Identity recognition and authentication system based on machine vision three-dimensional features

Abstract

The present application relates to the field of computer vision and biometric identification technology, specifically to an identity recognition and authentication system based on machine vision three-dimensional features, comprising: a data acquisition module: obtaining and preprocessing discrete three-dimensional point cloud data to generate a pure point cloud set; a manifold construction module: constructing a triangular mesh model and solving a generalized eigenvalue problem to obtain a feature sequence; a spectral domain decoupling module: performing frequency domain band-pass filtering on the feature sequence and reorganizing it into an identity feature spectrum; a fingerprint generation module: generating a hot kernel or wave kernel signature and converting it into a multi-dimensional spectrum fingerprint histogram; a matching decision module: calculating the spectral domain distance of the histogram and the standard spectrum fingerprint and judging the authentication condition; an adaptive updating module: if the condition is met, weighted updating is performed, otherwise, a rejection instruction is triggered; the present application improves the tolerance to non-rigid deformation while ensuring the identity structure recognition degree.

Description

Technical Field

[0001] This invention relates to the field of computer vision and biometric recognition technology, specifically to an identity recognition and authentication system based on machine vision three-dimensional features. Background Technology

[0002] As a core component of contactless access control systems, identity recognition and authentication systems in scenarios such as high-security data centers and security monitoring rely heavily on the reliability of 3D facial feature extraction for their accuracy and security. Therefore, accurate feature matching and anti-spoofing are crucial for ensuring the security of access control systems. Existing identity authentication methods have many problems. For example, many traditional spatial rigid registration algorithms are used, which are prone to falling into local extremum traps in spatial dimensions and are highly dependent on vertex index sequences. These inherent defects make it difficult to fully reflect the real and stable skeletal topology of the target. Especially when the object to be authenticated has complex appearance changes or is attacked by 3D physical material forgery, it is impossible to effectively extract high-fidelity identity features and perform accurate verification. Summary of the Invention

[0003] The purpose of this invention is to provide an identity recognition and authentication system based on machine vision 3D features, and to solve the following technical problems:

[0004] To address the issue of insufficient robustness of traditional spatial rigid registration under complex facial expressions, a spectral domain decoupling mechanism is used to effectively separate the tolerance for non-rigid deformation from the recognizability of the identity skeleton architecture. This eliminates the robustness dependency of feature descriptors on the order of mesh vertex indexes, while further enhancing the system's defense against physical forgery attacks using non-biological materials and its real-time engineering adaptability in unconstrained acquisition environments.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] An identity recognition and authentication system based on machine vision 3D features includes:

[0007] The data acquisition module is used to acquire discrete 3D point cloud data of the object to be authenticated, and to preprocess the discrete 3D point cloud data to remove outliers and generate a clean point cloud set.

[0008] The manifold construction module is used to construct a triangular mesh model based on a pure point cloud set, calculate the Riemannian metric tensor of the triangular mesh model, construct the Laplace-Beltramian operator based on the Riemannian metric tensor, solve the generalized eigenvalue problem through finite element analysis, and obtain the full-band eigenvalue sequence and the corresponding eigenfunction sequence.

[0009] The spectral domain decoupling module is used to introduce the spectral wavelet transform operator to perform multi-scale frequency domain bandpass filtering on the full-band feature value sequence, remove low-frequency feature components corresponding to large-scale non-rigid deformation, retain mid-to-high frequency feature components corresponding to the identity skeleton architecture, and reorganize the retained mid-to-high frequency feature components into an identity feature spectrum.

[0010] The fingerprint generation module is used to generate a hot kernel signature with isometry transformation invariance based on the identity feature spectrum and feature function sequence, and to convert the hot kernel signature into a multidimensional spectral fingerprint histogram.

[0011] The matching decision module is used to obtain the pre-stored standard identity spectrum fingerprint, calculate the spectral distance between the multidimensional spectral fingerprint histogram and the standard identity spectrum fingerprint, and determine whether the spectral distance meets the preset authentication pass conditions.

[0012] The adaptive update module is configured to: if the authentication pass conditions are met, perform a weighted update on the standard identity spectrum fingerprint based on the multidimensional spectral fingerprint histogram; if the authentication pass conditions are not met, trigger a rejection response command.

[0013] Alternatively, methods for solving generalized eigenvalue problems using finite element analysis include:

[0014] Based on the geometric topological relationships of the triangular mesh model, a sparse stiffness matrix and a sparse mass matrix are constructed.

[0015] Substitute the sparse stiffness matrix and sparse mass matrix into the generalized eigenvalue equation;

[0016] The generalized eigenvalue equation is decomposed using a preset iterative solution algorithm to calculate a preset number of eigenvalues ​​and their corresponding eigenvectors.

[0017] The eigenvalues ​​are sorted in ascending order of their numerical values ​​to form a full-band eigenvalue sequence, and the eigenvectors are used as a sequence of eigenfunctions.

[0018] Optionally, methods for generating identity feature genealogies include:

[0019] A preset low-frequency cutoff threshold corresponding to changes in facial appearance is set.

[0020] Traverse the full-band feature value sequence, identify feature values ​​with values ​​less than the low-frequency cutoff threshold, and mark them as deformation interference components;

[0021] Identify feature values ​​that are greater than or equal to the low-frequency cutoff threshold and mark them as structural feature components;

[0022] The deformation interference component is suppressed by using the spectral wavelet transform operator, and the structural feature component is enhanced by energy processing. The processed structural feature components are then recombined to obtain the identity feature spectrum.

[0023] Optionally, methods for generating heat kernel signatures with isometry invariance include:

[0024] Obtain the timescale parameters, which are used to control the range of thermonuclear diffusion;

[0025] For each feature value in the identity feature spectrum, calculate the exponential decay term with the natural logarithm as the base;

[0026] Square the feature function value at each vertex in the feature function sequence;

[0027] The exponential decay term and the squared feature function value are weighted and summed to generate the hot kernel signature value for each vertex.

[0028] Aggregate the hot core signature values ​​of all vertices to form a hot core signature mapping graph.

[0029] Optionally, methods for converting hotcore signatures into multidimensional spectral fingerprint histograms include:

[0030] The distribution range of all values ​​in the statistical hot core signature mapping graph;

[0031] The distribution range is divided into a preset number of numerical intervals;

[0032] Count the number of vertices falling into each numerical interval, and calculate the frequency density of each numerical interval;

[0033] Arrange the frequency density of all numerical intervals in interval order to construct one-dimensional or multi-dimensional feature vectors, and use the feature vectors as multi-dimensional spectral fingerprint histograms.

[0034] Optionally, methods for determining whether the spectral distance meets the preset authentication pass conditions include:

[0035] The chi-square distance algorithm or Bach distance algorithm is used to calculate the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint;

[0036] Obtain the preset strict authentication threshold and lenient authentication threshold, wherein the strict authentication threshold is less than the lenient authentication threshold;

[0037] If the difference metric is less than or equal to the strict authentication threshold, the authentication pass condition is met, and a high-confidence authentication success signal is output.

[0038] If the difference metric is greater than the strict authentication threshold but less than the lenient authentication threshold, the multi-frame fusion verification mechanism is activated to obtain the next frame of discrete 3D point cloud data for secondary verification.

[0039] If the difference metric is greater than or equal to the lenient certification threshold, the certification pass condition is not met.

[0040] Optionally, the adaptive update module is also used to perform defensive analysis, including:

[0041] Extract ultra-high frequency feature components whose values ​​are located in a preset high frequency range from the identity feature spectrum;

[0042] Calculate the energy distribution density of ultra-high frequency characteristic components;

[0043] Obtain a preset reference range for the energy density of non-biological materials;

[0044] Determine whether the energy distribution density falls within the reference range for the energy density of non-biological materials;

[0045] If the energy distribution density falls within the reference range of energy density for non-biological materials, an attack alarm command is generated, and the authentication pass condition is forcibly blocked.

[0046] Optionally, methods for initiating the multi-frame fusion verification mechanism include:

[0047] Continuously acquire supplementary point cloud data for a preset number of frames;

[0048] For each frame of supplementary point cloud data, manifold construction, spectral domain decoupling and fingerprint generation steps are performed independently to obtain the corresponding supplementary spectral fingerprint histogram.

[0049] Calculate the cosine similarity between the multidimensional spectral fingerprint histogram and all supplementary spectral fingerprint histograms, and use the arithmetic mean of the cosine similarity as the temporal consistency coefficient;

[0050] If the temporal consistency coefficient is greater than the preset stability threshold, the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint, as well as the average difference measure between all supplementary spectral fingerprint histograms and the standard identity spectral fingerprint, are calculated to obtain the average difference measure. Based on the average difference measure, it is re-determined whether the authentication pass condition is met.

[0051] If the temporal consistency coefficient is less than or equal to the stability threshold, it is determined to be a non-liveness attack or abnormal data collection, and an authentication failure signal is directly output.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention uses a spectral domain decoupling mechanism to remove low-frequency features corresponding to large-scale non-rigid deformations such as facial expressions by using the spectral wavelet transform operator, and accurately extracts the mid-to-high frequency components that represent the identity skeleton structure. This effectively solves the problem of insufficient robustness of traditional rigid registration algorithms under complex expressions such as talking and laughing, and ensures that the system can still extract high-fidelity stable identity features in a dynamic and unconstrained acquisition environment.

[0054] (2) The manifold construction module of this invention transforms three-dimensional coordinates to spectral domain processing, which avoids the defect of spatial dimension being prone to local extremum trap from the mathematical root. At the same time, it maps features with isometric transformation invariance to multidimensional spectral fingerprint histogram, eliminating the dependence of the verification process on the grid vertex index sequence and realizing a comparison method without vertex registration. Even when the acquisition distance changes or the point cloud resolution is scaled, the scale invariance of feature expression can still be maintained.

[0055] (3) By mining the ultra-high frequency components in the identity feature spectrum, this invention analyzes the energy distribution density of micron-level facial pores and skin microtextures, which can accurately identify the differences in physical properties between biological skin and non-biological materials such as silicone and resin. This defense mechanism does not require additional hardware such as multispectral sensors, and can effectively intercept high-imitation mask attacks by relying solely on geometric algorithms, which significantly enhances identity security in high-secrecy scenarios.

[0056] (4) The present invention adopts an iterative solution algorithm for large sparse matrices to ensure that the high-dimensional feature dimensionality reduction is completed within millisecond delay to meet the requirements of seamless passage; combined with a strict and lenient dual threshold decision mechanism, multi-frame fusion verification is started when the difference measurement value is at the fuzzy boundary, and environmental noise is filtered by using temporal consistency evaluation, which reduces the false rejection rate under extreme working conditions while ensuring the safety bottom line and optimizing the passage experience.

[0057] (5) The adaptive update module of the present invention can dynamically capture and adapt to the micro-facial drift of employees over time by weighting the updated features that have passed the authentication, ensuring that the system maintains high recognition accuracy in long-term operation; the whole solution achieves decoupling of deformation tolerance and recognition in the physical frequency domain dimension through frequency domain decoupling and multi-scale energy adjustment, which has strong engineering application value. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of a module for an identity recognition and authentication system based on machine vision 3D features, provided in an embodiment of this application. Detailed Implementation

[0059] 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.

[0060] Please see Figure 1 As shown, the identity recognition and authentication system based on machine vision 3D features includes:

[0061] The data acquisition module is used to acquire discrete 3D point cloud data of the object to be authenticated, and to preprocess the discrete 3D point cloud data to remove outliers and generate a clean point cloud set.

[0062] The manifold construction module is used to construct a triangular mesh model based on a pure point cloud set, calculate the Riemannian metric tensor of the triangular mesh model, construct the Laplace-Beltramian operator based on the Riemannian metric tensor, solve the generalized eigenvalue problem through finite element analysis, and obtain the full-band eigenvalue sequence and the corresponding eigenfunction sequence.

[0063] The spectral domain decoupling module is used to introduce the spectral wavelet transform operator to perform multi-scale frequency domain bandpass filtering on the full-band feature value sequence, remove low-frequency feature components corresponding to large-scale non-rigid deformation, retain mid-to-high frequency feature components corresponding to the identity skeleton architecture, and reorganize the retained mid-to-high frequency feature components into an identity feature spectrum.

[0064] The fingerprint generation module is used to generate a hot kernel signature with isometry transformation invariance based on the identity feature spectrum and feature function sequence, and to convert the hot kernel signature into a multidimensional spectral fingerprint histogram.

[0065] The matching decision module is used to obtain the pre-stored standard identity spectrum fingerprint, calculate the spectral distance between the multidimensional spectral fingerprint histogram and the standard identity spectrum fingerprint, and determine whether the spectral distance meets the preset authentication pass conditions.

[0066] The adaptive update module is configured to: if the authentication pass conditions are met, perform a weighted update on the standard identity spectrum fingerprint based on the multidimensional spectral fingerprint histogram; if the authentication pass conditions are not met, trigger a rejection response command.

[0067] This embodiment provides an identity recognition and authentication system based on machine vision 3D features, specifically applied to the identity verification scenario of a contactless access control system in a high-security data center, aiming to solve the problem of insufficient robustness of traditional spatial rigid registration under complex facial expressions;

[0068] The data acquisition module acquires discrete 3D point cloud data of the object to be authenticated and preprocesses the discrete 3D point cloud data to remove outliers and noise, thereby generating a clean point cloud set. This process ensures the continuity of the underlying topology.

[0069] The manifold construction module constructs a triangular mesh model based on a pure point cloud set, calculates the Riemannian metric tensor representing the intrinsic geometric properties of the surface, and constructs a Laplace-Beltramian operator based on the Riemannian metric tensor. It then solves the generalized eigenvalue problem through finite element analysis to obtain the full-band eigenvalue sequence and the corresponding eigenfunction sequence. This step transforms the three-dimensional spatial coordinate system to the spectral domain, avoiding the local extremum trap of the spatial dimension.

[0070] During this period, the spectral domain decoupling module introduces the spectral wavelet transform operator to perform multi-scale frequency domain bandpass filtering on the full-band feature value sequence, remove low-frequency feature components corresponding to large-scale non-rigid deformation, retain mid-to-high frequency feature components corresponding to the identity skeleton architecture, and reorganize the retained mid-to-high frequency feature components into the identity feature spectrum. For example, when an employee's facial muscles are stretched due to conversation, the system directly filters out such low-frequency deformation interference and only extracts the inherent skeletal structure features such as the cheekbone.

[0071] After feature decoupling is completed, the fingerprint generation module maps and generates a hot kernel signature with isometric transformation invariance based on the identity feature spectrum and feature function sequence, and transforms the hot kernel signature into a multidimensional spectral fingerprint histogram, which breaks the dependence on vertex index sequence.

[0072] The matching decision module obtains the pre-stored standard identity spectrum fingerprint, calculates the spectral distance between the multidimensional spectral fingerprint histogram and the standard identity spectrum fingerprint, and determines whether the spectral distance meets the preset authentication pass conditions; during the system initialization and input phase, the initial fingerprint of the authorized personnel is obtained through the same data collection and feature extraction process and stored in the database;

[0073] The adaptive update module is configured to perform a weighted update of the standard identity spectrum fingerprint based on the multidimensional spectral fingerprint histogram if the authentication conditions are met, in order to adapt to the microscopic facial drift of employees over time.

[0074] If the authentication pass conditions are not met, a rejection response command is triggered. This technical solution improves the tolerance to non-rigid deformation while ensuring the recognizability of the identity structure through a cascade mechanism of frequency domain decoupling. This method shows high tolerance to non-rigid deformation in dynamic interaction scenarios.

[0075] In a preferred embodiment of the present invention, the method for solving the generalized eigenvalue problem by finite element analysis includes: constructing a sparse stiffness matrix and a sparse mass matrix based on the geometric topological relationship of a triangular mesh model; and substituting the sparse stiffness matrix and the sparse mass matrix into the generalized eigenvalue equation.

[0076] The generalized eigenvalue equation is decomposed using a preset iterative solution algorithm to calculate a preset number of eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are then sorted in ascending order of their numerical values ​​to form a full-band eigenvalue sequence, and the eigenvectors are used as a sequence of eigenfunctions.

[0077] This embodiment is a further specification of the method by which the manifold construction module solves the generalized eigenvalue problem using the finite element analysis method. It aims to transform the dimensionality reduction of continuous manifold calculus into large-scale sparse matrix algebraic operations that can be efficiently processed by computers. When processing high-density facial point clouds collected by access control systems, the manifold construction module constructs sparse stiffness matrices and sparse mass matrices based on the geometric topological relationships of the triangular mesh model. The sparse stiffness matrix and sparse mass matrix are then substituted into the generalized eigenvalue equation.

[0078] The generalized eigenvalue equation is decomposed using a preset iterative solution algorithm to calculate a preset number of eigenvalues ​​and their corresponding eigenvectors; the eigenvalues ​​are arranged in ascending order according to their numerical values ​​to form a full-band eigenvalue sequence, and the eigenvectors are used as a sequence of eigenfunctions;

[0079] Specifically, the off-diagonal elements of the sparse stiffness matrix are calculated by taking half of the sum of the cotangent values ​​of the opposite corners of two adjacent triangles. The diagonal elements are set as the negative sum of all adjacent off-diagonal elements of the vertex, thus representing the local curvature change.

[0080] The sparse mass matrix adopts a lumped mass matrix, whose diagonal elements are calculated as 1 / 3 of the sum of the areas of all triangles adjacent to that vertex, thus representing the geometric measure. This in-situ construction method preserves the inherent geometric invariants of the manifold. At the same time, the specific mathematical expression of the generalized eigenvalue equation is: the sparse stiffness matrix multiplied by the eigenvector is equal to the eigenvalue multiplied by the product of the sparse mass matrix and the eigenvector.

[0081] The number of eigenvalues ​​extracted is set to ensure that high-dimensional feature reduction is completed within millisecond latency. In terms of algorithm selection, for the preset iterative solution algorithm, the system abandons the computationally complex full matrix eigenvalue decomposition, such as the QR decomposition algorithm, and instead adopts the Shift-and-Invert Lanczos iterative solution algorithm for large sparse matrices. This algorithm can transform the generalized eigenvalue problem close to the target frequency band into a standard eigenvalue problem for fast convergence by introducing a shift-invert transformation.

[0082] The core technical motivation for this selection is that the identity authentication system needs to focus on the feature components corresponding to macroscopic topology and mid-to-high frequency skeletal details, as well as the ultra-high frequency feature components used for subsequent material defense analysis. Therefore, a preset number is set before extraction, such as extracting the first 3,000 effective feature values, in order to fully cover the full-band information from macroscopic topological structure to millimeter-level facial feature details.

[0083] The Lanczos iterative algorithm can rapidly converge to the required wideband feature components by continuously approximating the extreme values ​​in the Krylov subspace. This iterative solution strategy for specific frequency band feature values ​​not only significantly reduces time complexity and avoids computing power overflow caused by global solutions, but also exhibits strong algorithm robustness when facing fluctuations in device computing power, meeting the hard constraints of real-time performance in seamless access scenarios.

[0084] In a preferred embodiment of the present invention, the method for generating an identity feature spectrum includes: presetting a low-frequency cutoff threshold corresponding to facial appearance changes; traversing the full-band feature value sequence, identifying feature values ​​with values ​​less than the low-frequency cutoff threshold, and marking them as deformation interference components; identifying feature values ​​with values ​​greater than or equal to the low-frequency cutoff threshold, and marking them as structural feature components; using a spectral wavelet transform operator to suppress the deformation interference components and perform energy enhancement processing on the structural feature components; recombining the processed structural feature components to obtain the identity feature spectrum.

[0085] This embodiment is a further specification of the method for generating identity feature spectrum by the spectral domain decoupling module, aiming to decouple non-rigid deformation tolerance and identity structure recognizability from the physical frequency domain dimension; the spectral domain decoupling module presets a low-frequency cutoff threshold corresponding to facial appearance changes;

[0086] Traverse the full-band feature value sequence, identify feature values ​​with values ​​less than the low-frequency cutoff threshold, and mark them as deformation interference components; identify feature values ​​with values ​​greater than or equal to the low-frequency cutoff threshold, and mark them as structural feature components; use the spectral wavelet transform operator to suppress the deformation interference components and perform energy enhancement processing on the structural feature components; reassemble the processed structural feature components to obtain the identity feature spectrum.

[0087] In the specific implementation logic, the spectral wavelet transform operator is executed by designing a piecewise bandpass filter kernel function: the filter kernel function assigns a constant attenuation weight close to zero to the deformation interference component in the frequency domain to achieve suppression processing;

[0088] For structural feature components, a nonlinear amplification weight greater than 1 is assigned to achieve energy enhancement processing. This nonlinear amplification weight is specifically constructed by calculating the natural logarithm of the ratio of the feature value to the low-frequency cutoff threshold and superimposing the basic gain constant. This logarithmic growth mechanism ensures that as the feature value extends to the high-frequency region, its enhancement amplitude gradually and smoothly converges, which not only highlights the skeletal topological information of the mid-to-high frequency range, but also prevents the ultra-high frequency noise from being over-amplified and causing feature distortion.

[0089] When an employee's laughter triggers a large-scale stretching of the cheek muscles, the low-frequency feature value fluctuations caused by this action are accurately captured and suppressed by the aforementioned near-zero weights. Through this multi-scale energy adjustment method, the system effectively smooths facial noise and highlights stable skeletal topological information. Even when the target object has complex facial expressions, it can still extract a high-fidelity identity description base, demonstrating the high cost-effectiveness and feature extraction stability of this solution in dynamic interaction scenarios.

[0090] In a preferred embodiment of the present invention, the method for generating a hot core signature with isochronous transformation invariance includes: obtaining a time scale parameter, which is used to control the range of hot core diffusion; calculating an exponential decay term with the natural logarithm as the base for each feature value in the identity feature spectrum; squaring the feature function value of each vertex in the feature function sequence; performing a weighted summation of the exponential decay term and the squared feature function value to generate a hot core signature value corresponding to each vertex; and aggregating the hot core signature values ​​of all vertices to form a hot core signature mapping map.

[0091] This embodiment is a further specification of the method for generating a heat kernel signature with isometric transformation invariance by the fingerprint generation module, aiming to construct a local geometric descriptor with mathematical invariance to isometric transformation in three-dimensional space by utilizing the thermal conduction dynamics characteristics;

[0092] The fingerprint generation module acquires timescale parameters, which control the range of heat core diffusion. Based on each feature value in the identity feature spectrum, it calculates an exponential decay term with the natural logarithm as the base. It then squares the feature function value of each vertex in the feature function sequence. Finally, it performs a weighted sum of the exponential decay term and the squared feature function value to generate the heat core signature value for each vertex. The heat core signature values ​​of all vertices are aggregated to form a heat core signature mapping. The specific formula for calculating the heat core signature is as follows:

[0093] ;

[0094] in, As vertex Time scale parameters The hot core signature value below; For summation, For each eigenvalue and its corresponding eigenfunction, there is an index sequence number. is the base of the natural logarithm; For the one-dimensional discrete node index integer in the triangular mesh topology graph, that is... This definition clarifies the mapping function. The input dimension structure is a discrete sequence of node indices rather than a continuous three-dimensional physical space coordinate vector. This transforms complex space mapping into efficient array addressing; The total number of vertices in the pure point cloud set; This refers to the total number of mid-to-high frequency feature components retained in the identity feature spectrum; The first in the identity feature spectrum A characteristic value, with the dimension of the reciprocal of the area. ,in It is a physical unit of length; For parameters on the real physical timescale, the dimensions are... ,in It is a physical quantity with the dimension of time; The preset thermal diffusivity has the following dimensions: This is used to balance the physical dimensions between eigenvalues ​​and time-scale parameters, ensuring the exponentiation... Satisfying the dimensionless mathematical constraint, such that As an exponential decay term with the natural logarithm as its base, it can characterize the rate of heat dissipation with time and frequency in accordance with the laws of physics. To address the issue of the retained first eigenvalue in finite element method solutions to generalized eigenvalue problems. eigenvalues Strictly rigidly bound corresponding characteristic functions at vertex indices The value at that location.

[0095] By using the preserved mid-to-high frequency eigenvalue indices, a subset of corresponding eigenfunctions is precisely extracted from the original full-band eigenfunction sequence, and then squared. It eliminates the symbolic ambiguity of the characteristic function in the solution process and avoids the misalignment of the eigenvalue and characteristic function dimensions and the data flow break caused by low-frequency elimination;

[0096] Time scale parameters The receptive field that determines local geometric features is smaller. Capturing micro-curvature changes, larger It reflects the macroscopic topological structure; during the access control recognition process, regardless of whether the user's head yaws or pitches, the mapping map can maintain a stable output of topological isomorphic features, verifying the excellent robustness of the feature mapping mechanism based on thermodynamic equations in dealing with spatial attitude disturbances.

[0097] In a preferred embodiment of the present invention, the method for converting a hot-core signature into a multidimensional spectral fingerprint histogram includes: statistically analyzing the distribution range of all values ​​in the hot-core signature mapping graph; dividing the distribution range into a preset number of value intervals; statistically analyzing the number of vertices falling into each value interval and calculating the frequency density of each value interval; arranging the frequency densities of all value intervals in interval order to construct a one-dimensional or multi-dimensional feature vector, and using the feature vector as a multidimensional spectral fingerprint histogram.

[0098] This embodiment is a further specification of the method by which the fingerprint generation module converts the hot core signature into a multidimensional spectral fingerprint histogram. It aims to eliminate the dependence of the feature descriptor on the grid vertex index sequence through statistical dimensionality reduction, thereby achieving zero-registration comparison. The fingerprint generation module statistically analyzes the distribution range of all values ​​in the hot core signature mapping graph and divides the distribution range into a preset number of value intervals. It counts the number of vertices falling into each value interval and calculates the frequency density of each value interval.

[0099] After the statistics are completed, the fingerprint generation module arranges the frequency density of all numerical intervals in interval order to construct one-dimensional or multi-dimensional feature vectors, and uses the feature vectors as multi-dimensional spectral fingerprint histograms. In order to achieve quantification and logical transparency in the frequency density calculation process, the system adopts the equal interval division method based on numerical extreme values ​​to set the numerical intervals. For example, the maximum and minimum values ​​of the distribution range are extracted, and the absolute numerical span is divided into discrete feature intervals at equal intervals to ensure that the span of each interval is consistent and to avoid the histogram flattening caused by equal frequency division, which would result in the loss of feature expression ability.

[0100] By dividing the number of vertices in a specific numerical range by the total number of vertices in the current point cloud set, the frequency density of that numerical range is obtained. This normalization design naturally possesses scale invariance to changes in point cloud resolution. In quantization, when the total number of vertices in the input high-fidelity facial point cloud is scaled due to changes in the acquisition distance, the number of vertices falling within a specific curvature range will also change proportionally.

[0101] Since the frequency density is calculated using the ratio of the number of vertices in the current interval to the total number of vertices, this proportional synchronous change is canceled out in the division operation, so that the final output frequency density value remains constant. In the actual verification scenario, the system extracts the frequency density distribution of the turning points of high curvature skeletons, directly compressing the high-dimensional spatial features into a fixed-length vector. This not only reduces storage overhead but also avoids the defect of traditional spatial registration algorithms that are prone to getting trapped in local optima, providing an efficient data foundation for subsequent feature retrieval.

[0102] In a preferred embodiment of the present invention, the method for determining whether the spectral distance meets the preset authentication pass conditions includes: using the chi-square distance algorithm or the Bach distance algorithm to calculate the difference metric between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint; obtaining preset strict authentication threshold and lenient authentication threshold, wherein the strict authentication threshold is less than the lenient authentication threshold;

[0103] If the difference metric is less than or equal to the strict authentication threshold, the authentication pass condition is met, and a high-confidence authentication success signal is output. If the difference metric is greater than the strict authentication threshold but less than the lenient authentication threshold, a multi-frame fusion verification mechanism is initiated to obtain the next frame of discrete 3D point cloud data for secondary verification. If the difference metric is greater than or equal to the lenient authentication threshold, the authentication pass condition is not met.

[0104] This embodiment is a further specification of the method by which the matching decision module determines whether the spectral distance meets the preset authentication pass conditions. It aims to establish a rigorous boundary condition processing logic to reduce the false rejection rate under extreme conditions. The matching decision module uses the chi-square distance algorithm or the Bach distance algorithm to calculate the difference metric between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint.

[0105] Obtain the preset strict authentication threshold and lenient authentication threshold, where the strict authentication threshold is less than the lenient authentication threshold; if the difference metric is less than or equal to the strict authentication threshold, the authentication pass condition is met, and a high-confidence authentication success signal is output; if the difference metric is greater than the strict authentication threshold but less than the lenient authentication threshold, the multi-frame fusion verification mechanism is started, and the next frame of discrete 3D point cloud data is obtained for secondary verification.

[0106] If the difference metric is greater than or equal to the lenient authentication threshold, the authentication pass condition is not met. It should be noted that the difference metric calculated by the chi-square distance algorithm or the Bach distance algorithm is the specific quantitative representation of the aforementioned spectral distance, and the two are consistent in physical meaning and judgment logic.

[0107] To eliminate the invisibility and uncertainty of algorithm logic caused by experience-based parameter tuning, the preset strict authentication threshold and lenient authentication threshold are obtained through offline sample statistics: during the system initialization and deployment phase, a large number of three-dimensional facial point clouds of employees with known identities under various daily expressions are collected as positive sample sets, and facial point clouds of unauthorized personnel are collected as negative sample sets, and the distribution of difference metric values ​​between all positive and negative sample pairs is calculated.

[0108] The system directly anchors the 95th percentile of the difference metric distribution of the positive sample set as the strict authentication threshold, ensuring that the vast majority of compliant requests are passed seamlessly in seconds; at the same time, it anchors the 5th percentile of the difference metric distribution of the negative sample set as the lenient authentication threshold. At the level of difference metric calculation, the core motivation for the system to choose the chi-square distance algorithm or the Bach distance algorithm is that the multidimensional spectral fingerprint histogram essentially reflects the frequency density distribution of local geometric features in each numerical interval.

[0109] Compared to the traditional Euclidean distance, the chi-square distance introduces the expected value of each interval as the denominator for dynamic normalization during calculation. In the identification of the skeletal structure of identity, some highly personalized bone point micro-protrusions have a low occurrence frequency. If Euclidean distance is used, these low-frequency key features are easily masked by high-frequency background information in large flat areas. However, the division penalty mechanism of the chi-square distance can non-linearly amplify the contribution weight of the distribution deviation of these low-frequency key intervals to the overall difference measure.

[0110] In a specific high-curvature feature region, due to the extremely low natural occurrence frequency of the micro-protrusions of this bone point, its expected frequency density in the standard identity spectrum fingerprint approaches zero. When the frequency density in this region shows a small absolute increase due to minor facial deformation of the test subject, the difference in Euclidean distance between the two is only a negligible linear deviation. However, under the chi-square distance measure, the square of this small absolute increase will be divided by the expected value approaching zero as the denominator, thus producing a significant penalty increment in the overall difference measure that is non-linearly amplified.

[0111] To avoid program crashes caused by an absolutely zero expected value, the system uniformly adds a very small positive smoothing constant to the denominator when calculating the chi-square distance, such as... This ensures the numerical stability of the algorithm under extreme conditions;

[0112] When an employee experiences slight facial swelling due to staying up late, resulting in a difference value greater than the strict authentication threshold but less than the lenient authentication threshold, the matching decision module does not directly trigger rejection. Instead, it initiates a multi-frame fusion verification mechanism to obtain the next frame of discrete 3D point cloud data for secondary verification. This multi-level decision mechanism, based on data-driven dual-threshold boundaries and in cases of boundary ambiguity, improves the user experience for requests with ambiguous boundaries while ensuring the system's basic security requirements, demonstrating the flexibility of the logical decision architecture.

[0113] In a preferred embodiment of the present invention, the adaptive update module is further configured to perform defense analysis, including: extracting ultra-high frequency feature components whose values ​​are located in a preset high frequency range from the identity feature spectrum; calculating the energy distribution density of the ultra-high frequency feature components; obtaining a preset reference range for the energy density of non-biological materials; determining whether the energy distribution density falls within the reference range for the energy density of non-biological materials; if the energy distribution density falls within the reference range for the energy density of non-biological materials, generating an attack alarm command and forcibly intercepting the authentication pass condition.

[0114] This embodiment further extends the functionality of the adaptive update module, specifying the logical steps for performing defense analysis. It aims to utilize spectral domain features to reflect the hidden characteristics of microscopic material physical properties, thereby achieving feature interception of three-dimensional physical forgery attacks. The adaptive update module extracts ultra-high frequency feature components whose values ​​are located in a preset high frequency range from the identity feature spectrum; and calculates the energy distribution density of the ultra-high frequency feature components.

[0115] The system clarifies the physical definition and acquisition source of the relevant parameters: the preset high frequency range is strictly defined as the extremely high frequency band in the last 10% of the full-band feature value sequence, which corresponds to the high-frequency geometric oscillation of micron-level facial pores and skin microtextures in manifold space.

[0116] In the process of calculating energy distribution density, in order to break the dependence on uninterpretable algorithm models, the system does not introduce complex prediction networks, but decomposes the index into the relative proportion of energy: specifically, by calculating the sum of the squares of the corresponding eigenvalues ​​of all characteristic components in the ultra-high frequency range, and dividing it by the total energy of the eigenvalues ​​of the entire frequency band, i.e. the sum of the squares of all eigenvalues, the normalized high-frequency energy proportion is obtained as the energy distribution density.

[0117] This normalization calculation method, where the numerator and denominator share the same physical dimension, not only effectively eliminates the physical interference of absolute distance and scale on the absolute energy value during point cloud acquisition, but also avoids the distortion of physical meaning caused by the interference of the number of features within the interval due to the use of the arithmetic mean; it obtains a preset reference interval for the energy density of non-biological materials; and determines whether the energy distribution density falls within the reference interval for the energy density of non-biological materials.

[0118] The preset reference range for the energy density of non-biological materials is obtained as follows: During the defense baseline configuration phase, the system imports point clouds of attack samples of common counterfeit materials such as silicone, 3D printed photosensitive resin and high-imitation latex, counts the ultra-high frequency energy distribution density of such non-biological materials, and extracts the maximum upper limit value of its statistical distribution, such as 0.05, as the upper limit boundary of the reference range, with the lower limit set to 0.

[0119] In situations where external personnel wearing high-fidelity resin masks attempt to breach access control, the resin material lacks the elastic modulus and micro-texture of real human skin at the micrometer scale, resulting in extremely weak ultra-high frequency characteristic components. Following this logic, the normalized energy distribution density of the high-fidelity resin mask is typically below 0.05, falling within the upper limit of the non-biological material energy density reference range of 0.05, thus significantly lower than the threshold. If the energy distribution density falls within the non-biological material energy density reference range, an attack alarm is generated, and the authentication pass condition is forcibly blocked. This defense mechanism requires no additional multispectral sensor hardware; it relies solely on microscopic acoustic vibration characteristics mined by pure geometric algorithms to detect and intercept physical forgery attacks using non-biological materials.

[0120] In a preferred embodiment of the present invention, the method for initiating a multi-frame fusion verification mechanism includes: continuously acquiring a preset number of supplementary point cloud data; independently performing manifold construction, spectral domain decoupling, and fingerprint generation steps on each frame of supplementary point cloud data to obtain the corresponding supplementary spectral fingerprint histogram; calculating the cosine similarity between the multidimensional spectral fingerprint histogram and all supplementary spectral fingerprint histograms, and using the arithmetic mean of the cosine similarity as the temporal consistency coefficient;

[0121] If the temporal consistency coefficient is greater than the preset stability threshold, the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint, as well as the average difference measure between all supplementary spectral fingerprint histograms and the standard identity spectral fingerprint, are calculated to obtain the average difference measure. Based on the average difference measure, it is re-evaluated whether the authentication pass condition is met. If the temporal consistency coefficient is less than or equal to the stability threshold, it is determined to be a non-liveness attack or abnormal data acquisition, and an authentication failure signal is directly output.

[0122] This embodiment is a further specification of the method for initiating a multi-frame fusion verification mechanism. It aims to filter transient environmental noise and discontinuous adversarial sample injection by evaluating the temporal consistency of features in the time dimension. After triggering multi-frame fusion, the system continuously collects a preset number of supplementary point cloud data. For each frame of supplementary point cloud data, manifold construction, spectral domain decoupling, and fingerprint generation steps are performed independently to obtain the corresponding supplementary spectral fingerprint histogram.

[0123] For the supplementary point cloud data of the preset number of frames, the system does not use arbitrary empirical constants, but dynamically constrains it based on the hardware sampling rate of the three-dimensional sensor and the physiological duration of human facial micro-expressions: for example, at a sampling rate of 30 frames per second, the subsequent 5 consecutive frames are extracted as supplementary data. This duration just covers a complete cycle of facial micro-muscle twitching, which can capture sufficient temporal topological changes and strictly prevent the infinite divergence of computational delay; the cosine similarity between the multidimensional spectral fingerprint histogram and all supplementary spectral fingerprint histograms is calculated, and the arithmetic mean of the cosine similarity is used as the temporal consistency coefficient;

[0124] When calculating the correlation between fingerprints between frames, the system explicitly chooses cosine similarity instead of Euclidean distance. The underlying technical motivation is to eliminate the spatial scale scaling interference during dynamic passage. Since the physical relative distance between the user and the 3D camera changes in real time during seamless passage, this spatial displacement will cause fluctuations in the overall scale of the facial point cloud captured in a single frame, resulting in an overall scaling of the absolute amplitude of the histogram frequency. However, as long as the identity of the target object is consistent, the relative proportions within each frequency band feature will remain constant.

[0125] Cosine similarity naturally shields the differences in the magnitude of multidimensional vectors, and measures the overlap of the distribution structure only by calculating the angle between the feature vectors in the multidimensional space, thereby decoupling the physical interference of distance fluctuations on the temporal consistency assessment.

[0126] In the process of determining temporal consistency, the stability threshold is not set arbitrarily, but is based on the quantitative statistical boundary of the natural physiological characteristics of living organisms: since the natural micro-movements of the human face in the real physical world will cause millisecond-level continuous inter-frame micro-fluctuations, but the core skeletal topology is still highly isomorphic, the stability threshold is anchored to 0.85.

[0127] The threshold is determined based on the quantitative statistics of thousands of natural physiological micro-movements of living faces, taking the lower limit of the confidence interval of their inter-frame similarity distribution, in order to distinguish physiological micro-movements from the rigid motion characteristics of non-biological materials.

[0128] If the temporal consistency coefficient is greater than the preset stability threshold, the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint, as well as the average difference measure between all supplementary spectral fingerprint histograms and the standard identity spectral fingerprint, are calculated to obtain the average difference measure. Based on the average difference measure, it is re-determined whether the authentication pass condition is met.

[0129] If the temporal consistency coefficient is less than or equal to the stability threshold, it is determined to be a non-liveness attack or abnormal data acquisition, and an authentication failure signal is directly output. This dynamic adaptive fusion strategy provides a secondary decision basis with statistical confidence for authentication requests with ambiguous boundaries.

[0130] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An identity recognition and authentication system based on machine vision 3D features, characterized in that, include: The data acquisition module is used to acquire discrete 3D point cloud data of the object to be authenticated, and to preprocess the discrete 3D point cloud data to remove outliers and generate a clean point cloud set. The manifold construction module is used to construct a triangular mesh model based on a pure point cloud set, calculate the Riemannian metric tensor of the triangular mesh model, construct the Laplace-Beltramian operator based on the Riemannian metric tensor, solve the generalized eigenvalue problem through finite element analysis, and obtain the full-band eigenvalue sequence and the corresponding eigenfunction sequence. The spectral domain decoupling module is used to introduce the spectral wavelet transform operator to perform multi-scale frequency domain bandpass filtering on the full-band feature value sequence, remove low-frequency feature components corresponding to large-scale non-rigid deformation, retain mid-to-high frequency feature components corresponding to the identity skeleton architecture, and reorganize the retained mid-to-high frequency feature components into an identity feature spectrum. generate The methods for establishing identity feature genealogies include: A preset low-frequency cutoff threshold corresponding to changes in facial appearance is set. Traverse the full-band feature value sequence, identify feature values ​​with values ​​less than the low-frequency cutoff threshold, and mark them as deformation interference components; Identify feature values ​​that are greater than or equal to the low-frequency cutoff threshold and mark them as structural feature components; The deformation interference component is suppressed by using the spectral wavelet transform operator, and the structural feature component is enhanced by energy. The processed structural feature components are then recombined to obtain the identity feature spectrum. The fingerprint generation module is used to generate a hot-core signature with isometry transformation invariance based on the identity feature spectrum and feature function sequence, and to convert the hot-core signature into a multidimensional spectral fingerprint histogram; the method for generating a hot-core signature with isometry transformation invariance includes: Obtain the timescale parameters, which are used to control the range of thermonuclear diffusion; For each feature value in the identity feature spectrum, calculate the exponential decay term with the natural logarithm as the base; Square the feature function value at each vertex in the feature function sequence; The exponential decay term and the squared feature function value are weighted and summed to generate the hot kernel signature value for each vertex. Aggregate the hot kernel signature values ​​of all vertices to form a hot kernel signature mapping graph; The matching decision module is used to obtain the pre-stored standard identity spectrum fingerprint, calculate the spectral distance between the multidimensional spectral fingerprint histogram and the standard identity spectrum fingerprint, and determine whether the spectral distance meets the preset authentication pass conditions. The adaptive update module is configured to: if the authentication pass conditions are met, perform a weighted update of the standard identity spectrum fingerprint based on the multidimensional spectral fingerprint histogram; if the authentication pass conditions are not met, trigger a denial response command; the adaptive update module is also used to perform defense analysis, including: Extract ultra-high frequency feature components whose values ​​are located in a preset high frequency range from the identity feature spectrum; Calculate the energy distribution density of ultra-high frequency characteristic components; Obtain a preset reference range for the energy density of non-biological materials; Determine whether the energy distribution density falls within the reference range for the energy density of non-biological materials; If the energy distribution density falls within the reference range of energy density for non-biological materials, an attack alarm command is generated, and the authentication pass condition is forcibly blocked.

2. The identity recognition and authentication system based on machine vision three-dimensional features according to claim 1, characterized in that, Methods for solving generalized eigenvalue problems using finite element analysis include: Based on the geometric topological relationships of the triangular mesh model, a sparse stiffness matrix and a sparse mass matrix are constructed. Substitute the sparse stiffness matrix and sparse mass matrix into the generalized eigenvalue equation; The generalized eigenvalue equation is decomposed using a preset iterative solution algorithm to calculate a preset number of eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are sorted in ascending order of their numerical values ​​to form a full-band eigenvalue sequence, and the eigenvectors are used as a sequence of eigenfunctions.

3. The identity recognition and authentication system based on machine vision three-dimensional features according to claim 2, characterized in that, Methods for converting hot nucleus signatures into multidimensional spectral fingerprint histograms include: The distribution range of all values ​​in the statistical hot core signature mapping graph; The distribution range is divided into a preset number of numerical intervals; Count the number of vertices falling into each numerical interval, and calculate the frequency density of each numerical interval; Arrange the frequency density of all numerical intervals in interval order to construct one-dimensional or multi-dimensional feature vectors, and use the feature vectors as multi-dimensional spectral fingerprint histograms.

4. The identity recognition and authentication system based on machine vision three-dimensional features according to claim 3, characterized in that, Methods for determining whether the spectral distance meets the preset authentication pass conditions include: The chi-square distance algorithm or Bach distance algorithm is used to calculate the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint; Obtain the preset strict authentication threshold and lenient authentication threshold, wherein the strict authentication threshold is less than the lenient authentication threshold; If the difference metric is less than or equal to the strict authentication threshold, the authentication pass condition is met, and a high-confidence authentication success signal is output. If the difference metric is greater than the strict authentication threshold but less than the lenient authentication threshold, the multi-frame fusion verification mechanism is activated to obtain the next frame of discrete 3D point cloud data for secondary verification. If the difference metric is greater than or equal to the lenient certification threshold, the certification pass condition is not met.

5. The identity recognition and authentication system based on machine vision three-dimensional features according to claim 4, characterized in that, Methods for initiating multi-frame fusion verification mechanisms include: Continuously acquire supplementary point cloud data for a preset number of frames; For each frame of supplementary point cloud data, manifold construction, spectral domain decoupling and fingerprint generation steps are performed independently to obtain the corresponding supplementary spectral fingerprint histogram. Calculate the cosine similarity between the multidimensional spectral fingerprint histogram and all supplementary spectral fingerprint histograms, and use the arithmetic mean of the cosine similarity as the temporal consistency coefficient; If the temporal consistency coefficient is greater than the preset stability threshold, the difference measure between the multidimensional spectral fingerprint histogram and the standard identity spectral fingerprint, as well as the average difference measure between all supplementary spectral fingerprint histograms and the standard identity spectral fingerprint, are calculated to obtain the average difference measure. Based on the average difference measure, it is re-determined whether the authentication pass condition is met. If the temporal consistency coefficient is less than or equal to the stability threshold, it is determined to be a non-liveness attack or abnormal data collection, and an authentication failure signal is directly output.

Images