A method and system for facial recognition and comparison in border inspection

Abstract

This invention relates to the field of facial image acquisition and comparison technology, and discloses a method and system for facial image acquisition and comparison for border inspection, comprising: acquiring an original intensity image with a preset polarization direction; based on the Stokes vector principle and a two-color reflection model, separating the acquired image into a specular reflection component characterizing surface texture and a diffuse reflection component characterizing subcutaneous tissue properties; calculating the spatial frequency domain energy distribution of the two components, using Boltzmann distribution weighting to determine the overlapping feature value of the diffuse specular frequency domain, and inverting the light scattering scale parameter of the current skin accordingly; constructing a tissue optical characteristic filter, modulating the diffuse component in the frequency domain to map it to a preset standard light scattering scale, and combining it with the specular component after energy whitening equalization to synthesize a normalized face image; generating a binarized final verification result by calculating the cosine similarity of the feature vectors of the normalized image and the document reference image, and combining it with an anomaly deviation penalty term.

Description

Technical Field

[0001] This invention relates to the field of facial image acquisition and comparison technology, and more specifically, to a method and system for facial image acquisition and comparison used in border inspection. Background Technology

[0002] With the increasing frequency of cross-border personnel flows globally, border control, as the first line of defense for national security, places extremely high demands on clearance efficiency and inspection accuracy. Currently, automated verification channels (E-gate) based on biometric image analysis and manual inspection counters are widely used at major ports of entry, aiming to confirm passenger identities by comparing the captured facial image with reference data within the document's chip. To ensure smooth operations at ports, the system needs to complete identity verification in a very short time, while also adapting to differences in image quality caused by passengers of varying heights and different types of data collection equipment.

[0003] However, in actual border inspection environments, lighting conditions are often complex and variable, and the difference in the standing position of passengers means that the imaging distance cannot be strictly fixed. This results in significant non-standardized fluctuations in the texture scale and sharpness of the acquired images, seriously affecting the accuracy of subsequent algorithms. Even more serious is the fact that, faced with increasingly sophisticated illegal border crossing methods, existing imaging technologies struggle to effectively defend against presentation attacks such as highly realistic silicone masks, high-definition printed photos, or electronic screen replays. Because traditional image acquisition methods primarily record the distribution of reflected light intensity on the object's surface and lack analysis of the object's material properties, the system is highly susceptible to misjudgment when faced with forged media that have a highly realistic appearance but lack the optical response of real skin, thus posing serious security risks. Summary of the Invention

[0004] This invention provides a method and system for capturing and comparing human images for border inspection, which solves the technical problems mentioned in the background art.

[0005] Firstly, a method for facial recognition and comparison in border inspection includes:

[0006] In the near-infrared acquisition environment of border inspection, a single-frame polarized face image is acquired, and the image is separated into a specular reflection component image that characterizes surface texture and a diffuse reflection component image that characterizes subcutaneous tissue characteristics by utilizing the difference in polarization state.

[0007] The spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image are calculated respectively, and the overlap characteristic value of the diffuse specular frequency domain is determined based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain.

[0008] Based on the overlapping feature values, the current skin light scattering scale parameters are inverted, a filter for tissue optical properties is constructed, and the diffuse reflection component image is mapped from the current light scattering scale to a preset standard light scattering scale using the filter.

[0009] A normalized face image is generated based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. The biometric similarity between the normalized face image and the document reference image is calculated, and the final verification result is generated by combining the abnormal deviation of the overlapping feature values.

[0010] Secondly, a facial recognition and comparison system for border inspection includes:

[0011] The data acquisition module acquires a single-frame polarized face image in the near-infrared acquisition environment of the border inspection, and uses the difference in polarization state to separate the image into a specular reflection component image that characterizes the surface texture and a diffuse reflection component image that characterizes the subcutaneous tissue characteristics.

[0012] The feature extraction module calculates the spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image respectively, and determines the overlapping feature value of the diffuse specular frequency domain based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain.

[0013] The normalization module inverts the current skin light scattering scale parameters based on the overlapping feature values, constructs a filter for normalizing tissue optical properties, and uses the filter to map the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale.

[0014] The comparison and verification module generates a normalized face image based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. It calculates the biometric similarity between the normalized face image and the document reference image, and generates the final verification result by combining the abnormal deviation of the overlapping feature values.

[0015] The beneficial effects of this invention are as follows: By analyzing polarization state differences and frequency domain energy distribution, the decoupling of skin surface texture information and subcutaneous tissue scattering characteristics is achieved. Furthermore, a tissue optical characteristic normalization filter based on a double exponential reflection model is introduced, fundamentally eliminating the nonlinear interference of shooting distance variations, lighting condition fluctuations, and individual skin translucency differences on imaging features. This invention not only generates standardized face images with a unified optical benchmark through reconvolution technology, significantly improving the robustness of feature extraction in complex scenarios for downstream deep learning recognition algorithms, but also accurately quantifies the coupling relationship between the photon mean free path and surface roughness unique to biological tissues using the overlapping feature values ​​of diffuse mirror frequency domains. This allows for highly sensitive identification and defense against highly realistic non-living attacks lacking true subcutaneous scattering characteristics, such as high-definition photos, silicone masks, and electronic screens, greatly enhancing the security and environmental adaptability of the identity verification system. Attached Figure Description

[0016] Figure 1 This is a flowchart of a facial recognition and comparison method for border inspection according to the present invention;

[0017] Figure 2 This is a block diagram of a facial recognition and comparison system for border inspection according to the present invention;

[0018] Figure 3 This is a schematic diagram of an implementation scenario of the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0020] Example 1: As Figure 1 As shown, a facial recognition and comparison method for border inspection includes:

[0021] In the near-infrared acquisition environment of border inspection, a single-frame polarized face image is acquired, and the image is separated into a specular reflection component image that characterizes surface texture and a diffuse reflection component image that characterizes subcutaneous tissue characteristics by utilizing the difference in polarization state.

[0022] The spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image are calculated respectively, and the overlap characteristic value of the diffuse specular frequency domain is determined based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain.

[0023] Based on the overlapping feature values, the current skin light scattering scale parameters are inverted, a filter for tissue optical properties is constructed, and the diffuse reflection component image is mapped from the current light scattering scale to a preset standard light scattering scale using the filter.

[0024] A normalized face image is generated based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. The biometric similarity between the normalized face image and the document reference image is calculated, and the final verification result is generated by combining the abnormal deviation of the overlapping feature values.

[0025] In a preferred embodiment, the image is separated into a specular reflection component image characterizing surface texture and a diffuse reflection component image characterizing subcutaneous tissue properties using differences in polarization states, including:

[0026] Get The original intensity images in the four polarization directions are denoted as follows: ;

[0027] The total light intensity image is calculated using the following formula. and polarization intensity image :

[0028] ;

[0029] ;

[0030] Setting the mirror polarization coefficient The initial specular reflection component image is calculated according to the following formula. and the initial diffuse reflection component image :

[0031] ;

[0032] The final diffuse reflection component image is calculated according to the following formula. And the final specular reflection component image :

[0033] ;

[0034] in, It is a preset constant and .

[0035] Preferably, based on the theory of light polarization transmission and the Stokes vector principle, the reflected light from skin or other semi-transparent media is decoupled.

[0036] First, you need to obtain The original intensity images in the four polarization directions are denoted as follows: These four images were acquired under the same field of view by rotating a polarizer or using a split-focus plane polarization camera, and represent the intensity distribution of linearly polarized light at four different phase angles.

[0037] Subsequently, based on the Stokes vector definition, the total light intensity image... This represents the sum of all light energy incident on the sensor surface, corresponding to the first component of the Stokes vector. Based on the principles of energy conservation and polarization orthogonality, the calculation formula is as follows: ,in for Pixel intensity in direction, for Pixel intensity in the direction;

[0038] Simultaneously, in order to quantify the polarization component in the reflected light, it is necessary to calculate the polarization intensity image. This image characterizes the intensity components of reflected light that retain polarization characteristics, and its calculation is based on the second component of the Stokes vector. With the third component The vector composition is calculated using the following formula: This formula calculates the modulus of the linear polarization vector using the Euclidean norm, effectively extracting polarization information dominated by surface specular reflection. and They are respectively and The original intensity image pixel value in the direction.

[0039] After obtaining the total light intensity and polarization intensity, in order to achieve the physical separation of diffuse reflection and specular reflection, the specular polarization coefficient needs to be set. This parameter is a dimensionless quantity used to characterize the degree of polarization of specular reflection in light reflected from a specular surface under a specific incident angle and material refractive index. Its value range is... Based on the prior of the dichromatic reflection model, diffuse reflected light becomes unpolarized due to multiple scatterings in subcutaneous tissue, while specular reflected light partially or completely retains the polarization state of the incident light. Therefore, the polarization intensity image... It originates solely from the specular reflection component, therefore it can be obtained through Inversely calculate the initial specular reflection component image The calculation formula is: By dividing by To include only the polarization portion This is restored to the complete specular reflection intensity, including both polarized and unpolarized components; furthermore, based on the principle of light intensity superposition, the initial diffuse reflection component image is obtained. It can be obtained by subtracting the specular reflection component from the total light intensity, i.e. .

[0040] However, in actual imaging processes, due to sensor noise, uneven illumination, or parameter issues... The effect of estimation bias is obtained through direct calculation. Negative values ​​may appear at some pixels, which violates the axiom that light energy is non-negative. Therefore, a nonlinear correction is needed to calculate the final diffuse reflection component image. And the final specular reflection component image Among them, the final diffuse reflection component image The calculation formula is: The rectification operation is implemented using algebraic methods, where... Equivalent to absolute value ,when hour, ,when hour, This filters out negative anomalies caused by noise, ensuring the effectiveness of the diffuse component; finally, to ensure the conservation of total energy, the final specular reflection component image... Through formula Obtain. Regarding the specifics, This is a preset constant, and its preferred value is usually determined based on Fresnel equations combined with the specific geometric optical path of the imaging system. For human skin tissue, when the incident angle is small (close to perpendicular incident), considering that the refractive index of the skin epidermis is approximately 1.43 and the refractive index of air is 1.0, The preferred value range is between 0.5 and 0.8. In a standard normal incidence polarization imaging system, the preferred setting is... A priori value of 0.5 is used, which reflects the energy distribution characteristics between cross-polarized and parallel-polarized channels, enabling the separated diffuse reflectance component image, which characterizes subcutaneous tissue properties, to have the highest contrast and tissue detail clarity.

[0041] In a preferred embodiment, the spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image are calculated respectively, including:

[0042] Calculate the interpupillary distance The specular reflection component image and the diffuse reflection component image are geometrically scaled so that the scaled interpupillary distance is equal to a preset constant. The normalized diffuse reflection component image is obtained. and specular reflection component image ;

[0043] Select the region of interest on the face and apply a two-dimensional window function. The windowed frequency domain signal is calculated according to the following formula. and :

[0044] ;

[0045] in, This represents a two-dimensional Fast Fourier Transform. Represents two-dimensional frequency coordinates;

[0046] Calculate power spectral density and ;

[0047] The radial frequency is calculated using the following formula. Spatial frequency domain energy distribution and :

[0048] ;

[0049] ;

[0050] in, Indicates that the condition is met. Calculate the average value of all pixels. Take a positive integer.

[0051] Preferably, it is necessary to eliminate the influence of changes in shooting distance on texture scale analysis and extract isotropic frequency domain features.

[0052] First, the scale invariance problem needs to be addressed, because different shooting distances cause the same skin texture to occupy different pixel widths in the image, thus altering its spatial frequency characteristics. Therefore, geometric normalization is necessary. Specifically, a facial landmark detection algorithm is used to identify the center positions of both eyes in the image and calculate the interpupillary distance. (Unit: pixels), then based on preset constants The specular reflection component image With the diffuse reflection component image Perform geometric scaling, with a scaling ratio of 100%. The image is resampled using a bicubic interpolation algorithm so that the scaled interpupillary distance is equal to Thus, the normalized diffuse reflection component image is obtained. and specular reflection component image Here, The optimal value is based on the standardized face analysis resolution requirements and the computational efficiency of the Fast Fourier Transform (FFT), and is usually preferably set to [value missing]. to Between pixels, preferably Pixels are used to ensure that skin texture details have sufficient resolution in the frequency domain without introducing excessive interpolation noise.

[0053] Next, in order to suppress spectral leakage caused by image edge truncation during frequency domain transformation, it is necessary to select the region of interest (flat area of ​​the cheek or forehead) and apply a two-dimensional window function. The window function The preferred method is a two-dimensional Hanning window or Hamming window, which smoothly transitions image edges to zero. Based on signal processing theory, the frequency domain signal after windowing is calculated using a two-dimensional fast Fourier transform. and The calculation formulas are as follows: and ,in This represents the two-dimensional Fast Fourier Transform operator. Indicates pixel-by-pixel multiplication. Representing two-dimensional frequency coordinates This is to transform the grayscale distribution in the spatial domain into a complex spectrum in the frequency domain.

[0054] Subsequently, the power spectral density was calculated based on the concept of energy density related to Parseval's Theorem. and ,in The modulus of the complex number is represented by the formula, which quantifies the energy intensity at different frequency components.

[0055] Finally, in order to extract statistical features independent of texture orientation, based on the prior knowledge that skin microtexture is statistically isotropic, the two-dimensional power spectrum is compressed into a one-dimensional radial distribution, and the radial frequency is calculated according to the formula. Spatial frequency domain energy distribution and : as well as In this formula, Represents the frequency vector The Euclidean norm (i.e., the frequency radius). Take positive integers, interval Defines a frequency domain with the origin as the center and a radius of . A circular region of unit width, This represents the arithmetic mean of all pixels falling within the annular region; specifically, it is achieved through a discretized annular integral average, by traversing all integer radii from low to high frequencies. The two-dimensional spectrum is projected into a one-dimensional energy-frequency curve, where The high-frequency attenuation rate of the curve directly characterizes the scattering properties of subcutaneous collagen fibers, while The distribution pattern reflects the smoothness and luster of the skin's surface oils and stratum corneum.

[0056] In a preferred embodiment, determining the overlap characteristic value of the diffuse mirror frequency domain based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain includes:

[0057] The radial frequency is calculated using the following formula. logarithmic spectrum difference at the location :

[0058] ;

[0059] in, and These represent the spatial frequency domain energy distributions of the diffuse reflection component image and the specular reflection component image, respectively. To prevent the use of tiny constants that are meaningless in logarithms;

[0060] The radial frequency is calculated using the following formula. Corresponding soft weights :

[0061] ;

[0062] in, The preset temperature coefficient, The maximum radial frequency;

[0063] The soft crossover frequency is calculated using the following formula. :

[0064] ;

[0065] The frequency domain overlap characteristic value of diffuse mirrors is calculated according to the following formula. :

[0066] ;

[0067] in, Interpupillary distance, These are preset system constants.

[0068] Preferably, it is necessary to find the critical frequency point where the surface texture (mainly dominated by specular reflection) and subcutaneous tissue scattering (mainly dominated by diffuse reflection) reach a balance in the frequency domain energy, and improve the noise robustness of this feature point through a soft thresholding method.

[0069] First, it is necessary to quantify the energy difference between the two components at different frequencies, and then calculate the radial frequency according to the formula. logarithmic spectrum difference at the location The logarithmic transformation is introduced to compress the large dynamic range of the power spectrum into a linear interval for easier comparison. and These represent the spatial frequency domain energy distributions of the diffuse reflection component image and the specular reflection component image, respectively. To prevent the use of infinitesimally small constants that are logarithmically meaningless, stability protection is used in numerical calculations to prevent mathematical singularities caused by zero energy in a certain frequency band. The preferred value range is to Preferred .

[0070] Secondly, in order to accurately locate the energy equilibrium point (i.e. To determine the location of the radial frequency and suppress random fluctuations at a single frequency point, a weighting mechanism based on the Boltzmann distribution is employed, and the radial frequency is calculated according to the formula. Corresponding soft weights This transforms the absolute value of the logarithmic spectrum difference into a probability density function, where the denominator is a normalization factor, ensuring that the sum of the weights of all frequency points is 1; in this formula, The maximum radial frequency is typically taken as half the minimum side length of the image. For a preset temperature coefficient, this parameter determines the sharpness or selectivity of the weight distribution. When the frequency approaches 0, the algorithm degenerates into finding the single frequency point with the smallest absolute difference (hard maximum value). When the weights are large, the weight distribution tends to be smooth. The optimal value needs to be determined based on the numerical range of the logarithmic spectrum, and the preferred range is usually [value missing]. to Preferred settings To achieve the best balance between positioning accuracy and noise resistance smoothness.

[0071] Subsequently, the frequencies are weighted and summed using the calculated weights, and the soft crossover frequency is calculated according to the formula. , The position of the center frequency that characterizes the competition between specular reflection and diffuse reflection energy reflects the relative proportion between skin surface gloss and subcutaneous scattering intensity.

[0072] Finally, in order to map the soft cross frequencies back to the true scale and make them comparable between individuals, the overlapping eigenvalues ​​of the diffuse mirror frequency domain need to be calculated according to the formula. ,in Interpupillary distance (in pixels). As a preset system constant (i.e., the target interpupillary distance for normalized scaling, preferably 200 pixels), this formula eliminates the scale effect brought about by the image resolution normalization process through a reverse scaling operation, making... It is a biometric quantity that is independent of shooting distance and is only related to the optical properties of the skin of the subject being tested.

[0073] In a preferred embodiment, the current skin light scattering scale parameters are inverted based on the overlapping feature values, and a filter for normalizing tissue optical properties is constructed, including:

[0074] Pre-establish mapping functions The current skin light scattering scale parameters are calculated according to the following formula. :

[0075] ;

[0076] in, This refers to the soft crossover frequency (i.e., the correlation quantity of overlapping eigenvalues ​​before geometric scaling). For offline fitting of a polynomial or lookup table based on a double exponential reflection model;

[0077] Define the frequency domain response model of the exponentially diffusing kernel. A tissue optical property normalization filter is constructed according to the following formula. :

[0078] ;

[0079] in, These are preset standard light scattering scale parameters. To prevent tiny constants with a denominator of zero, It is a two-dimensional frequency coordinate.

[0080] Preferably, deconvolution and reconvolution are needed to eliminate the different degrees of texture blurring caused by differences in skin translucency among different subjects, thereby standardizing the diffuse components of all images to a unified optical scattering benchmark.

[0081] First, a mapping function needs to be established in advance. The mapping function describes the soft crossover frequency in the frequency domain. Effective light scattering scale parameters in space The nonlinear correspondence between them is usually established based on Monte Carlo photon transport simulation using a double exponential reflection model or by calibration using a standard optical phantom, and then fitted to obtain a polynomial expression. ,in These are the fitting coefficients. The mean free path of photons after multiple scattering in subcutaneous tissue is characterized on the image plane at a projection scale (in pixels); based on this mapping function and soft crossover frequency... This allows for the calculation of the current skin light scattering scale parameters. .

[0082] Subsequently, in order to model the light scattering effect in the frequency domain, a frequency domain response model of the exponential diffusion kernel is defined. The model is based on the point spread function (PSF) of biological tissue, which approximates spatial exponential decay. In fact, according to the properties of the two-dimensional Fourier transform, its frequency domain expression is preferably a Lorentz-type distribution: ,in The modulus of the two-dimensional frequency coordinate system. The formula, which is the scattering scale parameter, describes how the texture details at different frequencies decay as the scattering distance increases, i.e., high-frequency information decays faster than low-frequency information.

[0083] Based on this, and according to the regularization concept of Wiener Deconvolution, a tissue optical property normalization filter is constructed. The calculation formula is: The formula consists of two parts: the denominator term. Acting on the original signal, it aims to remove the current skin-specific scattering blur (deconvolution), molecular terms The action applied to the signal aims to impose a standard scattering blur (reconvolution), thereby normalizing the optical properties; where, The preset standard light scattering scale parameter is selected based on the statistical analysis of the average scattering characteristics of a large number of healthy young skin samples, within the normalized interpupillary distance. In the pixel coordinate system, Preferred setting is to Between pixels, preferably Pixels represent the baseline for the translucency of standard skin.

[0084] In a preferred embodiment, mapping the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale using the filter includes:

[0085] The mapped diffuse reflection component image is calculated according to the following formula. :

[0086] ;

[0087] in, This is the normalized diffuse reflectance component image. This represents a two-dimensional Fast Fourier Transform. This represents the two-dimensional inverse fast Fourier transform. To organize the optical properties of the normalized filter, This indicates element-wise multiplication.

[0088] Preferably, the deconvolution and reconvolution operations in the spatial domain should be implemented using frequency domain multiplication based on the convolution theorem in linear system theory.

[0089] First, obtain the normalized diffuse reflection component image. and tissue optical property normalization filter Calculate the mapped diffuse reflection component image according to the formula. During this calculation process, the operator This refers to the two-dimensional Fast Fourier Transform (2DFastFourier Transform), which transforms spatial domain images... It decomposes into sinusoidal components of different frequencies to obtain its spectral representation, thereby transforming the image from a pixel coordinate system to a frequency coordinate system, so that operations on texture details are transformed into adjustments to frequency amplitudes; symbol This represents element-wise multiplication (Hadamard Product), which is the multiplication of complex values ​​at corresponding frequency coordinates. This operation shapes the spectrum of the original image based on the filter. The gain characteristics of the filter are such that if the current skin scattering is strong (causing image blurring), the filter will have a gain greater than 1 in the high-frequency part to compensate for the loss of detail; if the current skin scattering is weak, the gain will be less than 1 to introduce normalized blur, thereby forcibly correcting the optical transfer function (OTF) of the image to a standard state in the frequency domain; the operator This refers to the 2D Inverse Fast Fourier Transform, which restores the modulated complex spectrum back to the spatial domain, generating a real image with normalized optical properties. .

[0090] It should be noted that in actual algorithm implementation, due to the limitations of computer floating-point arithmetic precision, the result after the inverse transformation may contain an extremely small imaginary part. The final value should be the real part of the inverse transform result; furthermore, to ensure computational accuracy and prevent spectral aliasing and quantization errors, the above frequency domain operations are preferably performed in double-precision (64-bit) floating-point data format; the final obtained image is the mapped diffuse reflection component image. It eliminates the visual bias in texture caused by individual differences in skin translucency.

[0091] In a preferred embodiment, generating a normalized face image based on the mapped diffuse reflection component image and the energy-equalized specular reflection component image includes:

[0092] The energy whitening weight is calculated using the following formula. :

[0093] ;

[0094] in, The spatial frequency domain energy distribution of the specular reflection component image. It is a tiny constant;

[0095] The specular reflection component image after energy equalization is calculated according to the following formula. :

[0096] ;

[0097] in, This is the normalized image of the specular reflection component.

[0098] The normalized face image is calculated according to the following formula. :

[0099] ;

[0100] in, This is the mapped diffuse reflection component image. Preservation factor for preset surface details.

[0101] Preferably, the subcutaneous tissue information normalized by optical properties and the surface texture information enhanced by spectral whitening need to be weighted and recombined to synthesize a standard image that eliminates the differences in light scattering between individuals and preserves high-frequency skin details.

[0102] First, frequency domain energy equalization is performed on the specular reflection component to eliminate low-frequency light intensity fluctuations caused by uneven lighting or varying degrees of oil reflection, while simultaneously enhancing the contrast of fine textures (such as pores and fine lines). The energy whitening weight is then calculated using a formula. ,in The spatial frequency domain energy distribution of the specular reflection component image. The formula, which represents the magnitude of the frequency vector, constructs an isotropic whitening filter using the inverse square root of the power spectral density, thus flattening the spectrum from a statistical perspective.

[0103] Subsequently, the specular reflection component image after energy equalization was calculated according to the formula. This process first utilizes a two-dimensional fast Fourier transform. Normalized specular reflection component image Transform to the frequency domain, then with weights Perform element-wise multiplication, and finally perform inverse transformation. Restoring to the spatial domain; this operation is equivalent to enhancing the texture of the image, so that the surface details in different areas have a uniform energy response and are no longer affected by local highlights or shadows.

[0104] Finally, to synthesize the final visual effect, the normalized face image is calculated according to the formula. ,in This is the mapped diffuse reflection component image (representing normalized skin tone and light perception). The preset surface detail preservation coefficient is used to control the sharpness of surface texture in the synthesized image. Its value directly affects the texture of the final image. If the image size is too large, the image noise will be obvious; if it is too small, details will be lost and the image will appear to be smoothed. According to the characteristics of human visual vision and dermatological clinical image standards, The preferred value range is to To achieve a balance between clarity and realism, the optimal settings are... The final result This means that standardized face images are not only unified in geometric scale, but also standardized in light scattering characteristics and surface texture energy.

[0105] In a preferred embodiment, the biometric similarity between the normalized face image and the ID reference image is calculated, and the final verification result is generated by combining the abnormal deviation of the overlapping feature values, including:

[0106] Extract the feature vector of the normalized face image and the feature vector of the document reference image The cosine similarity is calculated according to the following formula. :

[0107] ;

[0108] in, This represents the transpose of a vector. Represents the magnitude of a vector;

[0109] The abnormal deviation penalty is calculated according to the following formula. :

[0110] ;

[0111] in, The frequency domain overlap characteristic value of the diffuse mirror is... This is the preset statistical mean of the population;

[0112] The final composite score is calculated using the following formula. :

[0113] ;

[0114] in, The default penalty weight;

[0115] The final verification result of binarization is output according to the following formula. :

[0116] ;

[0117] in, For the Sigmoid function, The preset system threshold, The preset hardening coefficient is used; an output of 1 indicates a pass, and an output of 0 indicates a fail.

[0118] Preferably, deep learning-based identity recognition and physical optics-based liveness detection are combined to eliminate attack media with non-real skin scattering characteristics (such as high-resolution photos, silicone masks, or electronic screens) using statistical anomaly detection mechanisms.

[0119] First, high-dimensional feature representations are extracted from the normalized face image generated in the previous step and the pre-stored document reference image in the database using a pre-trained deep convolutional neural network (such as ResNet or ArcFace architecture), respectively, to obtain feature vectors. and These two vectors typically reside on a hypersphere of 512 dimensions or higher; then, to quantify their geometric consistency in the feature space, cosine similarity is calculated according to the formula. ,in This represents the transpose of a vector. The Euclidean norm (modulus) of a vector is used to characterize the similarity of identities by calculating the cosine of the angle between two vectors. The value usually ranges from -1 to 1, and the closer the value is to 1, the more similar the identities are.

[0120] At the same time, in order to introduce an anti-counterfeiting dimension, it is necessary to calculate the abnormal deviation penalty term. According to the formula Perform the calculation; in this formula, The value represents the frequency domain overlap characteristic of diffuse specular surfaces. This characteristic reflects the coupling relationship between the mean free path of subcutaneous photons and the surface texture frequency of the imaged object. The mean value is a pre-defined population statistic, calculated based on a large-scale sample of real faces under the same normalization conditions. The expected value obtained through statistical analysis is based on the prior knowledge that the optical properties of real skin follow a log-normal distribution, for the normalized pupillary distance. Pixel system The preferred value range is to Preferred This penalty item The current object was measured using the square of the logarithmic distance. The degree of deviation from the actual population mean, if Significant deviation (e.g., paper-printed photos) Typically very small, and semi-transparent silicone masks (usually extremely large), then The value increases dramatically. Subsequently, the identity similarity score is combined with the material anomaly penalty, and the final composite score is calculated using the formula. ,in The preset penalty weight is used to adjust the influence of material inspection in the overall verification. Too high a threshold might mistakenly exclude legitimate users with specific skin types, while too low a threshold would fail to defend against highly realistic attacks. Based on the balance curve between the false acceptance rate (FAR) and the false rejection rate (FRR), The preferred value is to Preferred .

[0121] Finally, to output the pass / fail conclusion, the final binary verification result is output according to the formula. In this formula, It is the standard Sigmoid activation function. The preset system threshold is determined based on security level requirements. Usually set at to Between, preferably ; The preset hardening coefficient is used to control the slope of the activation function. As it approaches infinity, the continuous function converges to the unit step function (Heaviside Step Function). In actual software algorithm implementation, It is preferable to set it to a large finite constant (e.g.) To simulate the step effect; when hour, Output (Indicates verification passed) When hour, Output (Indicates verification failed), thus achieving high-precision and highly anti-counterfeiting automated identity verification.

[0122] Example 2: Figure 2 As shown, a facial recognition and comparison system for border inspection is characterized by comprising:

[0123] The data acquisition module acquires a single-frame polarized face image in the near-infrared acquisition environment of the border inspection, and uses the difference in polarization state to separate the image into a specular reflection component image that characterizes the surface texture and a diffuse reflection component image that characterizes the subcutaneous tissue characteristics.

[0124] The feature extraction module calculates the spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image respectively, and determines the overlapping feature value of the diffuse specular frequency domain based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain.

[0125] The normalization module inverts the current skin light scattering scale parameters based on the overlapping feature values, constructs a filter for normalizing tissue optical properties, and uses the filter to map the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale.

[0126] The comparison and verification module generates a normalized face image based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. It calculates the biometric similarity between the normalized face image and the document reference image, and generates the final verification result by combining the abnormal deviation of the overlapping feature values.

[0127] like Figure 3 As shown, the near-infrared image acquisition environment applied to the border inspection channel is as follows: Passengers stand in front of the acquisition terminal, which integrates a near-infrared supplementary lighting component, a polarization camera, and a document reader. Both the polarization camera and the document reader are communicatively connected to a computing terminal, which optionally connects to a document image library. The computing terminal outputs the verification results to the border inspection station display for staff to judge. This scenario is used to achieve single-frame polarization face acquisition, standardized face generation, and automatic verification closed loop with document reference images.

[0128] During the acquisition phase, the micro-polarization array camera is controlled to acquire four original intensity images at different polarization angles (e.g., 0°, 45°, 90°, 135°) in a single exposure. The total light intensity image and polarization intensity image are calculated based on the Stokes vector principle. The specular polarization coefficient is set, and the polarization intensity image is converted to obtain the initial specular reflection component image. The initial diffuse reflection component image is obtained by subtracting the specular reflection component from the total light intensity image. Subsequently, a non-negativity constraint operation is performed on the diffuse reflection component to obtain the final diffuse reflection component image. The final specular reflection component image is obtained from the difference between the total light intensity and the final diffuse reflection component, thus completing the physical decoupling of the specular reflection component characterizing surface texture and the diffuse reflection component characterizing subcutaneous tissue properties.

[0129] In the feature extraction stage, the interpupillary distance is first calculated based on the facial key point detection results, and the specular reflection component and the diffuse reflection component are geometrically normalized using a preset system constant to ensure that the interpupillary distances in the normalized images are consistent. A fixed region of interest for the face is extracted from the normalized image and a two-dimensional window function is applied. A two-dimensional fast Fourier transform is performed on the windowed signal and the power spectral density is calculated. In the frequency domain, multiple concentric ring regions are divided with zero frequency as the center, and the frequency components in each ring are averaged to obtain the radial spatial frequency domain energy distribution of the specular reflection component and the diffuse reflection component.

[0130] Furthermore, based on the idea that the two components achieve energy balance in the frequency domain, the logarithmic spectrum difference is calculated and soft weights are introduced to obtain the soft crossover frequency. Then, the frequency domain overlap characteristic value of the diffuse mirror is determined by combining the pupillary distance and the system constant. This characteristic value is used to characterize the coupling relationship between subcutaneous scattering characteristics and surface texture frequency, and to improve noise resistance robustness.

[0131] In the normalization stage, a mapping function between the soft crossover frequency and the effective light scattering scale parameter is pre-established. The current skin light scattering scale parameter is inverted based on the overlapping correlation quantity, and the frequency domain response model of the diffusion kernel is defined to construct a tissue optical property normalization filter. This filter is used to perform frequency domain mapping on the diffuse reflection component, so that the diffuse reflection component is mapped from the current light scattering scale to the preset standard light scattering scale, thereby eliminating the difference in texture blurring caused by the difference in skin translucency among different individuals.

[0132] In the synthesis and verification stage, the energy whitening weight is calculated based on the spatial frequency domain energy distribution of the specular reflection component. The normalized specular reflection component is subjected to frequency domain filtering and inverse transformation to obtain the specular reflection component after energy equalization. The specular reflection component is multiplied by a preset surface detail preservation coefficient and linearly superimposed with the mapped diffuse reflection component to generate a normalized face image, so as to achieve the standardization of subcutaneous scattering characteristics while preserving high-frequency surface details.

[0133] Subsequently, a pre-trained near-infrared feature extraction network is used to extract the first feature vector from the normalized face image, and a pre-trained visible light feature extraction network is used to extract the second feature vector from the document reference image. The cosine similarity between the two feature vectors is calculated as the identity similarity. At the same time, an abnormal deviation penalty term between the overlapping feature values ​​and the preset population statistical mean is calculated, and the identity similarity and penalty term are fused according to preset weights to obtain the final composite score. Finally, the final composite score is compared with the system threshold, and the binarized final verification result is output and displayed as pass / fail at the border inspection station.

[0134] It is important to note that all input data described in this solution is acquired in real-time through legal and compliant hardware interfaces with the user's full knowledge, explicit consent, and active cooperation. The preset parameters, prior constants, and statistical means are all derived from publicly available scientific literature data, de-identified general research datasets, or calibration data from laboratory environments, and do not contain any unauthorized sensitive third-party information. The system's data processing is limited to local or volatile memory computation transmitted via encrypted channels, aiming to complete the technical logic loop of identity verification and liveness detection. There is no situation of illegally collecting, stealing, or retaining user biometric data or infringing on user privacy without the user's knowledge. All parameter calls and generation comply with the principles of data minimization and legality, legitimacy, and necessity.

[0135] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for facial recognition and comparison in border inspection, characterized in that, include: In the near-infrared acquisition environment of border inspection, a single-frame polarized face image is acquired, and the image is separated into a specular reflection component image that characterizes surface texture and a diffuse reflection component image that characterizes subcutaneous tissue characteristics by utilizing the difference in polarization state. Calculate the spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image respectively, and determine the overlap feature value of the diffuse specular frequency domain based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain, including: Logarithmic operations are performed on the spatial frequency domain energy distribution of the diffuse reflection component image and the spatial frequency domain energy distribution of the specular reflection component image, and the absolute value of the difference between the two is calculated to obtain the logarithmic spectrum difference distribution. The logarithmic spectrum difference distribution is converted into a normalized soft-weighted distribution using an exponential decay function, where the frequency points with smaller differences correspond to larger weights. The soft-weighted distribution is used to perform a weighted summation of all radial frequencies to obtain the soft crossover frequency. The soft crossover frequency is multiplied by the interpupillary distance and divided by a preset system constant to obtain the final diffuse specular frequency domain overlap feature value. Based on the facial landmark detection results, the interpupillary distance is calculated. The interpupillary distance is then used to geometrically normalize the specular reflection component image and the diffuse reflection component image, ensuring that the interpupillary distance in the processed image equals a preset system constant. A fixed region of interest (ROI) is extracted from the normalized image, and a two-dimensional window function is applied to this ROI to generate a windowed signal. A two-dimensional fast Fourier transform is performed on the windowed signal, and the square of the modulus of the transform result is calculated to obtain the power spectral density. In the frequency domain, multiple concentric annular regions are divided with zero frequency as the center. The average value of all frequency components within each concentric annular region is calculated, and this average value is used as the spatial frequency domain energy distribution at the corresponding radial frequency. Based on the overlapping eigenvalues, the current skin light scattering scale parameters are inverted, and a filter for normalizing tissue optical properties is constructed, including: Pre-establish mapping functions The current skin light scattering scale parameters are calculated according to the following formula. : ； in, For soft crossover frequency, For offline fitting of a polynomial or lookup table based on a double exponential reflection model; Define the frequency domain response model of the exponentially diffusing kernel. A tissue optical property normalization filter is constructed according to the following formula. : ； in, These are preset standard light scattering scale parameters. To prevent tiny constants with a denominator of zero, Two-dimensional frequency coordinates; The filter is used to map the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale; A normalized face image is generated based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. The biometric similarity between the normalized face image and the document reference image is calculated, and the final verification result is generated by combining the abnormal deviation of the overlapping feature values.

2. The facial recognition and comparison method for border inspection according to claim 1, characterized in that, The image is separated into a specular reflection component image representing surface texture and a diffuse reflection component image representing subcutaneous tissue characteristics by utilizing differences in polarization states, including: A micro-polarization array camera is controlled to acquire four raw intensity images with different polarization angles in a single exposure. Based on the raw intensity images, the total light intensity image and polarization intensity image in the Stokes vector are calculated. A fixed specular polarization coefficient is set, and the polarization intensity image is divided by the specular polarization coefficient to obtain an initial specular reflection component image. The initial diffuse reflection component image is obtained by subtracting the initial specular reflection component image from the total light intensity image. A non-negativity constraint operation is performed on the initial diffuse reflection component image to obtain the final diffuse reflection component image. The final specular reflection component image is obtained by subtracting the final diffuse reflection component image from the total light intensity image.

3. The facial recognition and comparison method for border inspection according to claim 2, characterized in that, Mapping the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale using the filter includes: A two-dimensional fast Fourier transform is performed on the normalized diffuse reflection component image to obtain the diffuse radio frequency domain signal; the diffuse radio frequency domain signal is multiplied element-wise with the tissue optical property normalization filter to obtain the normalized diffuse radio frequency domain signal; a two-dimensional inverse fast Fourier transform is performed on the normalized diffuse radio frequency domain signal to obtain the mapped diffuse reflection component image.

4. The facial recognition and comparison method for border inspection according to claim 1, characterized in that, A normalized face image is generated based on the mapped diffuse reflection component image and the energy-equalized specular reflection component image, including: Energy whitening weights are calculated based on the spatial frequency domain energy distribution of the specular reflection component image, wherein the energy whitening weights are inversely proportional to the square root of the spatial frequency domain energy distribution value; a two-dimensional fast Fourier transform is performed on the normalized specular reflection component image, and the energy whitening weights are used to perform frequency domain filtering, followed by a two-dimensional fast Fourier inverse transform to obtain an energy-equalized specular reflection component image; the energy-equalized specular reflection component image is multiplied by a preset surface detail preservation coefficient and linearly superimposed with the mapped diffuse reflection component image to obtain a normalized face image.

5. A method for facial recognition and comparison in border inspection according to claim 1, characterized in that, Calculate the biometric similarity between the normalized face image and the ID reference image, and generate the final verification result by combining the abnormal deviation of the overlapping feature values, including: A first feature vector is obtained by extracting features from the normalized face image using a pre-trained near-infrared feature extraction network, and a second feature vector is obtained by extracting features from the document reference image using a pre-trained visible light feature extraction network. The cosine similarity between the first and second feature vectors is calculated. The square of the logarithm of the ratio between the diffuse specular frequency domain overlapping feature value and a preset population statistical mean is calculated as an anomaly deviation penalty term. The cosine similarity is subtracted from the product of the anomaly deviation penalty term and a preset weight to obtain the final composite score. The final composite score is compared with a preset system threshold using a continuously differentiable indicator function, and a binarized final verification result is output.

6. A facial recognition and comparison system for border inspection, comprising executing a facial recognition and comparison method for border inspection as described in any one of claims 1-5, characterized in that, include: The data acquisition module acquires a single-frame polarized face image in the near-infrared acquisition environment of the border inspection, and uses the difference in polarization state to separate the image into a specular reflection component image that characterizes the surface texture and a diffuse reflection component image that characterizes the subcutaneous tissue characteristics. The feature extraction module calculates the spatial frequency domain energy distribution of the specular reflection component image and the diffuse reflection component image respectively, and determines the overlapping feature value of the diffuse specular frequency domain based on the energy equilibrium point of the spatial frequency domain energy distribution in the frequency domain. The normalization module inverts the current skin light scattering scale parameters based on the overlapping feature values, constructs a filter for normalizing tissue optical properties, and uses the filter to map the diffuse reflection component image from the current light scattering scale to a preset standard light scattering scale. The comparison and verification module generates a normalized face image based on the mapped diffuse reflection component image and the specular reflection component image after energy equalization. It calculates the biometric similarity between the normalized face image and the document reference image, and generates the final verification result by combining the abnormal deviation of the overlapping feature values.

Images