Face recognition method and system for dynamic environment

Abstract

The present application relates to the technical field of face recognition, and a face recognition method and system for dynamic environment, the method comprising: collecting face video stream through a multispectral imaging device, and pre-processing through dynamic noise reduction, distortion correction, etc.; constructing a multiscale spatiotemporal fusion feature extraction network to extract dynamic spatiotemporal features and fuse cross-modal features; establishing an environment disturbance simulation generation model to generate virtual samples and perform domain self-adaptive alignment; designing an online incremental feature updating mechanism to optimize feature encoder parameters; deploying a heterogeneous graph neural network to perform multi-modal decision fusion; and using a hierarchical verification architecture to complete identity recognition. The system comprises multiple modules such as data acquisition and preprocessing, multiscale spatiotemporal fusion feature extraction, etc. The present application effectively solves the problem of face recognition in a dynamic environment, and can significantly improve recognition accuracy, robustness, real-time performance and reliability in complex lighting, posture and expression changes, occlusion and background noise scenarios, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of face recognition, in particular to a face recognition method and system for dynamic environment. BACKGROUND

[0002] In today's digital age, face recognition technology, as a key technology in the field of biometric recognition, is widely used in security monitoring, access control systems, intelligent payment and many other scenarios. However, traditional face recognition methods face serious challenges in dynamic environments and are difficult to meet the needs of practical applications.

[0003] The complexity of lighting conditions is one of the important difficulties. In outdoor environments, the intensity of light changes greatly from morning to evening, and the nature of light is also completely different under different weather conditions, such as sunny, cloudy, rainy, etc. In addition, the light in indoor environments is not uniformly distributed, such as in large shopping malls, warehouses, etc., there are obvious light and dark areas. In low-light areas, the face images obtained by traditional face recognition algorithms are often unclear and lose a lot of detail information, making feature extraction extremely difficult and severely affecting the accuracy of recognition. When the light is directly strong, the image is prone to overexposure, which will also destroy the key features of the face, leading to recognition failure.

[0004] The dynamic changes of face posture and expression also pose a great obstacle to recognition. In real-world scenarios, people's head posture changes constantly, such as nodding, shaking, and turning their heads, which makes the angle and position of the face in the image constantly change. Moreover, the rapid changes of micro-expressions cannot be ignored either. These subtle changes in expression may contain important emotional information, but due to their rapid changes and small amplitude, traditional algorithms are difficult to accurately capture and analyze these dynamic features. When the face posture and expression change dramatically, the recognition performance of traditional face recognition systems will be greatly reduced, and the identity cannot be accurately recognized.

[0005] Obstruction is also common in practical applications. Partial obstruction, such as wearing a mask, glasses, and a hat, will directly cover the key parts of the face, making it impossible for recognition algorithms based on complete face features to work properly. In some special scenarios, such as medical settings and industrial environments, people often need to wear protective equipment, which further increases the difficulty of recognition. At the same time, background noise interference can also affect face recognition, and complex background patterns, light and shadow interlacing, etc. may interfere with the accurate extraction of face features by algorithms, leading to misidentification.

[0006] In the face of the above problems, although some existing improvement methods can alleviate some problems to a certain extent, there are still many deficiencies. Some methods aim at the problem of illumination and use a simple brightness adjustment algorithm. This method cannot fundamentally solve the problem of non-uniform illumination and has poor effect in complex illumination scenes. When dealing with changes in posture and expression, some algorithms increase the feature dimension, but this brings the problem of dramatic increase in computational complexity, leading to poor real-time performance of the system and difficulty in meeting the demand for fast recognition in actual application. For the problems of occlusion and background noise, existing occlusion processing algorithms often rely on specific occlusion models, which have poor universality and cannot effectively cope with new occlusion situations. Therefore, it is urgent to develop a method and system for stable, accurate and efficient face recognition in dynamic environments. This not only can promote the in-depth application of face recognition technology in more fields, but also can provide stronger protection for the safety and convenience of society. SUMMARY

[0007] The purpose of the present application is to provide a face recognition method and system for dynamic environments to solve the problems raised in the background art.

[0008] To achieve the above purpose, the present application provides the following technical solution: a face recognition method for dynamic environments, the method comprising:

[0009] Step 1: Acquire face video stream containing visible light and near-infrared band through a multispectral imaging device, dynamically denoise and correct distortion of the original data, eliminate motion blur by using an adaptive smoothing algorithm based on spatiotemporal filtering, and enhance the brightness of low-illumination areas through a non-uniform illumination compensation model;

[0010] Step 2: Construct a multiscale spatiotemporal fusion feature extraction network, embed a bidirectional optical flow estimation module in a three-dimensional convolutional neural network, extract dynamic spatiotemporal features of posture changes and micro-expressions in the face sequence, and dynamically weight and fuse cross-modal features by using a channel attention mechanism;

[0011] Step 3: Establish an environmental disturbance simulation generation model, generate virtual samples with different illumination intensities, occlusion ratios and background noises by using a conditional generative adversarial network, and map the generated samples and real scene feature distributions to a shared hidden space by using a domain self-adaptive alignment algorithm;

[0012] Step 4: Design an online incremental feature update mechanism, collect scene change data in real time based on a sliding window strategy, dynamically optimize feature encoder parameters by using a contrastive self-supervised learning framework, and prevent historical feature degradation by using an elastic weight solidification algorithm;

[0013] Step 5: Deploying a heterogeneous graph neural network for multi-modal decision fusion, constructing visible light features, infrared thermal imaging features, and depth sensor point cloud features as heterogeneous nodes, modeling cross-modal feature correlation using edge attention mechanism, and generating robust joint representation;

[0014] Step 6: Using a hierarchical verification architecture to complete identity recognition, first calculating the face feature similarity threshold through a lightweight twin network, and then using a ternary verification module based on hyper-spherical metric learning to perform secondary discrimination on suspected matching results.

[0015] Preferably, the non-uniform illumination compensation model in step 1 uses an improved Retinex decomposition method to decompose the image into a reflection component and an illumination component, uses a variational energy function to constrain the smoothness of the illumination component, and predicts an illumination compensation coefficient matrix through a depth separable convolutional network.

[0016] Preferably, the bidirectional optical flow estimation module in step 2 uses the RAFT algorithm framework, which fuses ConvLSTM units in the feature pyramid to capture temporal dependencies, and uses a robust optimization strategy based on Huber loss for optical flow residual calculation.

[0017] Preferably, the domain adaptive alignment algorithm in step 3 uses Wasserstein distance to measure the distribution difference between the sample and the real sample, adds a spectral normalization layer in the generator network, and constructs an adversarial training mechanism through a gradient reversal layer.

[0018] Preferably, the contrastive self-supervised learning framework in step 4 uses a momentum contrast memory bank mechanism, introduces a temporal continuity constraint when constructing positive and negative sample queues, and uses a cosine similarity-based information noise contrast estimation loss function for feature similarity measurement.

[0019] Preferably, the edge attention mechanism in step 5 uses a multi-head graph attention network to achieve, defines the inter-modal correlation as a learnable parameter, and fuses a gated recurrent unit in graph convolution operation to model the temporal evolution law of cross-modal features.

[0020] Preferably, the hyper-spherical metric learning in step 6 uses a dynamic radius adjustment strategy, optimizes the inter-class interval in the training phase through an adaptive radius scaling algorithm, and uses an angular distance-based soft boundary classifier in the verification phase.

[0021] Preferably, the adaptive radius scaling algorithm constructs a differentiable radius prediction network, takes feature distribution tightness as input conditions, outputs the hyper-spherical radius adjustment coefficient of each class, and uses a weighted combination of margin loss and radius regularization term as the loss function.

[0022] Preferably, the multispectral imaging device comprises a polarized light imaging unit, a polarization state analysis sub-step is added before step 1, a mirror reflection component is separated by using a Stokes vector decomposition algorithm, and a depolarization feature map is constructed as an auxiliary input channel.

[0023] Preferably, the present application also includes a face recognition system for a dynamic environment, comprising:

[0024] The data acquisition and preprocessing module: the face video stream containing visible light and near-infrared waveband is collected by the multispectral imaging device, the original data is processed by using a dynamic noise reduction and distortion correction unit, the adaptive smoothing algorithm based on space-time domain joint filtering is used to eliminate motion blur, and the non-uniform illumination compensation model is used to enhance the brightness of the low-illumination area;

[0025] The multiscale space-time fusion feature extraction module: a multiscale space-time fusion feature extraction network is constructed, a bidirectional optical flow estimation module is embedded in a three-dimensional convolutional neural network, to extract dynamic space-time features of posture changes and micro-expressions in the face sequence, and a channel attention mechanism is used to dynamically weight and fuse the cross-modal features;

[0026] The virtual sample generation and domain adaptation module: an environmental disturbance simulation generation model is established, a conditional generative adversarial network is used to generate virtual samples of different light intensities, occlusion ratios and background noises, and a domain adaptation alignment algorithm is used to map the generated samples and real scene feature distributions to a shared hidden space;

[0027] The online incremental feature updating module: an online incremental feature updating mechanism is designed, scene change data is collected in real time based on a sliding window strategy, a contrastive self-supervised learning framework is used to dynamically optimize the feature encoder parameters, and an elastic weight solidification algorithm is used to prevent historical feature degradation;

[0028] The multi-modal decision fusion module: a heterogeneous graph neural network is deployed for multi-modal decision fusion, visible light features, infrared thermal imaging features and depth sensor point cloud features are constructed as heterogeneous nodes, an edge attention mechanism is used to model the correlation between cross-modal features, and a robust joint representation is generated;

[0029] The hierarchical verification module: the hierarchical verification architecture is used to complete identity recognition, the face feature similarity threshold is calculated by using a lightweight twin network first, and then a ternary tuple verification module based on hyperspherical metric learning is used to perform secondary discrimination on the suspected matching results.

[0030] Compared with the prior art, the present application has the following advantages:

[0031] In the data acquisition and preprocessing link, the multispectral imaging equipment collects face video stream containing visible light and near-infrared waveband, combines dynamic noise reduction, distortion correction, adaptive smoothing algorithm of space-time domain joint filtering and non-uniform illumination compensation model, and can effectively improve the image quality. Through the improved Retinex decomposition method for non-uniform illumination compensation, the reflection component and the illumination component are accurately separated, the compensation coefficient matrix is predicted by using the variational energy function and the smoothness constraint and the depth separable convolution network, and even in complex lighting environment, the face details can be clearly presented, providing a high-quality image basis for subsequent feature extraction, and greatly improving the recognition accuracy in low illumination and non-uniform illumination scene.

[0032] The construction of the multi-scale space-time fusion feature extraction network is a highlight. In the three-dimensional convolutional neural network, a bidirectional optical flow estimation module is embedded, a recurrent full convolutional transformation (RAFT) algorithm framework is adopted, a ConvLSTM unit is combined to capture the time sequence dependence, and a robust optimization strategy based on Huber loss is used for optical flow residual calculation, which can accurately extract dynamic space-time features such as facial posture changes and micro-expressions in the face sequence. At the same time, the channel attention mechanism dynamically weights and fuses the cross-modal features, fully excavates the value of different modal features, makes the system capture and analyze the facial dynamic features more accurately, improves the adaptability to different postures and expression changes, and effectively reduces the recognition errors caused by posture and expression changes.

[0033] The application of environment disturbance simulation generation model and domain adaptive alignment algorithm greatly enhances the robustness of the system. The conditional generative adversarial network is used to generate virtual samples with different light intensities, occlusion ratios and background noise, and the domain adaptive alignment algorithm based on Wasserstein distance measurement is used to map the generated samples and real scene feature distribution to a shared hidden space, so that the system can learn the features in various complex environments, simulate various adverse factors in real scenes, enhance the adaptability of the model to different environments, and reduce the interference of environmental factors on the recognition result.

[0034] The online incremental feature updating mechanism ensures the real-time and accuracy of the system. Based on the sliding window strategy, the scene change data is collected in real time, the momentum contrast memory library mechanism is used for contrast self-supervised learning framework to dynamically optimize the feature encoder parameters, and the elastic weight solidification algorithm is used to prevent the degradation of historical features, which can adapt to the dynamic changes of the scene in time, continuously update and optimize the model, and ensure that the recognition accuracy is still high under new environmental conditions.

[0035] The multi-modal decision fusion module constructs visible light features, infrared thermal imaging features and depth sensor point cloud features as heterogeneous nodes by deploying a heterogeneous graph neural network, and models the edge attention mechanism of cross-modal feature correlation by using a multi-head graph attention network to generate a robust joint representation.

[0036] The hierarchical verification architecture improves the accuracy and efficiency of identification. First, the similarity threshold of the face features is calculated by a lightweight twin network, and then a ternary verification module based on hyperspherical metric learning is used to make a secondary discrimination on the suspected matching result. The hyperspherical metric learning adopts a dynamic radius adjustment strategy. In the training stage, the adaptive radius scaling algorithm is used to optimize the inter-class interval. In the verification stage, a soft boundary classifier based on angular distance is used to quickly and accurately determine the face identity, reduce unnecessary calculation, and improve the overall identification efficiency of the system.

[0037] The present patent technology innovatively designs from data acquisition, feature extraction, environment adaptation, feature update, decision fusion to identity verification, effectively solves the problems of face recognition in dynamic environment such as illumination, posture and expression changes, occlusion and background noise, significantly improves the accuracy, robustness, real-time performance and reliability of identification, and has wide application prospect and important social and economic value. BRIEF DESCRIPTION OF DRAWINGS

[0038] Fig. 1 The working principle diagram of the face recognition method according to the present application;

[0039] Fig. 2 The working principle diagram of the non-uniform illumination compensation model;

[0040] Fig. 3 The working principle diagram of the bidirectional optical flow estimation module. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0042] Please refer to Figs. 1-3 The present application provides a technical solution: a face recognition method for dynamic environment, the method comprising:

[0043] Step 1: Collect a face video stream containing visible light and near-infrared wavebands through a multispectral imaging device. After collecting the raw data, process it using dynamic noise reduction and distortion correction techniques. Use a self-adaptive smoothing algorithm based on spatiotemporal filtering to eliminate motion blur, and use a non-uniform illumination compensation model to enhance the brightness in low-light areas.

[0044] Step 2: Construct a multiscale spatiotemporal fusion feature extraction network. Embed a bidirectional optical flow estimation module in a three-dimensional convolutional neural network to extract dynamic spatiotemporal features of pose changes and microexpressions in face sequences. Use a channel attention mechanism to dynamically weight and fuse cross-modal features.

[0045] Step 3: Establish an environmental disturbance simulation generation model. Use a conditional generative adversarial network to generate virtual samples with different light intensities, occlusion ratios, and background noise. Use a domain adaptive alignment algorithm to map the generated samples and real scene feature distributions to a shared hidden space.

[0046] Step 4: Design an online incremental feature update mechanism. Based on the sliding window strategy, collect scene change data in real time. Use a contrastive self-supervised learning framework to dynamically optimize feature encoder parameters. Use an elastic weight solidification algorithm to prevent historical feature degradation.

[0047] Step 5: Deploy a heterogeneous graph neural network for multi-modal decision fusion. Construct visible light features, infrared thermal imaging features, and depth sensor point cloud features as heterogeneous nodes. Use edge attention mechanisms to model cross-modal feature correlations and generate robust joint representations.

[0048] Step 6: Use a hierarchical verification architecture to complete identity recognition. First, calculate the face feature similarity threshold using a lightweight twin network. Then, use a three-way verification module based on hyperspherical metric learning to perform secondary discrimination on suspected matching results.

[0049] The application will be further described in conjunction with Examples 1 to 5:

[0050] Example 1:

[0051] In Step 1, the non-uniform illumination compensation model uses an improved Retinex decomposition method. Retinex theory states that an image can be decomposed into a reflection component R and an illumination component L, i.e., I = R·L. Traditional Retinex decomposition methods may not balance the relationship between illumination smoothness and detail preservation well when processing the illumination component.

[0052] This example uses a variational energy function to constrain the smoothness of the illumination component. The variational energy function E(L) can be expressed as:

[0053]

[0054] where λ1 and λ2 are weight parameters, denotes the gradient operator. The first term is used to constrain the smoothness of the illumination component, so that the illumination variation will not be too drastic; the second term ensures that the product of the decomposed reflectance component and the illumination component is as close to the original image as possible.

[0055] The illumination compensation coefficient matrix is predicted by a deep separable convolutional network. The deep separable convolutional network is composed of a deep convolutional layer and a pointwise convolutional layer. The deep convolutional layer is responsible for independent convolution of the features of each channel, and the pointwise convolutional layer is used to fuse channel information. Compared with traditional convolution, deep separable convolution can greatly reduce the amount of calculation and improve the training efficiency of the model. In the training process, a large number of images containing different illumination conditions are used as samples, which are input into the deep separable convolutional network. The network outputs the illumination compensation coefficient matrix, which is used to adjust the illumination component of the original image, thereby realizing the brightness enhancement of the low-illumination area and improving the overall quality of the image.

[0056] Embodiment 2:

[0057] In step 2, the bidirectional optical flow estimation module adopts the framework of the recurrent flow field transformation (RAFT) algorithm. The RAFT algorithm is based on the idea of feature matching and fuses ConvLSTM units in the feature pyramid to capture temporal dependencies. The feature pyramid can capture image features at different scales at the same time. Small-scale features contain more detailed information, and large-scale features contain more global information. By fusing features at different scales, the image content can be described more comprehensively.

[0058] ConvLSTM units have unique advantages when processing time series data. They can remember information from previous time steps and update based on current input, effectively capturing temporal dependencies in face sequences, such as continuous changes in facial pose and dynamic processes of micro-expressions.

[0059] In terms of optical flow residual calculation, a robust optimization strategy based on Huber loss is adopted. The Huber loss function is defined as:

[0060]

[0061] where x is the error of optical flow estimation, and δ is a hyperparameter. Compared with the traditional mean square error loss, Huber loss is similar to mean square error loss when the error is small, which can quickly converge. When the error is large, its growth rate is slower than that of mean square error loss, and it is more robust to outliers, which can effectively improve the accuracy of optical flow estimation and thus enhance the ability to capture dynamic features.

[0062] Embodiment 3:

[0063] In step 3, the domain adaptation alignment algorithm uses the Wasserstein distance metric to generate the distribution difference between the sample and the real sample. The Wasserstein distance W(P, Q) can measure the difference between two probability distributions P and Q, which is defined as:

[0064]

[0065] where Π(P, Q) is the set of all joint distributions of P and Q, represents expectation, and ||·|| represents distance metric. By minimizing the Wasserstein distance, the distribution of the generated sample can be made as close as possible to the distribution of the real sample.

[0066] A spectral normalization layer is added to the generator network. Spectral normalization normalizes the weight matrix to make the training of the generator network more stable. Specifically, for the weight matrix W, the spectral normalized weight matrix is: where σ(W) is the spectral norm of the weight matrix W.

[0067] An adversarial training mechanism is constructed through a gradient reversal layer. The gradient reversal layer is an identity mapping in forward propagation and multiplies the gradient by -1 in backward propagation. In the training process, the generator network generates virtual samples, and the discriminator network judges whether the sample is a real sample or a generated sample. Through the gradient reversal layer, the generator network learns how to generate virtual samples closer to real samples, so as to realize the mapping of generated samples and real scene feature distribution to a shared hidden space, and enhance the adaptability of virtual samples to real scenes.

[0068] Embodiment 4:

[0069] In step 4, the contrastive self-supervised learning framework uses a momentum contrast memory bank mechanism. The momentum contrast memory bank maintains a queue, which stores historical features. In the training process, the features of the current sample are compared and learned with the features in the queue. When constructing the positive and negative sample queues, a time continuity constraint is introduced, that is, samples at adjacent time points are more likely to belong to the same category. The samples at adjacent time points are taken as positive samples, and other samples are taken as negative samples.

[0070] The feature similarity metric uses a cosine similarity-based information noise contrast estimation loss function. The cosine similarity sim(x, y) is defined as: where x and y are two feature vectors. The information noise contrast estimation loss function L can be expressed as:

[0071]

[0072] where x is the current sample feature, q is the positive sample feature, k is the sample feature in the negative sample queue, K is the size of the negative sample queue, and τ is the temperature hyperparameter. By minimizing this loss function, the model can learn more discriminative feature representations. At the same time, in combination with the elastic weight solidification algorithm, when updating the feature encoder parameters, important historical weights are protected to prevent historical feature degradation, ensuring that the model does not lose previously learned useful information while continuously updating features.

[0073] Embodiment 5:

[0074] In step 5, the edge attention mechanism is implemented using a multi-head graph attention network. The multi-head graph attention network calculates attention weights in parallel through multiple heads, which can capture the relevance between cross-modal features from different angles. The inter-modal correlation is defined as a learnable parameter, and the importance between different modalities is learned through training. In the graph convolution operation, the gated recurrent unit is fused to model the temporal evolution rule of cross-modal features. The gated recurrent unit can adaptively update the cross-modal features according to the current input and the previous state, better capturing the changes of features over time.

[0075] In step 6, the hypersphere metric learning adopts a dynamic radius adjustment strategy. In the training phase, the adaptive radius scaling algorithm is used to optimize the inter-class interval. The adaptive radius scaling algorithm constructs a differentiable radius prediction network, taking the feature distribution tightness as the input condition and outputting the hypersphere radius adjustment coefficient for each class. The loss function uses a weighted combination of the margin loss and the radius regularization term. The margin loss L margin can be expressed as:

[0076] L margin = max(0, m + d(a, p) - d(a, n))

[0077] where m is the preset interval threshold, d(a, p) is the distance between the anchor sample and the positive sample, and d(a, n) is the distance between the anchor sample and the negative sample. The radius regularization term is used to prevent the hypersphere radius from being too large or too small, ensuring the stability of the model. In the verification phase, an angle distance-based soft boundary classifier is used, which can better measure the similarity between features. The soft boundary classifier can alleviate the overfitting problem to some extent and improve the accuracy of identity recognition.

[0078] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it should be taken in a descriptive sense and not a limiting sense.

[0079] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, it is to be understood that various modifications, changes, substitutions and alterations can be made to the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A face recognition method for dynamic environments, characterized in that, Includes the following steps: Step 1: Acquire face video streams containing visible and near-infrared bands using a multispectral imaging device, perform dynamic noise reduction and distortion correction on the raw data, eliminate motion blur using an adaptive smoothing algorithm based on spatiotemporal joint filtering, and enhance the brightness of low-light areas using a non-uniform illumination compensation model. Step 2: Construct a multi-scale spatiotemporal fusion feature extraction network, embed a bidirectional optical flow estimation module into a three-dimensional convolutional neural network, extract dynamic spatiotemporal features of pose changes and micro-expressions in face sequences, and use a channel attention mechanism to dynamically weight and fuse cross-modal features; Step 3: Establish an environmental disturbance simulation generation model, use conditional generative adversarial network to generate virtual samples with different light intensities, occlusion ratios and background noise, and use a domain adaptive alignment algorithm to map the generated samples and the feature distribution of the real scene to a shared latent space. Step 4: Design an online incremental feature update mechanism, collect scene change data in real time based on the sliding window strategy, dynamically optimize the feature encoder parameters using a contrastive self-supervised learning framework, and prevent historical feature degradation through an elastic weight solidification algorithm; Step 5: Deploy heterogeneous graph neural networks for multimodal decision fusion, construct heterogeneous nodes by combining visible light features, infrared thermal imaging features and depth sensor point cloud features, and use edge attention mechanism to model cross-modal feature correlation to generate robust joint representations; Step 6: Use a hierarchical verification architecture to complete identity recognition. First, calculate the facial feature similarity threshold through a lightweight twin network, and then use a triplet verification module based on hyperspherical metric learning to perform secondary discrimination on suspected matching results. In step 2, the bidirectional optical flow estimation module adopts the Cyclic All-Pair Transform (RAFT) algorithm framework, which integrates ConvLSTM units in the feature pyramid to capture temporal dependencies. The optical flow residual calculation adopts a robust optimization strategy based on Huber loss. In step 6, the hyperspherical metric learning adopts a dynamic radius adjustment strategy. During the training phase, the inter-class margin is optimized through an adaptive radius scaling algorithm, and during the verification phase, a soft boundary classifier based on angular distance is used. The adaptive radius scaling algorithm constructs a differentiable radius prediction network, takes the feature distribution density as an input condition, and outputs the hyperspherical radius adjustment coefficient for each category. The loss function is a weighted combination of interval loss and radius regularization term.

2. The face recognition method for dynamic environments according to claim 1, characterized in that, In step 1, the non-uniform illumination compensation model uses an improved Retinex decomposition method to decompose the image into reflection and illumination components. The smoothness of the illumination component is constrained by a variational energy function, and the illumination compensation coefficient matrix is ​​predicted through a depthwise separable convolutional network.

3. The face recognition method for dynamic environments according to claim 1, characterized in that, In step 3, the domain adaptive alignment algorithm uses the Wasserstein distance metric to measure the distribution difference between the generated samples and the real samples, adds a spectrum normalization layer to the generator network, and constructs an adversarial training mechanism through a gradient inversion layer.

4. The face recognition method for dynamic environments according to claim 1, characterized in that, In step 4, the contrastive self-supervised learning framework adopts a momentum contrastive memory mechanism, introduces time continuity constraints when constructing positive and negative sample queues, and uses an information noise contrastive estimation loss function based on cosine similarity to measure feature similarity.

5. The face recognition method for dynamic environments according to claim 1, characterized in that, In step 5, the side attention mechanism is implemented using a multi-head graph attention network. The intermodal correlation is defined as a learnable parameter. In the graph convolution operation, gated recurrent units are integrated to model the temporal evolution of cross-modal features.

6. The face recognition method for dynamic environments according to claim 1, characterized in that, The multispectral imaging device includes a polarization imaging unit. A polarization state analysis sub-step is added before step 1. The Stokes vector decomposition algorithm is used to separate the specular reflection component and construct a depolarization feature map as an auxiliary input channel.

7. A face recognition system for dynamic environments, characterized in that, include: Data acquisition and preprocessing module: Acquires face video streams containing visible and near-infrared bands through multispectral imaging equipment, processes the raw data using dynamic noise reduction and distortion correction unit, eliminates motion blur using an adaptive smoothing algorithm based on spatiotemporal joint filtering, and enhances brightness in low-light areas using a non-uniform illumination compensation model. Multi-scale spatiotemporal fusion feature extraction module: Construct a multi-scale spatiotemporal fusion feature extraction network, embed a bidirectional optical flow estimation module in a three-dimensional convolutional neural network to extract dynamic spatiotemporal features of pose changes and micro-expressions in face sequences, and use a channel attention mechanism to dynamically weight and fuse cross-modal features; Virtual Sample Generation and Domain Adaptation Module: Establish an environmental disturbance simulation generation model, use conditional generative adversarial network to generate virtual samples with different light intensities, occlusion ratios and background noise, and use a domain adaptive alignment algorithm to map the generated samples and the feature distribution of the real scene to a shared latent space. Online incremental feature update module: Design an online incremental feature update mechanism, collect scene change data in real time based on sliding window strategy, dynamically optimize feature encoder parameters using contrastive self-supervised learning framework, and use elastic weight solidification algorithm to prevent historical feature degradation; Multimodal decision fusion module: Deploy heterogeneous graph neural network for multimodal decision fusion, construct heterogeneous nodes by using visible light features, infrared thermal imaging features and depth sensor point cloud features, and use edge attention mechanism to model cross-modal feature correlation to generate robust joint representation; Hierarchical verification module: The hierarchical verification architecture is used to complete identity recognition. First, a lightweight twin network is used to calculate the facial feature similarity threshold. Then, a triplet verification module based on hyperspherical metric learning is used to perform secondary discrimination on suspected matching results. The bidirectional optical flow estimation module in the multi-scale spatiotemporal fusion feature extraction module adopts the Cyclic All-Pair Transform (RAFT) algorithm framework, which fuses ConvLSTM units in the feature pyramid to capture temporal dependencies. The optical flow residual calculation adopts a robust optimization strategy based on Huber loss. In the hierarchical verification module, the hyperspherical metric learning adopts a dynamic radius adjustment strategy. During the training phase, the inter-class margin is optimized through an adaptive radius scaling algorithm, and during the verification phase, a soft boundary classifier based on angular distance is used. The adaptive radius scaling algorithm constructs a differentiable radius prediction network, takes the feature distribution density as an input condition, and outputs the hyperspherical radius adjustment coefficient for each category. The loss function is a weighted combination of interval loss and radius regularization term.

Images