the edge-preserving and denoising face image space. The coordinates of unmatched keypoints in the reference template are mapped to the current image space as the estimated positions of the virtual keypoints. Otherwise, the complete set of keypoints and the 512-dimensional identity feature embedding vector are directly output. Occlusion threshold The acquisition steps are as follows: Step 1: Collect multiple sets of facial images under different dust concentrations at the mine site, and calculate the values ​​of each frame. Visible ratio compared to manually annotated keypoints; Step 2: Fitting The monotonic mapping relationship between the visibility ratio of keypoints and the keypoint visibility ratio is selected, corresponding to keypoint visibility ratios below 50%. As The initial value was finally determined after verification through no less than 5 rounds of on-site experiments.

[0116] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effectiveness of the invention, technicians set up a test environment at a fully mechanized coal mining face to simulate high-concentration dust conditions after tunneling and blasting. Facial feature extraction tests were conducted on 12 workers on site. The dust type was a mixture of coal dust and rock dust, with a particle size distribution range of 1–100 μm. The particle counter has a real-time sampling frequency of 1 Hz.

[0117] The test was divided into three dust concentration gradients: low, medium, and high. The dust concentration ranges for the three gradients are shown in Table 1.

[0118] Table 1 Dust Concentration Gradient Setting Table

[0119]

[0120] The polarization imaging sensor is fixedly installed at the entrance of the working surface. The exposure time is set to 8 ms, and a single exposure outputs four polarization intensity images with a resolution of 1280×1024 pixels. After step S01 is completed, the five-channel polarization feature map... The spatial distribution of the channels is as follows Figure 2 As shown, the face area can be observed intuitively. The value is significantly higher than that of the background dust scattering area, indicating a significant decoupling effect between the scattering background of the physical layer and the face reflection signal.

[0121] In step S02, the particle counter measures the dust concentration under a high concentration gradient. Approximately 1500 Calculate the volume scattering coefficient using empirical formulas. The theoretical transmittance was then obtained. Approximately 0.18, dark primary color prior estimate of transmittance The difference is approximately 0.32, indicating a significant overestimation bias in the dark primary color prior under highly absorbent dust conditions. After variational optimization, the final transmittance map is obtained. The mean value is 0.21, which is closer to the physical reality. The clear face image recovered by the atmospheric scattering inverse model still retains the facial contours and main textures under high concentrations of dust.

[0122] In step S03, the edge sensitivity parameter The fractal box dimension was set to 0.13 on the validation set, the prior fractal box dimension of the skin was set to 2.2–2.6, and the number of iterations was set to 35. Edge-preserving denoising of face images effectively protected the self-similar structure of skin texture while eliminating dust-like textures, and significantly improved the texture clarity of key point local areas. The fractal box dimension statistics for edge-preserving denoising effects under different concentration gradients are shown in Table 2.

[0123] Table 2. Statistical table of fractal box dimensions under different dust concentrations

[0124]

[0125] As shown in Table 2, the mean fractal box dimension of the processed skin region regressed to the skin prior interval of 2.2 to 2.6. The high fractal box dimension of the dust pseudo-texture was effectively suppressed, verifying the effectiveness of the anisotropic diffusion and fractal dimension regularization restoration algorithm.

[0126] In step S04, the Stokes vector sensing model is used under high dust concentration conditions ( Automatically The weight coefficient of the physical gating factor was amplified by 2.5 times, and the bypass gain of the polarization consistency check skip connection was activated. The model embedded 512-dimensional identity features from 12 workers at three concentration gradients. Cosine similarity calculation showed that the ratio of intra-class distance to inter-class distance remained low even at high concentration gradients, indicating that the identity discrimination features possess good robustness. Figure 3 As shown in the figure, the trend of the ratio of intra-class distance to inter-class distance of 512-dimensional identity feature embedding vectors with different dust concentrations shows that the present invention can effectively suppress feature vector degradation by dynamically adjusting the weights when the dust concentration increases.

[0127] In steps S05 and S06, technicians analyzed the keypoint matching results for each frame of images of 12 workers under high dust concentration. The average transmittance of some frames was lower than the occlusion threshold (reference value approximately 0.25), automatically enabling thin-plate spline interpolation for virtual keypoint completion. The keypoint matching and completion results are shown in Table 3. Even under high occlusion conditions, the completion rate of the complete keypoint set reached 100%, and the average deviation between the virtual keypoints completed by thin-plate spline interpolation and the manually labeled real positions remained within an acceptable range under geometric constraints.

[0128] Table 3. Statistical table of key point matching and completion under different dust concentrations

[0129]

[0130] like Figure 4 As shown, the monotonic mapping curves between the visible proportion of facial key points and the average transmittance under three dust concentration gradients intuitively verify the rationality of the selected occlusion threshold. The average transmittance corresponding to a key point visible proportion of less than 50% is approximately 0.25, which is consistent with the reference value of the occlusion threshold.

[0131] The advancements of this invention compared to traditional methods are reflected in the following aspects. Traditional methods rely solely on dark chromatic a priori data for dehazing at the image layer. Due to the lack of constraints from measured physical quantities, the prior assumptions fail in highly absorbent dust environments, and the transmittance estimation error accumulates continuously with increasing dust concentration, resulting in unreliable image restoration quality. This invention embeds the measured dust concentration from a particle counter into a variational optimized functional, ensuring that transmittance estimation remains constrained by physical laws even when the prior assumptions fail, fundamentally cutting off the path of error accumulation. Traditional feature extraction networks confuse dust artifacts with skin textures in dust-degraded images, lacking the ability to distinguish between the two types of signals at the physical layer. This invention, through polarization imaging and... The physical gating mechanism systematically suppresses the influence of scattering background nodes during the feature propagation stage, providing physical prior guarantees for the extraction of identity discrimination features. Traditional keypoint processing methods can only mark missing points or use linear interpolation in cases of high occlusion, ignoring the nonlinear deformation characteristics of facial geometry. This invention completes virtual keypoints under geometric constraints through thin-plate spline deformation fields, outputting a complete set of keypoints regardless of the degree of occlusion, thereby supporting the normal operation of subsequent identity recognition processes.

[0132] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0133] Table 4. Variable Explanation Table (Part 1)

[0134]

[0135] Table 5. Variable Explanation Table (Part Two)

[0136]

[0137] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting facial features in a dusty environment of an underground work space, characterized in that, Includes the following steps: A polarization imaging sensor is used to acquire face images, a five-channel polarization feature map is calculated, and the five-channel polarization feature map is output to subsequent steps; The dust concentration measured by the particle counter is read, and combined with the dark primary color prior, a clear face image is recovered through a physical adaptive dust removal algorithm. An anisotropic diffusion and fractal dimension regularization restoration algorithm is applied to a clear face image to output an edge-preserving and denoised face image. The five-channel polarization feature map and the edge-preserving and denoised face image are input into the Stokes vector perception model. The node weight allocation is adjusted according to the dust concentration exponential function, and a 512-dimensional identity feature embedding vector is output. Based on the dust concentration distribution map, the transmittance of each key point area is calculated, and a complete set of key points is constructed using a robust facial key point matching algorithm. Calculate the mean transmittance of the current frame. If the mean transmittance is lower than the occlusion threshold, enable the thin plate spline interpolation virtual key point completion processing. Otherwise, directly output the complete key point set and the 512-dimensional identity feature embedding vector.

2. The method for extracting facial features in a dusty environment in an underground work space according to claim 1, characterized in that, The polarization imaging sensor refers to an image sensor that integrates polarization filters in four directions (0°, 45°, 90°, and 135°) on the same focal plane, and outputs four polarization intensity images simultaneously in a single exposure.

3. The method for extracting facial features in a dusty environment in an underground work space according to claim 2, characterized in that, The occlusion threshold is obtained by collecting multiple sets of face images under different dust concentrations, fitting a monotonic mapping relationship between the average transmittance and the visible proportion of key points, selecting the average transmittance corresponding to a key point visible proportion of less than 50% as the initial value of the occlusion threshold, and finally determining it after verification by no less than 5 rounds of on-site experiments.

4. The method for extracting facial features in a dusty environment in an underground work space according to claim 3, characterized in that, The physical adaptive dust removal algorithm specifically involves: calculating the volume scattering coefficient from dust concentration using empirical formulas based on Mie scattering theory; constructing a variational optimization functional with the prior estimate of transmittance as the data fidelity term and the theoretical transmittance as the physical constraint term; solving the transmittance map using the Split Bregman iterative method; and then refining the image through guided filtering and recovering a clear face image using an inverse atmospheric scattering model.

5. The method for extracting facial features in a dusty environment in an underground work space according to claim 4, characterized in that, The anisotropic diffusion and fractal dimension regularization restoration algorithm is as follows: based on the Perona-Malik anisotropic diffusion equation, the fractal box dimension is calculated for local regions of the image. A regularization penalty term is constructed based on the deviation between the fractal box dimension and the prior fractal box dimension interval of the skin. The solution is accelerated by Gauss-Seidel iteration combined with multigrid. Finally, wavelet multi-scale adaptive thresholding is performed on the output image to remove residual high-frequency noise pulses.

6. The method for extracting facial features in a dusty environment in an underground work space according to claim 5, characterized in that, The Stokes vector-sensing graph model specifically comprises: an input layer receiving a concatenated five-channel polarization feature map and an edge-preserving, denoised face image, totaling eight channels; a graph structure is constructed using superpixel nodes; and an improved graph attention network is used as the graph convolution backbone, employing linear polarization degrees. The attention coefficients are adjusted as physical gating factors and then aggregated step by step through differentiable pooling layers to output a 512-dimensional identity feature embedding vector.

7. The method for extracting facial features in a dusty environment in an underground work space according to claim 6, characterized in that, The dust concentration exponential function In the middle, when When the concentration is above the high concentration threshold, the default attention weight allocation is maintained. When in the middle range, the degree of linear polarization The weighting coefficients of the physical gating factor are amplified to a low magnification value when The degree of linear polarization will be lower than the low concentration threshold. The weighting coefficients of the physical gating factor are amplified to a high magnification value and the bypass gain of the polarization consistency check jump connection is activated.

8. The method for extracting facial features in a dusty environment in an underground work space according to claim 7, characterized in that, The robust matching algorithm for facial key points specifically involves: using a set of 68 key points from a clear reference face template as the left vertex set of the bipartite graph, and using a set of candidate key points detected in the edge-preserving and denoising face image as the right vertex set. The edge weights are weighted and fused by appearance similarity, geometric consistency, and visibility probability. The Hungarian algorithm is used to solve the maximum weighted bipartite graph matching problem. For unmatched key points, thin-plate spline interpolation is used to infer their positions and form a complete set of key points.

9. The method for extracting facial features in a dusty environment in an underground work space according to claim 8, characterized in that, The visibility probability term is derived from the mean of the corresponding region in the transmittance map, and the geometric consistency term adopts the affine invariance constraint of triangle area ratio.

10. The method for extracting facial features in a dusty environment in an underground work space according to claim 9, characterized in that, The thin-plate spline interpolation refers to establishing a thin-plate spline deformation field from the reference clear face template space to the edge-preserving and denoising face image space, using the matched key point pairs as control point pairs. The thin-plate spline deformation field is then used to map the coordinates of unmatched key points in the reference clear face template to the edge-preserving and denoising face image space, serving as the estimated position of the virtual key points.

Images