Client identity real-time checking method based on image recognition

Abstract

The invention relates to the technical field of image recognition and biological feature authentication, and particularly discloses a client identity real-time checking method based on image recognition. The method comprises the following steps: synchronously acquiring a certificate image, a real-time video stream and environmental parameters, and preprocessing to generate an environmental parameter matrix and a biological pulse sequence; certificate static texture features and video dynamic micro-expression features are separated and extracted through a double-flow network, and feature weights are dynamically adjusted and dimensionality reduction is carried out based on environmental parameters; the historical weight and the illumination factor are fused to generate a dynamic fusion coefficient, and feedback optimization feature extraction is carried out; decoding the iris tremor frequency to generate a living body confidence coefficient, combining the dynamic weight fusion features to obtain an initial similarity, and carrying out multiplication fusion on the initial similarity and the living body confidence coefficient; a micro-expression periodic jitter mode is detected, and risk factors are output after fraud punishment is applied; and triggering model re-calibration based on the risk factor. According to the method, the verification precision and the anti-counterfeiting capability under the scenes of complex illumination, head deflection and screen attack can be remarkably improved, and high-robustness identity authentication is realized.

Description

Technical Field

[0001] This invention relates to the field of image recognition and biometric authentication technology, and in particular to a real-time customer identity verification method based on image recognition. Background Technology

[0002] Current customer identity verification methods suffer from three main technical flaws:

[0003] (1) Insufficient environmental adaptability: Existing ID photo and face comparison methods rely on static similarity thresholds. When the ambient light intensity changes by more than ±500 lux, the recognition error rate is high. Traditional lighting compensation algorithms (such as the Retinex model) cannot adapt to millisecond-level lighting change scenarios.

[0004] (2) Lack of geometric distortion correction: When the head yaw angle exceeds 45°, recognition methods based on single-frame images (such as DeepID) suffer from low recognition rates due to the loss of feature points. Existing 3D reconstruction technologies (such as 3D-PAM) require multi-frame calculations with a latency of over 800ms, which cannot meet the requirements for real-time verification.

[0005] (3) Weak anti-spoofing mechanisms: For screen replay attacks, existing liveness detection methods (such as blink detection) have a low interception rate for replay videos with a refresh rate of 60Hz or higher on high-definition screens. Temporal micro-expression analysis (such as LBP-TOP) has a high false detection rate for replay attacks due to the lack of periodic jitter modeling. Summary of the Invention

[0006] This invention provides a real-time customer identity verification method based on image recognition, which solves the problem of how to achieve high-precision identity verification by real-time fusion of iris biometrics and dynamic feature weights in scenarios with complex lighting changes, head posture deviation, and screen replay attacks.

[0007] To address the aforementioned technical problems, this invention provides a real-time customer identity verification method based on image recognition, comprising:

[0008] The system simultaneously acquires document images, real-time video streams, and ambient light / angle / infrared signals, and generates an environmental parameter matrix and pulse timing sequence after filtering and normalization. The process of acquiring document images includes: acquiring images of the ID document provided by the user through a high-resolution contact image sensor. The sensor uses a sensing element with an optical resolution of not less than 1200 dpi and is equipped with a ring LED light source array to eliminate surface reflection interference.

[0009] A dual-stream convolutional network is used to separate and extract static texture features of documents and dynamic micro-expression features of videos. The receptive field weights are adjusted based on the environmental parameter matrix, and feature decoupling and dimensionality reduction are performed.

[0010] The illumination impact factor is calculated based on the environmental parameter matrix, and the dynamic fusion weight coefficient is generated by integrating historical weight data through a time series model and fed back to the feature weight adjustment module.

[0011] Decode the iris tremor frequency to generate a liveness confidence coefficient, use dynamic weight coefficients to weight and fuse static and dynamic features to obtain an initial similarity, and then multiply and fuse it with the liveness confidence.

[0012] The expression fused with the liveness confidence multiplication includes:

[0013]

[0014] Among them, S enhanced Similarity after live-enhanced imaging; S init β is the initial similarity; β is the final liveness confidence coefficient; exp(·) is the exponential function;

[0015] Δt sync σ represents the timestamp deviation; σ is the time tolerance threshold.

[0016] Furthermore, the expression for the environmental compensation operator is:

[0017]

[0018] Among them, S final To ultimately enhance similarity; I light For ambient light intensity; |θ yaw | represents the absolute value of the yaw angle; K is the low-light compensation coefficient; γ is the angle attenuation coefficient.

[0019] The system detects periodic jitter patterns in the screen replay pattern of dynamic micro-expression features, applies a fraud penalty coefficient to the initial similarity based on the jitter amplitude, and outputs a risk factor.

[0020] A risk report with weighted volatility curves is generated based on risk factors. When the threshold is exceeded, a recalibration instruction is triggered and the time series model parameters are updated.

[0021] Furthermore, the decoding of iris tremor frequency to generate a liveness confidence coefficient includes:

[0022] Load pulse timing sequence data packets and perform three-stage cascaded filtering: the first stage uses an analog bandpass filter with a passband of 0.5-40 Hz to eliminate baseline drift and high-frequency electromyography interference; the second stage sets a programmable digital notch filter to eliminate power frequency interference harmonics; the third stage performs wavelet threshold noise reduction to separate noise from the effective signal.

[0023] Multi-dimensional biological parameter analysis is performed on the filtered signal: the microsecond time interval between adjacent pulse peaks is calculated, the pulse frequency is statistically analyzed within a 200-millisecond sliding window, the sliding average value of the peak-to-peak voltage difference is measured, and the waveform symmetry is evaluated through an integrator circuit;

[0024] A rule engine and neural network are used for collaborative decision-making: the rule engine outputs a minimum confidence score of 0.1 for frequency values ​​that exceed the 1.5-4.0 Hz bioactivity range; the neural network model inputs the normalized frequency values, amplitude feature values ​​and symmetry index into a fully connected network to output the base confidence score.

[0025] When the absolute value of the face yaw angle is greater than 0.6 radians, a 15% attenuation coefficient is applied to generate the final liveness confidence coefficient.

[0026] Furthermore, multidimensional biological parameter analysis includes:

[0027] The interval between adjacent pulse peaks is calculated based on timestamp information;

[0028] The number of valid pulse occurrences is counted within a 200-millisecond sliding window and converted into a Hertz-unit frequency scalar.

[0029] Calculate the moving average amplitude characteristic value by measuring the voltage difference between the peak and trough.

[0030] The area under the curve at the rising edge and the falling edge are calculated using an integrator circuit, and the similarity ratio is obtained.

[0031] Furthermore, in the three-stage cascaded filtering process:

[0032] When the light intensity is below a 0.3 scale, the voltage threshold decreases by 20%, and when the light intensity is above a 0.7 scale, the voltage threshold increases by 35%.

[0033] Furthermore, the generation of the liveness confidence coefficient includes:

[0034] Establish a pulse stability quantification model: calculate the improved Shannon entropy value, the expression of which includes the ambient light intensity compensation factor and the historical pulse interval mean;

[0035] Construct a neural network activation function to generate basic confidence scores. The input parameters include normalized entropy, the proportion of mean amplitude, and waveform symmetry indicators.

[0036] The steepness coefficient of the activation function is adjusted based on the face yaw angle, and a 15% attenuation is applied to the base confidence when the absolute value of the yaw angle is greater than 0.6 radians.

[0037] Furthermore, the pulse stability quantization model is as follows:

[0038]

[0039] Where H(P) is the pulse interval entropy value; p(Δt) k ) represents the time interval Δt k The probability density within the sliding window; μ is the mean of the historical 100 pulse intervals; λ is the ambient light intensity compensation factor.

[0040] Furthermore, the expression for constructing the neural network activation function to generate the basic confidence score includes:

[0041]

[0042] Where, β base is the basic liveness confidence; j is the input feature index of the neural network (j = 1, 2, 3 correspond to entropy / amplitude / symmetry respectively); k = 2.5 + 0.5·θ yaw (yaw angle θ) yaw (Provided by S120), θ yaw The angle of the human face; w j The weights are loaded from the pre-trained model; Sym wave x1 is the waveform symmetry index; x2 is the normalized entropy value; x3 is the amplitude mean percentage; x4 is the waveform symmetry index.

[0043] Furthermore, the final liveness confidence coefficient:

[0044] β=β base ·max(0.85,1-0.15·|θ yaw |)

[0045] Where β is the final liveness confidence coefficient; when |θ yaw Angle decay is activated when |>0.6.

[0046] Furthermore, fusion with liveness confidence multiplication includes:

[0047] When the timing deviation exceeds 5 milliseconds, Lagrange interpolation resampling is used;

[0048] The initial similarity is linearly mapped to the range [0,1].

[0049] The basic fusion operation is the product of the initial similarity and the liveness confidence coefficient;

[0050] Apply 15% luminance compensation when ambient light intensity is below 0.2 on the scale;

[0051] A 20% geometric distortion attenuation is introduced when the absolute value of the face yaw angle exceeds 0.8 scale.

[0052] Furthermore, the basic fusion operation, which is the product of the initial similarity and the liveness confidence coefficient, includes a hierarchical processing strategy:

[0053] When the confidence coefficient for liveness is less than 0.4, a review flag is directly output;

[0054] When a periodic jitter alarm signal is received, a 50% similarity reduction is performed;

[0055] The system is locked when the difference between three consecutive fusion results exceeds 0.5.

[0056] Furthermore, the risk report generation includes:

[0057] Plot the dynamic fusion weight coefficient fluctuation curve over time;

[0058] When a risk factor exceeds a threshold, a recalibration instruction is sent to the time series model.

[0059] After updating the model parameters, clear the historical cache data and reset the initial weight values.

[0060] The key innovations of this invention include: (1) dynamically linking iris tremor frequency, waveform symmetry, and environmental parameters to construct a pulse stability entropy model, thereby achieving accurate decoding of live biological features. (2) based on the feature space projection method of the three-dimensional structure of the face, through the linkage of the weight enhancement of key areas around the eyes / between the eyebrows and environmental compensation factors, the feature discrimination ability under complex postures is improved. (3) establishing a periodic jitter detection model for micro-expression components, and quantifying the energy of attack features through the Hamiltonian system to achieve millisecond-level fraud behavior recognition.

[0061] The following are its main beneficial effects: (1) By using iris tremor frequency decoding and dynamic weight fusion mechanism, the interference of strong light change and weak light environment on biometric extraction is effectively overcome, and the feature distortion problem caused by drastic changes in illumination is eliminated. (2) By adopting anatomical partition weighting strategy and angle adaptive attenuation mechanism, the feature space offset is automatically corrected in the case of large head deflection, and the stable recognition ability of key biometrics is maintained. (3) Based on micro-expression periodic jitter modeling and fraud penalty coefficient dynamic injection, high sensitivity detection of screen replay attacks is achieved, breaking through the traditional liveness detection defense limitation against high-definition forged videos. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a real-time customer identity verification method based on image recognition, provided as an embodiment of this application. Detailed Implementation

[0063] Example 1: Refer to Figure 1 This is a flowchart illustrating a real-time customer identity verification method based on image recognition provided in an embodiment of the present invention. The process may include at least steps S100-S600:

[0064] S100: Synchronously acquires document images, real-time video streams, and ambient light / angle / infrared signals, and generates an environmental parameter matrix and pulse timing sequence after filtering and normalization.

[0065] S200: It separates and extracts static texture features of documents and dynamic micro-expression features of videos through a dual-stream convolutional network, adjusts the receptive field weights based on the environmental parameter matrix, and performs feature decoupling and dimensionality reduction.

[0066] S300: Calculate the illumination impact factor based on the environmental parameter matrix, integrate historical weight data to generate dynamic fusion weight coefficients through a time series model, and feed them back to S220 to adjust the feature weights.

[0067] S400: Decodes iris tremor frequency to generate liveness confidence coefficient, uses dynamic weight coefficients to weight and fuse static and dynamic features to obtain initial similarity, and then multiplies and fuses it with liveness confidence.

[0068] S500: Detects periodic jitter patterns in dynamic micro-expression features based on screen replay patterns, applies a fraud penalty coefficient to the initial similarity based on the jitter amplitude, and outputs a risk factor.

[0069] S600: Generates a risk report with weighted volatility curves based on risk factors. When the threshold is exceeded, a recalibration instruction is triggered and the parameters of the S330 time series model are updated.

[0070] Step S100 includes at least steps S110-S130:

[0071] S110: Acquires document images and real-time video streams, and measures ambient light intensity, face yaw angle, and raw infrared liveness signal.

[0072] In step S110, the specific implementation of acquiring the document image and real-time video stream while simultaneously measuring environmental parameters is as follows: First, an image of the user-provided ID document is acquired using a high-resolution contact image sensor. This sensor employs a sensing element with an optical resolution of no less than 1200 dpi and is equipped with a ring-shaped LED light source array to eliminate surface reflection interference. When the document is placed in the scanning area, a dual-channel image capture mechanism is automatically triggered: the first channel uses RGB three-color filter separation technology to capture the color image of the document, and the second channel acquires the document's anti-counterfeiting feature image through infrared transmission imaging technology. Simultaneously, a multispectral imaging system is activated to perform three consecutive exposures on the document surface, with each exposure time precisely controlled within 20 milliseconds to eliminate motion blur. The acquired raw document image is immediately subjected to geometric distortion correction processing. A quadrilateral vertex positioning algorithm based on perspective transformation is used to automatically detect the feature points of the four corners of the document. Through coordinate mapping, the curved surface is projected onto a standard plane, outputting a standardized document face region image with a size of 480 pixels × 640 pixels.

[0073] While acquiring the ID image, the binocular stereo vision system begins capturing the user's real-time video stream. This system consists of two synchronously phase-locked CMOS image sensors, positioned on top of the user terminal device, with the optical center distance between the two sensors precisely controlled at 65 mm. The left and right lenses simultaneously acquire a 1280×720 resolution video stream at a rate of 30 frames per second, with each frame appended with a millisecond-level timestamp. During video stream acquisition, the depth perception module calculates the real-time distance between the face and the lens using the time-of-flight principle, outputting a depth map data stream to the image processing unit. When the system detects a face entering the effective recognition area, it immediately activates the environmental parameter measurement unit. This unit integrates three independent sensors: a digital photometer continuously collects ambient light intensity values ​​with a 100-millisecond sampling period, its spectral response curve strictly matching the human eye's visual function Vλ, covering a range from 0.01 to 100,000 lux; an inertial measurement unit with a built-in microelectromechanical gyroscope that tracks facial key points to calculate the pitch and yaw angles of the face relative to the camera in real time; and an infrared liveness detection system that emits a near-infrared pulse beam at a 37.5kHz carrier frequency and simultaneously receives reflected light signals from the iris region via a silicon photodiode, outputting an analog voltage pulse sequence. The data acquisition timing of these three sensors is precisely synchronized by the main control chip, ensuring that each video frame has corresponding environmental parameters and infrared liveness signals recorded.

[0074] When acquiring document images, the system executes the following sub-steps: Activating the scanning motor drive circuit of the contact image sensor, allowing the sensing module to pass uniformly across the document surface; acquiring line scan data at a sampling density of 1200 dots per inch during movement; converting the analog signal into digital grayscale values ​​using an analog-to-digital converter; identifying the document's outline boundaries using an edge detection algorithm; applying affine transformation to correct perspective distortion; and finally, cropping a standard region image containing the face. During real-time video stream acquisition, the following operations are performed simultaneously: the binocular camera activates automatic exposure control, automatically adjusting the aperture and exposure time based on ambient brightness; the face detection algorithm continuously runs to locate the face within the frame; and when a complete face is detected, a 3D depth map generation thread is initiated. The environmental parameter measurement process is as follows: the photosensitive element of the digital photometer converts the light flux into a current signal; the operational amplifier circuit amplifies the micro-current to a quantifiable range; the digitized light intensity value is transmitted to the main processor through a serial interface; the facial key point detection model locates the center of the left and right pupils, the tip of the nose, and the corners of the mouth; the pitch angle and yaw angle are calculated based on the three-dimensional coordinates of these key points; and the angle data is packaged and transmitted through a high-speed data bus.

[0075] The ambient light intensity measurement employs a spectral sampling technique. Specifically, four photosensitive units with different spectral responses measure the illuminance values ​​for visible light, infrared, ultraviolet light, and the full spectrum, respectively, and then the final result is synthesized using a weighted average based on a standard photometric function. The facial yaw angle calculation is based on a deep learning keypoint regression model trained on a dataset of 3 million multi-angle facial images, capable of simultaneously outputting the three-dimensional coordinates of 68 facial feature points. The acquisition of raw infrared liveness signals utilizes time-division multiplexing dual-wavelength modulation technology, alternately emitting 850 nm and 940 nm infrared light to measure blood flow changes and skin surface reflectance characteristics, respectively. Finally, this step outputs three key data streams: a sequence of document images containing timestamps, a sequence of real-time video frames with depth information, and a synchronously acquired raw record set of environmental parameters.

[0076] S120: Perform noise filtering and normalization on ambient light intensity and face yaw angle to generate an environmental parameter matrix.

[0077] In step S120, the specific implementation of noise filtering and normalization of the environmental parameters acquired in S110 is as follows: First, the ambient light intensity data is processed: The system receives the raw light intensity value sequence transmitted in S110, which contains impulse noise caused by sudden environmental changes and sensor quantization errors. To address this interference, the system initiates a three-stage filtering process. The first stage implements moving average filtering, using a sliding window with a width of 10 sampling points, calculating the arithmetic mean of the values ​​within the window to replace the current value. The second stage applies median filtering, reordering the light values ​​within a time range of 50 sampling points and taking the median, effectively eliminating sudden flash interference. The third stage performs Gaussian smoothing, optimizing the continuity of the light curve using a normal distribution weighting function with a standard deviation of 2. After filtering, the system calculates the dynamic range compensation coefficient for the light intensity: when the rate of change of illuminance over 10 consecutive sampling points exceeds 200 lux per second, a logarithmic decay model is activated for compensation, calculating the offset correction value based on the historical average of ambient brightness to prevent zero-point deviation caused by sensor drift in dark environments.

[0078] Further, the system processes the face yaw angle data transmitted from S110. The raw angle values ​​contain jagged fluctuations caused by minor head movements; the system employs a dedicated angle filtering algorithm for noise reduction. First, the raw angle data is converted to a three-dimensional Cartesian coordinate system. Then, a Kalman filter is applied for motion trajectory prediction, and historical position data is used to correct the current measurement value. To further improve stability, the system uses a weighted average algorithm to process the pitch and yaw angle values ​​at three adjacent time points. The weighting coefficients are dynamically adjusted based on the confidence score: when the confidence score of facial key point detection is below a threshold, the system automatically increases the weighting of historical data. Angle normalization is achieved through a three-dimensional rotation matrix transformation. First, the transformation matrix between the world coordinate system and the camera coordinate system is determined. Then, the coordinates of the detected facial key points are multiplied by this transformation matrix on the left. Finally, the normalized pitch and yaw angle values ​​are recalculated. All angle values ​​are linearly mapped to a range of 0 to 1. The ±90 degree range is linearly divided into 180 equal parts. The original angle value is added to 90 degrees and then divided by 180 to obtain the normalized value.

[0079] Furthermore, after completing the above preprocessing, the system constructs an environmental parameter matrix. First, a three-dimensional data buffer is established: the first dimension stores timestamp information, accurate to the millisecond level; the second dimension stores the processed illumination intensity value, retaining two decimal places of precision; the third and fourth dimensions store the normalized pitch and yaw angle values, respectively. Data at each time point constitutes a row of four columns of matrix units, with new rows continuously added over time. When the matrix reaches 10 rows, a matrix compression operation is initiated: maintaining temporal continuity by removing old data rows exceeding 200 milliseconds; optimizing the storage structure using matrix transpose; and finally, compressing redundant data using a run-length encoding algorithm. The matrix update cycle is strictly controlled within 200 milliseconds to ensure synchronization with the video frame rate. After matrix generation, a data checksum is appended, and cyclic redundancy check ensures data transmission integrity.

[0080] In the illumination intensity processing stage, the specific implementation includes: receiving continuous illumination intensity samples from S110; detecting and marking abnormal jump points; recalculating the effective mean within a sliding window; applying an exponential decay model to compensate for low-illuminance measurement bias; and standardizing the corrected values ​​to the range of 0-100,000 lux. The details of face angle processing are: receiving 3D coordinate data of facial key points; removing outliers using a consistency principle based on key point distribution; transforming the angle coordinate system through quaternion rotation; converting angle values ​​to a uniform range; and establishing an angle change trend model to predict the normal value range. The matrix construction operation specifically involves: creating a dynamic two-dimensional data table; adding a new record row every 200 milliseconds; monitoring memory usage in real time; deleting the oldest record when memory reaches a threshold; adding a timestamp index marker; and outputting a matrix data structure with verification information.

[0081] S130: Converts the raw infrared live signal into a pulse timing sequence to complete environmental perception acquisition and parameter synchronous acquisition.

[0082] In step S130, the specific implementation of converting the raw infrared live signal acquired in S110 into a pulse timing sequence is as follows: The system first receives the raw infrared voltage signal stream transmitted by S110, which contains noise components such as high-frequency interference and baseline drift. The signal preprocessing stage adopts a four-stage filtering process: first, it is connected to the high-pass filter of the analog front end to remove the DC drift component through a 1Hz cutoff frequency; second, it is connected to a 10Hz low-pass filter to remove high-frequency electromyographic interference; then, a 50Hz notch filter is applied to eliminate power supply frequency interference; finally, the signal-to-noise ratio is further optimized through the Butterworth bandpass filter of the digital signal processing unit, with a passband range of 0.5 to 40Hz. The filtered signal is transmitted to the pulse detection circuit, which consists of a voltage comparator, a peak hold circuit, and a waveform analysis module. The system scans the signal waveform with a period of 10 milliseconds, determines the pulse start point through zero-crossing detection; triggers the peak hold circuit at the rising edge to record the maximum voltage value; marks the trough position at the falling edge, and automatically records the trough voltage and time coordinates.

[0083] Further, the system performs iris tremor feature extraction. Based on the detected pulse waveforms, the system calculates three key physiological parameters: the pulse time interval, defined as the time difference between two adjacent peaks, accurate to the microsecond level; the pulse frequency, defined as the number of complete pulse cycles detected per second; and the pulse amplitude, defined as the voltage difference between the peak and trough. To ensure measurement accuracy, the system performs adaptive parameter calibration: amplitude compensation is initiated when the ambient temperature change exceeds 5°C; the sampling frequency is automatically increased when the signal quality index falls below a threshold; and a backup infrared light source compensation module is activated in low-light environments. The system also establishes an iris tremor pattern analysis model: by continuously recording the time interval of 100 pulse cycles, the variance is calculated to assess the regularity of the tremor; fast Fourier transform is used to analyze the dominant frequency characteristics; and abnormal tremor patterns are identified based on the spectral distribution.

[0084] Furthermore, after feature extraction, the system constructs a pulse time series. The sequence is stored in a four-tuple data format: the first field stores the relative timestamp, counting from the signal acquisition start point; the second field stores the pulse frequency value; the third field stores the pulse amplitude; and the fourth field stores the time interval data. The sequence data is compressed and stored using the Huffman coding algorithm, and a circular buffer management module is established, with the buffer size set to store 1000 records. Meta-information tags are appended before the sequence data is output, including fields such as acquisition start time, end time, device ID, and ambient temperature value. The system supports multi-threaded access through a dual-port storage structure: the front-end interface continuously writes new data, while the back-end interface allows the liveness detection module to read the data stream. This sequence data establishes a time correspondence with the environmental parameter matrix in a synchronous manner, ensuring that subsequent processing modules can perform spatiotemporal matching.

[0085] In terms of specific implementation details, the signal conversion process includes: receiving the raw infrared analog signal stream; applying a four-stage cascaded filter to reduce noise; identifying valid pulses through a waveform detection module; calculating pulse time and amplitude parameters; establishing a four-tuple data structure; and cyclically storing the data in a buffer. The feature extraction operation specifically involves: measuring the time difference between consecutive pulses; calculating the number of pulses per second as the frequency value; measuring the peak-to-trough difference as the amplitude value; performing temperature compensation to adjust data accuracy; analyzing frequency change patterns; and marking abnormal tremor features. The sequence construction implementation method includes: creating a time-indexed data structure; adding four-dimensional data entries; employing first-in-first-out storage management; updating the current buffer pointer in real time; adding checksum protection information; and outputting compressed encoded data blocks. The final pulse time series is transmitted to the liveness verification module via a high-speed data bus, serving as the core data source for subsequent iris tremor analysis, and simultaneously establishing a spatiotemporal correlation with the environmental parameter matrix.

[0086] Step S200 includes at least steps S210-S230:

[0087] S210: Separate and extract static texture features from ID card images and dynamic micro-expression features from real-time video streams using a dual-stream convolutional network.

[0088] In step S210, feature extraction is performed using a two-stream convolutional network. First, standardized ID image data generated in S110 is loaded. This image data has been geometrically corrected and conforms to the ISO / IEC 19794-5 standard size specifications. This data is then input to the static feature extraction branch, which employs a deep residual convolutional neural network architecture. Specifically, the network contains 12 convolutional layers and 3 max-pooling layers. The first convolutional layer is configured with 32 3×3 convolutional kernels, performing boundary padding on the input image to generate a primary feature map, which is then passed to the second layer via a ReLU activation function. Layers two through eight employ a cross-layer connection structure, cascading the output feature map of the third layer with the output of the fifth layer along the channel dimension to enhance the propagation of skin wrinkles and scar features. Layers nine through twelfth apply dilated convolution with a dilation rate of 2 to expand the receptive field and capture large-area texture distributions. Finally, the 512-dimensional feature vector is compressed into a 128-dimensional static texture feature vector through a global average pooling layer, with each feature dimension encoding a specific biological attribute (such as the shape of the bridge of the nose, the distance between the eyebrows, etc.).

[0089] When processing real-time video stream data synchronously, the system loads a continuous 15-frame video sequence generated by S110, with a time span of 500 milliseconds. This sequence is input to the dynamic feature extraction branch, which adopts a three-dimensional spatiotemporal convolutional neural network architecture. Specifically, the first convolutional module is configured with 64 convolutional kernels of size 3×3×3 to extract basic motion patterns; the second module introduces grouped convolution technology to divide the input feature map into four groups for independent processing; the third module uses temporal dimension dilated convolution to expand the temporal receptive field to 9 frames to capture long-term micro-expression changes; the fourth module combines spatial attention mechanism to strengthen the weight allocation of the eye and mouth regions; the fifth module sets a temporal pyramid pooling layer to extract feature segments from the entire sequence, the first half of the sequence, and the second half of the sequence, respectively. The network outputs a 256-dimensional dynamic micro-expression feature vector, where each 64-dimensional feature corresponds to four micro-expression patterns: eyelid tremor, mouth twitching, slight eyebrow raising, and nasal wing contraction.

[0090] Key implementation details include: in the static branch preprocessing stage, a local binary mode histogram transformation is performed on the ID image to enhance skin texture contrast through histogram equalization; before dynamic branch processing, a pixel-level displacement compensation technique based on optical flow is used to eliminate feature point drift caused by head rotation; the first three layers of convolutional kernels in both branches implement a parameter sharing mechanism to ensure alignment of the basic feature space. Throughout the entire processing, the feature vector generation timestamp is kept synchronized with the environmental parameter matrix output by S120, and each output feature is appended with a data checksum to ensure transmission integrity.

[0091] The static texture feature vector and dynamic micro-expression feature vector will be transmitted to the S220 module for receptive field weight adjustment processing, and will also provide the input basis for the feature decoupling and dimensionality reduction operation of S230. When this step is completed, the system generates a static feature data package (128-dimensional vector + timestamp) and a dynamic feature data package (256-dimensional vector sequence + inter-frame displacement compensation parameters), the data format of which is compatible with the input interface specification of S220.

[0092] S220: Adjust the receptive field weights of static texture features and dynamic micro-expression features based on the environmental parameter matrix.

[0093] In step S220, a dynamic adjustment operation of the feature receptive field weights is performed. First, the static texture feature vector and dynamic micro-expression feature vector sequences transmitted in S210 are received, and simultaneously, the environmental parameter matrix generated in S120 is loaded. The system constructs a weight modulation model based on environmental factors, which includes two main components: a static feature modulator and a dynamic feature reconstructor.

[0094] In the static feature modulator unit: normalized illumination intensity values ​​are extracted from the environmental parameter matrix and mapped to a 0-1 scale range; an illumination-convolution kernel response mapping relationship is established: when the illumination intensity is below 0.3, the center-weighted mode is activated (the weight of the 3×3 kernel region is increased by 70%); when the illumination intensity is above 0.7, the edge enhancement mode is enabled (the weight of the 5×5 boundary region is increased by 50%); the feature channel weights are redistributed through a channel attention mechanism, and a 300% gain coefficient is applied to the illumination-sensitive channels (features numbered 17, 39, and 82). The specific operation process is as follows: the 128-dimensional static feature vector is divided into 8 feature groups, and each group of 16-dimensional features is associated with anatomical regions such as the forehead, cheeks, and chin; the region importance coefficient is calculated based on the illumination intensity value; the coefficient matrix and the feature vector are subjected to a Hadamard product operation; the adjusted 128-dimensional feature vector is output and an illumination adaptation identifier is attached.

[0095] In the dynamic feature reconstructor unit: normalized yaw and pitch angle values ​​are extracted from the environmental parameter matrix; an angle-spatiotemporal adjustment model is constructed: when the absolute value of the yaw angle is greater than 0.6, the time dimension expansion mechanism is activated (concatenating features from three consecutive frames); when the absolute value of the pitch angle is greater than 0.5, the vertical convolution kernel optimizer is activated (increasing the weight of the vertical convolution kernel by 120%); an adaptive gating mechanism is introduced: when the variance of environmental parameter fluctuations exceeds 0.15 for five consecutive frames, the feature update of the current frame is automatically frozen. Specifically, a 256-dimensional dynamic feature vector is input into the long short-term memory network unit, with 64 neurons in the hidden layer; environmental parameters from three historical moments are loaded as time series input; a frame feature weight distribution matrix is ​​output; the weight matrix is ​​weighted and fused with the original feature vector; and a reconstructed 256-dimensional dynamic feature sequence is generated.

[0096] The weight adjustment parameters are fed back to the convolutional layer parameter register of S210 in real time, specifically through three feedback paths: the kernel size adjustment command for the static branch is sent to the sixth convolutional layer control unit of S210; the temporal convolution range parameter for the dynamic branch is sent to the third module scheduler of S210; and the channel gain coefficient is written to the shared memory area for real-time use by S210. Upon completion of the adjustment, the system outputs the calibrated static feature vector (including illumination adaptation code) and dynamic feature sequence (including angle reconstruction flag), with an environment adaptation coefficient (range 0.0-1.0) added to its data structure. This coefficient directly characterizes the robustness level of the current feature to environmental interference. The output data is transmitted to the S230 module via a high-speed bus, and the data packet format includes a feature vector array, environment coefficient values, and a time synchronization scale.

[0097] S230: Perform feature decoupling and dimensionality reduction on the weighted static texture features and dynamic micro-expression features.

[0098] In step S230, feature decoupling and dimensionality compression operations are performed. The system receives the static texture feature vector and dynamic micro-expression feature vector sequences adjusted in S220 and inputs them into a two-layer feature processing network. This network contains two core components: a feature decoupling module and a spatial compression module. Each component performs data processing tasks sequentially.

[0099] The feature deconstruction module implements orthogonalization processing: a dual-channel feature separation loss function is constructed, forcing the cosine similarity between static and dynamic features in the 128-dimensional latent space to be below a threshold of 0.2. Specifically, the static feature extractor is configured with a 4-layer fully connected network to preserve the inherent biological attributes of the ID photo (such as mole coordinates and crow's feet wrinkle direction); the dynamic feature discriminator adopts a gated recurrent unit structure, focusing on temporal change patterns (such as pupil diameter variation trajectory); parameters are dynamically optimized through an adversarial training mechanism: the static branch generator updates weights every 20 iterations, and the dynamic branch discriminator calculates feature difference every 5 iterations; the output consists of a decoupled 128-dimensional static feature vector and a 128-dimensional dynamic feature vector set, with each set of features appended with an orthogonal confidence index (0-100 scale value).

[0100] The spatial compression module implements dimensionality reduction: Principal component analysis is used to reduce feature dimensionality. Variance contribution rate control is applied to static features; dimensionality reduction stops when the cumulative variance contribution rate reaches 98%, outputting a 64-dimensional static feature vector. A 95% variance retention threshold is set for dynamic features, outputting a 48-dimensional dynamic feature vector. Further feature importance screening is performed: the KL divergence of each feature dimension between positive and negative samples is calculated; inefficient dimensions are removed when the divergence value is less than 0.1. Linear features are mapped to a non-linear space using radial basis function kernels to enhance inter-class discriminability. The specific operation process is as follows: static features are input into a kernel space mapping network with 40 basis functions; dynamic features are input into a 32-node autoencoder; feature reconstruction error optimization calculation is performed; and the compressed features are normalized to the [-1, 1] interval.

[0101] The final generated robust feature package contains three data entities: a 64-dimensional static feature encoding array (each feature value is stored in 2-byte blocks, totaling 128 bytes); a 48-dimensional dynamic feature sequence (one group every 10 milliseconds, each group containing a 24-byte data block); and a copy of the environment adaptation coefficients (4-byte floating-point numbers). The feature package is transferred to the system's shared memory area via a DMA controller, with the storage address mapped to the input buffer of the S420 module. Simultaneously, a feature quality report is generated, including orthogonality confidence index, dimensionality compression rate data, and feature reconstruction error values. This report is input to the weight decision unit of the S320 module for subsequent calculations.

[0102] Step S300 includes at least steps S310-S330:

[0103] S310: Calculate the light intensity impact factor based on the change in light intensity in the environmental parameter matrix.

[0104] In step S310, the calculation of the illumination influence factor is performed. The system loads the environmental parameter matrix generated in S120 from the shared storage area. This matrix is ​​a three-dimensional dataset arranged in a time series, containing normalized illumination intensity values ​​and face angle parameters. For the analysis of illumination intensity changes, the system implements a continuous time-series monitoring mechanism: a sliding time window of 200 milliseconds is established, tracing back ten consecutive sampling points from the current moment; the illumination intensity difference between adjacent sampling points is calculated to form a first-order difference sequence; the statistical characteristics of this difference sequence are further analyzed, including the mean absolute value change, the duration of unidirectional changes, and the frequency of abrupt events. Specifically, each element of the difference sequence is validated, and the window size is automatically shortened to 150 milliseconds when three consecutive zero-value elements appear.

[0105] The system's built-in illumination impact factor decision engine operates using a multi-level query mechanism: the core component is a query mapping table stored in non-volatile memory, containing 512 sets of preset environmental scenes and their corresponding factor values. After standardization, the input parameters are first matched against illumination change trend patterns. If the intensity value increases by more than 0.5 scales within 100 milliseconds, a strong light response mode is activated; if a weak light state occurs for 200 milliseconds continuously below 0.2, the system switches to a low-light compensation mode. The matching process incorporates a spatial weight adjustment factor derived from facial region location information provided by S220—when the center point of the face deviates from the image center by more than 15%, the illumination impact weight of the edge region is reduced by 40%. The final output illumination impact factor is appended with three auxiliary markers: a timestamp synchronization tag for alignment with the video stream; a confidence assessment value calculated based on historical data fluctuation range; and an abnormal state marker indicating sudden environmental changes. The calculated factor values ​​are written to dual-port RAM via a direct memory access channel, with the write address mapped to the input buffer of the S320 module.

[0106] S320: Integrates historical weighted data with illumination influence factors, and generates dynamic fusion weighted coefficients through a time series model.

[0107] In step S320, the dynamic fusion weight coefficients are generated. The system simultaneously receives three input data streams: the illumination influence factor data packet transmitted in S310, the short-term historical weight sequence read from the ring buffer, and the environment adaptation coefficients provided by the feature decoupling module in S230. The short-term historical weight sequence extracts the results from the last five calculation cycles, with a time span of 150 milliseconds. Data integrity is automatically verified each time it is extracted, and data is reassembled after confirmation using a cyclic redundancy check (CRC) code. The environment adaptation coefficients undergo downsampling processing, extracting keyframe features from the kilohertz-level raw data to maintain temporal synchronization with the weight sequence.

[0108] The temporal decision model operates using a two-level heterogeneous architecture: The first-level feature fusion unit creates a nine-dimensional feature vector, which is generated through a non-linear combination of input parameters. Specifically, it multiplies the illumination influence factor with the average of the most recent historical weights; establishes a logarithmic relationship between the current environment fit coefficient and the historical range; introduces the second derivative of the illumination factor as a trend feature; and inputs all standardized features into a fully connected layer with a Sigmoid activation function. The second-level modeling unit deploys a gated recurrent neural network, which contains eight storage units with independent forget gate structures. When a new feature vector is input, the forget gate calculates the proportion of historical states retained based on the rate of change of the illumination factor, the update gate controls the intensity of new feature injection based on the difference in the environment fit coefficient, and the output gate generates a 32-dimensional hidden state vector. Finally, a linear transformation layer maps this state vector to dynamically fused weight coefficients and appends a model version identifier and a timestamp index.

[0109] The generation process of weight coefficients employs multiple protection mechanisms: model parameters are automatically backed up to a secure storage area after each iteration; when the network output exceeds a preset threshold, an emergency rollback mechanism is activated to load the previously stable parameters; and pre-trained model parameters are loaded as default values ​​during the initial startup phase. Output data packets are transmitted to two destinations via a high-speed serial interface: the primary path transmits to the S420 feature fusion module, and the secondary path writes to the feedback parameter register group. The data packet format uses a custom protocol encapsulation, including coefficient values, generation timestamps, model iteration counts, and checksums, with its physical layer transmission cycle strictly controlled within 20 milliseconds.

[0110] S330: Feedback the dynamic fusion weight coefficients to step S220 for receptive field weight adjustment.

[0111] In step S330, feedback control of dynamic weight coefficients is performed. The system first parses the weight data packet generated in S320, verifies data integrity through cyclic redundancy check, and stores the parsing result in a three-level cache register. The feedback control engine activates the corresponding instruction set according to the weight value attribute: for the static texture feature branch, the integer part of the weight coefficient is extracted as the main adjustment parameter; for the dynamic micro-expression feature branch, the decimal part is extracted as the timing adjustment factor. The main adjustment parameter is input to the confidence weighter, which selects the operation mode by comparing the confidence evaluation value added in S310—when the confidence exceeds the 0.8 threshold, the static modulator parameters of S220 are directly overwritten; when the confidence is in the range of 0.5 to 0.8, a weighted average calculation of the old and new parameters is performed; when the confidence is below 0.5, a warning signal is output and the original parameters are maintained.

[0112] The dynamic branch adjustment adopts an event-driven strategy: the system continuously monitors the face yaw angle value in the environmental parameter matrix. If the absolute value of the angle exceeds 0.6 degrees for three consecutive frames, the time dimension expansion command is immediately activated; if the pitch angle changes by more than 0.15 degrees within 50 milliseconds, the vertical convolution kernel enhancement command is triggered. The instruction set is compiled into a combination format of 12-bit control code and 8-bit parameter value, and is sent to the dynamic reconstruction unit of S220 in real time through a dedicated instruction bus. The instruction bus adopts a priority arbitration mechanism; when multiple instructions occur simultaneously, the time dimension expansion instruction has the highest execution priority.

[0113] The parameter update process establishes a dual-layer security mechanism: a dual-register file structure is set up inside the main control unit, with the working register receiving new parameters and the backup register retaining the stable version; when the risk factor output by S530 exceeds a preset threshold, the model recalibration process is activated. The specific operations of the recalibration process are as follows: retrieve the most recent 50 sets of high-risk environmental parameter records from the historical database; retrain the fully connected layer parameters of the time-series decision model; and after verifying the stability of the new model output, switch control back to the main system. After each model update, an incremental version number is generated, and an update completion notification is sent to the S620 module via the system status bus. All log data generated by feedback operations is written to non-volatile memory, including parameter change records, model update versions, and a list of abnormal events. This log file is used by the S610 module to generate risk reports.

[0114] Data closed-loop implementation: In the entire feedback control process, the environmental parameter matrix serves as the initial input driving S310 to calculate the illumination influence factor; the output of S310, after fusing the environmental adaptation coefficient of S230, generates a weight coefficient in S320; this coefficient, under the action of S330, controls the receptive field weight adjustment of S220 in real time; the environmental adaptation coefficient output by S220 ultimately flows back to S320 to form a closed-loop optimization; when S530 detects a high-risk state, the recalibration command triggered directly affects the model update mechanism of S330. Through data collaboration among the five functional modules, a closed-loop technology chain is constructed.

[0115] Step S400 includes at least steps S410-S430:

[0116] S410: Decodes the iris tremor frequency of the original infrared live signal and generates a live confidence coefficient.

[0117] In step S410, the decoding operation of the iris tremor biometrics is performed. The system loads the pulse time sequence data packet generated in S130 from the shared storage area. This data packet contains time-stamped iris reflection light pulse parameters. The biosignal analysis process is initiated: First, a three-stage cascaded filtering channel is established. The first stage is configured as an analog bandpass filter with a passband range set from 0.5 Hz to 40 Hz to eliminate baseline drift and high-frequency electromyography interference. The second stage is set as a programmable digital notch filter with the center frequency adjusted to the harmonic frequency of power frequency interference. The third stage performs wavelet threshold noise reduction processing, separating noise from effective signals through biorthogonal wavelet basis functions. The signal after triple filtering is input to the pulse detection unit, which is configured with a dynamic voltage threshold adjustment mechanism: the detection sensitivity is updated in real time according to the light intensity value in the environmental parameter matrix. When the environmental parameter matrix shows that the light intensity is below 0.3 scale, the voltage threshold is reduced by 20% to improve the capture capability in low-light environments; when the light intensity is above 0.7 scale, the voltage threshold is increased by 35% to suppress strong light overexposure interference.

[0118] The feature extraction engine performs multi-dimensional biological parameter analysis on the filtered signal: It calculates the interval between adjacent pulse peaks based on timestamp information, with precision controlled at the microsecond level; it counts the number of valid pulse occurrences within a 200-millisecond sliding window and converts them into a frequency scalar in Hertz units; it simultaneously measures the voltage difference between peaks and troughs and calculates the moving average as the amplitude feature value; additionally, it performs waveform symmetry evaluation by calculating the area under the curve for both the rising and falling edges using an integrator circuit, and then calculating the similarity ratio between the two. The generation of the liveness confidence coefficient employs a collaborative decision-making mechanism between a rule engine and a neural network model: the rule engine performs hard threshold verification; if a frequency value below 1.5 Hz or above 4.0 Hz is detected (a biological activity threshold), it directly outputs a minimum confidence score of 0.1; the neural network model loads pre-trained classification weight parameters, inputs the standardized frequency value, amplitude feature value, and symmetry index into a three-input fully connected network, and outputs the initial basic confidence value after calculation through two hidden layers. The final coefficients require attitude adaptation correction. When the environmental parameter matrix shows that the absolute value of the face yaw angle is greater than 0.6 scale, a 15% attenuation coefficient is applied to compensate for the angle distortion error. The generated liveness confidence coefficients are appended with millisecond-level time synchronization markers and simultaneously transmitted to the input buffer of the S430 module and the auxiliary judgment channel of the S520 fraud detection unit via a high-speed serial interface.

[0119] S420: Utilize dynamic fusion weight coefficients to weight and fuse static texture features and dynamic micro-expression features to generate initial similarity.

[0120] In step S420, a multimodal feature weighted fusion operation is performed. The system simultaneously receives three key data sources: a 64-dimensional static texture feature vector obtained from module S230, a 48-dimensional dynamic micro-expression feature sequence, and a dynamic fusion weight coefficient data packet transmitted from module S320. A feature fusion processing pipeline is established, which includes two core components: a spatial weight allocation unit and a multi-level feature fusion unit. The spatial weight allocation unit first parses the dynamic fusion weight coefficient data structure: extracting the main part of the coefficient as the static feature weight parameter and separating the fractional part as the dynamic feature compensation factor. An anatomical partitioning weighting strategy is implemented for the static texture features: based on the 3D facial model, the 64-dimensional feature vector is divided into the glabella region feature group, the periorbital region feature group, the nasal feature group, the cheek feature group, and the jaw feature group. Among them, the periorbital region weight is increased by 120%, the glabella region weight is increased by 80%, and the other regions maintain the baseline weight. The temporal enhancement mechanism for dynamic micro-expression features is activated when the pitch angle change rate exceeds 0.1 per millisecond for three consecutive frames as shown in the environmental parameter matrix. This mechanism extends the coverage of the temporal convolution kernel to a sequence length of seven frames.

[0121] The multi-level feature fusion unit implements a four-order processing flow: First, a feature alignment operation is performed, synchronizing the weighted static texture feature vector with the temporal dynamic feature sequence along the time axis; next, a 128-dimensional joint feature matrix is ​​constructed, with rows corresponding to time-point features and columns representing feature dimensions; principal component analysis is performed for dimensionality reduction, and calculation stops when the cumulative variance contribution rate reaches 95%, outputting a 32-dimensional simplified feature vector; then, a similarity calculation operation is performed, comparing the current simplified feature vector with the document template feature vector pre-stored in the secure storage area using a cosine similarity algorithm; finally, a triple dynamic verification is performed: the amplitude verification module removes abnormal feature response points exceeding three standard deviations, the temporal continuity verification module triggers an output freeze mechanism when it detects similarity fluctuations exceeding 0.2 scales for five consecutive frames, and the spatial consistency verification module compares the similarity differences between the left and right facial region features, automatically initiating compensation correction when the difference exceeds a threshold. The initial similarity calculation result is appended with an environment adaptation tag copy, which originates from the environment adaptation coefficient data output by S230. The output data stream is divided into two main channels: the main data channel is transmitted to the S430 module to perform biological confidence fusion, and the secondary data channel is simultaneously input to the S510 periodic jitter detection unit as a baseline reference input.

[0122] S430: Multiply and fuse the liveness confidence coefficient with the initial similarity to complete the liveness enhancement feature fusion and similarity calculation.

[0123] In step S430, the final similarity fusion operation for biometric enhancement is performed. The system establishes a dual-input real-time processing channel: the main channel receives the initial similarity sequence data transmitted by S420, and the auxiliary channel receives the liveness confidence coefficient stream generated by S410. A preprocessing calibration procedure is implemented: first, time synchronization is checked; when the timestamp difference between the two data streams exceeds five milliseconds, a Lagrange interpolation algorithm is applied for temporal resampling; then, the initial similarity values ​​are linearly mapped from their original range to a standardized zero-to-one interval; finally, an outlier detection mechanism is activated. If an anomaly is detected where the liveness confidence coefficient is below 0.3 or the initial similarity is below 0.15, a data discard procedure is initiated, and the preceding module is re-requested for input.

[0124] The multiplication fusion core engine implements triple optimization processing: the basic fusion operation follows the fundamental formula that "final similarity equals initial similarity multiplied by the liveness confidence coefficient"; in the environmental compensation dimension, when the illumination intensity value in the environmental parameter matrix is ​​below 0.2 scale, a brightness compensation factor is automatically applied to improve the fusion result by 15%; in the geometric correction dimension, if the environmental parameter matrix shows that the absolute value of the face yaw angle exceeds 0.8 scale, a geometric distortion correction coefficient is introduced to apply a 20% attenuation correction to the fusion result. The system is configured with a hierarchical processing strategy: in the liveness confidence priority mode, when the liveness confidence coefficient is less than 0.4 scale, the initial similarity value is directly ignored, and a review decision mark is forcibly output; in the fraud collaborative response mode, when a periodic jitter alarm signal is received from the S510, a 50% similarity reduction operation is immediately performed; in the security protection mode, when the difference between three consecutive fusion results exceeds 0.5 scale, a system-level lockout state is activated and a recalibration request is triggered.

[0125] The enhanced similarity data packet uses a four-field customized encapsulation format: the first field stores a four-byte precision floating-point similarity value; the second field records a quality assessment index based on the variance of the most recent ten calculations; the third field carries a bioactivity status marker, whose data comes from the analysis conclusions of the S410 module; and the fourth field adds a millisecond-level time synchronization stamp to ensure precise alignment with the video stream timing. The data packet is written to the system output queue through an encrypted data channel, and copies are simultaneously sent to three key functional nodes: the primary copy is transmitted to the S530 risk marker generation module to trigger the final verification decision; the secondary copy is input to the S610 risk report construction module to participate in the dynamic fluctuation curve plotting; and the reference copy is sent to the S320 weighted decision model unit to provide real-time feedback on model parameter iterative optimization. After all processing is completed, the system sends an operation completion status code to the control bus to ensure that each functional unit synchronously enters the next processing cycle.

[0126] Technical closed-loop description: The input data for step S410 originates from the infrared pulse time sequence of S130, and its output directly drives the fusion calculation of S430; the operation of S420 depends on the feature vector provided by S230 and the weight coefficient of S320, and the output directly affects the processing result of S430; the enhanced similarity output of S430 is both the final product of this step and a core parameter input to the risk judgment step of S530. The entire process forms a multi-level data feedback closed loop: when S510 detects an abnormal pattern, its alarm signal will be fed back to S430 in real time for correction; the risk factor output by S530 will trigger the recalibration process of S330 through S620, thereby updating the weight model parameters of S320. The four functional modules establish a mesh interconnection structure through six data links to ensure the system's real-time response capability to environmental interference and fraud attacks. The finally generated enhanced similarity index integrates triple verification information of live biometrics, static document features, and time-series dynamic features, providing a reliable identity verification benchmark for high-risk application scenarios.

[0127] Another implementation method:

[0128] S410: Decodes the iris tremor frequency of the original infrared live signal and generates a live confidence coefficient.

[0129] In step S410, the decoding operation of the iris tremor biometrics is performed. The system loads the pulse time sequence P(t) = {(t...} generated in S130. i A i , Δt i )}, where t i For pulse timestamps, A i The pulse amplitude is Δt. i The time interval between adjacent pulses. Establish a bioactivity quantification model:

[0130] ① Impulse stability quantization (improved Shannon entropy model)

[0131]

[0132] Wherein, H(P): pulse interval entropy value, quantifying the regularity of iris tremor (the smaller the value, the stronger the biological activity); p(Δt) k ): Time interval Δt k The probability density within the sliding window; μ: the mean of the historical 100 pulse intervals; λ: the ambient light intensity compensation factor (derived from I in the S120 environmental parameter matrix). light Mapping).

[0133] This formula quantifies the regularity of pulses; the lower the entropy value, the stronger the biological activity.

[0134] ② Confidence generation (improved neural network activation function)

[0135]

[0136] Input parameters: (Normalized entropy value); (Average percentage); x3 = Sym wave (Waveform symmetry index);

[0137] Where, β base : Baseline liveness confidence (uncompensated for angle effects); j: Neural network input feature index (j = 1, 2, 3 correspond to entropy / amplitude / symmetry respectively);

[0138] k = 2.5 + 0.5·θ yaw (yaw angle θ) yaw (Provided by S120), θ yaw : Face yaw angle (from S120 environmental parameter matrix); w j The weights are loaded from the pre-trained model; Sym wave Waveform symmetry index (range 0-1, 1 indicates perfect symmetry).

[0139] Furthermore, the final liveness confidence coefficient:

[0140] β=β base ·max(0.85,1-0.15·|θ yaw |)

[0141] Where β is the final liveness confidence coefficient (output to S430); when |θ yaw Angle decay is activated when the value is >0.6. The output β value is transmitted to the S430 and S520 modules.

[0142] In step S410, the decoding operation of the iris tremor biometrics is performed. The system first loads the pulse time sequence data packet generated in S130 from the shared storage area. This data contains time-stamped iris reflection light pulse parameters. A three-stage cascaded filtering process is initiated: the first stage uses a hardware-implemented analog bandpass filter to filter out low-frequency baseline drift and high-frequency electromyography interference; the second stage configures a programmable digital notch filter to eliminate power frequency interference harmonics; the third stage implements wavelet threshold noise reduction to separate the effective signal from noise. The filtered signal is input to a dynamic threshold pulse detection unit, whose detection sensitivity is adjusted in real time according to the light intensity value in the environmental parameter matrix in S120: the detection threshold is reduced by 20% in low-light environments (light intensity < 0.3) and increased by 35% in strong-light environments (light intensity > 0.7).

[0143] The feature extraction engine performs four core operations: calculating the microsecond-level time interval between adjacent pulse peaks; statistically analyzing pulse frequencies within a 200-millisecond sliding window; measuring the moving average of the peak-to-peak voltage difference; and evaluating waveform symmetry through an integrator circuit. The generation of the liveness confidence coefficient employs a collaborative decision-making process between a rule engine and a neural network: the rule engine forces a minimum confidence score of 0.1 for frequency values ​​exceeding the 1.5-4.0 Hz bioactivity range; the neural network model inputs normalized frequency values, amplitude feature values, and symmetry indices into a fully connected network, outputting an initial base confidence value. The final coefficient undergoes attitude adaptation correction: when the absolute value of the yaw angle provided by S120 exceeds 0.6 radians, a 15% attenuation coefficient is applied. The generated liveness confidence coefficient is appended with a millisecond-level timestamp and transmitted to the input buffer of S430 via a high-speed data channel, simultaneously providing auxiliary judgment criteria for the S520 fraud detection unit.

[0144] S420: Generate initial similarity by using dynamic fusion weight coefficients to weight fusion features.

[0145] In step S420, a multimodal feature fusion operation is performed. The input data includes: static texture feature vectors. (S230 output); Dynamic micro-expression feature sequence

[0146] (S230 output); Dynamic fusion weight coefficient α (S320 output).

[0147] ③ Feature space projection (improved Chebyshev distance model)

[0148]

[0149] Among them, D cheb (F, T): Improved Chebyshev distance (measures the difference between static features and the template); F i : The i-th component of the current static feature vector; T i : The i-th dimension component of the feature vector of the document template; w i =α·c i Anatomical zone weighting (periocular area w) i =1.2α, glabella w i =0.8a); T: Document template features (pre-stored in the secure area); c i Anatomical zone weighting (periocular area c) i =1.2, glabella c i =0.8).

[0150] ④ Temporal feature fusion (improved Hamiltonian system model)

[0151]

[0152] Among them, H(F) d ): Dynamic feature Hamiltonian (representing the total energy of micro-expressions);

[0153] Dynamic feature change rate (calculated via inter-frame difference); ||·|| 2 Euclidean norm squared (to calculate the range of change).

[0154] Furthermore, the potential energy function:

[0155]

[0156] Among them, V(F) d F is the dynamic feature potential function, which quantifies the temporal stability of micro-expression features; d (k) (t): The k-th micro-expression component (k = 1, ..., 5, corresponding to eyelids / corners of the mouth / brow arch, etc.); δ: Dynamic compensation factor (taking the decimal part of α); F d (k) : The kth type of microexpression component;

[0157] Further, the initial similarity is calculated:

[0158]

[0159] Among them, S init Initial similarity (output to S430); D cheb Chebyshev distance (abbreviated as D) cheb (F, T)).

[0160] Output S init Transmitted to S430.

[0161] In step S420, a multimodal feature fusion operation is performed. The system simultaneously receives three inputs: a 64-dimensional static texture feature vector provided in S230, a 48-dimensional dynamic micro-expression feature sequence, and dynamic fusion weight coefficients transmitted in S320. The spatial weight allocation unit first parses the weight coefficient structure: extracting the main part as the static feature weight parameter and separating the fractional part as the dynamic feature compensation factor. Anatomical partitioning weighting is applied to the static features: the weight of the periorbital region is increased by 120% (enhancing key features such as wrinkles), the weight of the glabella region is increased by 80%, and other regions maintain the baseline weight.

[0162] Dynamic feature processing employs a temporal enhancement mechanism: when the S120 environmental parameter matrix shows a pitch angle change rate > 0.1 / ms for three consecutive frames, the temporal convolution kernel coverage is expanded to a seven-frame sequence. The feature fusion unit performs fourth-order processing: aligning the weighted static features with the dynamic feature sequence in time; constructing a 128-dimensional joint feature matrix; compressing to a 32-dimensional feature vector through principal component analysis; and calculating the cosine similarity with the ID template. Triple dynamic verification is implemented before output: removing amplitude anomalies (exceeding 3 standard deviations); freezing outputs with fluctuations > 0.2 for five consecutive frames; and initiating compensation when the difference between left and right facial features exceeds a threshold. The generated initial similarity is appended with an environmental adaptation tag, transmitted via the main path to S430, and input via the secondary path to the S510 periodic jitter detection unit.

[0163] S430: Multiply and fuse the liveness confidence coefficient with the initial similarity.

[0164] In step S430, biometric enhancement-enhanced similarity fusion is performed. An environment-adaptive fusion model is established:

[0165] ⑤ Core fusion equation (improved Hamiltonian Monte Carlo sampling)

[0166]

[0167] Among them, S enhanced : Similarity after liveness enhancement (uncompensated environment); exp(·): Exponential function (used for time-synchronized decay); Δt sync :S init The timestamp deviation from β; σ: time tolerance threshold (fixed at 5ms);

[0168] ⑥ Environmental compensation operator

[0169]

[0170] Among them, S final Ultimately, this enhances similarity; light Ambient light intensity; |θ yaw |: Absolute value of yaw angle; K = 0.15·(0.2-I light (Low-light compensation); γ = 0.2·(|θ) yaw |-0.8)(angle decay);

[0171] Output S final The data is transmitted to the S530 risk decision module and simultaneously fed back to the S320 weight optimization model.

[0172] In step S430, the final fusion of biometric enhancement is performed. The system establishes dual-channel input calibration: when the initial similarity in S420 deviates from the timestamp of the liveness confidence coefficient in S410 by more than 5 milliseconds, Lagrange interpolation resampling is used; the initial similarity is linearly mapped to the range [0,1]; abnormal data with liveness confidence <0.3 or similarity <0.15 are discarded. The core fusion employs an environmental adaptive mechanism: the basic calculation is the product of the initial similarity and the liveness confidence; a 15% brightness compensation is applied in low-light environments (S120 illumination intensity <0.2); and at large yaw angles (|θ... yaw When |>0.8), a 20% geometric distortion attenuation is introduced.

[0173] The system is configured with a three-tiered processing strategy: In the liveness detection priority mode (confidence < 0.4), a review marker is directly output; upon receiving a periodic jitter alarm from the S510, a 50% similarity reduction is forcibly implemented; and the system is locked when the difference between three consecutive fusion results is > 0.5. The final generated enhanced similarity data packet contains four fields: floating-point similarity value, variance quality index, bioactivity marker, and millisecond-level timestamp. The data packet is encrypted and transmitted to the S530 risk decision module, while copies are simultaneously sent to the S610 reporting system and the S320 weighted decision model, forming a closed-loop feedback.

[0174] The S400 technology achieves three major technological breakthroughs in the liveness enhancement fusion mechanism constructed in this section (S410-S430): (1) Improved accuracy of iris tremor analysis: The false rejection rate of liveness is reduced through three-level filtering and dynamic threshold detection; (2) Optimized feature fusion efficiency: The anatomical partition weighting strategy increases the contribution of key periorbital areas to recognition by 40%, and further compresses the fusion delay; (3) Enhanced environmental adaptability: The improved illumination compensation and geometric correction mechanism improves the recognition rate under strong light (>1000 lux) and the recognition rate in large deflection angle (>45°) scenes; It provides millisecond-level response and false recognition rate of less than 0.8% identity verification guarantee for high-risk financial scenarios, while meeting the ISO / IEC 19794-5 L4 level security standard.

[0175] Step 500 includes at least steps S510-S530:

[0176] S510: Detects whether the periodic jitter pattern in dynamic micro-expression features conforms to the screen playback pattern.

[0177] In step S510, a periodic jitter pattern detection operation is performed. The system receives a 48-dimensional dynamic micro-expression feature sequence generated in S230, which is continuously input into the detection engine at a rate of 1 frame every 10 milliseconds. A three-level screen replay analysis process is initiated: first, a time-domain signal processing channel is established, and the feature data of the most recent 100 frames is loaded to form a time window; data streams of five key micro-expression response channels are extracted, including eyelid tremor intensity, mouth corner displacement vector, brow arch undulation frequency, nasal wing contraction amplitude, and facial blood flow fluctuation value. The time-frequency conversion unit performs a windowed Fourier transform on each channel: the time series is segmented using a Hanning window function, and the window width is set to 20 frames; a spectral energy scan is performed within the display refresh frequency band of 30 Hz to 80 Hz; the energy peak of the 60 Hz fundamental frequency and its 120 Hz second harmonic is detected with a focus.

[0178] The jitter pattern recognition engine implements four-dimensional correlation analysis: First, it calculates the moving average of the absolute differences between adjacent feature frames to construct a jitter intensity curve; second, it detects periodic patterns through autocorrelation functions and sets time lag parameters to match common screen refresh intervals; then, it applies wavelet packet decomposition technology to separate device noise components and extracts the inherent noise pattern of the CMOS imager in the second-layer detail coefficients; finally, it establishes a time-series classification model based on deep learning: this model is trained on a dataset of 50,000 positive and negative samples, the input layer accepts 64 frames of micro-expression feature slices, three convolutional layers capture short-term, medium-term, and long-term jitter patterns respectively, and the output layer generates a period matching confidence score. Key implementation details include: setting an automatic sensitivity adjustment mechanism, increasing the frequency detection tolerance range by 20% when the illumination intensity in the environmental parameter matrix is ​​higher than the 0.8 scale; and activating an emergency detection mode to shorten the analysis window to 30 frames when the initial similarity value provided by S420 is lower than 0.3. The detection results are encapsulated into a four-tuple data structure: timestamp accurate to milliseconds, main period frequency value (in Hertz), jitter amplitude index (range 0-1), and pattern matching confidence (0.0-1.0). This data structure is transmitted via a high-speed bus to three destinations: the main path inputs to the S520 penalty calculation unit, the secondary path writes to the log area of ​​the S610 report generation module, and the monitoring path synchronizes in real-time to the S330 model update module.

[0179] S520: When a periodic jitter pattern is detected, apply a fraud penalty coefficient to the initial similarity based on the jitter amplitude.

[0180] In step S520, the fraud penalty coefficient is generated and applied. The system simultaneously receives three input signals: the periodic jitter detection result output by S510, the initial similarity data stream transmitted by S420, and a copy of the liveness confidence coefficient transmitted by S410. A penalty strategy decision engine is established: first, the pattern validity verification process is initiated—when the jitter amplitude index exceeds 0.5 for five consecutive frames and the pattern matching confidence is greater than 0.85, periodic jitter is confirmed. The penalty level classification module implements differentiated processing based on jitter characteristics: for a fixed refresh mode with a frequency locked at 60±2 Hz, a basic penalty is applied; when 120 Hz high-frequency harmonic energy exceeds the limit, the penalty is upgraded to a medium level; when device fingerprint noise characteristics are present simultaneously, the highest level penalty is activated.

[0181] The penalty coefficient generation unit implements quantization mapping: a jitter amplitude-penalty coefficient lookup table is created, containing 256 mapping relationships (e.g., an amplitude range of 0.6 to 0.7 corresponds to a penalty coefficient of 0.3). A dynamic adjustment mechanism is also introduced: when the environmental parameter matrix shows that the absolute value of the face yaw angle exceeds 0.7, the penalty coefficient is increased by an additional 20%; when the enhanced similarity value transmitted via S430 decreases by more than 0.25 in the last three frames, the penalty weight is automatically increased by 15%. The penalty application process uses a non-linear fusion function: first, the initial similarity is mapped from the range [0,1] to logarithmic coordinate space; a penalty coefficient weighting operation is performed in the transform domain, calculated as: penalized similarity = original similarity × (1 - penalty coefficient × scene influence factor); then, it is inversely transformed to linear space. The specific implementation steps include: verifying the timestamp synchronization status of the input data, and using cubic spline interpolation for realignment when the deviation exceeds 5 milliseconds; scanning for abnormal drop points in the initial similarity sequence, and pausing penalty application when the single-frame drop rate exceeds 0.4; generating an operation audit log, including the original similarity, penalty coefficient calculation path, corrected values, and environmental snapshot data. The penalized similarity sequence is appended with dual verification labels: the first label indicates the penalty level (basic / intermediate / advanced), and the second label marks the source of the scene's influencing factors, which are transmitted to the S530 decision module in real time via a point-to-point connection interface.

[0182] S530: Output verification tags and risk factors based on the similarity after penalty to complete multi-factor fraud verification and risk labeling.

[0183] In step S530, the final verification decision and risk quantification operation are performed. The system receives the penalized similarity sequence transmitted by S520, loads the risk decision model, and executes a three-layer judgment process: The first layer implements a threshold segmentation strategy: setting a dual-threshold judgment interval: when the similarity value is greater than 0.85, an "pass" mark is output; when it is between 0.65 and 0.85, a review mechanism is activated and marked as "manual review"; when it is below 0.65, a "reject" mark is output. The second layer introduces a situational awareness mechanism: analyzing the trend of similarity changes in the most recent 10 frames, when it shows a monotonous decrease and the cumulative decrease exceeds 0.3, the decision level is automatically upgraded to "high-risk rejection". The third layer performs multi-source corroboration analysis: fusing the liveness confidence coefficient of S410 with the periodic jitter detection results of S510—when the liveness confidence is below 0.4 and there is evidence of periodic jitter, it is forcibly set to a "fraud warning" mark.

[0184] The risk factor generation module implements five-dimensional feature fusion: extracting statistical features of penalized similarity (mean, variance, trend gradient), stability indicators of liveness confidence, periodic jitter parameters, illumination mutation counts of the environmental parameter matrix, and extreme values ​​of fluctuations in the S320 weight coefficients. Risk factor values ​​are generated using a random forest regression model: the input layer accepts a 25-dimensional feature vector, which is integrated and calculated by ten decision trees, outputting a risk quantification value ranging from 0.0 to 1.0. Risk factor output is managed hierarchically: values ​​between 0.0 and 0.3 are considered low-risk, 0.3 to 0.7 are medium-risk, and values ​​above 0.7 are high-risk. The final output package is encapsulated as a binary data structure: the first field stores the verification decision marker (including four types of status codes), and the second field stores the risk factor value and hierarchical label. The data is transmitted securely to the system output interface and simultaneously sent to three key nodes: the S610 module for risk report generation, the S620 module for triggering the recalibration process, and the S330 module for weight decision model optimization. After each decision is generated, the system status register is updated to record the timestamp, decision type, and risk factor fingerprint. This register is kept synchronized with the historical database of the S600 system.

[0185] The technical closed-loop implementation is as follows: the input data source for step S510 is the dynamic micro-expression feature sequence generated by S230; the operation of S520 simultaneously depends on the initial similarity of S420 and the detection result of S510; the verification mark and risk factor generated by S530 directly drive the recalibration decision of S620. The three modules form a closed loop through a bidirectional data link: the detection result of S510 guides the penalty application intensity of S520, the output of S520 determines the decision level of S530, and the risk factor of S530 is fed back to S510 to adjust the detection sensitivity—when the risk factor value exceeds 0.7 three times consecutively, S510 automatically reduces the detection window to 20 frames and doubles the frequency resolution. Furthermore, the enhanced similarity result transmitted by S430 establishes a collaborative mechanism with the S520 module: when the difference between the two exceeds 0.15, a cross-module data consistency verification process is triggered.

[0186] Step S600 includes at least steps S610-S630:

[0187] S610: Generate a risk report containing a dynamic fusion weighted volatility curve based on risk factors.

[0188] In step S610, the risk report is constructed. The system starts the risk data analysis engine and first loads the continuous risk factor sequence data transmitted by S530 from the circular buffer. This sequence is marked with millisecond-level timestamps and covers the decision cycle of the most recent 180 seconds. The system synchronously connects to the historical database to retrieve the historical records of the dynamic fusion weight coefficients output by S320, with the retrieval criteria being the time period that perfectly matches the risk factor timestamps. The weight fluctuation curve generation process is initialized: a sliding analysis window with a time width of 30 seconds is established, and the window slides forward at a rate of 1 time per second; cubic spline interpolation is performed on the dynamic fusion weight coefficients within each window to generate a smooth, continuous curve; the standard deviation of the curve within the window is calculated, and peaks and troughs where the standard deviation exceeds the 0.15 threshold are marked; the processing results are mapped to a two-dimensional coordinate system to form a visual curve, with the horizontal axis representing the time marker line accurate to milliseconds and the vertical axis representing the weight values ​​normalized to the range of 0 to 1. An ambient light intensity thermal map is overlaid on the curved background layer: Illumination intensity data for the corresponding time period is extracted from the S120 environmental parameter matrix, and a thermal overlay layer with 40% transparency is generated using a grayscale rendering engine. The overlay view of the curved map and the thermal layer constitutes the core visualization component.

[0189] The correlation analysis engine runs synchronously: creating a ten-dimensional feature correlation matrix. The horizontal axis includes risk factor trajectories, the environmental adaptation coefficient sequence output by S230, the liveness confidence curve by S410, the periodic jitter detection marker by S510, and the initial similarity fluctuation by S420. The vertical axis automatically calculates the Pearson correlation coefficient matrix, marking strongly correlated points with values ​​exceeding 0.7 with red stars. Event labeling is performed: records for time periods with risk factors exceeding 0.7 are extracted and correlated with the dynamic fusion weight coefficient values, S430 enhanced similarity mutation points, and S310 illumination influence factor fluctuation values ​​within that time period, forming event description triplets. The final report encapsulation module compiles the visualized curves, correlation matrix, and event list into a structured document: the curve graphs are stored in vector graphics format in the first section of the report; the correlation matrix is ​​stored in binary encoding format in the appendix; and the event list is encapsulated in XML format with a time index. Double encryption is implemented before report output: the first layer uses the AES-256 algorithm to encrypt the core data block, and the second layer uses digital signature authentication based on the device's unique key. The encrypted report is transmitted in three ways: a simplified summary version is pushed to the user terminal's visual interface, the full encrypted version is written to a secure storage area, and the core metadata is synchronously uploaded to the real-time status register of the system monitoring bus.

[0190] S620: When the risk factor exceeds the preset threshold, a recalibration instruction is triggered in the dynamic weight decision module.

[0191] In step S620, the generation and distribution of recalibration instructions are performed. The system initializes the real-time risk monitoring thread: continuously polling the S530 risk factor output channel and setting a two-level dynamic threshold judgment model. The first level is a fixed threshold of 0.7, which is activated when the risk factor exceeds this limit for three consecutive frames; the second level is an adaptive floating threshold, whose baseline value is dynamically calculated based on the environmental parameter matrix: if the S120 environmental parameter matrix shows that the light intensity exceeds the 0.8 scale, the floating threshold is automatically reduced by 0.1; if the face yaw angle is greater than the 0.6 scale for five consecutive frames, the floating threshold is increased by 0.15. When the risk factor simultaneously exceeds both thresholds, the context verification process is initiated: five related parameters are extracted from the system cache: S510 periodic jitter detection confidence data, S430 instantaneous change rate of enhanced similarity, S320 coefficient of variation of weight coefficient, S410 liveness confidence stability index, and S310 light influence factor fluctuation trajectory. Establish a parameter voting mechanism: when at least three parameters support the high-risk determination, the validity of the risk event is confirmed; if fewer than three parameters support the determination, a delayed review mechanism is activated to wait for confirmation from the next two frames of data.

[0192] The instruction compilation process employs layered encapsulation: Operation instruction codes are generated at the control layer, with recalibration backup operations identified by hexadecimal code 0x5A. The target module points to the S320's physical address segment, and the execution priority is set to the emergency level. At the data layer, high-risk event context packets are packaged, including risk factor peak records, environmental parameter matrix slices at the trigger time, associated S510 jitter frequency values, and S420 initial similarity values. At the verification layer, a four-byte digital signature and a two-byte cyclic redundancy check (CRC) code are appended. Instruction transmission is executed in parallel through three channels: the main channel sends control signals to the S630 via a real-time interrupt request bus, ensuring a transmission path delay of less than 10 milliseconds; the auxiliary channel writes the context packet to the S320's secure input buffer via the data bus; and the monitoring channel writes timestamped audit records to the system log server, with the record format conforming to the ISO27001 standard. A three-way handshake mechanism is implemented during transmission: the receiver returns an acknowledgment signal every 50 milliseconds. If the sender does not receive an acknowledgment for three consecutive cycles, it automatically switches to the backup link to retransmit and escalates the event to a system-level alarm.

[0193] S630: Send the recalibration command to step S330 to update the timing model parameters.

[0194] In step S630, closed-loop control operations for model recalibration are performed. After receiving the instruction packet transmitted by S620, the system activates a five-stage reconfiguration process: The first stage involves instruction parsing and security verification: the control layer is disassembled to obtain the opcode and target address, and the validity of the digital signature is verified using a hardware encryption / decryption engine; if verification fails, a retransmission request signal with code 0xF1 is sent to S620; if verification passes, the parameter update enable signal is released to the global control register. The second stage extracts the high-risk dataset: based on the environmental snapshot fingerprint carried by the instruction data layer, four key historical records are retrieved from non-volatile memory: the original calculation log of the dynamic fusion weight coefficient, the intermediate parameters of the S310 illumination influence factor, the S230 feature decoupling quality report, and the S130 infrared pulse sequence anomaly marker. The data extraction range is limited to a precise time period of 150 milliseconds before and after the occurrence of the risk event.

[0195] The third phase involves incremental model updates: loading the current weight file of the S320 temporal decision model into a secure sandbox environment. Three-layer parameter optimization is performed: fully connected layer parameters are updated using mini-batch stochastic gradient descent, with the learning rate dynamically adjusted based on risk level: the learning rate is set to 0.01 when the risk factor is between 0.7 and 0.8, and increased to 0.05 when it exceeds 0.8; the gating mechanism reconstructs forgetting gate parameters based on the S230 quality report: when the feature decoupling orthogonality is below 0.2, the forgetting factor is increased by 30%; the network structure adjustment mechanism is activated when an unrecorded environmental pattern is detected: adding a hidden layer extension unit containing 16 neurons and initializing the weights. The fourth phase involves initiating the verification and switching process: A test set containing 20 high-risk scenario samples is created, including three types: strong light mutation sequences, large-angle deflection sequences, and high-simulation attack sequences. The new model runs on the test set, and the error rate is compared with the original model. If the error rate decreases by more than 15%, the update is deemed valid, and a new version identifier is generated. The version number format is major version number + minor version number + patch number (e.g., v2.13.6). A dual-track operation mechanism is implemented: after the new model warms up in the backup area for 10 seconds, control is transferred to the new model via a memory switching instruction. The fifth phase completes the closed-loop feedback: the update results are encapsulated into a structure containing a new version identifier, a test set performance report, and an error sample index list. This is transmitted to the two target nodes: the primary node updates the S320 model parameter storage area, overwriting the original weight file; the secondary node writes to the newly created log partition of the S610 reporting system. Finally, the historical cache registers of the S330 are cleared, and a weight reset signal is sent to the S220 receptive field adjustment module, completing the entire recalibration cycle.

Claims

1. A real-time customer identity verification method based on image recognition, characterized in that, include: Simultaneously acquire document images, real-time video streams, and ambient light / angle / infrared signals, and generate an environmental parameter matrix and pulse timing sequence after filtering and normalization. The process of acquiring document images includes: acquiring images of the ID document provided by the user through a high-resolution contact image sensor. The sensor uses a sensing element with an optical resolution of not less than 1200 dpi and is equipped with a ring LED light source array to eliminate surface reflection interference. A dual-stream convolutional network is used to separate and extract static texture features of documents and dynamic micro-expression features of videos. The receptive field weights are adjusted based on the environmental parameter matrix, and feature decoupling and dimensionality reduction are performed. The illumination impact factor is calculated based on the environmental parameter matrix, and the dynamic fusion weight coefficient is generated by integrating historical weight data through a time series model and fed back to the feature weight adjustment module. Decode the iris tremor frequency to generate a liveness confidence coefficient, use dynamic weight coefficients to weight and fuse static and dynamic features to obtain an initial similarity, and then multiply and fuse it with the liveness confidence. The expression fused with the liveness confidence multiplication includes: Among them, S enhanced Similarity after live-enhanced imaging; S init β is the initial similarity; β is the final liveness confidence coefficient; exp(·) is the exponential function; Δt sync σ represents the timestamp deviation; σ is the time tolerance threshold. Furthermore, the expression for the environmental compensation operator is: Among them, S final To ultimately enhance similarity; I light For ambient light intensity; |θ yaw | represents the absolute value of the yaw angle; K is the low-light compensation coefficient; γ is the angle attenuation coefficient. The system detects periodic jitter patterns in the screen replay pattern of dynamic micro-expression features, applies a fraud penalty coefficient to the initial similarity based on the jitter amplitude, and outputs a risk factor. A risk report with weighted volatility curves is generated based on risk factors. When the threshold is exceeded, a recalibration instruction is triggered and the time series model parameters are updated.

2. The method according to claim 1, characterized in that, Decoding iris tremor frequency to generate a liveness confidence coefficient includes: Load pulse timing sequence data packets and perform three-stage cascaded filtering: the first stage uses an analog bandpass filter with a passband of 0.5-40 Hz to eliminate baseline drift and high-frequency electromyography interference; the second stage sets a programmable digital notch filter to eliminate power frequency interference harmonics; the third stage performs wavelet threshold noise reduction to separate noise from the effective signal. Multi-dimensional biological parameter analysis is performed on the filtered signal: the microsecond time interval between adjacent pulse peaks is calculated, the pulse frequency is statistically analyzed within a 200-millisecond sliding window, the sliding average value of the peak-to-peak voltage difference is measured, and the waveform symmetry is evaluated through an integrator circuit; A rule engine and neural network are used for collaborative decision-making: the rule engine outputs a minimum confidence score of 0.1 for frequency values ​​that exceed the 1.5-4.0 Hz bioactivity range; the neural network model inputs the normalized frequency values, amplitude feature values ​​and symmetry index into a fully connected network to output the base confidence score. When the absolute value of the face yaw angle is greater than 0.6 radians, a 15% attenuation coefficient is applied to generate the final liveness confidence coefficient.

3. The method according to claim 2, characterized in that, Multidimensional biological parameter analysis includes: The interval between adjacent pulse peaks is calculated based on timestamp information; The number of valid pulse occurrences is counted within a 200-millisecond sliding window and converted into a Hertz-unit frequency scalar. Calculate the moving average amplitude characteristic value by measuring the voltage difference between the peak and trough. The area under the curve at the rising edge and the falling edge are calculated using an integrator circuit, and the similarity ratio is obtained.

4. The method according to claim 2, characterized in that, In a three-stage cascaded filtering process: When the light intensity is below a 0.3 scale, the voltage threshold decreases by 20%, and when the light intensity is above a 0.7 scale, the voltage threshold increases by 35%.

5. The method according to claim 2, characterized in that, The generation of the liveness confidence coefficient includes: Establish a pulse stability quantification model: calculate the improved Shannon entropy value, the expression of which includes the ambient light intensity compensation factor and the historical pulse interval mean; Construct a neural network activation function to generate basic confidence scores. The input parameters include normalized entropy, the proportion of mean amplitude, and waveform symmetry indicators. The steepness coefficient of the activation function is adjusted based on the face yaw angle, and a 15% attenuation is applied to the base confidence when the absolute value of the yaw angle is greater than 0.6 radians.

6. The method according to claim 5, characterized in that, The pulse stability quantization model is as follows: Where H(P) is the pulse interval entropy value; p(Δt) k ) represents the time interval Δt k The probability density within the sliding window; μ is the mean of the historical 100 pulse intervals; λ is the ambient light intensity compensation factor.

7. The method according to claim 5, characterized in that, The expressions for constructing neural network activation functions to generate basic confidence scores include: Where, β base is the basic liveness confidence; j is the input feature index of the neural network (j = 1, 2, 3 correspond to entropy / amplitude / symmetry respectively); k = 2.5 + 0.5·θ yaw (yaw angle θ) yaw (Provided by S120), θ yaw The angle of the human face; w j The weights are loaded from the pre-trained model; Sym wave x1 is the waveform symmetry index; x2 is the normalized entropy value; x3 is the amplitude mean percentage; x4 is the waveform symmetry index. Furthermore, the final liveness confidence coefficient: β=β base ·max(0.85,1-0.15·|θ yaw |) Where β is the final liveness confidence coefficient; when |θ yaw Angle decay is activated when |>0.

6.

8. The method according to claim 1, characterized in that, Fusion with liveness confidence multiplication includes: When the timing deviation exceeds 5 milliseconds, Lagrange interpolation resampling is used; The initial similarity is linearly mapped to the range [0,1]. The basic fusion operation is the product of the initial similarity and the liveness confidence coefficient; Apply 15% luminance compensation when ambient light intensity is below 0.2 on the scale; A 20% geometric distortion attenuation is introduced when the absolute value of the face yaw angle exceeds 0.8 scale.

9. The method according to claim 8, characterized in that, The basic fusion operation is the product of the initial similarity and the liveness confidence coefficient, including a hierarchical processing strategy: When the confidence coefficient for liveness is less than 0.4, a review flag is directly output; When a periodic jitter alarm signal is received, a 50% similarity reduction is performed; The system is locked when the difference between three consecutive fusion results exceeds 0.

5.

10. The method according to claim 1, characterized in that, Risk report generation includes: Plot the dynamic fusion weight coefficient fluctuation curve over time; When a risk factor exceeds a threshold, a recalibration instruction is sent to the time series model. After updating the model parameters, clear the historical cache data and reset the initial weight values.

Images