A method for determining whether a person is a minor based on facial recognition

Abstract

The application discloses a method for judging whether a person is a minor based on face recognition, and belongs to the technical field of identity recognition and computer vision. The method synchronously collects face videos through a visible light camera and a near-infrared camera, extracts static development structure feature vectors reflecting craniofacial bone and skin texture and dynamic autonomic nervous response feature vectors reflecting blood vessel elasticity and autonomic nervous response sensitivity under a sudden auditory stimulus; a cross-modal attention fusion network is used to perform bidirectional cross attention calculation on the two types of features, generate a synergistically constrained fusion representation and output a minor probability. The application utilizes development coupling, so that a counterfeit attack cannot simultaneously satisfy the matching of static and dynamic features, and has inherent anti-counterfeiting capability while improving the critical age recognition accuracy.

Description

Technical Field

[0001] This invention relates to the fields of identity recognition and computer vision technology, and in particular to a method for determining whether someone is a minor based on facial recognition. Background Technology

[0002] With the strengthening of legislation protecting minors in online services, accurately determining whether a user is a minor has become a core requirement in identity authentication. Facial image-based age recognition technology is widely used due to its advantages of being non-contact and easy to collect. Current mainstream solutions use deep convolutional neural networks to directly extract features from facial images and regress age values ​​or age ranges, then output a binary judgment of whether the user is a minor based on a preset age threshold, such as 18 years old. However, such solutions have the following prominent drawbacks.

[0003] First, age estimation relying solely on two-dimensional facial textures and geometry is extremely vulnerable to sophisticated deception. Attackers can easily bypass the system using high-resolution photos, pre-recorded videos, 3D masks, or deepfakes because the texture distribution of these forgeries in image space is highly similar to the age characteristics of a real face, making it impossible for the system to distinguish between a live person and a forgery from the image content itself. Existing countermeasures typically involve adding an independent liveness detection module, such as requiring the user to perform actions like blinking or turning their head. However, this combination is merely a logical AND operation between two independent judgments and does not establish an intrinsic correlation between age characteristics and physiological features of a live person. Attackers can still deceive both modules simultaneously by synthesizing forged videos that match the instructions in real time.

[0004] Second, facial image-based age regression models show a significant drop in accuracy for the critical age group of 16 to 18 years old. Individuals in this age range have facial structures approaching early adulthood, and simple geometric proportions and texture details are insufficient to provide stable distinguishing information. Furthermore, they are highly susceptible to the effects of makeup, hairstyle, shooting angle, and lighting intensity, leading to misidentification of adults as minors or minors approaching adulthood as adults. To address the critical age issue, some solutions have incorporated X-ray imaging of the human skeleton, such as wrist bone age assessment. However, such technologies involve ionizing radiation, making them unsuitable for large-scale civilian deployment in everyday non-medical settings. They also cannot be integrated with facial image acquisition, resulting in a poor user experience.

[0005] Third, in existing technologies, static facial structural features and dynamic physiological features related to the development of the human autonomic nervous system have never been collaboratively modeled within the same judgment framework. During human development from adolescence to early adulthood, static indicators such as facial bone proportions, cartilage morphology, and skin microstructure, as well as dynamic indicators such as the autonomic nervous system's response patterns and vascular elasticity, are all synchronously regulated by the endocrine axis, exhibiting an inherent coupling in their developmental sequence. Relying on only one type of feature is insufficient to simultaneously achieve high-precision minor identification and immunity to spoofing attacks. Therefore, there is an urgent need for a minor identification technology that can deeply integrate facial developmental structural features and dynamic autonomic nervous system response features, and that, mechanistically, allows these two heterogeneous features to mutually corroborate and constrain each other. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides a method for determining whether someone is a minor based on facial recognition, comprising the following steps:

[0007] Step 1: Use a synchronous triggering device to control the visible light camera and the near-infrared camera. With the help of a speaker, synchronously capture video of the face of the object being identified to obtain visible light facial video stream and near-infrared facial video stream. At the end of the 3rd second after the acquisition process starts, trigger the speaker to play white noise pulse as a sudden auditory stimulus signal and record the timestamp of the stimulus start.

[0008] Step 2: Extract static developmental structure feature vectors from the visible light facial video stream. The static developmental structure feature vectors include facial vertical proportion index, cheekbone-to-mandibular width ratio, nasal bone protrusion curvature, and texture anisotropy feature values.

[0009] Step 3: Calculate the relative amplitude of the diphtheria wave from the near-infrared facial video stream; starting from the timestamp of the stimulus start, extract the facial optical flow amplitude sequence from the visible light facial video stream and determine the maximum optical flow peak amplitude and recovery half-life, and extract the iPPG signal from the near-infrared facial video stream to calculate the heart rate acceleration peak. Combine the relative amplitude of the diphtheria wave, the maximum optical flow peak amplitude, the recovery half-life, and the heart rate acceleration peak into a dynamic autonomic nervous system response feature vector.

[0010] Step 4: Input the static developmental structural feature vector and the dynamic autonomic neural response feature vector into the developmental stage joint reasoning model; the developmental stage joint reasoning model includes a structural encoding branch, a physiological encoding branch, and a cross-modal attention fusion network. The cross-modal attention fusion network uses structural semantic representation as the query and physiological semantic representation as the key and value to perform cross-attention calculation to obtain physiologically enhanced structural representation, and uses physiological semantic representation as the query and structural semantic representation as the key and value to perform cross-attention calculation to obtain structurally enhanced physiological representation. The physiologically enhanced structural representation and the structurally enhanced physiological representation are concatenated to output the probability value of minor.

[0011] Step 5: When the probability value of being a minor is greater than a preset threshold, the identified object is determined to be a minor.

[0012] Preferably, step 2, which involves extracting the static developmental structure feature vector from the visible light facial video stream, includes:

[0013] Frames with an absolute value of facial deflection angle of less than 5 degrees and an eye aspect ratio of greater than 0.8 are selected from the visible light facial video stream as standard frontal face frames.

[0014] Facial key points are located on the standard frontal face frame, and a three-dimensional deformable model is fitted using the located facial key points to reconstruct a normalized three-dimensional face mesh.

[0015] The vertical distance from the glabella to the chin and the Euclidean distance between the center points of the left and right pupils are extracted from the three-dimensional face mesh. The ratio of the vertical distance from the glabella to the chin to the Euclidean distance between the center points of the left and right pupils is calculated as the facial vertical proportion index.

[0016] Extract the width between the left and right cheekbone points and the width between the left and right mandibular angle points, and calculate the ratio of the width between the left and right cheekbone points to the width between the left and right mandibular angle points as the cheekbone-mandibular width ratio;

[0017] The sagittal contour curve between the nasal root point and the nasal tip point in the three-dimensional face mesh is fitted with a circle to obtain the radius of the best fitted circle. The reciprocal of the radius of the best fitted circle is calculated as the curvature of the nasal bone protrusion.

[0018] The gray-level co-occurrence matrix is ​​calculated from the regions of interest of the skin on both cheeks of the standard frontal face frame. Based on the length of the major axis and the length of the minor axis of the moment of inertia ellipse of the gray-level co-occurrence matrix, the texture anisotropic eigenvalue is calculated. The texture anisotropic eigenvalue is 1 minus the ratio of the minor axis length to the major axis length.

[0019] The facial vertical proportion index, the cheekbone-to-mandibular width ratio, the nasal bone protrusion curvature, and the texture anisotropy feature value are concatenated to form the static developmental structure feature vector.

[0020] Preferably, the formula for calculating the facial vertical proportion index is:

[0021]

[0022] in, This refers to the facial vertical proportion index. The vertical distance from the glabella point to the chin point is [missing information]. The distance between the center points of the left and right pupils is the Euclidean distance.

[0023] The formula for calculating the cheekbone-mandibular width ratio is:

[0024]

[0025] in, The ratio of the cheekbone to the width of the mandible is given. The width between the left and right cheekbone points. The width between the left and right mandibular angle points;

[0026] The formula for calculating the curvature of the nasal bone protrusion is:

[0027]

[0028] in, The curvature of the nasal bone protrusion. The radius of the optimal fitting circle is obtained by performing a circle fit on the sagittal contour curve between the nasal root point and the nasal tip point.

[0029] The formula for calculating the anisotropic texture feature value is as follows:

[0030]

[0031] in, The anisotropic feature value of the texture. and These are the major axis length and minor axis length of the ellipse of inertia obtained from the gray-level co-occurrence matrix, respectively. .

[0032] Preferably, the process of calculating the relative amplitude of the diphtheria wave from the near-infrared facial video stream in step 3 includes:

[0033] In the near-infrared facial video stream, regions of interest are defined on both cheeks, and the time series of average pixel intensity within the regions of interest on both cheeks is extracted.

[0034] Bandpass filtering is applied to the average pixel intensity time series to obtain the volumetric pulse wave signal;

[0035] The maximum amplitude during the systolic phase of the volumetric pulse wave signal is detected as the peak amplitude during the systolic phase, and the first local maximum amplitude that appears after the diphthoplasty notch of the volumetric pulse wave signal is detected as the peak amplitude after the diphthoplasty notch.

[0036] The relative amplitude of the diphthalamic wave is calculated using the following formula:

[0037]

[0038] in, The relative amplitude of the diphthical wave. The peak amplitude during the contraction period, The peak amplitude after the diphthous wave notch is given.

[0039] Preferably, step 3, which involves extracting the facial optical flow amplitude sequence from the visible light facial video stream and determining the maximum optical flow peak amplitude and recovery half-life, includes:

[0040] Starting from the timestamp of the stimulus, dense optical flow is calculated for the visible light facial video stream within a time window of 0.5 to 2.0 seconds after the stimulus.

[0041] Extract the facial region optical flow amplitude sequence from the dense optical flow;

[0042] The maximum peak amplitude of optical flow is determined in the optical flow amplitude sequence of the facial region;

[0043] The time taken for the optical flow amplitude sequence of the facial region to decay from the maximum optical flow peak amplitude to half of its amplitude is calculated as the recovery half-life.

[0044] Preferably, step 3, which involves extracting the iPPG signal from the near-infrared facial video stream and calculating the peak heart rate acceleration, includes:

[0045] The iPPG signal is extracted from the forehead region of interest in the near-infrared facial video stream;

[0046] The average heart rate baseline was calculated based on the iPPG signal within 2 seconds prior to the timestamp of the stimulation start.

[0047] Within 1 second after the timestamp of the stimulus start, the maximum instantaneous heart rate is calculated based on the iPPG signal;

[0048] The difference between the maximum instantaneous heart rate and the average heart rate baseline is taken as the peak heart rate acceleration, and the formula for calculating the peak heart rate acceleration is as follows:

[0049]

[0050] in, The peak heart rate acceleration is mentioned. The average heart rate baseline, The maximum instantaneous heart rate is denoted as .

[0051] Preferably, the structural encoding branch of the developmental stage joint reasoning model in step 4 is composed of a first fully connected network, which maps the static developmental structural feature vector to a structural semantic representation; the physiological encoding branch is composed of a second fully connected network, which maps the dynamic autonomic neural response feature vector to a physiological semantic representation; the cross-modal attention fusion network performs the following operations:

[0052] The structural semantic representation is multiplied by the first learnable weight matrix to obtain the structural query vector. The physiological semantic representation is multiplied by the second and third learnable weight matrices to obtain the physiological key vector and physiological value vector, respectively. The scaled dot product attention of the structural query vector and the physiological key vector is calculated to obtain the physiologically enhanced structural representation.

[0053] The physiological semantic representation is multiplied by the fourth learnable weight matrix to obtain the physiological query vector. The structural semantic representation is multiplied by the fifth and sixth learnable weight matrices to obtain the structural key vector and structural value vector, respectively. The scaled dot product attention of the physiological query vector and the structural key vector is calculated to obtain the structure-enhanced physiological representation.

[0054] After concatenating the physiologically enhanced structural representation and the structurally enhanced physiological representation into vectors, the input is given to the output layer consisting of a fully connected layer and a sigmoid activation function, and the output is the underage probability value.

[0055] Preferably, the scaling dot product attention formula used to calculate the structural representation of the physiological enhancement is:

[0056]

[0057] in, The structure is the query vector. Let be the physiological bond vector. The physiological value vector, Let be the dimension of the physiological bond vector;

[0058] The scaling dot product attention formula used to calculate the physiological representation of the enhanced structure is as follows:

[0059]

[0060] in, Let be the physiological query vector. The structure's key vector, For the structure value vector, Let be the dimension of the structure key vector.

[0061] Preferably, the sudden auditory stimulus signal in step 1 is an 85 dB white noise pulse with a duration of 100 milliseconds; the synchronization triggering device provides a synchronization triggering signal to the visible light camera and the near-infrared camera to synchronize the frames of the visible light facial video stream and the near-infrared facial video stream, and the timestamp of the speaker playing the sudden auditory stimulus signal is aligned with the timestamp of the video frame.

[0062] Preferably, the wavelength of the near-infrared camera is 850 nanometers, and the visible light camera is an RGB camera; the acquisition frame rate of both the visible light camera and the near-infrared camera is not less than 30 frames / second; the system clock determines the end of the 3rd second after the acquisition process starts.

[0063] The beneficial effects of this invention are:

[0064] 1. This invention extracts static developmental structural feature vectors and dynamic autonomic neural response feature vectors by simultaneously acquiring visible light and near-infrared facial videos. It also utilizes a cross-modal attention fusion network to establish a bidirectional cross-attention constraint between structural semantic representation and physiological semantic representation. This enables forged media to be effectively identified because they cannot simultaneously reproduce the autonomic neural responses coupled with the static developmental level. This overcomes the vulnerability of simply superimposing traditional age estimation and liveness detection, and achieves inherent immunity to deception attacks such as photos, videos, and 3D masks.

[0065] 2. This invention integrates static developmental structural features such as facial vertical ratio index, cheekbone-to-mandibular width ratio, nasal bone protrusion curvature, and texture anisotropy with dynamic autonomic nervous system response features such as relative amplitude of dichroic waves, peak amplitude of micro-expression optical flow induced by sudden stimulation, recovery half-life, and peak heart rate acceleration. Through the combination of multidimensional developmental indicators, it significantly enhances the recognition and differentiation ability of individuals aged 16 to 18, improving the accuracy of recognition for minors and the rejection rate for adults.

[0066] 3. This invention induces autonomic nerve responses through sudden auditory stimulation signals and extracts pulse wave elasticity and heart rate variability using near-infrared imaging. The entire process requires the identified subject to perform prescribed actions, and the data collection process is non-invasive, covert, and natural. It ensures both high security and anti-counterfeiting performance and a good user experience, making it suitable for large-scale online identity verification scenarios. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0069] Figure 2 This is a flowchart of step 2 of the method of the present invention;

[0070] Figure 3 This is a flowchart illustrating the steps involved in extracting the facial optical flow amplitude sequence and determining the maximum optical flow peak amplitude and recovery half-life in the method of this invention. Detailed Implementation

[0071] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0072] Please see Figures 1-3 This invention provides a method for determining whether someone is a minor based on facial recognition. The core concept is that it does not use static facial image features as the sole age criterion, nor does it use liveness detection as an independent additional module. Instead, it incorporates static developmental structural features reflecting craniofacial bone development and skin microtexture with dynamic autonomic nervous system response features reflecting autonomic nervous system development and vascular elasticity into the same judgment framework. Through a cross-modal attention fusion network, it achieves deep interaction and bidirectional constraints between the two types of heterogeneous features, thereby obtaining a collaborative judgment effect that far exceeds that of simple feature superposition.

[0073] Specifically, during development, an individual's static structural features, such as facial bone proportions, nasal bone morphology, and skin texture anisotropy, are intrinsically coupled with dynamic physiological features, such as the micro-expression responses of the autonomic nervous system to sudden environmental stimuli and the pulse waveforms reflected by cardiovascular elasticity, all regulated by the endocrine axis. A genuine minor simultaneously exhibits static structural parameters and dynamic response patterns matching their age stage; however, no forgery medium, whether high-resolution photographs, replay videos, 3D masks, or deepfake content, can simultaneously and accurately reproduce the coupling relationship between these two heterogeneous modalities of a genuine minor. This invention utilizes this principle, fusing static developmental structural feature vectors and dynamic autonomic nervous system response feature vectors through cross-modal attention. This ensures that the model only outputs a high-confidence probability of being a minor when the two types of features mutually corroborate and align with each other. Thus, without requiring the user to perform specific actions, it achieves inherent immunity to forgery attacks and high-precision differentiation of the 16-18 age threshold.

[0074] The implementation process of this invention is described in detail below with a specific embodiment. The application scenario of this embodiment is online identity verification for minors by network service providers.

[0075] Example

[0076] In this embodiment, the hardware deployment includes a 1920×1080 visible light RGB camera, a 640×480 near-infrared camera with a wavelength of 850 nanometers, a synchronization triggering device, a speaker for playing auditory stimulus signals, and a computing terminal for running algorithms. The visible light camera and the near-infrared camera are mounted side-by-side with a horizontal optical center spacing of 40 mm, and the speaker is located 0.8 meters directly in front of the object being identified. The synchronization triggering device simultaneously sends frame synchronization signals to both cameras to ensure that the timestamps of each frame of the visible light facial video stream and the near-infrared facial video stream are aligned. The synchronization triggering device is also connected to the speaker and can trigger audio playback at a specified time.

[0077] Step 1: Using a synchronous triggering device, control the visible light camera and near-infrared camera to synchronously acquire video of the subject's face with the assistance of a speaker, obtaining visible light and near-infrared facial video streams. At the end of the 3rd second after the acquisition process begins, trigger the speaker to play an 85 dB white noise pulse. The duration of the white noise pulse is 100 milliseconds, serving as a sudden auditory stimulus signal, and record the timestamp of the stimulus's start. The acquisition frame rate of both the visible light and near-infrared cameras is set to 30 frames per second.

[0078] The subject sits facing both cameras, keeping their face centered in the camera's field of view. At second 0, the synchronization trigger starts both cameras recording simultaneously. At the end of second 3, i.e., at the completion of the 30th frame of second 3, the synchronization trigger outputs a trigger signal to the speaker, which immediately plays a 100-millisecond 85-decibel white noise pulse, and simultaneously records the timestamp of the stimulus's initiation. Second.

[0079] Step 2: Extract static developmental structure feature vectors from the visible light facial video stream. These vectors include the facial vertical proportion index, the ratio of cheekbone to mandible width, the nasal bone protrusion curvature, and texture anisotropy feature values. The specific extraction process is as follows:

[0080] First, frames with an absolute facial deflection angle of less than 5 degrees and an eye aspect ratio greater than 0.8 are selected from the visible light facial video stream as standard frontal face frames. The facial deflection angle refers to the angle of deflection of the facial axis of symmetry relative to the camera's optical axis, calculated using the plane normal vector formed by the detected nose tip and the inner corners of the eyes. The eye aspect ratio is defined as the ratio of the vertical distance between the upper and lower eyelids to the horizontal distance between the inner and outer corners of the eyes, reflecting the degree of eye opening. In this embodiment, 90 visible light images are acquired within the first 3 seconds of acquisition. After screening, 67 standard frontal face frames that meet the criteria are obtained, and the first frame of these is selected as the input for subsequent processing.

[0081] Next, a multi-task convolutional neural network was used to locate facial landmarks in the standard frontal face frame, resulting in the localization of 68 landmarks. Using these 68 landmarks, an iterative nearest-point algorithm was employed to fit the Basel face model, a 3D deformable model, to reconstruct a normalized 3D face mesh. During the fitting process, shape and expression parameters were optimized to minimize the weighted error between the projection of the 3D mesh onto the image plane and the positions of the 2D landmarks. The normalized 3D face mesh eliminated differences in head pose and size.

[0082] Then, the following features are extracted from the reconstructed 3D face mesh:

[0083] Facial vertical proportion index: Extract the 3D coordinates of the glabella (keypoint 22) and chin (keypoint 9) from the 3D mesh, and calculate the vertical distance between the glabella and chin. mm; extract the center of the left pupil (the center of the ellipse fitted by keypoints 37 to 42) and the center of the right pupil (the center of the ellipse fitted by keypoints 43 to 48), and calculate the Euclidean distance between the centers of the left and right pupils. mm. Facial vertical proportion index. .

[0084] Cheekbone-to-mandibular width ratio: Extract the left cheekbone point (keypoint 2) and the right cheekbone point (keypoint 16), and calculate the width between the left and right cheekbone points. mm; Extract the left mandibular angle point (key point 7) and the right mandibular angle point (key point 11), and calculate the width between the left and right mandibular angle points. mm. Cheekbone-to-jaw width ratio .

[0085] Nasal bone protrusion curvature: A sagittal contour curve from the nasal root point (keypoint 28) to the nasal tip point (keypoint 31) is extracted from the 3D face mesh, i.e., the contour line obtained by cutting the mesh with the facial symmetry plane. Least-squares circle fitting is performed on 50 sampling points on this contour curve to obtain the radius of the best-fit circle. mm. Nasal bone protrusion curvature .

[0086] Texture anisotropic eigenvalues: On a standard frontal face frame, regions of interest (ROIs) for the skin of both cheeks are delineated based on keypoint localization results. Specifically, the inner rectangular region of the left cheek is bounded by keypoints 2, 3, 4, and 5, and the inner rectangular region of the right cheek is bounded by keypoints 14, 15, 16, and 17. The pixel grayscale values ​​of both regions are converted to grayscale images, and grayscale co-occurrence matrices are calculated for each region. The distance parameter is set to 1 pixel, and the direction is taken as the average of four directions: 0°, 45°, 90°, and 135°. The moment of inertia ellipse is calculated for the grayscale co-occurrence matrix to obtain the length of the major axis. and minor axis length Texture anisotropic eigenvalues .

[0087] Finally, the above , , and By concatenating the features, a 4-dimensional static developmental structure feature vector is obtained. .

[0088] Step 3: Calculate the relative amplitude of the diphthoplasty from the near-infrared facial video stream; using the timestamp of the stimulus start. Starting from a second, the facial optical flow amplitude sequence is extracted from the visible light facial video stream, and the peak amplitude of the maximum optical flow and the recovery half-life are determined. The iPPG signal is extracted from the near-infrared facial video stream to calculate the peak heart rate acceleration. The relative amplitude of the dicrotic wave, the peak amplitude of the maximum optical flow, the recovery half-life, and the peak heart rate acceleration are combined to form a dynamic autonomic nervous system response feature vector. The specific process is as follows:

[0089] (3.1) Calculation of the relative amplitude of the diphthical wave:

[0090] In each frame of the near-infrared facial video stream, using the facial key point locations determined in step 2, a region of interest (ROI) is delineated on both cheeks. This region is a 40×40 pixel square area in the middle of the left and right cheeks, avoiding strong light reflection. The average grayscale value of all pixels within this region is extracted to obtain the average pixel intensity time series. Bandpass filtering is applied to the 90 frames of signal from second 0 to second 3, with a passband frequency of 0.75 Hz to 4.0 Hz, corresponding to a heart rate range of 45 bpm to 240 bpm, to obtain the volumetric pulse wave signal. A peak detection algorithm is used to identify the maximum amplitude of systole in each cardiac cycle, and the peak amplitude of systole in the last complete cycle before the 3rd second is taken. (Pixel intensity unit). The first local maximum amplitude value appearing after detecting the diphtheria wave notch, i.e., the peak amplitude after the diphtheria wave notch. Relative amplitude of diphtheria wave .

[0091] (3.2) Calculation of the peak amplitude of maximum optical flow and recovery half-life:

[0092] The timestamp of the stimulus Starting at a certain second, dense optical flow was calculated over a 45-frame visible light video stream (3.5 to 5.0 seconds) within a time window from 0.5 to 2.0 seconds after the stimulus. The Farnebäck optical flow algorithm was used, with a pyramid layer of 3 and a window size of 15×15 pixels. The optical flow amplitude sequence for the facial region was extracted, defined as an elliptical region surrounded by keypoints. Within this time window, the first significant peak in the facial region optical flow amplitude sequence was identified as the micro-expression response induced by the stimulus, and this peak was determined as the maximum optical flow peak amplitude. Pixels / frame. The optical flow amplitude then gradually decays. The time it takes for the amplitude to decrease from the peak to half its current value (0.17 pixels / frame) is calculated to obtain the recovery half-life. Second.

[0093] (3.3) Calculation of peak heart rate acceleration:

[0094] iPPG signals were extracted from the region of interest (a 25×25 pixel square area above the brow, avoiding hair) on the forehead in the near-infrared facial video stream. Within 2 seconds prior to the stimulation start timestamp (i.e., from second 1.0 to second 3.0), Fourier transforms were performed on the iPPG signals to extract the heart rate value corresponding to the dominant frequency as the average heart rate baseline. bpm. Within one second after the stimulation begins (i.e., from second 3.0 to second 4.0), the instantaneous heart rate time series is calculated. The instantaneous heart rate is obtained by calculating the reciprocal of the time interval between two consecutive systolic peaks. The maximum instantaneous heart rate within this time period is taken as the maximum instantaneous heart rate. bpm. Peak heart rate. .

[0095] Finally, , pixels / frame Second, bpm combination forms a 4-dimensional dynamic autonomic neural response feature vector. .

[0096] Step 4: Convert the static developmental structure feature vector and dynamic autonomic nervous system response feature vector A joint reasoning model for input developmental stages. This model includes structural encoding branches, physiological encoding branches, and a cross-modal attention fusion network.

[0097] The structural coding branch is the first fully connected network, containing 4 neurons in the input layer, 16 neurons in the hidden layer, and 8 neurons in the output layer, using the ReLU activation function. Mapped to an 8-dimensional structural semantic representation .

[0098] The physiological coding branch is a second fully connected network, containing 4 neurons in the input layer, 16 neurons in the hidden layer, and 8 neurons in the output layer, using the ReLU activation function. Mapped to 8-dimensional physiological semantic representation .

[0099] Cross-modal attention fusion networks perform the following operations:

[0100] Structural semantic representation With the first learnable weight matrix Multiplication yields the structure query vector The dimension is 8. Physiological semantic representation Each with the second learnable weight matrix and the third learnable weight matrix Multiplication yields the physiological bond vector and physiological value vector All dimensions are 8. Scaled dot product attention is calculated to obtain physiologically enhanced structural representations. :

[0101]

[0102] Physiological semantic representation With the fourth learnable weight matrix Multiplication yields the physiological query vector Structural semantic representation Each with the fifth learnable weight matrix and the sixth learnable weight matrix Multiplication yields the structural bond vector and structure value vector Calculate the scaled dot product attention to obtain the physiological representation of structural enhancement. :

[0103]

[0104] Will and Vector concatenation yields a 16-dimensional fused representation, which is then input to an output layer consisting of a fully connected layer (16 input neurons, 1 output neuron) and a sigmoid activation function to obtain the probability value of a minor. In this embodiment, the model outputs a probability value of 0.846.

[0105] All the learnable weight matrices mentioned above were determined through pre-training. During training, a large number of labeled samples of minors and adults with real ages were used. The loss function was binary cross-entropy, and the optimization objective was to enable the model to make judgments based on both static and dynamic features.

[0106] Step 5: Compare the probability value of being a minor with the preset threshold of 0.5. In this embodiment, the output probability value of 0.846 is greater than 0.5, therefore the identified object is determined to be a minor.

[0107] To demonstrate the beneficial effects of the method in this embodiment, a comparative experiment was conducted. The comparative data consisted of a test set containing 300 real individuals (150 of whom were minors aged 16 to 18 and 150 of whom were adults aged 19 to 25) and 300 fake attack samples (including high-definition photo re-enactments, replays of recorded facial videos, and 3D silicone masks).

[0108] Comparative Example 1 uses a pure facial image age estimation scheme: the age value is directly regressed using an age estimation network based on Inception-ResNet-v1. If the age value is less than 18 years old, the person is judged to be a minor, without additional liveness detection.

[0109] Comparative Example 2 uses a scheme that combines age estimation with independent liveness detection: the age estimation network is the same as that in Comparative Example 1, and the liveness detection uses interactive liveness detection based on action instructions, requiring the identified object to complete blinking and turning its head to the left as prompted. Only when the age estimation is that of a minor and the liveness detection passes is the object determined to be a minor.

[0110] The method of this embodiment (static developmental structural features + dynamic autonomic nervous response features + cross-modal attention fusion) was compared with the two comparative examples mentioned above, and the results are shown in Table 1.

[0111] Table 1 Performance Comparison of Different Methods

[0112] Correct identification rate of borderline minors (16-18 years old) 78.0% 79.3% 93.3% Correct rejection rate for adults (19-25 years old) 82.7% 86.0% 92.7% Acceptance rate of photo reposting attacks 62.0% 3.3% 0.7% Video replay attack acceptance rate 71.3% 8.0% 1.3% 3D mask attack acceptance rate 54.7% 12.7% 2.0%

[0113] As shown in Table 1, Comparative Example 1, lacking liveness detection, struggles to resist various forgery attacks. Comparative Example 2, by adding interactive liveness detection, significantly reduces the acceptance rate of photo and video attacks, but still exhibits a high acceptance rate against 3D masks, and the improvement in recognition accuracy within the critical age range is limited. The method of this invention significantly improves both the correct recognition rate within the critical age range and the rejection rate among adults, while simultaneously reducing the acceptance rate against all three types of forgery attacks to a very low level. This is because forged samples cannot simultaneously provide autonomic neural response features coupled with static developmental structural characteristics. The cross-modal attention fusion network automatically reduces the probability value of minors when it detects a mismatch between structural and physiological representations, thereby achieving robust defense against forgery attacks and accurate identification of critical age groups.

[0114] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0115] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for determining whether someone is a minor based on facial recognition, characterized in that, Includes the following steps: Step 1: Use a synchronous triggering device to control the visible light camera and the near-infrared camera, and with the help of a speaker, synchronously capture video of the face of the object being identified to obtain a visible light facial video stream and a near-infrared facial video stream. At the end of the 3rd second after the acquisition process starts, the speaker is triggered to play a white noise pulse as a sudden auditory stimulus signal, and the timestamp of the stimulus start is recorded. Step 2: Extract static developmental structure feature vectors from the visible light facial video stream. The static developmental structure feature vectors include facial vertical proportion index, cheekbone-to-mandibular width ratio, nasal bone protrusion curvature, and texture anisotropy feature values. Step 3: Calculate the relative amplitude of the diphtheria wave from the near-infrared facial video stream; starting from the timestamp of the stimulus start, extract the facial optical flow amplitude sequence from the visible light facial video stream and determine the maximum optical flow peak amplitude and recovery half-life, and extract the iPPG signal from the near-infrared facial video stream to calculate the heart rate acceleration peak. Combine the relative amplitude of the diphtheria wave, the maximum optical flow peak amplitude, the recovery half-life, and the heart rate acceleration peak into a dynamic autonomic nervous system response feature vector. Step 4: Input the static developmental structural feature vector and the dynamic autonomic neural response feature vector into the developmental stage joint reasoning model; the developmental stage joint reasoning model includes a structural encoding branch, a physiological encoding branch, and a cross-modal attention fusion network. The cross-modal attention fusion network uses structural semantic representation as the query and physiological semantic representation as the key and value to perform cross-attention calculation to obtain physiologically enhanced structural representation, and uses physiological semantic representation as the query and structural semantic representation as the key and value to perform cross-attention calculation to obtain structurally enhanced physiological representation. The physiologically enhanced structural representation and the structurally enhanced physiological representation are concatenated to output the probability value of minor. Step 5: When the probability value of being a minor is greater than a preset threshold, the identified object is determined to be a minor.

2. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, Step 2, which involves extracting the static developmental structure feature vector from the visible light facial video stream, includes: Frames with an absolute value of facial deflection angle of less than 5 degrees and an eye aspect ratio of greater than 0.8 are selected from the visible light facial video stream as standard frontal face frames. Facial key points are located on the standard frontal face frame, and a three-dimensional deformable model is fitted using the located facial key points to reconstruct a normalized three-dimensional face mesh. The vertical distance from the glabella to the chin and the Euclidean distance between the center points of the left and right pupils are extracted from the three-dimensional face mesh. The ratio of the vertical distance from the glabella to the chin to the Euclidean distance between the center points of the left and right pupils is calculated as the facial vertical proportion index. Extract the width between the left and right cheekbone points and the width between the left and right mandibular angle points, and calculate the ratio of the width between the left and right cheekbone points to the width between the left and right mandibular angle points as the cheekbone-mandibular width ratio; The sagittal contour curve between the nasal root point and the nasal tip point in the three-dimensional face mesh is fitted with a circle to obtain the radius of the best fitted circle. The reciprocal of the radius of the best fitted circle is calculated as the curvature of the nasal bone protrusion. The gray-level co-occurrence matrix is ​​calculated from the regions of interest of the skin on both cheeks of the standard frontal face frame. Based on the length of the major axis and the length of the minor axis of the moment of inertia ellipse of the gray-level co-occurrence matrix, the texture anisotropic eigenvalue is calculated. The texture anisotropic eigenvalue is 1 minus the ratio of the minor axis length to the major axis length. The facial vertical proportion index, the cheekbone-to-mandibular width ratio, the nasal bone protrusion curvature, and the texture anisotropy feature value are concatenated to form the static developmental structure feature vector.

3. The method for determining whether someone is a minor based on facial recognition according to claim 2, characterized in that, The formula for calculating the facial vertical proportion index is as follows: in, This refers to the facial vertical proportion index. The vertical distance from the glabella point to the chin point is [missing information]. The distance between the center points of the left and right pupils is the Euclidean distance. The formula for calculating the cheekbone-to-mandibular width ratio is: in, The ratio of the cheekbone to the width of the mandible is given. The width between the left and right cheekbone points. The width between the left and right mandibular angle points; The formula for calculating the curvature of the nasal bone protrusion is: in, The curvature of the nasal bone protrusion. The radius of the optimal fitting circle is obtained by performing a circle fit on the sagittal contour curve between the nasal root point and the nasal tip point. The formula for calculating the anisotropic texture feature value is as follows: in, The anisotropic feature value of the texture. and These are the major axis length and minor axis length of the ellipse of inertia obtained from the gray-level co-occurrence matrix, respectively. .

4. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, Step 3, which involves calculating the relative amplitude of the diphtheria wave from the near-infrared facial video stream, includes: In the near-infrared facial video stream, regions of interest are defined on both cheeks, and the time series of average pixel intensity within the regions of interest on both cheeks is extracted. Bandpass filtering is applied to the average pixel intensity time series to obtain the volumetric pulse wave signal; The maximum amplitude during the systolic phase of the volumetric pulse wave signal is detected as the peak amplitude during the systolic phase, and the first local maximum amplitude that appears after the diphthoplasty notch of the volumetric pulse wave signal is detected as the peak amplitude after the diphthoplasty notch. The relative amplitude of the diphthalamic wave is calculated using the following formula: in, The relative amplitude of the diphthical wave. The peak amplitude during the contraction period, The peak amplitude after the diphthous wave notch is given.

5. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, Step 3, which involves extracting the facial optical flow amplitude sequence from the visible light facial video stream and determining the maximum optical flow peak amplitude and recovery half-life, includes: Starting from the timestamp of the stimulus, dense optical flow is calculated for the visible light facial video stream within a time window of 0.5 to 2.0 seconds after the stimulus. Extract the facial region optical flow amplitude sequence from the dense optical flow; The maximum peak amplitude of optical flow is determined in the optical flow amplitude sequence of the facial region; The time taken for the optical flow amplitude sequence of the facial region to decay from the maximum optical flow peak amplitude to half of its amplitude is calculated as the recovery half-life.

6. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, Step 3, which involves extracting the iPPG signal from the near-infrared facial video stream and calculating the peak heart rate acceleration, includes: The iPPG signal is extracted from the forehead region of interest in the near-infrared facial video stream; The average heart rate baseline was calculated based on the iPPG signal within 2 seconds prior to the timestamp of the stimulation start. Within 1 second after the timestamp of the stimulus start, the maximum instantaneous heart rate is calculated based on the iPPG signal; The difference between the maximum instantaneous heart rate and the average heart rate baseline is taken as the peak heart rate acceleration, and the formula for calculating the peak heart rate acceleration is as follows: in, The peak heart rate acceleration is mentioned. The average heart rate baseline, The maximum instantaneous heart rate is denoted as .

7. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, The structural encoding branch of the developmental stage joint reasoning model in step 4 is composed of a first fully connected network, which maps the static developmental structural feature vector to a structural semantic representation; the physiological encoding branch is composed of a second fully connected network, which maps the dynamic autonomic neural response feature vector to a physiological semantic representation; the cross-modal attention fusion network performs the following operations: The structural semantic representation is multiplied by the first learnable weight matrix to obtain the structural query vector. The physiological semantic representation is multiplied by the second and third learnable weight matrices to obtain the physiological key vector and physiological value vector, respectively. The scaled dot product attention of the structural query vector and the physiological key vector is calculated to obtain the physiologically enhanced structural representation. The physiological semantic representation is multiplied by the fourth learnable weight matrix to obtain the physiological query vector. The structural semantic representation is multiplied by the fifth and sixth learnable weight matrices to obtain the structural key vector and structural value vector, respectively. The scaled dot product attention of the physiological query vector and the structural key vector is calculated to obtain the structure-enhanced physiological representation. After concatenating the physiologically enhanced structural representation and the structurally enhanced physiological representation into vectors, the input is given to the output layer consisting of a fully connected layer and a sigmoid activation function, and the output is the underage probability value.

8. The method for determining whether someone is a minor based on facial recognition according to claim 7, characterized in that, The scaling dot product attention formula used to calculate the structural representation of the physiological enhancement is as follows: in, The structure is the query vector. Let be the physiological bond vector. The physiological value vector, Let be the dimension of the physiological bond vector; The scaling dot product attention formula used to calculate the physiological representation of the enhanced structure is as follows: in, Let be the physiological query vector. The structure's key vector, For the structure value vector, Let be the dimension of the structure's key vector.

9. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, The sudden auditory stimulus signal mentioned in step 1 is an 85 dB white noise pulse with a duration of 100 milliseconds; the synchronization triggering device provides a synchronization triggering signal to the visible light camera and the near-infrared camera to synchronize the frames of the visible light facial video stream and the near-infrared facial video stream, and the timestamp of the speaker playing the sudden auditory stimulus signal is aligned with the timestamp of the video frame.

10. The method for determining whether someone is a minor based on facial recognition according to claim 1, characterized in that, The near-infrared camera has a wavelength of 850 nanometers, and the visible light camera is an RGB camera; the acquisition frame rate of both the visible light camera and the near-infrared camera is not less than 30 frames / second; the system clock determines the end of the 3rd second after the acquisition process starts.

Images