Biological recognition method and system based on palm print and palm vein information fusion

Abstract

The present application relates to the field of image-based biometric identification, and discloses a biometric identification method and system based on palmprint and palm vein information fusion, which acquires a user's visible light palmprint image sequence and near-infrared palm vein image sequence, and combines user age information to perform non-rigid deformation compensation and age adaptive enhancement processing, effectively solving the problem of image sequence deformation accumulation and poor feature consistency caused by skin sagging in the elderly group; a pre-constructed residual network and convolution network are used to extract palmprint and palm vein deep feature vectors respectively, and an adaptive fusion strategy driven by mutual information and age information is used to generate a final identity feature vector, which is combined with multi-dimensional similarity weighting and dynamic authentication threshold comparison to output an authentication result; the present application effectively improves the identity recognition accuracy of the elderly group through multi-modal data fusion and age adaptive processing mechanism.

Description

Technical Field

[0001] This invention relates to the field of image-based biometrics, and more specifically, to a biometrics method and system based on the fusion of palm print and palm vein information. Background Technology

[0002] With the continuous upgrading of information security and identity authentication needs, traditional identity recognition methods based on passwords and IC cards are gradually being limited due to their susceptibility to theft and difficulty in uniquely identifying individuals. Biometric recognition technology, with its uniqueness and stability, has become a core research direction in this field. Among them, palmprint recognition has the advantages of convenient collection and stable features, but single palmprint features are easily affected by changes in lighting, surface stains, and angular deviations. Palm vein recognition relies on near-infrared imaging to obtain the distribution of subcutaneous blood vessels, which has strong anti-forgery capabilities and is less affected by the external environment, but it suffers from high feature extraction difficulty and slow recognition speed. Therefore, fusing palmprint and palm vein features to complement each other's advantages has become an important path to improve the accuracy and robustness of identity authentication.

[0003] Existing technologies include research on the fusion of palm print and palm vein features. For example, Chinese patent application CN119942600A discloses an intelligent identity verification method and system that fuses palm vein and palm print features. This method collects palm images and identity data of the user to be authenticated, encodes the palm images based on the identity data to obtain a set of palm identity-encoded images, obtains multidimensional features of palm veins and palm prints through region labeling and feature extraction, generates a palm encoding fusion feature dataset through spatial alignment and fusion, and then obtains an adaptive palm identity recognition network through encryption processing and network training. Finally, this network is used to complete identity verification, solving the problems of insufficient security, low accuracy, and poor anti-interference ability caused by single biometric features, and improving the security, accuracy, and anti-interference ability of identity verification. For example, the Chinese patent with authorization announcement number CN101251889B discloses a near-infrared imaging device and identity recognition method based on palm veins and palm prints. It uses a near-infrared imaging device to acquire a single palm image, extracts a central sub-block sample and inputs it into a feature extraction module, matches the features and calculates the similarity, determines the optimal weighted combination of palm print and vein structure based on training samples, and makes a decision based on a preset threshold after fusing the similarity. This overcomes the defects of few image features and simple processing, and improves the system recognition rate and stability.

[0004] However, while existing studies have attempted to fuse the two, most have remained at the level of simple feature-level stitching or decision-level weighted averaging, failing to fully consider the complementary relationship between palm prints and palm veins at both the image and feature levels, resulting in limited fusion effects. In particular, with age, the loss of collagen in the dermis leads to a decrease in the skin's elastic modulus. The deformation of the palm skin under pressure changes from elastic deformation in youth to viscoelastic deformation, with recovery time increasing from milliseconds to seconds. This viscoelastic deformation causes non-uniform displacement of the palm print ridges, especially at the edge of the palm, where displacement can reach 2-3 mm. Furthermore, the relaxation of subcutaneous tissue reduces the stability of veins, causing even slight palm tilting to cause vascular slippage. In the process of identifying elderly users, due to the prolonged recovery time of skin deformation in the continuously acquired image sequence, the position of palm print ridges and veins undergoes non-uniform displacement between different frames. This cumulative displacement effect causes misalignment of palm print and palm vein features when they are spatially aligned. Traditional fusion methods cannot effectively eliminate deformation interference, resulting in a significant decrease in the discriminative power of feature fusion. This makes the identification accuracy of elderly users lower than that of younger groups. Furthermore, the consistency of image features in continuously captured images during the same acquisition process is poor, which cannot meet the reliable identity authentication needs of elderly people and other groups with significant changes in physiological characteristics. Summary of the Invention

[0005] This invention is applicable to scenarios requiring high-precision identity authentication, such as financial transactions, access control systems, and smart terminal logins, and is particularly suitable for applications with a high concentration of elderly users, such as community elderly care service centers and hospital health management systems. To overcome the aforementioned shortcomings of existing technologies, this invention provides a biometric identification method and system based on the fusion of palmprint and palm vein information. By acquiring visible light palmprint image sequences and near-infrared palm vein image sequences from users, and combining this with user age information for non-rigid deformation compensation and age-adaptive enhancement processing, it effectively solves the problem of image sequence deformation accumulation and poor feature consistency caused by skin laxity in the elderly. It utilizes pre-constructed residual networks and convolutional networks to extract palmprint and palm vein depth feature vectors respectively, and generates the final identity feature vector through an adaptive fusion strategy driven by mutual information and age information. The authentication result is output by combining multi-dimensional similarity weighting and dynamic authentication threshold comparison. This invention effectively improves the accuracy of identity recognition for the elderly through multi-modal data fusion and age-adaptive processing mechanisms.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Biometric methods based on the fusion of palmprint and palm vein information include:

[0008] The visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated are acquired, and the user's age information is obtained at the same time; non-rigid deformation compensation and age-adaptive enhancement processing are performed on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain the enhanced palmprint image sequence and enhanced palm vein image sequence.

[0009] A pre-built palmprint feature extraction residual network is used to extract palmprint depth feature vectors from the enhanced palmprint image sequence; a pre-built palm vein feature extraction convolutional network is used to extract palm vein depth feature vectors from the enhanced palm vein image sequence.

[0010] Calculate the mutual information between the palm print depth feature vector and the palm vein depth feature vector, and adaptively fuse the palm print depth feature vector and the palm vein depth feature vector according to the mutual information and the user's age information to generate the final identity feature vector to be authenticated.

[0011] Calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registration template feature vector in the database. Calculate the weighted average of the multidimensional similarity scores to obtain the comprehensive similarity score. Compare the comprehensive similarity score with the dynamic authentication threshold to output the identity authentication result.

[0012] The method for non-rigid deformation compensation of visible light palmprint image sequences and near-infrared palm vein image sequences is as follows: determine whether there is non-rigid deformation in the visible light palmprint image sequences and near-infrared palm vein image sequences, and define the image sequences with non-rigid deformation as deformation sequences; for the deformation sequences, use a non-rigid registration method based on thin plate spline interpolation to generate deformation-compensated palmprint image sequences and deformation-compensated palm vein image sequences.

[0013] The method for determining whether non-rigid deformation exists in visible light palmprint image sequences and near-infrared palm vein image sequences includes:

[0014] The first frame of the visible light palmprint image sequence and the first frame of the near-infrared palm vein image sequence are selected as palmprint reference frames, respectively. The palmprint optical flow field and palm vein optical flow field between the subsequent frames of each image sequence and the corresponding reference frames are calculated.

[0015] The divergence of the palmprint optical flow field and the palm vein optical flow field are calculated separately. A deformation threshold is set based on the user's age information. The standard deviations of the palmprint optical flow field divergence and the palm vein optical flow field divergence are determined to exceed the deformation threshold. When the standard deviation of the palmprint optical flow field divergence exceeds the deformation threshold, the visible light palmprint image sequence is determined to have non-rigid deformation. When the standard deviation of the palm vein optical flow field divergence exceeds the deformation threshold, the near-infrared palm vein image sequence is determined to have non-rigid deformation.

[0016] The non-rigid registration method based on thin plate spline interpolation is as follows: a control point grid is set for the deformation sequence, and the correspondence between the control points is obtained. Based on the correspondence between the control points, the deformation sequence is deformed by thin plate spline interpolation.

[0017] The methods for generating the enhanced palmprint image sequence and the enhanced palm vein image sequence include:

[0018] Multimodal registration is performed on the deformed palmprint image sequence and the deformed palm vein image sequence to obtain spatially aligned palmprint image sequence and spatially aligned palm vein image sequence. Temporally stable regions of interest are extracted from the spatially aligned palmprint image sequence and spatially aligned palm vein image sequence.

[0019] Based on the user's age information, age-adaptive enhancement processing is performed on spatially aligned palm print image sequences and spatially aligned palm vein image sequences within temporally stable regions of interest, respectively, to generate enhanced palm print image sequences and enhanced palm vein image sequences.

[0020] The multimodal registration method includes:

[0021] Common physiological landmarks were extracted from the deformed palm print image sequence and the deformed palm vein image sequence, and the correspondence between the landmarks was established by SIFT feature descriptors.

[0022] If the number of marker correspondences is sufficient, the affine transformation parameters are calculated based on the marker correspondences. If insufficient, the mutual information registration method is used to determine the affine transformation parameters. The deformation-compensated palmprint image sequence and the deformation-compensated palm vein image sequence are then registered into a spatially aligned palmprint image sequence and a spatially aligned palm vein image sequence using the affine transformation parameters.

[0023] The method for extracting the temporally stable region of interest includes:

[0024] Calculate the standard deviation of grayscale values ​​for each pixel in spatially aligned palm print image sequences and spatially aligned palm vein image sequences to generate visible light stability maps and near-infrared stability maps.

[0025] Binarization thresholds are set based on user age information. Visible light stability maps and near-infrared stability maps are binarized, and temporally stable regions of interest are extracted through logical AND operations.

[0026] The method for performing age-adaptive enhancement processing on the spatially aligned palmprint image sequence includes:

[0027] The temporally stable region of interest is divided into multiple sub-blocks, and the coefficient of variation of gray values ​​between sub-blocks is calculated. Based on the coefficient of variation, it is determined whether there is uneven illumination in the spatially aligned palm print image sequence.

[0028] For palmprint image sequences with uneven illumination and spatial alignment, adaptive gamma correction and texture enhancement are performed based on user age information to generate enhanced palmprint image sequences.

[0029] The method for adaptively fusing palm print depth feature vectors and palm vein depth feature vectors includes:

[0030] The fusion weights of the palm print depth feature vector and the palm vein depth feature vector are determined based on the mutual information and the user's age information. Based on the fusion weights of the two vectors, a fusion feature vector sequence is generated by the bilinear fusion method.

[0031] The statistical features of the fused feature vector sequence are calculated. The fused feature vector sequence is input into a long short-term memory network to output a temporal feature representation. The temporal feature representation is then concatenated with the statistical features of the fused feature vector sequence to generate the final identity feature vector to be authenticated.

[0032] A biometric system based on the fusion of palmprint and palm vein information, used to implement the aforementioned biometric method based on the fusion of palmprint and palm vein information, the system comprising:

[0033] Dual-modal image preprocessing module: used to acquire visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated, and at the same time obtain the user's age information; perform non-rigid deformation compensation and age-adaptive enhancement processing on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain enhanced palmprint image sequence and enhanced palm vein image sequence.

[0034] Feature extraction module: Extracts palmprint depth feature vectors from the enhanced palmprint image sequence through a pre-built palmprint feature extraction residual network; extracts palm vein depth feature vectors from the enhanced palm vein image sequence through a pre-built palm vein feature extraction convolutional network.

[0035] Feature Adaptive Fusion Module: This module calculates the mutual information between the palmprint depth feature vector and the palm vein depth feature vector. Based on the mutual information and the user's age information, it adaptively fuses the palmprint depth feature vector and the palm vein depth feature vector to generate the final identity feature vector to be authenticated.

[0036] Identity authentication module: It is used to calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registered template feature vector in the database, and to obtain the comprehensive similarity score by weighted averaging of the multidimensional similarity scores. The comprehensive similarity score is compared with the dynamic authentication threshold to output the identity authentication result.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention utilizes the collaborative acquisition of dual-modal image sequences and age information to provide fundamental data for subsequent processing that combines modal complementarity with age-specificity. By combining non-rigid deformation compensation and age-adaptive enhancement processing, it can specifically optimize the image feature quality for users of different age groups. In particular, it can alleviate the interference of non-rigid deformation accumulation effects caused by skin laxity in the elderly on feature consistency, ensuring that the enhanced palmprint and palm vein features more closely match the user's actual biological characteristics. Furthermore, by extracting depth feature vectors for both modalities through a pre-constructed deep network, it can fully exploit the texture discrimination information of palmprints and the vascular structure of palm veins. By constructing discrimination information and then using an adaptive fusion strategy based on mutual information and age information, the complementary advantages of the two modalities can be effectively integrated, avoiding the loss of intermodal correlation information caused by traditional simple splicing or weighted averaging. Finally, by combining multidimensional similarity calculation with dynamic authentication thresholds, the matching degree between the features to be authenticated and the features of the registered template can be accurately evaluated, avoiding the limitations of fixed thresholds or single similarity measures that cannot adapt to the differences in feature distribution across different age groups. Overall, this improves the accuracy of identity authentication and enhances the robustness of the system, especially meeting the reliable identity authentication needs of elderly people and other groups with significant changes in physiological characteristics. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating the principle of the biometric identification method based on the fusion of palm print and palm vein information provided in this embodiment of the invention;

[0041] Figure 2 A schematic diagram illustrating the principle of the dual-modal image acquisition and pressure detection mechanism provided in this application embodiment;

[0042] Figure 3 A flowchart illustrating the principle of determining whether a visible light palmprint image sequence and a near-infrared palm vein image sequence exhibit non-rigid deformation, provided for embodiments of the present invention.

[0043] Figure 4 This is a schematic diagram of multimodal image physiological landmark extraction and spatial alignment provided in an embodiment of this application;

[0044] Figure 5 A flowchart illustrating the principle of calculating multidimensional similarity scores provided in an embodiment of the present invention;

[0045] Figure 6 A functional block diagram of a biometric system based on the fusion of palm print and palm vein information provided in an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] Please see Figure 1 As shown, this embodiment provides a biometric identification method based on the fusion of palmprint and palm vein information, including:

[0049] Step S10: Collect the visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated, and obtain the user's age information at the same time; perform non-rigid deformation compensation and age adaptive enhancement processing on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain the enhanced palmprint image sequence and enhanced palm vein image sequence.

[0050] Further, step S10 includes:

[0051] Step S11: Simultaneously acquire visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated, and obtain the user's age information at the same time.

[0052] See Figure 2 This is a schematic diagram illustrating the principle of the dual-modal image acquisition and pressure detection mechanism provided in this application embodiment. Figure 2 As shown, this embodiment uses a dual-modal imaging device to acquire visible light palmprint image sequences and near-infrared palm vein image sequences of the user's palm at the same frame rate. The visible light palmprint image sequence and the near-infrared palm vein image sequence have the same number of frames. Figure 2As shown, the dual-modal imaging device includes a visible light camera, a near-infrared camera, and a hardware synchronization signal generator. The hardware synchronization signal generator sends trigger signals to both the visible light camera and the near-infrared camera based on the same clock source to ensure that the acquisition timing of the visible light palmprint image and the near-infrared palm vein image is completely aligned, avoiding subsequent multimodal registration deviations due to inter-frame misalignment. See also... Figure 2 A pressure sensor array is embedded in the surface of the acquisition platform. The detection area of ​​the pressure sensor array coincides with the imaging field of view of the camera, enabling real-time acquisition of the pressure distribution in different areas of the palm. Figure 2 As indicated in the label, when the pressure sensor detects that the average pressure of the palm exceeds the set force threshold, the system activates a delay mechanism. The delay duration is determined through prior experiments on the recovery of skin deformation after palm pressure from users of different ages. The delay duration allows the temporary skin deformation caused by excessive pressure to return to its natural state. This process avoids morphological distortion of blood vessels caused by compression and texture deformation caused by excessive stretching of the skin, ensuring that the acquired images reflect the true physiological characteristics of the palm.

[0053] After the palm reaches a stable state, the visible light camera and near-infrared camera continuously acquire multiple frames of images, forming an image sequence. User age information is obtained through two methods: first, the user actively inputs it through the input interface provided with the dual-modal imaging device; second, the system automatically estimates it. Automatic estimation relies on a pre-trained age estimation network. The input to the age estimation network is selected as the first frame of the visible light palmprint image sequence because the palm has just reached a stable state when the first frame is acquired, resulting in minimal deformation and the clearest age-related features such as skin texture, wrinkles, and pigmentation, thus reducing the interference of deformation on age estimation. The age estimation network learns from a large number of palm print image samples from users of different ages, extracting skin texture roughness (reflecting the degree of aging of the stratum corneum), wrinkle depth (which deepens with age), and pigment deposition characteristics (the distribution pattern of pigments such as age spots) as criteria to estimate the user's age. Based on the user's age, the user is divided into three age groups: youth group (18-35 years old), middle-aged group (36-55 years old), and elderly group (56 years old and above). The classification result is the user's age information, providing basic parameters for all subsequent processing steps that need to take into account age differences.

[0054] The synchronous acquisition of dual-modal images ensures a one-to-one spatial correspondence between palm print and palm vein images, laying a spatial consistency foundation for subsequent multimodal registration and avoiding registration errors caused by temporal misalignment. The pressure detection and delay mechanism eliminates non-physiological deformation interference caused by manual pressing operations, making the acquired image features closer to the user's real biometric features and reducing the impact of feature distortion on the recognition results. The acquisition of multi-frame image sequences provides temporal dimension data support for subsequent inter-frame deformation analysis, capturing dynamic deformation information of the palm during the acquisition period, rather than relying solely on the limitations of a single static image. The acquisition of age information enables the entire system to adapt to the physiological differences of different age groups, breaking the limitations of traditional biometric methods that adopt a uniform processing strategy, and improving the system's adaptability and recognition robustness to users of different age groups. Furthermore, the reliance of the age estimation network on the first frame image not only simplifies the network input selection logic but also improves the accuracy of age estimation by reducing deformation interference. This ensures that all subsequent age-based processing steps can accurately adapt to the user's physiological characteristics, forming a complete technical chain of "age information - adaptive processing - accurate recognition". Without this step of obtaining age information, subsequent steps such as age adaptive enhancement, feature extraction, and similarity weight adjustment will lack data support, making it impossible to achieve differentiated processing for users of different age groups, resulting in a decrease in the recognition accuracy of groups with significant changes in physiological characteristics, such as the elderly.

[0055] Step S12: Determine whether there is non-rigid deformation in the visible light palmprint image sequence and the near-infrared palm vein image sequence. Define the image sequence with non-rigid deformation as a deformed sequence. For the deformed sequence, use a non-rigid registration method based on thin plate spline interpolation to generate deformed palmprint image sequence and deformed palm vein image sequence.

[0056] Further, step S12 includes:

[0057] Step S121, see Figure 3 The first frame of the visible light palmprint image sequence and the first frame of the near-infrared palm vein image sequence were selected as palmprint reference frames, respectively, and the palmprint optical flow field and palm vein optical flow field between the subsequent frames of each image sequence and the corresponding reference frames were calculated.

[0058] Step S122, see Figure 3The divergence of the palmprint optical flow field and the palm vein optical flow field are calculated separately. A deformation threshold is set based on the user's age information. The standard deviations of the palmprint optical flow field divergence and the palm vein optical flow field divergence are determined to exceed the deformation threshold. If the standard deviation of the palmprint optical flow field divergence exceeds the deformation threshold, the visible light palmprint image sequence is determined to have non-rigid deformation; otherwise, no non-rigid deformation exists. If the standard deviation of the palm vein optical flow field divergence exceeds the deformation threshold, the near-infrared palm vein image sequence is determined to have non-rigid deformation. Image sequences with non-rigid deformation are defined as deformation sequences; otherwise, no non-rigid deformation exists.

[0059] Step S123: Set control point grid for the deformation sequence and obtain the correspondence of control points. Based on the correspondence of control points, use thin plate spline interpolation to perform deformation correction on the deformation sequence. If the deformation sequence is a visible light palmprint image sequence, the deformation-compensated palmprint image sequence is obtained; if the deformation sequence is a near-infrared palm vein image sequence, the deformation-compensated palm vein image sequence is obtained.

[0060] Specifically, in step S121, the first frame of both the visible light palmprint image sequence and the near-infrared palm vein image sequence is selected as the reference frame. The selection criteria for the reference frame are: the first frame is acquired when the palm has just entered a stable state, at which time the palm has not yet undergone significant minor deformations, such as slight muscle relaxation leading to positional shifts or natural skin peristalsis, etc., and thus has the highest feature stability, serving as a benchmark for inter-frame deformation comparison. After the reference frame is determined, the optical flow field between each subsequent frame and the reference frame is calculated. This calculation is performed using the Lucas-Kanade optical flow algorithm. This algorithm assumes that the pixel grayscale values ​​are constant within a local region and uses the least squares method to solve for the pixel motion vectors, efficiently obtaining the displacement information of pixels between adjacent frames and forming an optical flow field. Each optical flow vector in the optical flow field represents the direction and distance of motion of the corresponding pixel from the reference frame to the subsequent frame, which can intuitively reflect the dynamic deformation trend of the palm during the acquisition period, providing quantitative data support for subsequent deformation determination.

[0061] In step S122, the divergence of the optical flow field is calculated to analyze deformation characteristics. The optical flow field divergence is calculated by taking the partial derivatives of the x-component and y-component of the optical flow vector in the x-direction and then summing them. It characterizes the degree of expansion or contraction of the image region. A positive divergence value indicates that the region has expansion deformation; a negative divergence value indicates that the region has contraction deformation; the larger the absolute value of the divergence, the more severe the deformation. By calculating the standard deviation of the optical flow field divergence, the dispersion of deformation across the entire image range can be statistically analyzed. The larger the standard deviation, the more significant the deformation differences between different regions in the image, i.e., the existence of non-rigid deformation. Rigid deformation manifests as overall translation, rotation, or scaling, with a smaller divergence standard deviation. The deformation threshold is set based on the user age information obtained in step S11. This is because there are significant differences in skin elasticity among users of different age groups: young people have good skin elasticity and experience less non-rigid deformation during acquisition, so a lower deformation threshold, such as 0.3, is set; elderly people have loose skin and low elastic modulus, making them prone to larger non-rigid deformations, so a higher deformation threshold, such as 0.8, is set; middle-aged people fall between the two. By comparing the standard deviation of the optical flow field divergence with the deformation threshold for the corresponding age group, it is possible to accurately determine whether the image sequence exhibits non-rigid deformation, avoiding unnecessary compensation processing for sequences with rigid deformation or no significant deformation, and reducing computational resource consumption.

[0062] In step S123, compensation processing is performed on sequences (deformation sequences) determined to have non-rigid deformation. First, a control point grid is set, uniformly distributed to cover the entire image area, ensuring a comprehensive description of the image's deformation distribution. Simultaneously, the control point density is adaptively adjusted based on the user's age information. The control point density is higher in the edge region of the palm of elderly users than in the central region. This is because the loose skin of elderly users leads to more drastic deformation at the palm edges, such as the base of the fingers and the edges of palm lines. Higher density control points can more precisely capture the deformation details of the edge regions, avoiding inaccurate deformation descriptions due to sparse control points. The correspondence between control points is obtained by minimizing the bending energy function. This function measures the deformation difference of control points between the reference frame and subsequent frames. An age-related smoothing constraint term is introduced into the bending energy function, with a higher smoothing constraint weight for the elderly group than for the younger group. This is because the skin deformation of the elderly is more gradual, and excessive pursuit of precise control point matching can easily lead to local distortion in the corrected image. A higher smoothing constraint weight makes the deformation correction process more closely match the deformation patterns of elderly skin. By finding the minimum value of the bending energy function, the corresponding position of each control point in subsequent frames relative to the control point in the reference frame can be determined, i.e., the control point correspondence. Based on the control point correspondence, thin-plate spline interpolation is used to correct the deformation sequence. Thin-plate spline interpolation constructs a smooth interpolation function to map the deformation of the control points to the entire image region. This ensures the overall smoothness of the image while accurately restoring the correction effect of local deformation, avoiding image edge blurring or feature distortion caused by traditional linear interpolation. If the deformation sequence is a visible light palmprint image sequence, the corrected result is a deformation-compensated palmprint image sequence; if it is a near-infrared palm vein image sequence, the corrected result is a deformation-compensated palm vein image sequence; for sequences that do not require deformation compensation, the original sequence is directly output to ensure processing efficiency.

[0063] Step S12 establishes a stable deformation comparison benchmark through reference frame selection, avoiding deformation analysis errors caused by unstable benchmark frames; optical flow field calculation and divergence analysis realize the quantitative judgment of non-rigid deformation, breaking through the limitation of traditional rigid registration methods that cannot handle non-rigid skin deformation, and accurately identifying the sequence that needs compensation; age-adaptive deformation threshold setting enables the judgment criteria to fit the physiological characteristics of different age groups, improving the accuracy of judgment; density adjustment of control point grid and age-related constraint terms of bending energy function enable deformation compensation to be differentiated for the skin deformation patterns of different age groups, especially for fine compensation of the edge region of the elderly group, solving the problem of insufficient handling of severe edge deformation by traditional compensation methods; thin plate spline interpolation ensures the feature integrity of the corrected image, avoiding feature loss or distortion during the compensation process. Step S12 follows the image sequence and age information from step S11. By using deformation compensation, it eliminates the problem of poor feature consistency caused by skin laxity in the elderly, providing a feature-stable image sequence for the subsequent multimodal registration in step S13. Without this step, multimodal registration would be based on images with non-rigid deformation, resulting in the palm print and palm vein features not being accurately aligned in space, thus affecting the accuracy of subsequent region of interest extraction. At the same time, the image sequence after deformation compensation ensures that the features obtained in the subsequent feature extraction stage are more consistent, reducing feature shifts caused by deformation, and laying a reliable foundation for feature fusion in step S20 and similarity calculation in step S30. Furthermore, the deep integration of each sub-step in step S12 with age information forms an adaptive processing logic of "age information - deformation judgment - differential compensation". This not only improves the accuracy of deformation compensation, but also significantly enhances the adaptability of the entire system to users of different age groups. For example, the elderly group can achieve accurate compensation for severe deformation by using the synergy of high-density control points at the edges and high smoothness constraint weights, while avoiding image distortion caused by overcorrection. This refined processing for specific groups significantly improves the robustness of the system in recognizing the elderly.

[0064] Step S13: Perform multimodal registration on the deformed palm print image sequence and the deformed palm vein image sequence to obtain spatially aligned palm print image sequence and spatially aligned palm vein image sequence, and extract temporally stable regions of interest from the spatially aligned palm print image sequence and spatially aligned palm vein image sequence.

[0065] Further, step S13 includes:

[0066] Step S131: Extract common physiological landmarks from the deformed palm print image sequence and the deformed palm vein image sequence, and establish the landmark correspondence through SIFT feature descriptors;

[0067] Step S132: Determine whether the number of marker correspondences is sufficient. If sufficient, calculate the affine transformation parameters based on the marker correspondences. If insufficient, use the mutual information registration method to determine the affine transformation parameters, and use the affine transformation parameters to register the deformed palmprint image sequence and the deformed palm vein image sequence into a spatially aligned palmprint image sequence and a spatially aligned palm vein image sequence.

[0068] Step S133: Calculate the standard deviation of grayscale values ​​of each pixel in the spatially aligned palm print image sequence and the spatially aligned palm vein image sequence, and generate a visible light stability map and a near-infrared stability map.

[0069] Step S134: Set a binarization threshold based on the user's age information, perform binarization processing on the visible light stability map and the near-infrared stability map, extract the temporally stable region of interest through logical AND operation, and calculate the area ratio of the temporally stable region of interest to the entire palm area.

[0070] See Figure 4 This is a schematic diagram illustrating the extraction and spatial alignment of physiological landmarks from multimodal images provided in this application embodiment. In step S131, common physiological landmarks are extracted from the two modalities of the images. These common physiological landmarks must be clearly visible in both visible light palmprint and near-infrared palm vein images and possess physiological stability. Combined with... Figure 4 As shown, for example, common physiological landmarks include the finger valley point and the palmar contour feature point; wherein, the finger valley point is the depression point where the base of the finger connects to the palm, and the palmar contour feature point refers to the contour inflection point where the palmar contour connects to the wrist. Figure 4 The diagram also illustrates the overlap of these markers in the fused image after spatial alignment. The extraction of common physiological markers is achieved through edge detection and morphological operations: first, Gaussian filtering is applied to both modal images after deformation compensation to remove noise; then, the Canny edge detection algorithm is used to extract image edges; subsequently, morphological closing operations are used to connect broken edges; finally, based on preset geometric features of the markers, such as the "V"-shaped edge structure of the finger valley and the circular concave contour of the palm depression, candidate markers are matched. After removing false markers caused by noise or local deformation, the final set of common physiological markers is determined. Figure 4The dashed line connecting the visible light palmprint image and the near-infrared palm vein image shows that the correspondence between marker points is established using SIFT (Scale-Invariant Feature Transform) feature descriptors. SIFT feature descriptors construct a 16×16 neighborhood window around each marker point, dividing the window into 4×4 sub-windows. For each sub-window, gradient histograms in eight directions are calculated, generating a 128-dimensional SIFT feature vector for the marker point. This vector possesses scale invariance and rotation invariance, effectively resisting slight changes in image acquisition angle and distance. By calculating the Euclidean distance between the SIFT feature vectors of marker points in the two modalities, marker point pairs with a distance less than a preset matching threshold are identified as marker point correspondences. The matching threshold is determined through prior experimental statistics to ensure a correct matching rate higher than a preset proportion, exemplarily 90%, avoiding excessive incorrect matching that could affect subsequent registration accuracy. Finally, based on the determined marker point correspondences, the following is achieved: Figure 4 The spatial alignment effect shown.

[0071] In step S132, it is determined whether the number of marker point correspondences is sufficient. The sufficiency criterion is determined based on the registration accuracy requirements: if the number of marker point correspondences is greater than or equal to the minimum number required to stably solve for the affine transformation parameters, it is considered sufficient. The minimum number required to stably solve for the affine transformation parameters is determined by the degrees of freedom of the affine transformation. For example, if the affine transformation has 6 degrees of freedom, at least 3 pairs of non-collinear marker points are required. When sufficient, the affine transformation parameters are calculated based on the marker point correspondences. The affine transformation parameters include translation, rotation angle, scaling factor, and shearing factor. The marker point coordinate error function is constructed using the least squares method, and the parameter values ​​that minimize the error function are solved. This process ensures that the spatial deviation between the corresponding points of the transformed palm vein image and the palm print image is minimized. When the number of marker point correspondences is insufficient, the mutual information registration method is used to determine the affine transformation parameters. Mutual information is used to measure the statistical correlation between the two modal images. The larger the value, the higher the information overlap between the images and the better the registration effect. Mutual information registration calculates the mutual information values ​​of two modalities within a preset parameter search space (covering possible translation, rotation, and scaling ranges) using a sliding window. The parameter corresponding to the maximum mutual information value is determined as the optimal affine transformation parameter. This method does not rely on marker points and can achieve high-precision registration even when marker points are insufficient. By switching between the two methods, reliable affine transformation parameters are ensured regardless of the number of marker points, converting the deformed palmprint and palm vein image sequences into spatially aligned image sequences and eliminating spatial misalignment caused by differences in imaging perspective.

[0072] In step S133, the standard deviation of the grayscale value of each pixel in the spatially aligned image sequence is calculated to generate a stability map. The calculation of the standard deviation of the grayscale value is performed on the set of grayscale values ​​of the same pixel in multiple frame sequences, and the formula is as follows: Where σ is the standard deviation of gray values, h i Let be the grayscale value of the pixel in the i-th frame, μ be the mean grayscale value of the pixel in the n-th frame, and n be the number of frames in the image sequence. The standard deviation of the grayscale value reflects the degree of temporal fluctuation of the grayscale value of the pixel. The smaller the standard deviation, the less affected the pixel is by transient deformation, noise, and other interference, and the more stable the feature; the larger the standard deviation, the worse the temporal stability of the pixel. Based on the standard deviation of the grayscale values ​​of all pixels, visible light stability maps and near-infrared stability maps of the corresponding palmprint image sequences are generated respectively. The grayscale value of each pixel in the stability map is the standard deviation of the grayscale value of the corresponding pixel in the original image, which intuitively presents the temporal stability of each region of the image.

[0073] In step S134, a binarization threshold is set based on the user's age information. The setting of the binarization threshold is based on the differences in skin physiological characteristics among users of different age groups: Elderly users have loose skin, and even after deformation compensation, some areas may still have slight temporal fluctuations. If the same binarization threshold as the youth group is used, it is easy to misjudge stable areas with slight fluctuations as unstable areas, resulting in an excessively small effective area. Therefore, the binarization threshold for the elderly group is set to k' times that of the youth group, where k' < 1, for example, 0.7. For the youth group, a large number of synchronized, deformation-compensated, and spatially aligned image sequences are first collected. The standard deviation of the pixel grayscale values ​​in stable areas is calculated to form a sample set, and the 95th percentile of the sample set is used as the binarization threshold setting for the youth group. For the middle-aged group, the binarization thresholds of the youth and elderly groups are used as interpolation endpoints, and a linear interpolation method is used to determine the binarization threshold, ensuring that the binarization threshold can adapt to the grayscale fluctuation characteristics of stable areas in different age groups. The stability map is binarized, with pixels whose grayscale standard deviation is less than the binarization threshold marked as 1 (stable region) and pixels whose standard deviation is greater than or equal to the threshold marked as 0 (unstable region), resulting in binarized stable maps for both modalities. Temporally stable regions of interest (ROIs) are extracted using a logical AND operation, retaining only the pixel regions marked as 1 in both modal binarized stable maps. These regions simultaneously satisfy the temporal stability requirements of palm prints and palm veins, eliminating interference from regions that are stable in one modality but unstable in the other, ensuring the reliability of both modal features within the region. The area ratio of the temporally stable ROI to the entire palm region is calculated by dividing the number of pixels in that region by the total number of pixels in the palm region. The palm region is determined through skin color detection and contour extraction: skin color thresholding is performed on the visible light palm print image, and then morphological opening operations are used to remove noise such as hair, resulting in a binary mask of the palm region. The total number of pixels in the palm is then counted based on the binary mask of the palm region. If the area ratio is less than the preset minimum effective ratio, for example 40%, it indicates that there is insufficient stable region in the currently acquired image. The binarization threshold needs to be lowered and recalculated until the area ratio meets the requirements. This is to avoid insufficient data for subsequent feature extraction due to the stable region being too small, which would affect the recognition effect.

[0074] Step S13 provides a physiologically meaningful matching basis for multimodal registration by combining common physiological markers with SIFT feature descriptors. The scale and rotation invariance of SIFT ensures the robustness of registration to slight changes in imaging conditions. The switching between marker quantity determination and mutual information registration solves the problem of failure of traditional single registration methods when there are insufficient markers, ensuring the universality and reliability of registration. The calculation of gray value standard deviation realizes the quantitative evaluation of temporal stability, avoiding the error of subjective judgment of stable regions. The combination of age-adaptive binarization threshold and logical "AND" operation enables the extracted temporally stable region of interest to adapt to the characteristics of different age groups and take into account the stability requirements of dual modes, eliminating the interference of unstable regions of single mode. Step S13 follows the deformation-compensated image from step S12, eliminating spatial misalignment of the bimodal model through registration. This provides a precise region range for the subsequent enhancement processing in step S14. Without step S13, the enhancement in step S14 would apply to the entire image, including temporally unstable regions. Enhancement of unstable regions would introduce false features, leading to a decrease in the discriminative power of subsequent feature extraction. Simultaneously, the extraction of temporally stable regions of interest (ROIs) limits the high-quality data source for feature extraction in step S20, reducing the interference of noise from unstable regions on feature vectors and improving the stability and discriminative power of feature vectors. Furthermore, the age-adaptive adjustment of the binarization threshold ensures the effectiveness of stable region extraction for users of different age groups, avoiding the loss of effective regions due to inappropriate thresholds in the elderly group. This further enhances the system's adaptability to the elderly population, overcoming the limitation of traditional stable region extraction ignoring age differences.

[0075] Step S14: Based on the user's age information, age-adaptive enhancement processing is performed on the spatially aligned palm print image sequence and the spatially aligned palm vein image sequence within the temporally stable region of interest, respectively, to generate enhanced palm print image sequence and enhanced palm vein image sequence.

[0076] Further, step S14 includes:

[0077] Step S141: Divide the temporally stable region of interest into multiple sub-blocks, calculate the coefficient of variation of gray values ​​between sub-blocks, and determine whether there is uneven illumination in the spatially aligned palm print image sequence based on the coefficient of variation.

[0078] Step S142: For the palm print image sequence with uneven illumination and spatial alignment, adaptive gamma correction and texture enhancement are performed based on the user's age information to generate an enhanced palm print image sequence.

[0079] Step S143: Calculate the Weber contrast of the spatially aligned palm vein image sequence within the temporally stable region of interest, and determine whether contrast enhancement is needed based on the Weber contrast.

[0080] Step S144: For the spatially aligned palm vein image sequence that requires contrast enhancement, adaptive histogram equalization is performed by setting the window size of the sliding window according to the user's age information, and the blood vessel edges are enhanced to generate an enhanced palm vein image sequence.

[0081] Step S14 involves age-adaptive enhancement of the spatially aligned palmprint and palm vein image sequences for the temporally stable region of interest extracted in step S13. By adjusting differential parameters, the feature separability of the two modalities is strengthened, providing high-quality image data for subsequent feature extraction. In step S141, the illumination non-uniformity of the spatially aligned palmprint image sequence is evaluated: firstly, the temporally stable region of interest is divided into multiple sub-blocks. The size of each sub-block is determined based on the palmprint image resolution and the area of ​​the region of interest. It is necessary to ensure that each sub-block contains complete local palmprint texture, for example, 16×16 pixels, to avoid sub-blocks that are too small resulting in incomplete texture information or too large resulting in the averaging of local illumination differences. The mean and standard deviation of the grayscale value of each sub-block are calculated. Based on the mean and standard deviation of the grayscale value of each sub-block, the coefficient of variation (Cv) between sub-blocks is calculated as Cv = σ' / μ', where σ' is the standard deviation of the mean grayscale value of all sub-blocks, and μ' is the average of the mean grayscale value of all sub-blocks. The coefficient of variation eliminates the influence of the absolute level of grayscale value and only reflects the degree of difference in grayscale distribution between sub-blocks. If the coefficient of variation exceeds a preset judgment threshold, the palmprint image is determined to have uneven illumination. The judgment threshold is determined through previous experimental statistics: when the coefficient of variation exceeds the judgment threshold, the grayscale contrast of the palmprint ridges and valleys will decrease significantly, resulting in blurred texture features and affecting the accuracy of subsequent feature extraction. An exemplary judgment threshold is set to 0.3, which corresponds to the critical value of the impact of uneven illumination on palmprint feature recognition.

[0082] In step S142, age-adaptive enhancement is performed on palmprint images with uneven illumination: first, adaptive gamma correction is performed, which is achieved through nonlinear transformation of grayscale values, with the transformation formula being O=I. γWhere I is the pixel grayscale value before correction, normalized to [0,1], O is the pixel grayscale value after correction, and γ is the gamma coefficient. The gamma coefficient is adjusted according to the user's age information: elderly users have deeper skin folds on their palms, and the folded areas are prone to shadows, resulting in lower local grayscale values. A larger gamma coefficient needs to be set, for example, 1.2, to brighten the shadow areas through non-linear transformation while avoiding overexposure in non-shadow areas; young users have smooth skin and less uneven lighting, so a smaller gamma coefficient is set, for example, 1.0, and only slight correction is needed; the gamma coefficient for middle-aged users is determined by linear interpolation of the coefficients for young and elderly users. After gamma correction, a Gabor filter bank is used for texture enhancement. The Gabor filter has direction selectivity and scale selectivity, which can effectively extract palm print texture features in different directions. The direction parameters of the filter are adjusted according to the user's age information: for elderly users, the palm print wrinkles are more numerous and complex, so a filter with more directions is used, for example, 8 directions, one every 45°, to ensure that wrinkle textures with different directions can be captured; for young and middle-aged users, the palm print textures are relatively simple, so a filter with fewer directions is used, for example, 4 directions, one every 90°, to reduce the amount of computation while ensuring the texture extraction effect. The scale parameter of the Gabor filter is determined based on the width of the palm print ridges. By statistically analyzing the average width of the palm print ridges in the region of interest, the filter scale is set to 1.5 times the ridge width to ensure that the filter can accurately match the ridge scale and enhance the contrast between ridges and valleys.

[0083] In step S143, the contrast of the spatially aligned palm vein image sequence is evaluated: the Weber contrast ratio within the temporally stable region of interest is calculated. The Weber contrast ratio measures the brightness difference between blood vessels and surrounding tissues, and is calculated using the formula C = (Lt − Lb) / Lb, where C is the Weber contrast ratio, Lt is the average gray value of the blood vessel region, and Lb is the average gray value of the surrounding tissue. Under near-infrared light, blood vessels exhibit low gray values ​​due to light absorption, while surrounding tissues exhibit high gray values ​​due to light reflection. A higher Weber contrast ratio indicates a higher degree of differentiation between blood vessels and surrounding tissues, and clearer vascular features. If the Weber contrast ratio is lower than a preset evaluation threshold, contrast enhancement is required. The evaluation threshold is determined experimentally: when the contrast ratio is lower than the evaluation threshold, the edges of blood vessels are blurred, and some small blood vessels easily merge with surrounding tissues, resulting in incomplete extraction of vascular features. An exemplary evaluation threshold is set to 0.15.

[0084] In step S144, age-adaptive enhancement is performed on the palm vein image that needs enhancement: First, adaptive histogram equalization is used. This method adjusts the grayscale histogram within a local sliding window of the image to avoid local over-enhancement or loss of detail caused by global histogram equalization. The window size is set according to the user's age information: for elderly users, the subcutaneous fat layer is thinner and the blood vessel distribution is relatively sparse. If a small window is used, it is easy to result in a small number of blood vessel pixels within the window, leading to amplification of local noise after equalization; therefore, a larger window is set, for example, 64×64 pixels, to achieve smooth enhancement through a larger range of grayscale statistics and avoid blood vessel breakage; for young users, the blood vessel distribution is dense, so a smaller window is used, for example, 32×32 pixels, which can more finely enhance the contrast of local blood vessels and highlight the features of small blood vessels; for middle-aged users, the window size is determined by linear interpolation. After adaptive histogram equalization, blood vessel edge enhancement is performed using a directional high-pass filter. The direction of the directional high-pass filter is consistent with the direction of the blood vessel, which can effectively enhance the blood vessel edge while suppressing noise perpendicular to the direction of the blood vessel.

[0085] Step S14 accurately determines palm print illumination unevenness through the coefficient of variation, avoiding the judgment error of traditional subjective vision-based methods. Age-adaptive gamma coefficients and Gabor direction parameters specifically address the differences in illumination and texture of palm prints in different age groups. The large gamma coefficient for the elderly group can effectively brighten wrinkle shadows, and the multi-directional Gabor filter can fully capture complex wrinkle textures. The parameters for the young group avoid texture distortion caused by over-enhancement. The calculation of Weber contrast provides a quantitative basis for palm vein enhancement, avoiding over-processing of images that do not need enhancement. The age-adaptive histogram equalization window and directional high-pass filter adapt to the characteristics of sparse blood vessels in the elderly group and dense blood vessels in the young group, ensuring the continuity and clarity of blood vessels after enhancement. Step S14 follows the temporally stable region of interest identified in step S13, enhancing only this region to prevent noise in unstable regions from being amplified simultaneously. It also adjusts parameters based on the age information from step S11, forming a processing logic of "age information - stable region - differentiated enhancement." Without step S14, the enhanced palmprint might suffer from uneven illumination and blurred texture, while palm veins might have insufficient contrast and unclear vessel edges, leading to a decrease in the discriminative power of the feature vector extracted in step S20 and an inability to effectively distinguish different users. Furthermore, step S14 employs a differentiated enhancement strategy for the two modalities: palmprints emphasize illumination and texture, while palm veins emphasize contrast and edges. This fully leverages the feature advantages of both modalities, laying a high-quality foundation for feature fusion in the subsequent step S20. The fused features can simultaneously encompass clear palmprint texture and vascular structure information, improving the uniqueness and robustness of identity authentication.

[0086] Step S10 overcomes the limitations of traditional rigid registration in handling the viscoelastic deformation of elderly skin through multi-frame temporal acquisition and non-rigid deformation compensation, significantly improving the feature consistency of different frames of the same user. Through multimodal registration and stable region extraction, it solves the spatial misalignment and transient noise interference caused by the two imaging principles, ensuring that the region used for feature extraction possesses both spatial alignment and temporal stability. Age-adaptive enhancement addresses the problem that uniform parameters cannot adapt to the physiological differences of different age groups, ensuring that image features of users of different ages are clearly presented. The method in step S10 ensures that the bimodal data after front-end image preprocessing retains the unique texture of palm prints and the unique vascular structure information of palm veins, while effectively suppressing deformation, illumination, and contrast interference specific to the elderly population. This provides a "clean" data source for the subsequent deep feature extraction in step S20. Without step S10, subsequent feature extraction would be based on deformed, misaligned, and noise-interfered images, resulting in feature vectors that cannot accurately reflect the uniqueness of the user's biometrics, leading to low accuracy and high false rejection rates in elderly group identification.

[0087] Step S20: Extract palmprint depth feature vector from the enhanced palmprint image sequence using a pre-constructed palmprint feature extraction residual network; extract palm vein depth feature vector from the enhanced palm vein image sequence using a pre-constructed palm vein feature extraction convolutional network; calculate the mutual information between the palmprint depth feature vector and the palm vein depth feature vector; adaptively fuse the palmprint depth feature vector and the palm vein depth feature vector based on the mutual information and user age information to generate the final identity feature vector to be authenticated.

[0088] Further, step S20 includes:

[0089] Step S21: The first frame of the enhanced palmprint image sequence is used as the main feature extraction frame. The user's age information is encoded into an age embedding vector. The age embedding vector and the main feature extraction frame are input into the palmprint feature extraction residual network. The palmprint depth feature vector is generated through the age-aware attention module embedded in the palmprint feature extraction residual network.

[0090] Specifically, step S21 extracts palmprint depth feature vectors from the enhanced palmprint image sequence using a residual network with an embedded age-aware mechanism. The selection of the main feature extraction frame is based on the results of deformation compensation in step S12 and enhancement processing in step S14. The first frame is not affected by minor deformations in subsequent frames, and features such as palmprint ridges and wrinkles have the highest stability, minimizing the impact of deformation on feature extraction. The palmprint feature extraction residual network is an improved residual network, including an age encoding layer, a fully connected layer, a convolutional layer, residual blocks, and an age-aware attention module. The age encoding layer encodes the three age groups (youth, middle-aged, and elderly) obtained in step S11. The encoding method employs a strategy combining one-hot encoding and embedding mapping: First, one-hot encoding is performed on three age information categories: young group corresponds to vector [1,0,0], middle-aged group corresponds to vector [0,1,0], and elderly group corresponds to vector [0,0,1]. Then, the one-hot encoded vectors are input into a fully connected layer for embedding mapping. The weights of the fully connected layer are learned during network training, transforming the encoding result into a low-dimensional, dense age embedding vector. Convolutional layers are responsible for extracting the basic texture features of the main feature extraction frame, resulting in a palmprint feature map. Residual blocks alleviate the gradient vanishing problem in deep networks through cross-layer connections, ensuring the depth and effectiveness of feature extraction. An age-aware attention module is embedded after the residual blocks. Its input consists of the palmprint feature map output by the residual blocks and the age embedding vector. The age-aware attention module first maps the age embedding vector to a weight vector matching the number of channels in the palmprint feature map through a fully connected layer, and then generates a spatial weight matrix through a spatial attention mechanism. Both modules work together to adjust the attention distribution of different channels and regions in the palmprint feature map. For the elderly group, the age-aware attention module increases the attention weight of the central area of ​​the palm print and deep texture, while weakening the interference from easily deformable edge areas. For the younger group, attention is evenly distributed to preserve complete texture details. This mechanism allows the palm print feature extraction residual network to dynamically adjust its receptive field, performing differentiated extraction based on the differences in palm print feature distribution across different age groups. The generated palm print depth feature vector can accurately capture core discriminative information appropriate for age. User age information provides a basis for attention allocation, and the enhanced image provides clear input for feature extraction, enabling the feature vector to effectively resist texture distortion caused by skin laxity in the elderly group, avoiding the problem of decreased discriminative power of features extracted by traditional unified networks in the elderly group. Without this step, palm print feature extraction will ignore age differences, and the feature vectors of the elderly group are easily mixed with deformation noise, leading to reduced reliability of subsequent feature fusion.

[0091] Step S22: Input the enhanced palm vein image sequence into the palm vein feature extraction convolutional network to obtain the palm vein feature map, calculate the vascular stability mask, perform feature weighting on the palm vein feature map according to the vascular stability mask, and generate the palm vein depth feature vector through temporal aggregation.

[0092] The depthwise separable convolutional structure of the palm vein feature extraction convolutional network comprises two consecutive stages: depthwise convolution and pointwise convolution. Depthwise convolution uses a single-channel kernel to perform spatial convolution on each channel of the enhanced palm vein image sequence, capturing local spatial features such as the edge orientation and branching morphology of the blood vessels, and outputting a basic palm vein feature map. Pointwise convolution uses a 1×1 kernel to fuse cross-channel information from the basic palm vein feature map output by depthwise convolution, generating a palm vein feature map containing global spatial correlations. This structure, by separating the spatial extraction and channel fusion functions of standard convolution, significantly reduces the number of parameters and computational cost while maintaining feature extraction capabilities. It effectively adapts to the real-time processing requirements of palm vein image sequences and solves the processing delay problem caused by the excessive computational complexity of traditional standard convolution. The calculation of the vascular stability mask requires two steps: temporal analysis and age adaptation. First, for the enhanced palm vein image sequence, the temporal standard deviation of the gray value of each pixel in multiple frames is calculated, which is the gray value standard deviation σ calculated in step S133. This value directly reflects the temporal stability of the vascular region. The smaller the value, the less the vascular is affected by slippage. Then, the temporal standard deviation is normalized to the [0,1] interval through linear transformation to obtain the preliminary mask. Then, it is smoothed by combining age information. Gaussian filtering is used to achieve smoothing for the elderly group. The size of the filter kernel is set according to the average amplitude of vascular slippage in the elderly group. Through the stability of feature extraction under different filter kernels in the previous experiment, the kernel size that minimizes the intraclass variance of the features is selected. Mean filtering is used for slight processing for the young group to retain the small fluctuations of vascular details. In the final generated vascular stability mask, the larger the value, the more reliable the vascular features of the corresponding region. Feature weighting is achieved by multiplying the palm vein feature map with the vascular stability mask pixel by pixel, which strengthens the feature signal of stable vascular regions and suppresses noise interference in unstable regions, resulting in a weighted palm vein feature map. Temporal aggregation uses temporal average pooling to take the average of the weighted multi-frame palm vein feature maps in the time dimension, integrating the stability information of multiple frames into a single feature vector, further reducing feature fluctuations caused by vascular slippage in a single frame, and obtaining a palm vein depth feature vector.

[0093] Enhancement processing provides high-contrast vascular input to the convolutional network, ensuring that the basic feature map clearly depicts the vascular structure. Deformation compensation eliminates the influence of static deformation, and temporal aggregation filters out dynamic slippage interference. The combination of these two methods gives the palm vein depth feature vector both spatial integrity and temporal stability. Traditional palm vein feature extraction often relies on single-frame images or unified convolutional networks, which cannot distinguish between stable and unstable regions of blood vessels. Feature fluctuations caused by vascular slippage in the elderly group are directly mixed into the feature vector. However, this step quantifies the temporal stability of vascular features into weights through stability masking and temporal aggregation, achieving accurate extraction of reliable features. Without this step, the palm vein depth feature vector would contain a large amount of noise information from unstable regions, and the recognition accuracy in the elderly group would further decrease due to the vascular slippage problem, and it would not be able to effectively complement the palmprint features.

[0094] Step S23: Calculate the mutual information between the palm print depth feature vector and the palm vein depth feature vector. Determine the fusion weight of the two vectors based on the mutual information and the user's age information. Based on the fusion weight of the two vectors, generate a fusion feature vector sequence using the bilinear fusion method.

[0095] The mutual information calculation employs the k-nearest neighbor estimation method. The value of k was determined through prior experiments: different k values ​​were selected for mutual information calculation on a large number of bimodal feature samples. The ratio of inter-class to intra-class distances of the fused features under different k values ​​was statistically analyzed, and the k value with the largest ratio was selected as the optimal parameter. An exemplary k value can be set to 5. Mutual information is used to quantify the statistical correlation between the palmprint depth feature vector and the palm vein depth feature vector. The larger the value, the more overlapping information the two features contain, and the stronger the correlation. The determination of the vector fusion weights consists of two steps: basic weight calculation and age adjustment. In the basic weights, the basic weight w of the palmprint features... palm The normalized value of the arctangent function of mutual information is w. palm =arctan(I) / π×2, where I is the mutual information. This function maps the mutual information to the interval [0,1] and increases gradually with the mutual information, ensuring smoother weight adjustment when the correlation changes; the basic weight of palm vein features is... Age adjustment is based on the physiological differences across age groups: in the elderly group, the subcutaneous fat layer is thinner, and blood vessels are less affected by skin laxity, resulting in higher feature stability than palm prints. Therefore, an additional weight Δw is added to the palm vein feature weights on top of the base weights. The value of Δw is determined by comparing the intra-class variance of the bimodal features across different age groups, selecting the Δw value that minimizes the intra-class variance of the features in the elderly group. In the youth group, the stability of palm print and palm vein features is similar, so the base weights remain unchanged. For the middle-aged group, linear interpolation is used to determine the adjusted weights. The fusion weights of the palm print depth feature vector are multiplied by the palm print depth feature vector to obtain the weighted palm print depth feature vector; similarly, the fusion weights of the palm vein depth feature vectors are multiplied by the palm vein depth feature vector to obtain the weighted palm vein depth feature vector. Bilinear fusion is performed by performing an outer product operation on the weighted palm print depth feature vector and the weighted palm vein depth feature vector to generate a feature matrix. The outer product operation can capture the second-order interaction information between the two modal features, preserving the intermodal correlation features that are easily lost in simple concatenation. Since the input is a time sequence, a fusion operation is performed on each frame to finally generate a fused feature vector sequence.

[0096] Step S23 quantifies the correlation of bimodal features through mutual information calculation, providing an objective basis for the fusion weight allocation of the two vectors and solving the problem of subjective setting of traditional fusion weights; arctangent function normalization ensures the rationality and stability of the basic weights; age-adaptive weight adjustment makes the fusion strategy adaptable to the feature characteristics of different age groups, fully utilizing stable vein features in the elderly group and balancing the use of bimodal features in the young group; bilinear fusion retains the second-order interaction information between modalities, improving the discriminative power of the fused features; the generation of the fused feature vector sequence provides multi-frame fused feature input for subsequent temporal modeling. Step S23 follows the single-modal feature vectors of steps S21 and S22 to achieve deep fusion of bimodal features, generating fused features that combine the advantages of both modalities; without this step, the fusion will remain at the level of simple splicing or weighted averaging, failing to capture the interaction information between modalities. In the elderly group, due to the low reliability of palmprint features, the discriminative power of the fused features is insufficient, leading to a decline in recognition performance.

[0097] Step S24: Calculate the statistical features of the fused feature vector sequence, input the fused feature vector sequence into the long short-term memory network to output the temporal feature representation, and concatenate the temporal feature representation with the statistical features of the fused feature vector sequence to generate the final identity feature vector to be authenticated.

[0098] Long Short-Term Memory (LSTM) networks consist of four core components: input gate, forget gate, output gate, and cell state. The fused feature vector sequence is input to the LSTM. First, it enters the forget gate, which outputs a weight vector between 0 and 1 through a sigmoid activation function. This weight vector is multiplied by the cell state from the previous time step, filtering out irrelevant transient noise in the sequence, such as feature fluctuations caused by accidental skin movements during data acquisition. The input gate consists of a sigmoid layer and a tanh layer. The sigmoid layer determines the feature information to be updated, and the tanh layer generates candidate feature vectors to be updated. The two are multiplied and then superimposed on the cell state processed by the forget gate to complete the cell state update. The output gate filters the information to be output from the cell state through the sigmoid layer, and then compresses it through the tanh layer to output the hidden state at the current time step. This state integrates the current input and historical information. The fused feature vector sequence is input into the LSTM frame by frame, and the hidden state at the last time step serves as a temporal feature representation. This temporal feature representation captures the dynamic changes of features over time, filtering out transient interference. The calculation of statistical features targets the fused feature vector sequence, selecting parameters such as mean and variance to characterize the overall distribution of the sequence. The mean reflects the static baseline information of the features, while the variance reflects the degree of feature fluctuation. The temporal feature representation and statistical features are concatenated dimensionally to generate a spatiotemporal joint feature vector, which simultaneously encompasses the temporal dynamic consistency and spatial static distribution information of the features. Finally, a fully connected layer maps the spatiotemporal joint feature vector to a fixed-dimensional final identity feature vector. The weights of the fully connected layer are determined through training, with the identity recognition accuracy as the objective function. Optimizing the weights ensures the vector accurately characterizes the uniqueness of the identity. The fused feature vector sequence provides temporal data containing bimodal information. LSTM mines the temporal correlations, and statistical features supplement the overall distribution information. The combination of these two aspects gives the final feature vector both spatiotemporal characteristics. Traditional feature processing often ignores temporal information, relying only on single-frame fused features, making it susceptible to transient noise. This problem is more pronounced in the elderly group due to greater feature fluctuations. This step integrates multi-frame information into a stable feature representation through LSTM temporal modeling and statistical feature concatenation, significantly improving the feature's resistance to transient deformation. Without this step, transient noise in the fused feature vector sequence cannot be effectively filtered out, resulting in poor temporal consistency of features. Feature fluctuations in the elderly group will directly affect the authentication results, leading to an increased false rejection rate.

[0099] Step S20 addresses the poor recognition performance caused by differences in the distribution of bimodal features across different age groups through differential feature extraction and adaptive fusion. The age-aware attention mechanism and vascular stability mask optimize the age-related characteristics of palm prints and palm veins, respectively. Weight adjustments combining mutual information and age achieve accurate bimodal fusion, while LSTM temporal modeling further enhances feature robustness. Step S20, in deep collaboration with preprocessed age information, deformation compensation, and image enhancement results, forms a complete technical chain of "preprocessing-differential extraction-adaptive fusion-temporal modeling." This ensures that the final generated identity feature vector effectively resists the cumulative effect of non-rigid deformation caused by skin laxity in the elderly, preserving the complementary advantages of the bimodal approach while adapting to the physiological differences across different age groups. This multi-layered deep coupling breaks through the limitations of traditional single feature extraction and simple fusion. It not only improves the recognition accuracy of the elderly group, but also significantly enhances the system's adaptability to users of different age groups. At the same time, the joint modeling of temporal and spatial information enables the feature vector to capture both static physiological structure and dynamic stability patterns. Even if the elderly group has slight deformation, it can still achieve accurate recognition through the aggregation of multi-frame temporal information, avoiding the false rejection problem caused by the dependence on single-frame static features in traditional methods.

[0100] Step S30: Calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registration template feature vector in the database. Calculate the weighted average of the multidimensional similarity scores to obtain the comprehensive similarity score. Compare the comprehensive similarity score with the dynamic authentication threshold and output the identity authentication result.

[0101] Step S30 calculates the multidimensional similarity between the final identity feature vector to be authenticated and the feature vector of the registered template in the database, and combines it with an age-adaptive dynamic authentication threshold to output the identity authentication result. This solves the problem that traditional single similarity measurement and fixed threshold cannot adapt to the differences in feature distribution among different age groups, and in particular improves the false rejection phenomenon that is prone to occur in the elderly group due to feature fluctuations.

[0102] See Figure 5 Furthermore, step S30 includes:

[0103] Step S31: Retrieve the registration template feature vector from the database, and calculate the Euclidean distance, cosine similarity, and Mahalanobis distance between the final identity feature vector to be authenticated and the registration template feature vector;

[0104] Specifically, step S31 first retrieves the registration template feature vector from the database. The retrieval logic is based on the user's declared identity information. The registration template feature vectors stored in the database are all generated through the same extraction process as the final identity feature vector to be authenticated, ensuring consistency in feature dimensions and distribution characteristics, and avoiding matching deviations caused by differences in extraction processes. If multiple registration templates for the same user exist in the database, such as templates registered by the user at different age groups, the template with the registration time closest to the current user's age is selected as the main template. This selection is based on the fact that the physiological characteristics of users at different ages change gradually, and the distribution difference between template features of similar ages and the current features to be authenticated is smaller, which can reduce the interference of feature shift caused by age growth on the matching results and improve the accuracy of the initial matching benchmark. Subsequently, the Euclidean distance, cosine similarity, and Mahalanobis distance between the final identity feature vector to be authenticated and the registration template feature vector are calculated. The calculation of the Euclidean distance requires L2 normalization of the two vectors. L2 normalization is achieved by dividing each dimension value of the vector by the L2 norm of the vector, making the vector magnitude uniform to 1, eliminating the influence of vector magnitude differences on distance calculation, and ensuring that the distance only reflects the absolute positional difference of the features in high-dimensional space. The physical meaning of Euclidean distance is to measure the straight-line distance between two feature vectors in space. The smaller the distance, the closer the absolute positions of the two features are. It is suitable for young people with relatively concentrated feature distribution, but it is sensitive to feature position shifts caused by deformation in the elderly group and needs to be used in combination with other similarity.

[0105] Cosine similarity is calculated directly based on the cosine of the angle between two vectors. This metric focuses on the directional consistency of feature vectors, ignoring amplitude differences. Its core logic is: if two feature vectors are similar in direction, even if there is an absolute positional shift, they can still be identified as having the same identity feature. This characteristic makes it robust to feature amplitude fluctuations caused by changes in lighting and slight deformation, especially suitable for feature amplitude changes caused by skin laxity in the elderly, compensating for the sensitivity of Euclidean distance to positional shifts. For non-negative feature vectors, cosine similarity naturally lies in the [0,1] interval and can be directly used for subsequent calculations; if the vector has negative values, it needs to be normalized to map it to the [0,1] interval. The normalization formula is "normalized cosine similarity = (cosine similarity + 1) / 2", ensuring that the cosine similarity of different types of feature vectors has a uniform numerical range and avoiding interference from negative values ​​in subsequent fusion. The calculation of Mahalanobis distance relies on a pre-constructed intra-class covariance matrix. The construction of this matrix must cover user samples from different age groups: Before system deployment, a large number of identity feature vector samples from users of different age groups are collected and categorized into youth, middle-aged, and elderly groups. For each group, an intra-class covariance matrix is ​​calculated. The elements in the matrix reflect the correlation between different dimensions of the feature vectors for that age group, such as the correlation between palm print texture and palm vein structure. During calculation, the feature vectors of each group are first decentralized, i.e., each vector is subtracted from the mean vector of that group. Then, the covariance matrix is ​​obtained by summing the outer product of the sample vector and the decentralized vector, and dividing by the sample size minus one. When using this method, based on the user age information obtained in step S11, select the intraclass covariance matrix of the corresponding age group and substitute it into the Mahalanobis distance formula. The core advantage of this distance is that it considers the correlation between the various dimensions of the features and can correct the distance deviation caused by the difference in variance of different dimensions. For example, the variance of the palm print dimension is larger in the elderly group and the variance of the palm vein dimension is smaller. Mahalanobis distance can balance this difference through the covariance matrix, making the distance calculation more in line with the actual distribution of the features of the elderly group and avoiding the excessive influence of the large variance of a single dimension on the overall distance.

[0106] Step S31 comprehensively evaluates the feature matching degree from three perspectives—absolute position, directional pattern, and statistical distribution—through multi-dimensional similarity calculation. Traditional single similarity measures can only reflect one dimension of a feature's characteristics. For example, Euclidean distance alone cannot ignore the positional shift of features in the elderly group, and cosine similarity alone cannot distinguish between different identity features with similar directions but significantly different statistical distributions. This step uses three similarity measures in synergy, retaining the advantages of each measure while overcoming the limitations of a single measure, providing a comprehensive matching basis for subsequent comprehensive score calculation. Simultaneously, the selection logic of the registration template is combined with age information to ensure the age adaptability of the matching benchmark to the features to be authenticated, reducing interference from cross-age feature shifts. The age group division of the intra-class covariance matrix allows Mahalanobis distance to accurately adapt to the feature relevance of different age groups, further improving the accuracy of distance calculation.

[0107] Step S32: Calculate the comprehensive similarity score between the final identity feature vector to be authenticated and the registration template feature vector based on Euclidean distance, cosine similarity, and Mahalanobis distance.

[0108] Euclidean distance is converted into a first similarity score, cosine similarity is converted into a second similarity score, and Mahalanobis distance is converted into a third similarity score. The fusion weight of the three similarity scores is determined based on the user's age information. Based on the fusion weight of the three similarity scores, the three similarity scores are weighted and averaged to generate a comprehensive similarity score.

[0109] Step S32 unifies the three metrics into similarity scores within the [0,1] interval, and determines the fusion weight by combining age information, calculating the comprehensive similarity score. The Euclidean distance conversion uses a Gaussian kernel function, and the formula is: First similarity score ,in, The similarity score is calculated using an exponential function with base e of the natural logarithm, where d is the Euclidean distance and σ1 is an age-related parameter. σ1 is set based on the degree of positional shift of features in different age groups: For young people, features are stable with small shifts, so a smaller σ1 is set to ensure that small changes in the Euclidean distance cause significant fluctuations in the first similarity score, ensuring sensitivity to positional differences; for older people, features are prone to shifts with large shifts, so a larger σ1 is set to make the first similarity score respond more smoothly to distance changes, avoiding a sharp drop in score due to slight positional shifts, thus preserving the possibility of matching older people who, despite shifts, still belong to the same identity. For example, σ1 can be set to 0.5 for young people, 0.6 for middle-aged people, and 0.7 for older people. The cosine similarity normalization process addresses potential negative values ​​by converting the original cosine values ​​in the [-1,1] interval to similarity scores in the [0,1] interval through linear mapping, ensuring consistency with the numerical ranges of the other two similarity scores and avoiding weight imbalances during fusion due to differences in numerical ranges. For non-negative feature vectors, the cosine similarity is already in the [0,1] interval, so no additional conversion is needed. It can be directly used as the cosine similarity score. This processing logic simplifies the calculation and ensures that the cosine similarity of different types of feature vectors can participate in the fusion.

[0110] The Mahalanobis distance conversion uses an exponential decay function, with the formula "Third Similarity Score = exp(-Mahalanobis Distance / σ²)". Here, σ² is an adjustment parameter used to adapt to the "intra-class distribution differences" of palmprint-palm vein fusion features across different age groups. The determination of σ² is related to the eigenvalues ​​of the intra-class covariance matrix for each age group: first, all eigenvalues ​​of the covariance matrix for the corresponding age group are calculated, and the average value is taken as the base parameter. Then, it is multiplied by the age correlation coefficient k² to obtain σ². The age correlation coefficient k² is set based on the degree of dispersion of the statistical distribution of features in different age groups: for the youth group, the statistical distribution of features is concentrated, and the Mahalanobis distance is small, so a smaller k² is set to ensure that changes in the Mahalanobis distance are effectively reflected in the third similarity score; for the elderly group, the statistical distribution of features is dispersed, and the Mahalanobis distance is large, so a larger k² is set to slow down the decay rate of the third similarity score as the Mahalanobis distance increases, avoiding excessive reduction of the third similarity score due to the increase in Mahalanobis distance caused by the dispersed distribution. For example, k² is 2.0 for the youth group, 2.5 for the middle-aged group, and 3.0 for the elderly group.

[0111] After the three similarity scores are converted, the fusion weights are determined based on the user's age information. The base weights are set based on the characteristics of the three similarities: cosine similarity is more stable in judging directional patterns and is suitable as the primary reference, hence its highest base weight; Euclidean distance reflects absolute position, and Mahalanobis distance reflects statistical distribution, serving as supplements, hence their relatively lower base weights. An example base weight allocation is Euclidean similarity weight 0.3, cosine similarity weight 0.4, and Mahalanobis distance weight 0.3. The adjustment logic based on age information is as follows: For the elderly group, the absolute position of features is prone to shift, reducing the reliability of Euclidean distance; therefore, the Euclidean similarity weight is reduced. Simultaneously, the feature patterns of the elderly group (such as blood vessel orientation and overall palm print structure) remain stable, increasing the reliability of cosine similarity; for the young group, both feature position and statistical distribution are relatively stable, so the base weights remain unchanged; for the middle-aged group, feature characteristics fall between the two, and the weights are adjusted using linear interpolation, with the interpolation endpoints being the weight values ​​of the young and elderly groups. The overall similarity score is calculated using a weighted average. This weighted average process organically integrates three similarity metrics assessed from different dimensions, preserving the advantages of each metric while adjusting weights to suit the characteristics of different age groups. This allows the overall score to more accurately reflect the true degree of matching. Step S32 addresses the inconsistency between different metric ranges and semantics through a unified conversion of similarity scores, ensuring the rationality of the fusion calculation. Age-adaptive weight adjustments allow the fusion strategy to dynamically optimize based on the reliability differences of features across different age groups, avoiding the distortion in the assessment of matching degree for the elderly group caused by traditional fixed-weight fusion.

[0112] Step S33: Determine the dynamic authentication threshold based on the user's age information and the proportion of the temporally stable region of interest to the entire palm area. If the overall similarity score is greater than or equal to the dynamic authentication threshold, output "Authentication successful"; if the overall similarity score is less than the dynamic authentication threshold, output "Authentication failed".

[0113] In step S33, a basic authentication threshold is first set, and then adjusted according to the user's age information and the proportion of the temporally stable region of interest to the entire palm area to obtain a dynamic authentication threshold. The setting of the basic authentication threshold is based on the overall recognition accuracy requirements of the system and is determined through extensive sample testing in the early stage: matching and non-matching sample pairs of users of different age groups are collected, the comprehensive similarity score is calculated, and the score that minimizes the system's error rate is selected as the basic authentication threshold. For example, the basic authentication threshold can be set to 0.75. The first adjustment factor of the basic authentication threshold is the user's age information. The adjustment logic stems from the difference in feature stability among different age groups: the youth group has high feature stability, the comprehensive similarity score of matching samples is generally high, and the score of non-matching samples is generally low, so maintaining the basic authentication threshold is sufficient to meet the accuracy requirements; the middle-aged group has slightly decreased feature stability, and the matching sample score is slightly lower, so the basic authentication threshold needs to be appropriately reduced to reduce false rejections; the elderly group has the lowest feature stability, and the matching sample score is further reduced, so the basic authentication threshold needs to be reduced more significantly to compensate for the score drop caused by feature fluctuations and avoid the false rejection of legitimate users, thus obtaining the age-adjusted threshold. For example, the age-adjusted threshold is 0.75 for the youth group, 0.72 for the middle-aged group, and 0.68 for the elderly group.

[0114] The second adjustment factor for the basic authentication threshold is the proportion of the temporally stable region of interest (ROI) to the entire palm area. When the proportion is lower than the preset area proportion threshold, it indicates that there are insufficient effective stable regions in the currently acquired image, the reliability of feature extraction decreases, and the overall similarity score of the matched samples may decrease. If the age-adjusted threshold is still used, it may lead to false rejection of legitimate users. Therefore, the age-adjusted threshold needs to be further reduced to obtain the dynamic authentication threshold. When the area proportion is higher than or equal to the preset area proportion threshold, it indicates that the effective image area is sufficient and the feature reliability is high. There is no need to adjust the age-adjusted threshold, and the age-adjusted threshold is used as the dynamic authentication threshold. The area proportion threshold is determined by statistically analyzing the relationship between the area of ​​the stable region and the recognition accuracy to ensure that the area of ​​the stable region is sufficient to support feature extraction and matching. An exemplary preset area proportion threshold is 40%. The overall similarity score is compared with the dynamic authentication threshold: when the overall similarity score is greater than or equal to the dynamic authentication threshold, it is determined that the identity matches, and "authentication successful" is output; when the overall similarity score is less than the dynamic authentication threshold, it is determined that the identity does not match, and "authentication failed" is output. Simultaneously, the system records authentication logs, including user age information, three similarity scores, comprehensive similarity score, dynamic authentication threshold, and final authentication result. This information can be used for subsequent system optimization, such as analyzing the rationality of threshold adjustments for different age groups and optimizing similarity weight allocation. Step S33, through dynamic threshold adjustment, ensures authentication accuracy for young people and high-quality images while avoiding false rejections for older people and low-quality images. Without step S33, using a fixed threshold would lead to an increased false rejection rate for older people or low-quality images, or an increased false positive rate for young people, failing to meet the authentication accuracy requirements in different scenarios.

[0115] Example 2

[0116] This embodiment, based on Embodiment 1, provides a biometric system based on the fusion of palmprint and palm vein information, such as... Figure 6 As shown, it includes:

[0117] Dual-modal image preprocessing module: used to acquire visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated, and at the same time obtain the user's age information; perform non-rigid deformation compensation and age-adaptive enhancement processing on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain enhanced palmprint image sequence and enhanced palm vein image sequence.

[0118] Feature extraction module: Extracts palmprint depth feature vectors from the enhanced palmprint image sequence through a pre-built palmprint feature extraction residual network; extracts palm vein depth feature vectors from the enhanced palm vein image sequence through a pre-built palm vein feature extraction convolutional network.

[0119] Feature Adaptive Fusion Module: This module calculates the mutual information between the palmprint depth feature vector and the palm vein depth feature vector. Based on the mutual information and the user's age information, it adaptively fuses the palmprint depth feature vector and the palm vein depth feature vector to generate the final identity feature vector to be authenticated.

[0120] Identity authentication module: It is used to calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registered template feature vector in the database, and to obtain the comprehensive similarity score by weighted averaging of the multidimensional similarity scores. The comprehensive similarity score is compared with the dynamic authentication threshold to output the identity authentication result.

[0121] Furthermore, in the dual-modal image preprocessing module, the method for non-rigid deformation compensation of the visible light palmprint image sequence and the near-infrared palm vein image sequence includes:

[0122] Step S121: Select the first frame of the visible light palmprint image sequence as the palmprint reference frame and the first frame of the near-infrared palm vein image sequence as the palm vein reference frame, respectively, and calculate the palmprint optical flow field and palm vein optical flow field between the subsequent frames of each image sequence and the corresponding reference frame.

[0123] Step S122: Calculate the divergence of the palmprint optical flow field and the palm vein optical flow field respectively. Set a deformation threshold based on the user's age information. Determine whether the standard deviation of the palmprint optical flow field divergence and the standard deviation of the palm vein optical flow field divergence exceed the deformation threshold. If the standard deviation of the palmprint optical flow field divergence exceeds the deformation threshold, it is determined that the visible light palmprint image sequence has non-rigid deformation; otherwise, there is no non-rigid deformation. If the standard deviation of the palm vein optical flow field divergence exceeds the deformation threshold, it is determined that the near-infrared palm vein image sequence has non-rigid deformation. Define the image sequence with non-rigid deformation as a deformation sequence; otherwise, there is no non-rigid deformation.

[0124] Step S123: Set control point grid for the deformation sequence and obtain the correspondence of control points. Based on the correspondence of control points, use thin plate spline interpolation to perform deformation correction on the deformation sequence. If the deformation sequence is a visible light palmprint image sequence, the deformation-compensated palmprint image sequence is obtained; if the deformation sequence is a near-infrared palm vein image sequence, the deformation-compensated palm vein image sequence is obtained.

[0125] The age-adaptive enhancement processing method includes:

[0126] Step S141: Divide the temporally stable region of interest into multiple sub-blocks, calculate the coefficient of variation of gray values ​​between sub-blocks, and determine whether there is uneven illumination in the spatially aligned palm print image sequence based on the coefficient of variation.

[0127] Step S142: For the palm print image sequence with uneven illumination and spatial alignment, adaptive gamma correction and texture enhancement are performed based on the user's age information to generate an enhanced palm print image sequence.

[0128] Step S143: Calculate the Weber contrast of the spatially aligned palm vein image sequence within the temporally stable region of interest, and determine whether contrast enhancement is needed based on the Weber contrast.

[0129] Step S144: For the spatially aligned palm vein image sequence that requires contrast enhancement, adaptive histogram equalization is performed by setting the window size of the sliding window according to the user's age information, and the blood vessel edges are enhanced to generate an enhanced palm vein image sequence.

[0130] Furthermore, in the identity authentication module, the calculation method for the multidimensional similarity score includes:

[0131] Retrieve the registration template feature vector from the database, and calculate the Euclidean distance, cosine similarity, and Mahalanobis distance between the final identity feature vector to be authenticated and the registration template feature vector;

[0132] Euclidean distance is converted into a first similarity score, cosine similarity is converted into a second similarity score, and Mahalanobis distance is converted into a third similarity score. The fusion weights of the three similarity scores are determined based on the user's age information. Based on the fusion weights of the three similarity scores, a weighted average of the three similarity scores is calculated to generate a comprehensive similarity score.

[0133] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0134] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A biometric identification method based on the fusion of palmprint and palm vein information, characterized in that, The method includes: The visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated are acquired, and the user's age information is obtained at the same time; non-rigid deformation compensation and age-adaptive enhancement processing are performed on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain the enhanced palmprint image sequence and enhanced palm vein image sequence. The method for non-rigid deformation compensation of visible light palmprint image sequences and near-infrared palm vein image sequences is as follows: determine whether there is non-rigid deformation in the visible light palmprint image sequences and near-infrared palm vein image sequences, and define the image sequences with non-rigid deformation as deformation sequences; for the deformation sequences, use a non-rigid registration method based on thin plate spline interpolation to generate deformation-compensated palmprint image sequences and deformation-compensated palm vein image sequences. The method for determining whether a visible light palmprint image sequence and a near-infrared palm vein image sequence have non-rigid deformation includes: selecting the first frame of the visible light palmprint image sequence as the palmprint reference frame and the first frame of the near-infrared palm vein image sequence as the palm vein reference frame, respectively; calculating the palmprint optical flow field and the palm vein optical flow field between subsequent frames of each image sequence and the corresponding reference frame; calculating the divergence of the palmprint optical flow field and the palm vein optical flow field, respectively; setting a deformation threshold based on the user's age information; and determining whether the standard deviation of the palmprint optical flow field divergence and the standard deviation of the palm vein optical flow field divergence exceed the deformation threshold. When the standard deviation of the palmprint optical flow field divergence exceeds the deformation threshold, it is determined that the visible light palmprint image sequence has non-rigid deformation; when the standard deviation of the palm vein optical flow field divergence exceeds the deformation threshold, it is determined that the near-infrared palm vein image sequence has non-rigid deformation. A pre-built palmprint feature extraction residual network is used to extract palmprint depth feature vectors from the enhanced palmprint image sequence; a pre-built palm vein feature extraction convolutional network is used to extract palm vein depth feature vectors from the enhanced palm vein image sequence. Calculate the mutual information between the palm print depth feature vector and the palm vein depth feature vector, and adaptively fuse the palm print depth feature vector and the palm vein depth feature vector according to the mutual information and the user's age information to generate the final identity feature vector to be authenticated. Calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registration template feature vector in the database. Calculate the weighted average of the multidimensional similarity scores to obtain the comprehensive similarity score. Compare the comprehensive similarity score with the dynamic authentication threshold to output the identity authentication result.

2. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 1, characterized in that, The non-rigid registration method based on thin plate spline interpolation is as follows: a control point grid is set for the deformation sequence, and the correspondence between the control points is obtained. Based on the correspondence between the control points, the deformation sequence is deformed by thin plate spline interpolation.

3. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 2, characterized in that, The methods for generating the enhanced palmprint image sequence and the enhanced palm vein image sequence include: Multimodal registration is performed on the deformed palmprint image sequence and the deformed palm vein image sequence to obtain spatially aligned palmprint image sequence and spatially aligned palm vein image sequence. Temporally stable regions of interest are extracted from the spatially aligned palmprint image sequence and spatially aligned palm vein image sequence. Based on the user's age information, age-adaptive enhancement processing is performed on spatially aligned palm print image sequences and spatially aligned palm vein image sequences within temporally stable regions of interest, respectively, to generate enhanced palm print image sequences and enhanced palm vein image sequences.

4. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 3, characterized in that, The multimodal registration method includes: Common physiological landmarks were extracted from the deformed palm print image sequence and the deformed palm vein image sequence, and the correspondence between the landmarks was established by SIFT feature descriptors. If the number of marker correspondences is sufficient, the affine transformation parameters are calculated based on the marker correspondences. If insufficient, the mutual information registration method is used to determine the affine transformation parameters. The deformation-compensated palmprint image sequence and the deformation-compensated palm vein image sequence are then registered into a spatially aligned palmprint image sequence and a spatially aligned palm vein image sequence using the affine transformation parameters.

5. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 4, characterized in that, The method for extracting the temporally stable region of interest includes: Calculate the standard deviation of grayscale values ​​for each pixel in spatially aligned palm print image sequences and spatially aligned palm vein image sequences to generate visible light stability maps and near-infrared stability maps. Binarization thresholds are set based on user age information. Visible light stability maps and near-infrared stability maps are binarized, and temporally stable regions of interest are extracted through logical AND operations.

6. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 5, characterized in that, The method for performing age-adaptive enhancement processing on the spatially aligned palmprint image sequence includes: The temporally stable region of interest is divided into multiple sub-blocks, and the coefficient of variation of gray values ​​between sub-blocks is calculated. Based on the coefficient of variation, it is determined whether there is uneven illumination in the spatially aligned palm print image sequence. For palmprint image sequences with uneven illumination and spatial alignment, adaptive gamma correction and texture enhancement are performed based on user age information to generate enhanced palmprint image sequences.

7. The biometric identification method based on the fusion of palmprint and palm vein information according to claim 6, characterized in that, The method for adaptively fusing palm print depth feature vectors and palm vein depth feature vectors includes: The fusion weights of the palm print depth feature vector and the palm vein depth feature vector are determined based on the mutual information and the user's age information. Based on the fusion weights of the two vectors, a fusion feature vector sequence is generated by the bilinear fusion method. The statistical features of the fused feature vector sequence are calculated. The fused feature vector sequence is input into a long short-term memory network to output a temporal feature representation. The temporal feature representation is then concatenated with the statistical features of the fused feature vector sequence to generate the final identity feature vector to be authenticated.

8. A biometric system based on the fusion of palmprint and palm vein information, used to implement the biometric method based on the fusion of palmprint and palm vein information as described in any one of claims 1-7, characterized in that, The system includes: Dual-modal image preprocessing module: used to acquire visible light palmprint image sequence and near-infrared palm vein image sequence of the palm of the user to be authenticated, and at the same time obtain the user's age information; perform non-rigid deformation compensation and age-adaptive enhancement processing on the visible light palmprint image sequence and near-infrared palm vein image sequence to obtain enhanced palmprint image sequence and enhanced palm vein image sequence. Feature extraction module: Extracts palmprint depth feature vectors from the enhanced palmprint image sequence through a pre-built palmprint feature extraction residual network; extracts palm vein depth feature vectors from the enhanced palm vein image sequence through a pre-built palm vein feature extraction convolutional network. Feature Adaptive Fusion Module: This module calculates the mutual information between the palmprint depth feature vector and the palm vein depth feature vector. Based on the mutual information and the user's age information, it adaptively fuses the palmprint depth feature vector and the palm vein depth feature vector to generate the final identity feature vector to be authenticated. Identity authentication module: It is used to calculate the multidimensional similarity score between the final identity feature vector to be authenticated and the registered template feature vector in the database, and to obtain the comprehensive similarity score by weighted averaging of the multidimensional similarity scores. The comprehensive similarity score is compared with the dynamic authentication threshold to output the identity authentication result.

Images