A face recognition method and system based on cross-age features

Abstract

This invention discloses a face recognition method and system based on cross-age features, belonging to the field of face recognition technology. It addresses the technical problem that traditional face recognition is slow, and prone to errors and slow recognition for elderly people who need to use facial recognition for medical treatment, affecting their efficiency. The method includes: labeling the initial facial image with dermal texture features and variable epidermal features to determine the facial cortical feature image; dynamically stripping the variable epidermal features from the facial cortical feature image and applying physical optical constraints to the dermal texture features; extracting facial feature points based on topological structure and shape from the optimized facial image; performing age attribute editing diffusion processing on the optimized facial image; and performing feature dimensionality reduction and restoration processing on key facial features to determine the social security identity information of the patient.

Description

Technical Field

[0001] This application relates to the field of face recognition, and in particular to a face recognition method and system based on cross-age features. Background Technology

[0002] With the rapid development of medical automation, seeing a doctor and buying medicine have become more convenient. Using technologies such as facial recognition, the entire process, such as registration, picking up medicine, and payment, can be completed. Even without an ID card or social security card, one can still complete "face-scanning medical treatment".

[0003] A large proportion of those seeking medical care are elderly. Complex medical procedures may require more paperwork, and the elderly are also more likely to forget to bring their social security cards or ID cards. In emergencies, such as sudden illness, they often rush to seek medical attention and neglect to bring valid identification documents. Therefore, "facial recognition for medical care" is an effective way to solve this problem, greatly reducing the number of procedures and the need for documents; the elderly only need to verify their identity information by scanning their face.

[0004] However, the identity information entered by the elderly is often quite old. Social security or identity information is often facial information from a long time ago, which differs significantly from the facial features at the current time of medical treatment. When performing facial recognition on the elderly, it is easy to cause facial recognition errors or failure to recognize facial information. Traditional facial recognition algorithms are slow and may require multiple attempts to correctly recognize facial information, which can easily lead to slow medical treatment or delays in registration time. In other words, traditional facial recognition relies heavily on static texture features and has certain limitations. In cross-age recognition, the changes in facial features brought about by age changes will affect the accuracy of facial recognition of the current user to a certain extent, thereby affecting the user's medical treatment efficiency. Summary of the Invention

[0005] This application provides a face recognition method and system based on cross-age characteristics to solve the following technical problem: For elderly people who need to use facial recognition for medical treatment, traditional face recognition is relatively slow, and it is easy to cause slow recognition and recognition errors for people with a large age range, which affects the efficiency of users' medical treatment.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] On one hand, this application provides a face recognition method based on cross-age features, including: labeling the initial facial image of a medical patient with dermal texture features and variable epidermal features to determine a facial cortical feature image; dynamically stripping the variable epidermal features in the facial cortical feature image and applying physical optical constraints to the dermal texture features to obtain an optimized facial image; extracting facial feature points related to topology and shape from the optimized facial image using a multi-scale temporal deformation model to obtain facial deformation features; performing age attribute editing diffusion processing on the optimized facial image to determine facial growth constraint features; performing attention fusion processing on the facial deformation features and facial growth constraint features using a preset feature enhancement extraction network and outputting key facial features; performing feature dimensionality reduction and restoration processing on the key facial features related to the real facial image to obtain cross-age facial features; and performing identity information query processing on the cross-age facial features to determine the social security identity information of the medical patient.

[0008] This application's embodiments, through labeling and processing dermal texture features and variable epidermal features, can more accurately extract age-related facial features, which is crucial for cross-age face recognition. Utilizing dynamic stripping of variable epidermal features and physically optically constrained dermal texture features helps optimize facial images, reduce interference from non-age-related features, and improve recognition accuracy. Furthermore, using a multi-scale temporal deformation model to extract facial feature points can capture changes in facial shape at different ages, thereby improving recognition robustness. Editing age attributes can also determine facial growth constraint features, aiding in the identification of individuals of different age groups, especially when there are significant age variations. The feature enhancement extraction network, through attention fusion processing of facial deformation features and growth constraint features, can improve feature quality and enhance the performance of the recognition system. Dimensionality reduction of key facial features reduces computational complexity while retaining key information, improving recognition efficiency. Simultaneously, by extracting cross-age facial features, effective identification between individuals of different age groups can be achieved, expanding the application scope of face recognition. Ultimately, by querying and processing identity information, the social security identity information of patients can be quickly and accurately determined, thereby improving the efficiency and security of medical services.

[0009] In one feasible implementation, the initial facial image of the patient is labeled with dermal texture features and variable epidermal features to determine the facial dermal feature image. Specifically, this includes: sequentially illuminating the patient's facial area with light of different polarization angles using a multi-band polarized light source pre-installed in the smart medical mutual aid machine; and simultaneously acquiring multi-angle polarized images of the facial area using a preset high-resolution polarization camera; based on the infrared band of the multi-band polarized light source, performing penetration imaging processing on the anisotropic texture in the multi-angle polarized images to determine the dermal texture features; wherein, the dermal texture features include: Facial pore distribution characteristics and facial blood vessel distribution characteristics; using a non-polarized reference light source, ambient light interference is compensated for on the multi-angle polarized image, and feature marking processing under high scattering characteristics is performed on the multi-angle polarized image to determine the variable epidermal layer characteristics; the variable epidermal layer characteristics include: wrinkle distribution characteristics, spot distribution characteristics, and pigmentation distribution characteristics; based on the anisotropic texture and the multi-angle polarized image processed with high scattering characteristics, the initial facial image is determined; according to the RGB color channels, the dermal texture characteristics and the variable epidermal layer characteristics in the initial facial image are subjected to dimensional layering processing to generate the facial skin layer feature image.

[0010] This application's embodiments achieve age-robust dermal biometric extraction by decoupling polarized light physical property analysis with neural network features. This solves problems such as feature drift caused by interference from age-related epidermal features (wrinkles, age spots, etc.) in traditional facial recognition, the inability of existing optical imaging methods to separate the multi-layered scattering characteristics of skin tissue, and the lack of physical model constraints in conventional adversarial networks, leading to uncontrollable feature decoupling processes. Furthermore, by utilizing RGB color channels, it achieves dimensional layering of dermal texture features and variable epidermal features, facilitating the generation of facial dermal feature images with multi-dimensional textural characteristics.

[0011] In one feasible implementation, before dynamically stripping the variable epidermal features from the facial cortical feature image and applying physical optical constraints to the dermal texture features to obtain an optimized facial image, the method further includes: inputting a historical set of facial surface feature images into the generator's U-Net architecture; wherein the historical facial surface images are the input, and the dermal texture features and variable epidermal features are both outputs; based on the dimensional hierarchical characteristics between the dermal texture features and the variable epidermal features, a cortical discriminator with a dual-branch structure is constructed; wherein the cortical discriminator is used to distinguish the physiological texture authenticity of the dermal texture features and the dynamic noise distribution of the variable epidermal features; performing adversarial network generation processing on the U-Net architecture and the cortical discriminator to obtain an adversarial decomposition network; according to Ladv =E[logD(G(x))]+E[log(1-D(x))], to obtain the adversarial loss L of the adversarial decomposition network. adv Wherein, D is the cortical discriminator; G(x) are the dermal texture features and variable epidermal features output by the generator; E is a mathematical symbol representation; using a preset polarization transfer function as a regularization term, the dermal texture features and variable epidermal features are subjected to dimensional level constraint loss calculation related to the RGB color channels to obtain the physical constraint loss; the cyclic consistency between the dermal texture features and the variable epidermal features is controlled, and the cyclic consistency loss is determined; the adversarial decomposition network is adversarially trained using the adversarial loss, the physical constraint loss, and the cyclic consistency loss to obtain the optimized adversarial decomposition network.

[0012] This application's embodiments achieve the physical interpretability separation of dynamic noise in the facial epidermis (such as light reflection and makeup interference) from biological features in the dermis (such as capillary distribution and collagen fiber orientation) through the collaborative design of the U-Net architecture and a dual-branch discriminator. By using the polarization transfer function as a regularization term, the physiological authenticity of dermal texture features is more accurately determined. Introducing a dynamic noise distribution discrimination module in the epidermis during adversarial training, and through joint optimization of adversarial loss and physical constraint loss, effectively suppresses the impact of environmental lighting changes (such as highlights / shadows) and short-term epidermal state fluctuations (such as sweat and oil) on core biological features. By implementing closed-loop reconstruction of epidermal-dermal dual-channel features through cyclic consistency constraints, a mathematical mapping relationship between the light field transfer equation and biological tissue characteristics is established, which is beneficial for adversarial training of the resistance decomposition network and increases the network's adversarial performance.

[0013] In one feasible implementation, the variable epidermal features in the facial skin feature image are dynamically stripped, and the dermal texture features are physically optically constrained to obtain an optimized facial image. Specifically, this includes: calculating the reflectivity of the current facial skin feature image based on polarization angle and degree of polarization to obtain a polarization-reflection relationship based on high scattering characteristics and anisotropic texture; using an optimized adversarial decomposition network and based on the polarization-reflection relationship, providing real-time feedback on the signal-to-noise ratio of the variable epidermal features in the facial skin feature image, and dynamically adjusting the attention weights of the skin discriminator to complete the dynamic stripping of the variable epidermal features; based on non-negative matrix factorization, suppressing cross-interference in the dynamic stripping results to obtain the optimized stripped dermal texture features; applying physical optical constraints of Fresnel reflection boundaries to the dermal texture features, and performing adaptive histogram equalization on the image features corresponding to the dermal texture features to obtain the optimized facial image with invariant texture features.

[0014] This application's embodiments, by dynamically stripping volatile epidermal layer features, can remove unstable features caused by factors such as illumination and facial expression changes, thereby improving the stability and accuracy of face recognition. Calculating the reflectivity of polarization angle and degree of polarization, and analyzing the polarization-reflection relationship between high scattering characteristics and anisotropic textures, helps to gain a deeper understanding of the optical properties of the facial surface, providing a basis for subsequent processing. It can provide real-time feedback on the signal-to-noise ratio of volatile epidermal layer features and dynamically adjust the attention weights of the cortical discriminator, enabling real-time adaptation to image changes and improving the real-time performance and adaptability of the recognition system. Furthermore, an optimized adversarial decomposition network can be used to more effectively separate and extract facial features, enhancing the performance of the recognition system.

[0015] In one feasible implementation, a multi-scale temporal deformation model is used to extract facial feature points from the optimized facial image based on its topological structure and shape to obtain facial deformation features. Specifically, this includes: extracting feature points from the optimized facial image based on key facial feature points in a face recognition model to obtain original facial feature points; wherein the key facial feature points include at least: mouth feature points, eye feature points, nose feature points, facial contour feature points, cheekbone contour feature points, and jaw angle feature points; using the short-term dynamic scale in the multi-scale temporal deformation model, filtering the neuromuscular control-related muscular feature points from the original facial feature points based on age-related factors to obtain muscular feature points unrelated to age-related factors; based on the multi-scale... In the intermediate dynamic scale of the time-series deformation model, a non-rigid algorithm is used to calculate the distance between the original facial feature points and the skeletal development feature points to obtain the feature point distance term. Based on the recombination transformation of the feature point distance term, the key point term is determined. Through the arc-node correlation matrix, a rigid calculation of the topology of the skeletal development feature points is performed on the original facial feature points to obtain the feature point rigid term. The feature point distance term, the feature point rigid term, and the key point term are subjected to a double-layer loop iterative calculation based on the energy function to determine the facial bone deformation trajectory. Through the topology-preserving manifold learning algorithm, the facial evolution model is performed on the texture feature points and the facial bone deformation trajectory to obtain the facial evolution model. The facial deformation features in the facial evolution model are then extracted.

[0016] This application's embodiments effectively decouple age-related deformation from individual inherent characteristics by integrating a dual mechanism of short-term dynamic scale screening and mid-term skeletal development modeling. Employing dynamic screening techniques that control muscular texture features through neuromuscular means (such as LBP texture analysis) eliminates the interference of age factors on skin texture, improving the robustness of the face recognition system in cross-age scenarios. Through rigid constraint calculation of the arc-node correlation matrix and bi-layer iterative optimization of the energy function, the stability of the facial topology is maintained while capturing nonlinear skeletal deformation.

[0017] In one feasible implementation, the optimized facial image undergoes age attribute editing diffusion processing to determine facial growth constraint features. Specifically, this includes: identifying existing target attribute regions in the optimized facial image using a GAN model and a diffusion model; and identifying supplementary attribute regions in the optimized facial image that are missing facial features and require attribute additions through complementary attention branches. The target attribute regions are attribute regions whose facial features do not change significantly with age; the supplementary attribute regions are attribute regions whose facial features change significantly with age. Using the attention mask of the GAN model, the supplementary attribute regions undergo cross-age region feature generation processing to obtain a first mask image. Based on a color mask, the target attribute regions undergo cross-age region age change effect processing to obtain a second mask image. The first mask image and the second mask image are combined using a masking mechanism to generate a facial growth image after age editing. Key facial growth features in the facial growth image are constrained under an attention mechanism, and facial growth constraint features are extracted.

[0018] This application's embodiments, through the combined use of GAN and diffusion models, can more efficiently complete age editing processing, generating higher-quality, more natural facial growth images, and better utilize attention masks to capture age-related attributes. Simultaneously, by combining attribute regions with insignificant and significant facial feature changes, it can more accurately predict facial features of individuals in historical age groups. Furthermore, by incorporating constraint processing under the attention mechanism, key facial growth features can be accurately constrained, thereby reducing computational overhead and improving computer processing speed.

[0019] In one feasible implementation, a preset feature enhancement extraction network is used to perform attention fusion processing on the facial deformation features and the facial growth constraint features, and output key facial features. Specifically, this includes: using a spatial attention network to extract spatial features from the facial deformation features and the facial growth constraint features respectively, obtaining a first spatial feature and a second spatial feature; using a channel attention network to extract channel features from the first spatial feature and the second spatial feature respectively, obtaining a first channel feature and a second channel feature; and based on the channel importance determined by expert scoring, the first channel feature and the second channel feature are... Weights are assigned to the two-channel features, and cross-age channel features with cross-age facial feature maps are labeled. Based on the fusion loss function and the cross-age channel features, the facial deformation features and the facial growth constraint features are merged under multi-head attention. The distributed facial features, learned and trained by a fully connected layer, are predicted and mapped into the output space to generate a cross-age facial image corresponding to the patient. The cross-age facial image is a predicted facial image of the patient's historical age range. Based on the key facial feature points in the face recognition model, feature points are extracted from the cross-age facial image to obtain the key facial features.

[0020] This application embodiment performs attention fusion processing on facial deformation features and facial growth constraint features. On the basis of multi-head attention, a fusion loss forced attention mechanism is added. The facial deformation features that do not change with age after adversarial decomposition and the facial growth constraint features after age editing are integrated into the feature fusion processing, thereby generating key facial features across age stages. Combining these key facial features, the corresponding person's facial image can be quickly and accurately identified in the facial information of subsequent historical age groups. At the same time, through spatial attention and channel attention, as well as the response of the feature enhancement extraction network to key feature information, multiple different attention regions can be captured, realizing the spatiotemporal fusion of facial features, thereby outputting more realistic cross-age facial images.

[0021] In one feasible implementation, the key facial features are subjected to feature dimensionality reduction and restoration processing based on a real facial image to obtain cross-age facial features, specifically including: according to

[0022] Obtain the cross-age facial features for each age group. Where k is the feature type in the facial key features, N is the cross-age change attribute parameter corresponding to the facial deformation feature in the facial key features, C is the cross-age change attribute parameter corresponding to the facial growth constraint feature in the facial key features, and i and j are both mathematical constants; The pooling window size is the size of the real facial image; D age V represents the age feature size parameter for facial images across different ages; age For age ranges.

[0023] In one feasible implementation, the cross-age facial features are processed to query identity information and determine the social security identity information of the medical patient. Specifically, this includes: extracting cross-age facial features corresponding to the medical patient; generating identity codes from the cross-age facial features to obtain the identity code information to be queried; inputting the identity code information to be queried into the social security patient information database, and arranging the identity information within a matching threshold range according to the matching degree; sequentially displaying the social security identity information to be determined according to the order of the identity information; and obtaining the social security identity information matching the medical patient based on the manually confirmed social security identity information.

[0024] On the other hand, this application embodiment also provides a face recognition system based on cross-age features. The system includes: a texture recognition module, used to perform labeling processing on the initial facial image of the medical patient, involving dermal texture features and variable epidermal features, to determine a facial cortical feature image; dynamically stripping the variable epidermal features in the facial cortical feature image and applying physical optical constraints to the dermal texture features to obtain an optimized facial image; a facial topology recognition module, used to extract facial feature points related to topological structure and shape from the optimized facial image to obtain facial deformation features; a facial growth prediction module, used to perform age attribute editing diffusion processing on the optimized facial image to determine facial growth constraint features; a feature fusion module, used to perform attention fusion processing on the facial deformation features and the facial growth constraint features, and output key facial features; and an identity information query module, used to perform feature dimensionality reduction and restoration processing on the key facial features related to the real facial image to obtain cross-age facial features; and perform identity information query processing on the cross-age facial features to determine the social security identity information of the medical patient.

[0025] This application provides a face recognition method and system based on cross-age features. Compared with the prior art, the embodiments of this application have the following beneficial technical effects:

[0026] 1. Feature labeling and processing: By labeling and processing the texture features of the dermis and the variable features of the epidermis, age-related facial features can be extracted more accurately, which is crucial for cross-age face recognition.

[0027] 2. Dynamic peeling and physical optical constraints: Dynamic peeling of variable epidermal features and physical optical constraints on dermal texture features help optimize facial images, reduce interference from non-age-related features, and improve recognition accuracy.

[0028] 3. Multi-scale temporal deformation model: Using a multi-scale temporal deformation model to extract facial feature points can capture changes in facial shape at different ages, thereby improving the robustness of recognition.

[0029] 4. Diffusion processing of age attribute editing: By editing the age attribute, facial growth constraint features can be determined, which helps to identify individuals of different age groups, especially when there are large age variations.

[0030] 5. Feature Enhancement Extraction Network: The feature enhancement extraction network processes facial deformation features and growth constraint features through attention fusion, which can improve the quality of features and enhance the performance of the recognition system.

[0031] 6. Feature Dimensionality Reduction and Restoration Processing: Dimensionality reduction processing of key facial features can reduce computational complexity while retaining key information and improving recognition efficiency.

[0032] 7. Cross-age facial feature extraction: By extracting cross-age facial features, this method can effectively identify individuals of different ages, expanding the application scope of face recognition.

[0033] 8. Identity Information Inquiry and Processing: Ultimately, through identity information inquiry and processing, the social security identity information of medical patients can be quickly and accurately determined, improving the efficiency and security of medical services. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0035] Figure 1 A flowchart of a face recognition method based on cross-age features is provided for embodiments of this application;

[0036] Figure 2 A schematic diagram of a recognition network structure with fusion feature enhancement provided in an embodiment of this application;

[0037] Figure 3 This is a schematic diagram of the structure of a face recognition device based on cross-age characteristics, provided as an embodiment of this application. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0039] This application provides a face recognition method based on cross-age features, such as... Figure 1 As shown, the face recognition method based on cross-age features specifically includes steps S101-S106:

[0040] S101. The initial facial images of the patients are labeled with relevant dermal texture features and variable epidermal features to determine the facial cortical feature image.

[0041] It should be noted that in applications of polarized light systems, facial skin computer-aided imaging systems utilize both cross-polarized and parallel-polarized light sources. Parallel-polarized light is primarily used to enhance detailed features on the skin surface, including fine lines, pores, and blemishes. Cross-polarized light is used to enhance tissue features beneath the skin's surface, including pigmentation and blood vessels. Facial skin computer-aided imaging systems typically analyze white light and polarized light images using specific algorithms, ultimately presenting the calculation results as images and data in multiple modes.

[0042] Specifically, the first step is to use a multi-band polarized light source pre-installed in a smart medical mutual aid machine (such as a smart self-service window) to sequentially illuminate the face of the patient with light of different polarization angles, and then use a pre-set high-resolution polarization camera to simultaneously acquire multi-angle polarized images of the face.

[0043] Furthermore, based on the infrared band of a multi-band polarized light source, the anisotropic textures in the multi-angle polarized images are subjected to penetration imaging processing to determine the dermal texture features. These dermal texture features include: facial pore distribution features and facial blood vessel distribution features.

[0044] Furthermore, it is necessary to compensate for ambient light interference in the multi-angle polarized images using a non-polarized reference light source, and to perform feature marking processing on the multi-angle polarized images under high scattering characteristics to determine the variable epidermal layer features. The variable epidermal layer features include: wrinkle distribution features, spot distribution features, and pigmentation distribution features.

[0045] Furthermore, the initial facial image was determined based on the multi-angle polarized image processed with anisotropic texture and high scattering characteristics.

[0046] Furthermore, based on the RGB color channels, the dermal texture features and variable epidermal features in the initial facial image are subjected to dimensional layering processing to generate a facial cortical feature image. This means that the facial cortical image can be more consistently linked to the patient's identity information, while also enabling deep recognition of detailed facial features, preventing malicious use of other people's faces.

[0047] In one embodiment, the multi-angle polarization light source array can employ a 450nm / 630nm dual-wavelength LED ring light source, configured with a 0° / 45° / 90° / 135° four-way linear polarizer. A high-resolution polarization camera can be equipped with a rotating polarizer module (0-180° continuous rotation) to acquire 16-bit RAW format polarization image sequences. An optical pre-processing module can also be pre-loaded: obtaining the degree of polarization (DoP) map and angle of polarization (AoP) map through Stokes vector calculation. Then, a layered reflection model of skin tissue can be established. This model can utilize the infrared band of the multi-band polarization light source and a non-polarized reference light source, combined with the epidermal / dermal layer reflectance coefficient matrix and dermal layer scattering coefficient, to identify dermal texture features and variable epidermal features respectively. Then, using RGB color channels, the identified dermal features located at different dimensional levels are layered and labeled to generate a labeled facial dermal feature image.

[0048] S102. Dynamically peel off the volatile epidermal features in the facial skin feature image and apply physical optical constraints to the dermal texture features to obtain an optimized facial image.

[0049] Specifically, the historical facial surface feature image set needs to be input into the generator's U-Net architecture first. The historical facial surface images are the input, while the dermal texture features and variable epidermal features are the outputs.

[0050] In one embodiment, a set of historical facial surface feature images (RGB images captured under polarized light, 1024×1024 resolution) can be input first to complete the preliminary training of the adversarial decomposition network. In the generator (G), a U-Net architecture is adopted, with the facial surface image as input and two-branch features as output: dermal texture features, including physiological textures (such as capillary distribution and pore structure); and variable epidermal features, including dynamic noise (such as sweat, oil, and short-term wrinkles). For example, given a multi-angle facial image with polarization angles of 0°, 45°, 90°, and 135° as input, the generator outputs a dermal feature map (512×512 grayscale image) and an epidermal noise map (dynamically changing RGB noise distribution).

[0051] Furthermore, based on the dimensional hierarchical characteristics between dermal texture features and variable epidermal features, a cortical discriminator with a dual-branch structure is constructed. The cortical discriminator is used to distinguish between the physiological texture authenticity of dermal texture features and the dynamic noise distribution of variable epidermal features.

[0052] In one embodiment, regarding the construction of the two-branch cortical discriminator (D), a dermal discriminator branch can be configured first: this branch determines whether the generated dermal texture conforms to physiological realism (e.g., the continuity and fractal characteristics of capillaries). Then, an epidermal discriminator branch is configured: this branch determines whether the dynamic noise distribution conforms to the statistical regularity of natural epidermal noise (e.g., a Gaussian-Poisson mixture distribution). An adversarial design is then incorporated: the generator and discriminator are optimized through adversarial training (GAN), with the generator attempting to deceive the discriminator, and the discriminator improving its discrimination ability. Then, in subsequent cortical discrimination, the dermal discriminator extracts texture orientation consistency features through a convolutional network, while the epidermal discriminator verifies the noise spectrum distribution through frequency domain analysis.

[0053] Furthermore, the U-Net architecture and cortical discriminator are subjected to adversarial network generation processing to obtain the adversarial decomposition network. Based on L... adv =E[log D(G(x))]+E[log(1-D(x))], which gives the adversarial loss L of the adversarial decomposition network. adv Where D is the cortical discriminator; G(x) are the dermal texture features and variable epidermal features output by the generator; and E is a mathematical symbol representation.

[0054] Furthermore, using a pre-defined polarization transfer function as a regularization term, the dimensional constraint loss related to the RGB color channels of the dermal texture features and the variable epidermal features is calculated to obtain the physical constraint loss. Cyclic consistency is then maintained between the dermal texture features and the variable epidermal features, and the cyclic consistency loss is determined. Finally, the adversarial decomposition network is adversarially trained using the adversarial loss, physical constraint loss, and cyclic consistency loss to obtain the optimized adversarial decomposition network.

[0055] As a feasible implementation method, in order to constrain the reflection characteristics of the dermis and epidermis in the RGB channels based on the polarization transfer function:

[0056] The physical constraint loss (Physics-based Loss) L can be obtained. phy , where Φ c This refers to the channel constraint corresponding to the polarization transfer function. To ensure that the original input image can be recovered after the dermis and epidermis features are reconstructed: L cycle =||G -1(G(x))-x||1, we can obtain the Cycle Consistency Loss.

[0057] Furthermore, reflectivity calculations related to polarization angle and degree of polarization are performed on the current facial surface feature image to obtain the polarization reflection relationship based on high scattering characteristics and anisotropic texture.

[0058] Furthermore, through an optimized adversarial decomposition network and based on polarization reflection relationship, the signal-to-noise ratio of volatile epidermal features in facial surface feature images is fed back in real time, and the attention weight of the cortical discriminator is dynamically adjusted to complete the dynamic stripping of volatile epidermal features.

[0059] Furthermore, based on nonnegative matrix factorization, the dynamic stripping results are subjected to cross-interference suppression processing to obtain optimized dermal texture features after stripping.

[0060] In one embodiment, polarization reflection parameters of high scattering characteristics (dermis) and anisotropic texture (epidermis) are calculated based on the polarization angle and degree of polarization of the facial surface. For example, if the degree of polarization of the epidermis is lower than that of the dermis (due to oil scattering), the discriminator weights are dynamically adjusted using the polarization reflection coefficient. Then, nonnegative matrix factorization (NMF) is performed to suppress cross-interference, which involves performing nonnegative decomposition on the feature matrix of the peeled dermis to eliminate residual epidermal noise.

[0061] Furthermore, the Fresnel reflection boundary is physically and optically constrained for the dermal texture features, and adaptive histogram equalization is performed on the image features corresponding to the dermal texture features to obtain an optimized facial image with invariant texture features. Here, invariant texture features are facial texture features that do not change with age.

[0062] In one embodiment, the skin refractive index in Fresnel constraints can be used to constrain the physical reflection boundaries of the dermal texture, suppressing optical artifacts. Then, adaptive histogram equalization is performed, which locally enhances the contrast of the dermal image, highlighting facial texture features that do not change significantly with age.

[0063] As a feasible implementation method, suppressing cross-interference through non-negative matrix factorization ensures purer dermal texture features after peeling, reducing the impact of noise and interference. Furthermore, applying physical optical constraints on Fresnel reflection boundaries to the dermal texture features simulates optical behavior in real-world environments, improving the robustness of the recognition system under complex lighting conditions. Simultaneously, adaptive histogram equalization of image features improves image contrast, making the image visually more uniform, which is beneficial for subsequent feature extraction and recognition. The resulting optimized facial image possesses immutable textural features, meaning that these features remain stable even as facial expressions or lighting conditions change, contributing to improved accuracy in face recognition.

[0064] S103. Using a multi-scale temporal deformation model, facial feature points under the relevant topological structure and shape are extracted from the optimized facial image to obtain facial deformation features.

[0065] Specifically, based on the key facial feature points in the face recognition model, feature points are extracted from the optimized facial image to obtain the original facial feature points. These key facial feature points include at least: mouth feature points, eye feature points, nose feature points, facial contour feature points, cheekbone contour feature points, and jaw angle feature points.

[0066] In one embodiment, an optimized facial image (dermal texture feature map output from a pre-order adversarial decomposition network, resolution 1024×1024) is first input. Then, based on a pre-trained model using HRNet (High Resolution Network), 68 key feature points are located. Examples of key point classification: Mouth: 20 points (lip contour and corners of the mouth). Eyes: 12 points (upper and lower eyelids and corners of the eyes). Nose: 9 points (bridge of the nose, alar, tip of the nose). Facial contour: 17 points (cheekbone, mandibular angle, chin). Finally, the original set of facial feature points, P = {p1, p2, ..., p68}, is output. For example, for a facial image of a 50-year-old male, the coordinates of the mandibular angle point (p16, p17) are extracted, and the change in mandibular width is calculated.

[0067] Furthermore, by using the short-term dynamic scale in the multi-scale temporal deformation model, the texture features related to neuromuscular control in the original facial feature points are screened under the age-related change factor to obtain texture features that are independent of the age-related change factor.

[0068] In one embodiment, a short-term dynamic scale is used to screen age-independent textural features. This can be achieved by analyzing the short-term dynamic changes of micro-movements (such as blinking and smiling) of these features using a neuromuscular control model. If the variance of the feature's displacement in the time series is less than a threshold, it is considered a stable textural feature independent of age. For example, wrinkles around the eyes (highly affected by age) are removed, while nasal bone features (lowly affected by age) are retained.

[0069] Furthermore, based on the mid-term dynamic scale in the multi-scale temporal deformation model, and using a non-rigid algorithm, the distances between skeletal development feature points of the original facial feature points are calculated to obtain the feature point distance term. Then, based on the recombination transformation of the feature point distance term, the key point term is determined.

[0070] In one embodiment, it is also necessary to quantify the change in feature point distance due to skeletal development. The feature points are non-rigidly aligned using Thin Plate Spline (TPS), and the change in skeletal point spacing is calculated. Then, the feature point distance term is used: Determine the feature point p i With feature point p j The distance term between the baseline age t0 and the target age t is used. Then, point pairs that are sensitive to skeletal deformation (such as the mandibular angle-zygomatic bone point pair) are selected to construct a set of key point terms.

[0071] Furthermore, by using the Arc-Node Matrix, a rigidity calculation is performed on the topological map of skeletal development feature points of the original facial feature points to obtain the feature point rigidity term. Then, the feature point distance term, feature point rigidity term, and key point term are subjected to a two-level loop iterative calculation based on the energy function to determine the facial bone deformation trajectory.

[0072] In one embodiment, regarding the arc-node matrix, skeletal feature points need to be connected into a rigid structure (such as a triangular mesh of cheekbone-bridge of the nose-mandible), and then the rigidity term is calculated: through Procrustes analysis, the rotation and translation invariance errors of the topological graph are calculated. Finally, the feature point distance term, feature point rigidity term, and key point term are combined to perform a two-layer loop iterative calculation based on the energy function, that is, using a two-layer loop optimization of the energy function (outer layer iterates the age step size, inner layer iterates the deformation parameters), using the L-BFGS algorithm to minimize the energy function, and outputting the skeletal deformation trajectory (such as the rate of cheekbone forward movement and the mandibular angle blunting curve).

[0073] Furthermore, using a topology-preserving manifold learning algorithm, facial evolution modeling is performed on the texture feature points and facial bone deformation trajectories to obtain a facial evolution model and extract facial deformation features from it. Specifically, in topology-preserving manifold learning, t-SNE or UMAP algorithms can be used to map texture feature points and bone trajectories to a low-dimensional manifold space, preserving local neighborhood relationships. Combining the bone deformation subspace (representing bone development, such as the opening angle of the mandibular angle with age) and the texture aging subspace (representing skin elasticity decay, such as the depth of nasolabial folds) in the facial evolution model, data support processing for the facial evolution model is completed. Finally, principal component analysis (PCA) is extracted from the manifold space to obtain cross-age deformation basis vectors, which are the facial deformation features in the facial evolution model.

[0074] As a feasible implementation method, multi-scale modeling combining short-term muscle dynamics (<5 years) and medium-term skeletal development dynamics (5-15 years) can be used. The facial evolution model constructed through manifold learning can predict facial deformation trajectories over the previous 10-20 years. Feature point distance term reorganization transformation and keypoint term dynamic pruning techniques are employed to reduce computational complexity while maintaining the accuracy of non-rigid algorithms. Through a topology-preserving manifold learning algorithm, spatiotemporal alignment of skeletal development features (rigid terms) and muscle movement features (non-rigid terms) is achieved, which is beneficial for facial evolution model generation.

[0075] S104. Perform diffusion processing on the optimized facial image to edit the age attribute and determine the facial growth constraint features.

[0076] Specifically, using a GAN model and a diffusion model, existing target attribute regions in facial images are identified and processed for optimization. A complementary attention branch is then used to optimize supplementary attribute regions in facial images that are missing facial features and require additional attributes. Target attribute regions are those whose facial features do not change significantly with age. Supplementary attribute regions are those whose facial features change significantly with age.

[0077] In one embodiment, the optimized facial image (from the immutable texture feature map output from the previous step) is first input into the GAN model and the diffusion model. Then, attribute region identification is performed, where the target attribute region (without significant age changes) is defined as: skeletal contour, eye shape, and nasal bridge structure (segmented using a pre-trained UNet, masked as Mtarget). The complementary attribute region (with significant age changes) is defined as: skin texture, wrinkles, and hairline (masked as Msupp). Next, using a complementary attention branch, a StyleGAN3 generator can be employed, guided by Msupp, to generate age features for the complementary regions (e.g., removing wrinkles, skin texture, etc.). Then, a Stable Diffusion implicit denoising network is used, constrained by the Mtarget mask, to maintain the stability of the skeletal structure. For example, the target regions (cheekbones, jawline) of a 60-year-old woman's image are preserved in shape, and the diffusion model is used to generate the dermal texture of a 30-year-old (Msupp region).

[0078] Furthermore, by using the attention mask of the GAN model, feature generation processing is performed on the supplementary attribute region across age regions to obtain the first mask image, and based on the color mask, age change effect processing is performed on the target attribute region across age regions to obtain the second mask image.

[0079] In one embodiment, the first mask image (generated by GAN) is: the current facial image + target age label (e.g., "-20 years old"). Then, mask multiplication is used to generate youthful features for the supplementary region (e.g., removing crow's feet and nasolabial folds). The second mask image (generated by diffusion model) is: the same image + age-related keywords (e.g., "wrinkles, eye bags"). Color mask constraints are also required, i.e., skin color correction, to conform to the pigmentation patterns of young people's skin.

[0080] Furthermore, the first mask image and the second mask image are combined using a masking technique to generate a facial growth image after age editing. That is, the first mask image and the second mask image can be fused in the gradient domain at the mask boundaries to eliminate seams, thereby achieving Poisson image fusion. The age-edited facial growth image is then output, preserving the original identity features while adding age-related changes.

[0081] Furthermore, key facial growth features in the facial growth image are subjected to attention-based constraint processing, and facial growth constraint features are extracted.

[0082] As a feasible implementation, Vision Transformer (ViT) can be used to extract multi-scale feature maps from facial growth images and calculate attention weight matrices. Then, key growth regions (such as nasolabial fold depth, drooping corners of the eyes, hairline, and mouth position) are constrained to maintain consistency with a biomechanical model. The constrained features are then compressed into low-dimensional vectors using an autoencoder and used as a description of cross-age deformation, thus extracting facial growth constraint features.

[0083] In other words, by using the collaborative generation of GAN and diffusion model, the physical realism of Poisson fusion, and the attention mechanism of biomechanical constraints, high-precision and identity-preserving cross-age facial editing was achieved, solving the problems of excessive deformation or identity loss in traditional methods. Moreover, it can accurately predict the facial growth constraint features of a person when they are young under cross-age deformation.

[0084] S105. Through a preset feature enhancement extraction network, facial deformation features and facial growth constraint features are fused with attention and the key facial features are output.

[0085] Specifically, through a spatial attention network, spatial features are extracted from facial deformation features and facial growth constraint features respectively, resulting in first spatial features and second spatial features.

[0086] Furthermore, through a channel attention network, channel features are extracted from the first spatial features and the second spatial features respectively, resulting in first channel features and second channel features.

[0087] Furthermore, based on the importance of the channels after expert scoring, weight values ​​are assigned to the first channel features and the second channel features, and cross-age channel features with cross-age facial feature maps are marked.

[0088] In one embodiment, Figure 2 This is a schematic diagram of a recognition network structure with fusion feature enhancement provided in an embodiment of this application, as shown below. Figure 2 As shown, a pre-trained spatial attention network (such as SENet or CBAM) is first used to process facial deformation features and facial growth constraint features. Then, spatial features are extracted from the two feature sets respectively, resulting in the first spatial feature (F1) and the second spatial feature (F2). Next, a channel attention network (such as CBAM or SENet) is applied to F1 and F2 to extract channel features, resulting in the first channel feature (C1) and the second channel feature (C2). Then, the importance of the channel features needs to be determined based on expert scoring: First, weight values ​​are assigned to C1 and C2 to reflect their importance in cross-age facial feature recognition. Second, channel features with cross-age facial features are labeled, forming cross-age channel features (CCF).

[0089] Furthermore, based on the fusion loss function and cross-age channel features, facial deformation features and facial growth constraint features are merged under multi-head attention. The distributed facial features, learned and trained through a fully connected layer, are then predicted and mapped into the output space to generate cross-age facial images corresponding to the patients. These cross-age facial images are predicted facial images for the patients' historical age ranges.

[0090] Furthermore, based on the key facial feature points in the face recognition model, feature points are extracted from cross-age facial images to obtain key facial features.

[0091] In one embodiment, such as Figure 2 As shown, in conjunction with the aforementioned embodiments, a multi-head attention mechanism is also required. Combined with a fusion loss function, F1 and F2 are merged with CCF features. Through fully connected layer learning, the merged features are mapped to the distributed facial feature space. Furthermore, the trained distributed facial features are used to predict and map into the output space, generating cross-age facial images corresponding to the patient's historical age group. Finally, using key facial feature point detection algorithms in the face recognition model (such as a deep learning-based facial landmark detector), key facial features are extracted from the generated cross-age facial images.

[0092] S106. Perform dimensionality reduction and restoration processing on the key facial features based on the real facial image to obtain cross-age facial features. Then, perform identity information query processing on the cross-age facial features to determine the social security identity information of the medical patient.

[0093] Specifically, using Obtain cross-age facial features for each age group Where k is the feature type in the key facial features, N is the cross-age change attribute parameter corresponding to the facial deformation feature in the key facial features, C is the cross-age change attribute parameter corresponding to the facial growth constraint feature in the key facial features, and i and j are both mathematical constants; The pooling window size is the size of the real facial image; D age V represents the age feature size parameter for facial images across different ages; age For age ranges.

[0094] Furthermore, it is necessary to extract cross-age facial features corresponding to the patients. These cross-age facial features are then processed to generate identity codes, resulting in the identity code information to be queried. This identity code information is then input into the social security personnel information database, and the identity information within the matching threshold range is sorted according to the matching degree.

[0095] Furthermore, based on the order in which the identity information is arranged, the social security identity information to be determined is displayed sequentially. Based on the social security identity information confirmed manually, the social security identity information matching the medical patient is obtained.

[0096] In one embodiment, the aforementioned face recognition model is used to extract cross-age facial features from the facial images of medical patients. The extracted features are then standardized to facilitate subsequent encoding and comparison. A deep learning autoencoder or variational autoencoder (VAE) model is then used to encode the standardized cross-age facial features, generating identity encoding information. This identity encoding information should contain sufficient features to distinguish different individuals while being compact enough to reduce storage requirements. A pre-created social security personnel information database, containing each person's facial features and corresponding social security identity information, is also required to ensure that the facial features in the database have been converted to the same encoding format as the query features.

[0097] In one embodiment, the generated identity code information to be queried also needs to be input into the social security personnel information database. Similarity metrics (such as cosine similarity, Euclidean distance, etc.) can be used to calculate the matching degree between the query code and the codes in the database. The social security identity information is sorted according to the matching degree, with the information having the highest matching degree displayed first. A matching degree threshold is set, and only identity information with a matching degree higher than this threshold is displayed. Next, the system displays the sorted social security identity information to be determined in sequence. Then, through the user interface, a human operator or medical patient confirms the displayed information. Finally, based on the result of the human confirmation, the social security identity information matching the medical patient is determined.

[0098] As a feasible implementation method, by extracting and encoding facial features across ages, the system can process images of medical patients of different ages, improving the accuracy of identity verification. The automated identity encoding and matching degree calculation process reduces manual intervention and increases processing speed. Furthermore, by setting a matching degree threshold, the accuracy of the matching results is ensured, reducing the risk of erroneous matches. The system can also display the matching results through a clear interface, facilitating identity verification by human operators or medical patients. Simultaneously, the optimized identity recognition process reduces query time and manual workload, improving overall work efficiency. The security of the identity information database is enhanced, ensuring personal privacy and data security through a rigorous matching process.

[0099] In addition, embodiments of this application also provide a face recognition system based on cross-age characteristics, such as... Figure 3 As shown, the face recognition system 300 based on cross-age characteristics includes:

[0100] The texture recognition module 310 is used to label the dermal texture features and variable epidermal features of the initial facial images of medical patients to determine the facial skin feature image. The variable epidermal features in the facial skin feature image are dynamically stripped, and the dermal texture features are subjected to physical optical constraints to obtain an optimized facial image.

[0101] The facial topology recognition module 320 is used to extract facial feature points related to topological structure and shape from the optimized facial image to obtain facial deformation features.

[0102] The facial growth prediction module 330 is used for diffusion processing of age attribute editing of optimized facial images to determine facial growth constraint features.

[0103] The feature fusion module 340 is used to perform attention fusion processing on facial deformation features and facial growth constraint features, and output key facial features.

[0104] The identity information query module 350 is used to perform feature reduction and restoration processing on key facial features based on real facial images to obtain cross-age facial features. Then, it performs identity information query processing on these cross-age facial features to determine the social security identity information of the medical patient.

[0105] This application's embodiments, through labeling and processing dermal texture features and variable epidermal features, can more accurately extract age-related facial features, which is crucial for cross-age face recognition. Utilizing dynamic stripping of variable epidermal features and physically optically constrained dermal texture features helps optimize facial images, reduce interference from non-age-related features, and improve recognition accuracy. Furthermore, using a multi-scale temporal deformation model to extract facial feature points can capture changes in facial shape at different ages, thereby improving recognition robustness. Editing age attributes can also determine facial growth constraint features, aiding in the identification of individuals of different age groups, especially when there are significant age variations. The feature enhancement extraction network, through attention fusion processing of facial deformation features and growth constraint features, can improve feature quality and enhance the performance of the recognition system. Dimensionality reduction of key facial features reduces computational complexity while retaining key information, improving recognition efficiency. Simultaneously, by extracting cross-age facial features, effective identification between individuals of different age groups can be achieved, expanding the application scope of face recognition. Ultimately, by querying and processing identity information, the social security identity information of patients can be quickly and accurately determined, thereby improving the efficiency and security of medical services.

[0106] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0107] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0108] The above description is merely an embodiment of this application and is not intended to limit this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of this application should be included within the scope of the claims of this application.

Claims

1. A face recognition method based on cross-age features, characterized in that, The method includes: The initial facial images of patients were labeled with relevant dermal texture features and variable epidermal features to determine the facial cortical feature images. The variable epidermal features in the facial skin feature image are dynamically stripped, and the dermal texture features are subjected to physical optical constraints to obtain an optimized facial image. By employing a multi-scale temporal deformation model, facial feature points related to topological structure and shape are extracted from the optimized facial image to obtain facial deformation features, specifically including: Based on the key facial feature points in the face recognition model, feature points are extracted from the optimized facial image to obtain the original facial feature points; wherein, the key facial feature points include at least: mouth feature points, eye feature points, nose feature points, facial contour feature points, cheekbone contour feature points, and jaw angle feature points. By using the short-term dynamic scale in the multi-scale temporal deformation model, the texture feature points related to neuromuscular control in the original facial feature points are screened under the age change factor to obtain texture feature points that are unrelated to the age change factor. Based on the intermediate dynamic scale in the multi-scale temporal deformation model, and through a non-rigid algorithm, the distance between the original facial feature points and the relevant skeletal development feature points is calculated to obtain the feature point distance term; and based on the recombination transformation of the feature point distance term, the key point term is determined. By using the arc-node correlation matrix, the rigidity of the topological map of the original facial feature points related to skeletal development feature points is calculated to obtain the feature point rigidity term. The facial bone deformation trajectory is determined by performing a two-level loop iterative calculation based on the energy function on the feature point distance term, the feature point rigidity term, and the key point term. A topology-preserving manifold learning algorithm is used to model facial evolution by comparing the texture feature points with the facial bone deformation trajectory, resulting in a facial evolution model; and facial deformation features are extracted from the facial evolution model. The optimized facial image is subjected to age attribute editing diffusion processing to determine facial growth constraint features; The facial deformation features and facial growth constraint features are fused together by a preset feature enhancement extraction network, and key facial features are output. The key facial features are subjected to feature reduction and restoration based on the real facial image to obtain cross-age facial features; and the cross-age facial features are then used to query identity information to determine the social security identity information of the medical patient.

2. The face recognition method based on cross-age features according to claim 1, characterized in that, The initial facial images of patients were labeled with relevant dermal texture features and variable epidermal features to determine the facial cortical feature image, specifically including: The facial area of ​​the patient is sequentially irradiated with light of different polarization angles by a multi-band polarized light source pre-installed in the smart medical mutual aid machine, and multi-angle polarized images of the facial area are simultaneously acquired by a preset high-resolution polarization camera. Based on the infrared band of a multi-band polarized light source, the anisotropic texture in the multi-angle polarized image is subjected to penetration imaging processing to determine the dermal texture features; wherein, the dermal texture features include: facial pore distribution features and facial blood vessel distribution features. Using a non-polarized reference light source, ambient light interference is compensated for in the multi-angle polarized image, and feature marking processing under high scattering characteristics is performed on the multi-angle polarized image to determine the variable epidermal layer features; the variable epidermal layer features include: wrinkle distribution features, spot distribution features, and color spot distribution features. The initial facial image is determined based on the anisotropic texture and the multi-angle polarized image processed with the high scattering characteristics. Based on the RGB color channels, the dermal texture features and the variable epidermal features in the initial facial image are subjected to dimensional layering processing to generate the facial skin feature image.

3. The face recognition method based on cross-age features according to claim 1, characterized in that, Before dynamically stripping the variable epidermal features in the facial skin feature image and applying physical optical constraints to the dermal texture features to obtain an optimized facial image, the method further includes: The historical facial surface feature image set is input into the generator's U-Net architecture; where the historical facial surface image is the input, and the dermal texture features and variable epidermal features are the outputs; Based on the dimensional hierarchical characteristics between the dermal texture features and the variable epidermal features, a cortical discriminator with a dual-branch structure is constructed; wherein, the cortical discriminator is used to distinguish the physiological texture authenticity of the dermal texture features and the dynamic noise distribution of the variable epidermal features. The U-Net architecture and the cortical discriminator are subjected to adversarial network generation processing to obtain an adversarial decomposition network; according to The adversarial loss of the adversarial decomposition network is obtained. ;in, D For the cortical discriminator; The dermal texture features and variable epidermal features output by the generator; E Represented by mathematical symbols; Using a preset polarization light transfer function as a regularization term, the dermal texture features and the variable epidermal features are subjected to dimensional level constraint loss calculation related to the RGB color channels to obtain the physical constraint loss. The cyclic consistency between the dermal texture features and the variable epidermal features is controlled, and the cyclic consistency loss is determined. The adversarial decomposition network is adversarially trained using the adversarial loss, the physical constraint loss, and the cycle consistency loss to obtain an optimized adversarial decomposition network.

4. The face recognition method based on cross-age features according to claim 3, characterized in that, The variable epidermal features in the facial skin feature image are dynamically stripped, and the dermal texture features are subjected to physical optical constraints to obtain an optimized facial image, specifically including: The reflectivity of the current facial surface feature image is calculated based on the polarization angle and polarization degree to obtain the polarization reflection relationship between high scattering characteristics and anisotropic texture. The optimized adversarial decomposition network, based on the polarization reflection relationship, provides real-time feedback on the signal-to-noise ratio of the volatile epidermal features in the facial surface feature image, and dynamically adjusts the attention weight of the cortical discriminator to complete the dynamic stripping of the volatile epidermal features. Based on nonnegative matrix factorization, the dynamic stripping results are subjected to cross-interference suppression processing to obtain the optimized dermal texture features after stripping. Physical optical constraints of Fresnel reflection boundaries are applied to the dermal texture features, and adaptive histogram equalization is performed on the image features corresponding to the dermal texture features to obtain the optimized facial image with invariant texture features.

5. The face recognition method based on cross-age features according to claim 1, characterized in that, The optimized facial image undergoes diffusion processing to edit its age attribute, thereby determining facial growth constraint features, specifically including: The existing target attribute regions in the optimized facial image are identified and processed using a GAN model and a diffusion model. The supplementary attribute regions that are missing facial features and need to be added are identified through a complementary attention branch. The target attribute regions are attribute regions whose facial features do not change significantly with age, while the supplementary attribute regions are attribute regions whose facial features change significantly with age. The attention mask of the GAN model is used to perform cross-age region feature generation processing on the supplementary attribute region to obtain a first mask image, and the target attribute region is processed with cross-age region age change effect processing according to the color mask to obtain a second mask image. The first mask image and the second mask image are combined using a mask to generate a facial growth image after age editing. The key facial growth features in the facial growth image are subjected to attention-based constraint processing, and facial growth constraint features are extracted.

6. The face recognition method based on cross-age features according to claim 1, characterized in that, The facial deformation features and facial growth constraint features are fused together using a pre-defined feature enhancement extraction network, and key facial features are output, including: Using a spatial attention network, spatial features are extracted from the facial deformation features and the facial growth constraint features respectively, resulting in a first spatial feature and a second spatial feature. The first spatial feature and the second spatial feature are processed by channel attention network to extract channel features, respectively, to obtain the first channel feature and the second channel feature. Based on the importance of the channels after expert scoring, weight values ​​are assigned to the first channel features and the second channel features, and cross-age channel features with cross-age facial feature maps are marked. Based on the fusion loss function and cross-age channel features, the facial deformation features and facial growth constraint features are merged under multi-head attention. The distributed facial features learned by the fully connected layer are then predicted and mapped into the output space to generate a cross-age facial image corresponding to the patient. The cross-age facial image is a predicted facial image of the patient in the patient's historical age range. Based on the key facial feature points in the face recognition model, feature points are extracted from the cross-age facial images to obtain the key facial features.

7. The face recognition method based on cross-age features according to claim 1, characterized in that, The key facial features are subjected to dimensionality reduction and restoration based on real facial images to obtain cross-age facial features, specifically including: according to The cross-age facial features for each age group are obtained. ;in, k The feature type is one of the key facial features. N These are the cross-age change attribute parameters corresponding to facial deformation features among the key facial features; C These are the cross-age change attribute parameters corresponding to the facial growth constraint features among the key facial features; i and j All are mathematical constants; The pooling window size is the size of the real facial image; The age feature size parameter for facial images across different ages; For age ranges.

8. The face recognition method based on cross-age features according to claim 1, characterized in that, The cross-age facial features are processed to determine the social security identity information of the medical patient, specifically including: Extract cross-age facial features corresponding to the patients; The cross-age facial features are processed to generate identity codes, resulting in the identity code information to be queried. The identity code information to be queried is input into the social security personnel information database, and the identity information within the matching degree threshold is sorted according to the identity information matching degree. Based on the order of the identity information, the social security identity information to be determined is displayed sequentially. Based on the social security identity information confirmed by manual verification, social security identity information matching the medical patient is obtained.

9. A face recognition system based on cross-age characteristics, characterized in that, The system includes: The texture recognition module is used to label the dermal texture features and variable epidermal features of the initial facial images of medical patients to determine the facial skin feature image; the variable epidermal features in the facial skin feature image are dynamically stripped, and the dermal texture features are subjected to physical optical constraints to obtain an optimized facial image. The facial topology recognition module is used to extract facial feature points related to topological structure and shape from the optimized facial image using a multi-scale temporal deformation model to obtain facial deformation features. Specifically, this includes: extracting feature points from the optimized facial image based on key facial feature points in a face recognition model to obtain original facial feature points; wherein the key facial feature points include at least: mouth feature points, eye feature points, nose feature points, facial contour feature points, cheekbone contour feature points, and jaw angle feature points; using the short-term dynamic scale in the multi-scale temporal deformation model, filtering the neuromuscular control-related muscular feature points from the original facial feature points under age-related factors to obtain muscular feature points unrelated to age-related factors; and based on the multi-scale... In the intermediate dynamic scale of the time-series deformation model, a non-rigid algorithm is used to calculate the distance between the original facial feature points and related skeletal development feature points to obtain the feature point distance term. Based on the recombination transformation of the feature point distance term, the key point term is determined. Through the arc-node correlation matrix, a rigid calculation of the topology of the original facial feature points and related skeletal development feature points is performed to obtain the feature point rigid term. The feature point distance term, the feature point rigid term, and the key point term are then subjected to a double-layer loop iterative calculation based on an energy function to determine the facial bone deformation trajectory. A topology-preserving manifold learning algorithm is used to model the facial evolution of the texture feature points and the facial bone deformation trajectory to obtain the facial evolution model. Finally, facial deformation features are extracted from the facial evolution model. The facial growth prediction module is used to perform age attribute editing diffusion processing on the optimized facial image to determine facial growth constraint features; The feature fusion module is used to perform attention fusion processing on the facial deformation features and the facial growth constraint features, and output the key facial features; The identity information query module is used to perform feature dimensionality reduction and restoration processing on the key facial features based on the real facial image to obtain cross-age facial features; and to perform identity information query processing on the cross-age facial features to determine the social security identity information of the medical patient.

Images