Image domain adaptive calibration method and device for cross-terminal face scanning health detection

Abstract

The present application relates to the technical field of health detection, and discloses an imaging domain adaptive calibration method and device for cross-terminal face scanning health detection, which comprises the following steps: constructing an imaging domain evidence vector in combination with the equipment archives, acquisition parameters and original imaging data of the face scanning terminal; calling a reference color demixing matrix, generating a bounded residual through a calibration super network, and obtaining an initial calibration matrix; mapping the original imaging data to a standard imaging domain to obtain a standardized face timing signal; extracting a candidate pulse signal and a contrast region noise signal, and iteratively carrying out conversation self-calibration to obtain a self-calibration matrix; calculating calibration distinguishability, and outputting a detection result according to a threshold grading shunt. The calibration mechanism separates the equipment reference and the conversation residual, combines counterfactual training and high-trust conversation anchor point updating, improves the consistency of cross-terminal detection, adapts to the complex on-site environment of mobile phones and self-service health terminals, and ensures the reliability of the face scanning health detection result.

Description

Technical Field

[0001] This invention relates to the field of health detection technology, and more specifically to an imaging domain adaptive calibration method and apparatus for cross-terminal facial recognition health detection. Background Technology

[0002] Facial health checks analyze subtle color fluctuations in facial skin caused by changes in blood volume to extract remote photoplethysmography (rPPG) waveforms, and further estimate heart rate, respiratory rate, heart rate variability, blood pressure risk, or other health indicators.

[0003] Slight color fluctuations are usually much smaller than pixel changes caused by static skin reflection, automatic exposure, white balance, camera noise, video compression, and ambient light variations. Even when shooting the same user, different terminals may produce different RGB color sequences due to differences in sensor spectral response, lens and filter, image signal processor, gamma curve, exposure time, analog gain, white balance gain, frame rate control, and encoding strategies. There are also dynamic changes within the same device. The automatic exposure of a mobile phone will adjust the exposure time and gain during the measurement process, and the location of the self-service terminal may use light-emitting diodes (LEDs) with power frequency flicker. The user's skin color, facial oil reflection, differences in light exposure on the left and right cheeks, and local occlusion will also change the projection of the pulse signal in each color channel. Simply adding light enhancement or fine-tuning the model on the target device during training cannot guarantee that every actual acquisition is within the training distribution, nor can it explain whether a certain calibration has sufficient evidence. Summary of the Invention

[0004] This invention provides an imaging domain adaptive calibration method and apparatus for cross-terminal facial recognition health detection, in order to solve the problem that terminal imaging characteristics and session imaging state cannot be modeled separately.

[0005] In a first aspect, the present invention provides an imaging domain adaptive calibration method for cross-terminal facial recognition health detection, the method comprising: An imaging domain evidence vector is constructed based on the device profile, acquisition parameters, and raw imaging data of the facial recognition terminal. The reference color unmixing matrix is ​​extracted from the equipment file, combined with the imaging domain evidence vector, and a bounded residual is generated using a calibration supernetwork. An initial calibration matrix is ​​then generated based on the bounded residual. Using the initial calibration matrix, the raw imaging data is mapped to the standard imaging domain to obtain the standardized face temporal signal; Candidate pulse signals corresponding to multiple preset skin regions are extracted from the standardized face time-series signal. At the same time, noise signals of preset comparison regions are extracted. Self-calibration is performed iteratively based on the candidate pulse signals and noise signals to correct the initial calibration matrix and obtain the self-calibration matrix. The calibration discriminability is calculated based on the imaging domain evidence vector and the self-calibration matrix, and the detection result is output based on the relationship between the calibration discriminability and the discriminability threshold.

[0006] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided by this invention constructs an imaging domain evidence vector to comprehensively represent the device, environment, and acquisition status through an imaging calibration mechanism that separates device benchmarks from session bounded residuals. It utilizes multi-skin region candidate pulse synchronization constraints combined with contrast region noise signals to complete session self-calibration. This effectively suppresses false pulse interference caused by light flicker and automatic camera exposure, reducing the impact of hardware imaging differences between different facial recognition terminals on rPPG signals. Through calibration discriminability gating, it performs hierarchical and tiered processing, performing limited calibration or stopping output when quality is insufficient to avoid outputting false health values. Combined with counterfactual training and high-confidence session anchor point updates, it improves the consistency of cross-terminal measurement results, adapts to complex on-site acquisition environments of various devices such as mobile phones and self-service health terminals, and ensures the reliability of facial recognition health detection results.

[0007] In one alternative implementation, the training process for calibrating the supernetwork includes: Obtain the original face video sample, apply counterfactual imaging transformation to the original face video sample to obtain the counterfactual video sample, and the physiological timeline of the counterfactual video sample is consistent with that of the original face video sample; Based on the original face video samples and counterfactual video samples, a counterfactual consistency loss is constructed using calibrated photoplethysmography pulse waves, health indicator vectors, and standard imaging domain features. A total loss function is constructed based on counterfactual consistency loss, contact reference loss, session self-calibration loss, and calibration boundary loss. The parameters of the calibration supernetwork are then optimized based on the total loss function to obtain the final calibration supernetwork.

[0008] This invention provides an imaging domain adaptive calibration method for cross-device facial recognition health detection. It simulates imaging biases across various devices through a counterfactual imaging transformation that maintains the physiological time axis, constructing a multi-dimensional total loss optimization calibration supernetwork to isolate imaging interference caused by devices and lighting. It eliminates the need for simple feature alignment based on the dataset, ensuring consistency between pulse waveforms and health indicator outputs under different imaging conditions, and improving the model's cross-device generalization ability.

[0009] In one alternative implementation, the method further includes: High-trust sessions are selected based on calibration identifiability, and consistency checks are performed on the high-trust sessions. The consistency checks include: cross-terminal consistency checks and / or single-device repeated measurement consistency checks. When a high-trust session passes the consistency check and meets the calibration matrix constraints, the session residuals of the high-trust session are extracted. Based on session residuals, the device residual anchor points in the device file are iteratively updated according to a preset small step update rate.

[0010] The imaging domain adaptive calibration method for cross-terminal face scanning health detection provided by this invention selects only high-confidence sessions for updating device residual anchor points, filters invalid measurement data through consistency verification and matrix constraints, and updates device files in a small-step iterative manner to avoid contaminating file data with inferior sessions. This enables continuous iterative optimization of device files, constantly corrects inherent imaging deviations of the terminal, and improves the calibration accuracy and overall reliability of subsequent face scanning detection.

[0011] In one optional implementation, the raw imaging data is mapped to a standard imaging domain using an initial calibration matrix to obtain a standardized face temporal signal, including: Based on the camera response curve in the device file, inverse gamma or piecewise inverse response is performed on the raw imaging data, and the channel bias and exposure gain are compensated using the initial calibration matrix to obtain compensated imaging data. The compensated imaging data is mapped to the standard imaging domain to obtain a standardized face temporal signal.

[0012] The imaging domain adaptive calibration method for cross-terminal facial scanning health detection provided by this invention restores the camera imaging link through inverse response, and maps it to a unified standard imaging domain by combining channel bias and exposure gain compensation. This eliminates color differences caused by gamma characteristics and exposure parameters of different terminals, does not pursue visual appeal, and focuses on preserving the weak physiological fluctuation signals of the face, providing a standardized input basis for the subsequent extraction of reliable pulse signals.

[0013] In one optional implementation, self-calibration is performed iteratively based on candidate pulse signals and noise signals to correct the initial calibration matrix, resulting in a self-calibration matrix, including: Construct a session self-calibration loss function, including multi-skin region pulse synchronization loss, candidate pulse and noise signal coherence suppression loss, physiological frequency band out-of-band energy suppression loss, session residual amplitude constraint loss, and calibration parameter time smoothing loss. Based on the candidate pulse signal and noise signal, the bounded residual is iteratively optimized using the self-calibration loss function to obtain the self-calibration matrix.

[0014] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided by this invention constructs a multi-dimensional session self-calibration loss function, which not only ensures the synchronization of pulse signals in multiple skin regions, but also suppresses pseudo pulses caused by background noise. At the same time, it constrains the residual amplitude and parameter mutations, and optimizes the calibration parameters through multi-constraint joint iteration, reduces signal interference caused by light flicker and automatic exposure, avoids overfitting, and improves the accuracy of pulse signal extraction and the stability of calibration results.

[0015] In one optional implementation, a calibration discriminability is calculated based on the imaging domain evidence vector and the self-calibration matrix, and a stream is split according to the relationship between the calibration discriminability and the discriminability threshold, outputting the detection result, including: The calibration identifiability is calculated based on the reliability of device metadata, the effective redundancy of multiple skin regions, the observability of physiological frequency bands, the uncertainty of calibration parameters, and the burden of imaging quality. When the calibration discernibility is higher than the first threshold, perform a complete dynamic calibration and output the health test result. When the calibration discernibility is higher than the second threshold but not higher than the first threshold, a restricted calibration is performed, where the second threshold is lower than the first threshold. When the calibration discernibility falls below the second threshold, stop outputting numerical results and guide the user to re-collect data.

[0016] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided by this invention calculates calibration discriminability across multiple dimensions, implements hierarchical and triage processing based on thresholds, executes complete calibration output results under high confidence, enables restricted calibration under medium confidence, and stops output and prompts for resampling if the quality is too low. This avoids false health data generated by low-quality videos, prevents over-correction of the calibration model, and ensures the reliability of facial recognition health detection output results.

[0017] In one optional implementation, if the same user completes the authorization measurement through different facial recognition terminals within a preset time interval, a cross-terminal consistency index is calculated, and the source of measurement deviation is located based on the cross-terminal consistency index.

[0018] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided by this invention calculates cross-terminal consistency indicators for authorized measurements of the same user across multiple terminals within a short period of time. It objectively compares the measurement outputs of different terminals, quickly locates the source of deviations from device file mismatch, abnormal ambient light, or changes in the user's physiological state, avoids simply averaging the measurement results, facilitates the investigation of terminal problems, and further ensures that the facial recognition health detection results are auditable and traceable.

[0019] Secondly, the present invention provides an imaging domain adaptive calibration device for cross-terminal facial recognition health detection, the device comprising: The evidence vector construction module is used to construct imaging domain evidence vectors based on the device profile, acquisition parameters, and raw imaging data of the face scanning terminal. The bounded residual calculation module is used to extract the reference color unmixing matrix from the equipment file, combine it with the imaging domain evidence vector, use the calibration supernetwork to generate the bounded residual, and generate the initial calibration matrix based on the bounded residual. The standard mapping module is used to map the raw imaging data to the standard imaging domain using the initial calibration matrix to obtain the standardized face temporal signal; The image self-calibration module is used to extract candidate pulse signals corresponding to multiple preset skin regions from the standardized face time-series signal, and at the same time extract noise signals from the preset comparison region. Iterative self-calibration is performed based on the candidate pulse signals and noise signals to correct the initial calibration matrix and obtain the self-calibration matrix. The detection result output module is used to calculate the calibration discriminability based on the imaging domain evidence vector and the self-calibration matrix, and to perform stream splitting based on the relationship between the calibration discriminability and the discriminability threshold, and output the detection result.

[0020] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the overall process of an imaging domain adaptive calibration method for cross-terminal facial recognition health detection according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first process of an imaging domain adaptive calibration method for cross-terminal facial scanning health detection according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the second process of the imaging domain adaptive calibration method for cross-terminal facial scanning health detection according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the session self-calibration, gating and anchor point update process in the imaging domain adaptive calibration method for cross-terminal face scanning health detection according to an embodiment of the present invention. Figure 5 This is a structural block diagram of an imaging domain adaptive calibration device for cross-terminal facial recognition health detection according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0026] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0027] This invention provides an imaging domain adaptive calibration method for cross-terminal facial recognition health detection. It employs an imaging calibration mechanism that separates device references from session bounded residuals for adaptive calibration, ensuring that health detection results are unaffected by terminal devices or external lighting conditions. Figure 1 The diagram illustrates the overall workflow of the imaging domain adaptive calibration method. First, the device and acquisition parameters are read. Then, ambient light, skin color, and face region imaging features are estimated from the video. A constrained calibration supernetwork generates color demixing, bias, and exposure compensation parameters for the current session. Subsequently, the low-dimensional residuals are optimized by leveraging the relationship that multiple skin regions share a pulse while the background region does not. Finally, full calibration, constrained calibration, or reacquisition is selected based on discernibility.

[0028] According to an embodiment of the present invention, an imaging domain adaptive calibration method for cross-terminal facial scanning health detection is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0029] This embodiment provides an imaging domain adaptive calibration method for cross-terminal facial recognition health detection, which can be used in the aforementioned computer system. Figure 2 This is a flowchart of an imaging domain adaptive calibration method for cross-terminal facial recognition health detection according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process includes the following steps: Step S201: Construct an imaging domain evidence vector based on the device profile, acquisition parameters, and raw imaging data of the face scanning terminal.

[0030] Specifically, the imaging domain evidence vector is constructed according to time windows, rather than using only a device number or a single frame's average brightness to represent the domain. The evidence vector for each time window is:

[0031] in, The device profile includes the terminal model, camera module, sensor or lens identifier, firmware and image signal processing version, nominal spectral response, gamma and encoding configuration; This represents the acquisition parameters for the t-th time window, including exposure time, ISO or analog gain, white balance gain, frame rate, resolution, focus status, compression bitrate, and timestamp jitter. The ambient light description includes luminance, RGB chromaticity, spatial non-uniformity, 50 Hz or 60 Hz flicker and its aliasing components, which are jointly estimated by the ambient light sensor, the background area and the highlight area. The user's skin color and static skin optical description are obtained from the chroma, melanin approximation index and saturation statistics of non-overexposed, low-reflectivity skin areas, without saving the identity template; It represents the imaging characteristics of areas such as the forehead and left and right cheeks, including mean and variance, channel correlation, local noise, overexposure ratio, specular reflection ratio, motion and effective skin pixel ratio; The available mask and trust level of each piece of metadata are obtained by cross-verification of the values ​​reported by the client and the values ​​derived from the video.

[0032] Step S202: Extract the reference color unmixing matrix from the device file, combine it with the imaging domain evidence vector, use the calibration supernetwork to generate bounded residuals, and generate the initial calibration matrix based on the bounded residuals.

[0033] Specifically, the device profile includes the baseline color unmixing matrix A0(d) and inverse response curve stored by the registered cameras of the scanning terminal. For unregistered devices, the nearest device profile is selected based on the camera module, resolution, exposure behavior, and visual domain characteristics, and the uncertainty is passed to the gating controller. Exposure, lighting, skin tone, and local area differences in the current session do not directly rewrite the device baseline; instead, bounded residuals are generated by the calibration supernetwork Hθ, allowing only limited adjustments within a small range of the device baseline to generate temporary correction parameters, preventing the algorithm from arbitrarily changing colors and creating false heartbeats. The inherent bias of the device and the temporary bias of the current environment are separated, and the initial calibration parameters are output.

[0034]

[0035] Where, ΔA t For color channel demixing residuals; b t For black level, encoding, and environment bias vectors; g t ρ represents the exposure gain compensation; ρ limits the residual amplitude, ensuring the current session can only adjust near the device reference, preventing the network from generating periodic signals using arbitrary color changes. Calibration Matrix A t It also needs to meet preset constraints on the number of conditions, sign, and time smoothness. When metadata is missing, Hθ uses a visual encoding branch to fill in the gaps, but... It will reduce subsequent recognizability.

[0036] Step S203: Using the initial calibration matrix, the original imaging data is mapped to the standard imaging domain to obtain the standardized face temporal signal.

[0037] Specifically, regardless of the terminal device used to collect the facial images, the original imaging data is uniformly converted into a unified standard image space through inverse camera response, offset cancellation, gain compensation, and matrix demixing using an initial calibration matrix, resulting in a standardized facial temporal signal.

[0038] Step S204: Extract candidate pulse signals corresponding to multiple preset skin regions from the standardized face time sequence signal, and extract noise signals from preset comparison regions. Then, perform self-calibration iteratively based on the candidate pulse signals and noise signals to correct the initial calibration matrix and obtain the self-calibration matrix.

[0039] Specifically, the forehead and both cheeks of the same user should share a common basic pulse frequency and phase trend within the allowable small blood flow propagation difference. However, ambient light flicker, automatic exposure, and encoding changes may also generate a global periodicity, leading to exposure interference. To prevent mistaking the illumination periodicity for a pulse, non-physiological noise trajectories are extracted from the background, hair, collar, or high specular reflection areas. During self-calibration optimization, the pulse waveforms extracted from the forehead and both cheeks are required to be as synchronized as possible; the facial pulse signal is forced to be consistent with the periodic signal from the background light flickering, thereby filtering out false heartbeats caused by the light; the signal is limited to appearing only within a reasonable heart rate frequency range; and the correction parameters are limited to a small amplitude, with no jumps between consecutive frames.

[0040] Step S205: Calculate the calibration discriminability based on the imaging domain evidence vector and the self-calibration matrix, and perform stream splitting based on the relationship between the calibration discriminability and the discriminability threshold, and output the detection result.

[0041] Specifically, the calibration recognizable value is calculated by considering factors such as the completeness of the integrated equipment parameters, the amount of usable facial skin area, the clarity of the heartbeat signal, the uncertainty, and the degree of overexposure, motion, or lighting interference. Based on the relationship between the calibration recognizable value and the recognizable value threshold, three cases are identified and output.

[0042] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided in this embodiment constructs an imaging domain evidence vector that comprehensively represents the device, environment, and acquisition status by separating the device benchmark and the session bounded residual through an imaging calibration mechanism. It utilizes multi-skin region candidate pulse synchronization constraints combined with contrast region noise signals to complete session self-calibration. This effectively suppresses false pulse interference caused by light flicker and automatic camera exposure, reducing the impact of hardware imaging differences between different facial recognition terminals on rPPG signals. Through calibration discriminability gating, it performs hierarchical and tiered processing, performing limited calibration or stopping output when quality is insufficient to avoid outputting false health values. Combined with counterfactual training and high-confidence session anchor point updates, it improves the consistency of cross-terminal measurement results, adapts to complex on-site acquisition environments of various devices such as mobile phones and self-service health terminals, and ensures the reliability of facial recognition health detection results.

[0043] In some alternative implementations, the training process for calibrating the supernetwork includes: Step a1: Obtain the original face video sample, apply counterfactual imaging transformation to the original face video sample to obtain a counterfactual video sample, and the physiological timeline of the counterfactual video sample is consistent with that of the original face video sample.

[0044] Specifically, original facial video samples are acquired, which contain real physiological information (heartbeat, blood flow changes) and can be used in conjunction with contact rPPG / electrocardiogram (ECG) as a ground truth reference. Counterfactual imaging transformation simulates the transformation operations of the imaging chain, modifying only imaging-level factors while completely preserving the physiological timeline: that is, blood volume fluctuations on the face and the timing of heartbeats remain unchanged. It only simulates the changes in image quality caused by different cameras' spectral responses, gamma curves, exposure times, sensitivity gains, white balance, ambient light chromaticity, and video compression to generate counterfactual video samples.

[0045]

[0046] in, Indicates counterfactual imaging transformation, This represents a counterfactual video sample.

[0047] Step a2: Based on the original face video samples and counterfactual video samples, construct the counterfactual consistency loss using the calibrated photoplethysmography pulse wave, health indicator vector, and standard imaging domain features.

[0048] Specifically, the original face video samples and counterfactual video samples are input into the calibration link of this invention, and imaging domain calibration is performed separately. Each yields: the calibration output rPPG photoplethysmography pulse wave; a health indicator vector (heart rate, heart rate variability, etc.) inferred from the pulse waveform; and mapped standard imaging domain intermediate features. The original face video samples and counterfactual video samples should recover the same standard pulse and health indicators.

[0049] The counterfactual consistency loss is constrained in three layers: the rPPG pulse waveforms of the original sample and the counterfactual sample are as correlated as possible; the difference between the health index vectors output by the two is as small as possible; and the features of the two in the standard imaging domain are as aligned as possible.

[0050]

[0051] in, T represents the counterfactual consistency loss, where p and p′ are the two standard rPPG waveforms, h and h′ are the health index vectors, and z and z′ are the standard imaging domain features. c → c Only the cause of the imaging is changed, not the timing of the heartbeat.

[0052] If the device interference introduced by the imaging transformation is not eliminated by the calibration network, the pulse waveform, health indicators, and standard domain features will deviate, and the loss function value will increase, thereby forcing the calibration network to eliminate the offset caused by the imaging.

[0053] Step a3: Construct a total loss function based on counterfactual consistency loss, contact reference loss, session self-calibration loss, and calibration boundary loss, and optimize the parameters of the calibration supernetwork based on the total loss function to obtain the final calibration supernetwork.

[0054] Specifically, during training, the following loss methods are used simultaneously: contact PPG or ECG reference loss (Lref), self-calibration loss (Lself), counterfactual consistency loss (Lcf), and calibration boundary loss (Lbound). The counterfactual consistency loss is responsible for cross-terminal generalization, ensuring consistent output results under different imaging conditions. The contact reference loss compares the calibrated rPPG waveform with the synchronized contact PPG / ECG ground truth, ensuring the output closely matches real physiological signals. The session self-calibration loss reproduces the self-calibration constraints during the inference phase (multi-region synchronization, background noise suppression, residual smoothing, etc.), allowing the network during training to adapt to the self-calibration logic during inference, narrowing the gap between training and inference. The calibration boundary loss imposes constraints on the residuals and calibration matrix of the network output, limiting the matrix condition number, residual amplitude, and response monotonicity, preventing unreasonable color changes in the network output, and avoiding the generation of false pulse signals.

[0055] The expression for the total loss function is:

[0056] Where, Lbound is paired with A t Constraints are imposed on the condition number, residual magnitude, response monotonicity, and time variation. This training method enables the calibration network to learn to eliminate imaging causes, rather than simply performing uninterpreted alignment with the feature distribution of the training dataset.

[0057] This embodiment provides an imaging domain adaptive calibration method for cross-device facial recognition health detection. It simulates imaging biases across various devices using a counterfactual imaging transformation that maintains the physiological time axis, constructing a multi-dimensional total loss optimization calibration supernetwork to isolate imaging interference caused by devices and lighting. It eliminates the need for simple alignment of dataset features, ensuring consistency between pulse waveforms and health indicator outputs under different imaging conditions, thus improving the model's cross-device generalization ability.

[0058] In some alternative implementations, the method further includes: Step b1: Filter high-trust sessions based on calibration identifiability and perform consistency verification on the high-trust sessions. The consistency verification includes: cross-terminal consistency verification and / or single-device repeated measurement consistency verification.

[0059] Specifically, each camera module, firmware, and terminal version maintains the device calibration anchor point. First, the calibration discernibility calculated from a single measurement session is read. Only sessions with a calibration discernibility higher than a preset high threshold are marked as high-trust sessions; sessions with insufficient discernibility are directly excluded and do not participate in the device anchor point update.

[0060] For the highly trusted sessions selected through screening, either one of two types of consistency checks or both types of checks are performed simultaneously: (1) Cross-terminal consistency verification: Within the allowed time window, the same user completes authorized measurements using the terminal under test and a reference trusted terminal respectively. Calculate the cross-terminal consistency index, compare the rPPG waveforms and health indicators output by the two sets of devices, and determine whether the two measurement results are consistent. This is used to verify whether the output of the terminal under test is aligned with the reference device.

[0061] (2) Single-device repeated measurement consistency verification: The same user completes multiple independent measurements on the same terminal under test within a short period of time. The pulse waveforms output from the multiple measurements are compared with health indicators to determine whether the results are stable and consistent. This is suitable for deployment scenarios without a reference terminal. Either type of verification is sufficient; both are not required to be used simultaneously.

[0062] Step b2: When the high-trust session passes the consistency check and meets the calibration matrix constraints, extract the session residuals of the high-trust session.

[0063] Specifically, the current session residual will only participate in the anchor point update if the calibration identifiability reaches a high threshold, cross-terminal consistency or same-device repeatability passes, and the calibration matrix constraints are satisfied.

[0064] Among them, Ā d - and Ā d + These represent the device residual anchor points before and after the update; ζ represents the small step update rate; Iupd represents the update condition indicator value; mean t (ΔA t ) represents the mean residual value of the effective time window.

[0065] If consistency fails or matrix constraints are not met, the session is abandoned and the session will not participate in anchor point updates.

[0066] Step b3: Based on the session residuals, iteratively update the device residual anchor points in the device file according to the preset small step update rate.

[0067] Specifically, the device file stores the residual anchor points of the old device corresponding to the camera module. A weighted iterative update is performed using a preset small step update rate of less than 1: New anchor point = Old anchor point × (1 - Effective update weight) + Average residual of the current session × Effective update weight. The effective update weight is determined by the small step update rate, the recognizability score, and the update indicator flag.

[0068] The system employs weighted smooth updates, rather than directly overwriting old anchor points with the residuals from the current session, limiting the extent of changes to the profile in a single session. The system only stores de-identified device statistical parameters and quality summaries, without saving original facial videos, thus protecting user privacy. It continuously monitors the cumulative offset of device anchor points, and when the cumulative offset exceeds a threshold, a new version of the device profile is generated, supporting playback and rollback of historical versions. The updated device residual anchor points will serve as the device benchmark for subsequent measurements of this terminal, used for generating calibration matrices in subsequent sessions, forming a closed-loop optimization.

[0069] The imaging domain adaptive calibration method for cross-terminal face scanning health detection provided in this embodiment only selects high-confidence sessions for updating device residual anchor points, filters invalid measurement data through consistency verification and matrix constraints, and updates device files in a small-step iterative manner to avoid contaminating file data with inferior sessions. This enables continuous iterative optimization of device files, constantly corrects inherent imaging deviations of the terminal, and improves the calibration accuracy and overall reliability of subsequent face scanning detection.

[0070] This embodiment provides an imaging domain adaptive calibration method for cross-terminal facial recognition health detection, which can be used in the aforementioned computer system. Figure 3 This is a flowchart of an imaging domain adaptive calibration method for cross-terminal facial recognition health detection according to an embodiment of the present invention, as shown below. Figure 3 As shown, the process includes the following steps: Step S301: Construct an imaging domain evidence vector based on the device profile, acquisition parameters, and raw imaging data of the facial recognition terminal. For details, please refer to [link to relevant documentation]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0071] Step S302: Extract the reference color unmixing matrix from the device file, combine it with the imaging domain evidence vector, use the calibration supernetwork to generate bounded residuals, and generate the initial calibration matrix based on the bounded residuals. For details, please refer to [link to details]. Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0072] Step S303: Using the initial calibration matrix, the original imaging data is mapped to the standard imaging domain to obtain the standardized face temporal signal.

[0073] Specifically, step S303 includes: Step S3031: Based on the camera response curve in the device file, perform inverse gamma or piecewise inverse response on the original imaging data, and use the initial calibration matrix to compensate for channel bias and exposure gain to obtain compensated imaging data.

[0074] Specifically, the device file pre-stores the camera response curve of the corresponding camera. This curve describes the mapping relationship between the sensor's raw image format (RAW) signal and the camera's image signal processor (ISP) after processing, resulting in an output RGB image.

[0075] Inverse gamma / piecewise inverse response processing performs an inverse transformation on the raw imaging data output by the terminal. Ordinary mobile phones output gamma-compressed coded images; if the raw RAW data cannot be obtained, a piecewise monotonic approximation curve of the same model is retrieved for inverse response calculation to counteract the nonlinear color compression introduced by the camera's ISP and restore the linear relationship of the signal.

[0076] Channel offset compensation uses the offset terms in the initial calibration matrix to eliminate fixed offsets in each color channel caused by black level, video encoding offset, and slow-changing ambient light.

[0077] Exposure gain compensation is used to compensate for the overall brightness scaling differences caused by exposure time and ISO analog gain in different sessions, and a very small anti-zero constant is added to avoid numerical anomalies.

[0078] After inverse response, bias compensation, and exposure gain compensation, compensated imaging data is output. This data has been stripped of interference from the camera's own nonlinearity, exposure, and bias, but cross-device color demixing and alignment have not yet been completed.

[0079] Step S3032: Map the compensated imaging data to the standard imaging domain to obtain the standardized face temporal signal.

[0080] Specifically, the compensated imaging data is input into the initial calibration matrix to perform matrix operations, complete the color demixing between channels, and map it to the preset standard imaging domain.

[0081] Standardized facial timing signals do not aim for visual aesthetics or optimize image quality; they primarily preserve the subtle relative temporal fluctuations in RGB values ​​caused by changes in blood flow in facial skin.

[0082] The standardized facial temporal signal obtained after mapping is comparable to the same physiological facial signal collected by different models of terminals within the standard imaging domain, and can be directly used to extract candidate pulse signals from multiple skin regions.

[0083] For example, the RGB observation x of the k-th face region within time window t. ktThe system first performs inverse gamma or segmented inverse response L based on the device file. d Then eliminate the bias and exposure gain, and finally pass A t Mapped to the standard imaging domain.

[0084]

[0085] L d Obtained from the camera response curve in the device profile; when the video has been encoded and linear RAW values ​​cannot be recovered, L d Use a piecewise monotonically similar approximation with the same model of equipment. ε is a small positive number to prevent division by zero. Output y. kt Preserving relative physiological changes, not prioritizing visual appeal. Standard rPPG extractor B for y k Time series generation of candidate pulses p k =B(y k ).

[0086] The imaging domain adaptive calibration method for cross-terminal facial scanning health detection provided in this embodiment restores the camera imaging link through inverse response and maps it to a unified standard imaging domain by combining channel bias and exposure gain compensation. This eliminates color differences caused by gamma characteristics and exposure parameters of different terminals. It does not pursue visual appeal but focuses on preserving the weak physiological fluctuation signals of the face, providing a standardized input basis for the subsequent extraction of reliable pulse signals.

[0087] Step S304: Extract candidate pulse signals corresponding to multiple preset skin regions from the standardized face time sequence signal, and extract noise signals from preset comparison regions. Then, perform self-calibration iteratively based on the candidate pulse signals and noise signals to correct the initial calibration matrix and obtain the self-calibration matrix.

[0088] Specifically, step S304 includes: Step S3041: Construct the session self-calibration loss function, including multi-skin region pulse synchronization loss, candidate pulse and noise signal coherence suppression loss, physiological frequency band out-of-band energy suppression loss, session residual amplitude constraint loss, and calibration parameter time smoothing loss.

[0089] Specifically, the session self-calibration loss function is:

[0090] in, To improve the pulse synchronicity of different skin areas, the actual physiological pulse rhythms of effective skin areas such as the forehead and left and right cheeks should be similar. This loss penalizes the lack of correlation between candidate pulse signals in different skin areas, and promotes the synchronization of output pulse waveforms in each skin area. To suppress the coherence between candidate pulses and non-physiological lighting trajectories, noise signals are taken from non-skin contrast areas such as background, hair, and collar, representing non-physiological interferences such as light flicker and exposure fluctuations. This loss suppresses the frequency domain coherence between candidate pulses and noise signals, avoiding misidentification of environmental periodic interference as pulses. Used to suppress energy outside the physiological frequency band. The human heart rate is within a limited physiological frequency band. This loss constrains candidate pulse signals, reduces signal energy outside the physiological frequency band, and suppresses non-physiological cycle components. Used to limit the amplitude and size of session residuals, prevent the network from generating excessive color changes, avoid the algorithm from artificially faking periodic physiological signals, and ensure that calibration is only made with limited adjustments near the device reference. This is used to limit abrupt changes in calibration parameters between adjacent windows, constraining the calibration matrix from drastic jumps between adjacent time windows, ensuring stable calibration results within continuous video windows, and suppressing waveform disturbances caused by parameter abrupt changes. Region weight w ij It is derived from the effective skin ratio, motion, overexposure, and signal-to-noise ratio. Only ΔA is updated during optimization. t b t and g t The low-dimensional residuals are not updated to maintain the health model, thus avoiding overfitting of short videos.

[0091] Step S3042: Based on the candidate pulse signal and noise signal, the bounded residual is iteratively optimized using the self-calibration loss function to obtain the self-calibration matrix.

[0092] Specifically, candidate pulse signals from each skin region and noise signals from the contrast region are input into the session self-calibration loss function to iteratively update the bounded residuals of the session. The optimization process only updates the low-dimensional session residuals, fixing the health indicator model parameters to prevent overfitting of short video samples. Boundary constraints such as matrix condition number and residual magnitude are continuously applied during iteration. Iteration stops when the loss converges or the number of iterations reaches a preset iteration limit. The optimized bounded residuals are superimposed onto the device baseline matrix to obtain the corrected self-calibration matrix, which is used for subsequent calculations of calibration discriminability and health indicator inference. If the constraints cannot be met even after reaching the iteration limit, the process enters a verification branch, and no valid detection results are output.

[0093] like Figure 4 The diagram illustrates the process of session self-calibration, gating, and anchor point update. First, the forehead, left and right cheek skin regions, and background reference regions are selected to generate initial dynamic calibration parameters. Candidate pulses and illumination noise are extracted from each region, and the self-calibration objective function is calculated. The low-dimensional session residuals are iteratively updated; if convergence is not achieved, iteration continues until the upper limit is reached. Next, calibration discriminability is calculated, and branching is performed based on the interval where the calibration discriminability falls: full dynamic calibration is performed in the high interval; restricted diagonal calibration is performed in the middle interval; and the output stops and a prompt for re-acquisition is displayed in the low interval.

[0094] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided in this embodiment constructs a multi-dimensional session self-calibration loss function, which not only ensures the synchronization of pulse signals in multiple skin regions, but also suppresses pseudo pulses caused by background noise. At the same time, it constrains residual amplitude and parameter mutations, and optimizes calibration parameters through multi-constraint joint iteration, reduces signal interference caused by light flicker and automatic exposure, avoids overfitting, and improves the accuracy of pulse signal extraction and the stability of calibration results.

[0095] Step S305: Calculate the calibration discriminability based on the imaging domain evidence vector and the self-calibration matrix, and perform stream splitting based on the relationship between the calibration discriminability and the discriminability threshold, and output the detection result.

[0096] Specifically, step S305 includes: Step S3051: Calculate calibration identifiability based on device metadata reliability, effective redundancy of multiple skin regions, physiological frequency band observability, calibration parameter uncertainty, and imaging quality burden.

[0097] Specifically, dynamic calibration is only reliable when metadata, regional redundancy, and physiological band information are sufficient. The formula for calculating calibration discriminability Qcal is:

[0098] Wherein, M represents the integrity and reliability of device and exposure metadata, characterizing whether the reported metadata such as device files, exposure, white balance, and frame rate are complete and effective, and also verifying whether the reported parameters are consistent with the video back-inference information. Missing or inconsistent metadata will reduce the score of this item; R represents the effective redundancy of multiple skin regions, evaluating the effective pixel ratio, tracking stability, and light complementarity of skin regions such as the forehead and left and right cheeks; the more usable skin regions and the better the quality, the higher the redundancy; P represents the observability of the physiological frequency band after removing background noise. After removing background noise interference, it evaluates the strength of pulse signal energy in the physiological frequency band and measures whether the real physiological signal can be effectively observed; U represents the uncertainty obtained from the calibrator integration model or parameter perturbation, reflecting the perturbation sensitivity of the calibration network parameters. When the device file is not matched accurately or the metadata is missing, the uncertainty will increase; O represents the comprehensive burden of overexposure, underexposure, motion, occlusion, and inseparable flicker, comprehensively statistically analyzing the negative impact of adverse interferences such as overexposure, underexposure, face motion, face occlusion, and power frequency light flicker; η1 to η5 are the non-negative coefficients of the validation set calibration; σ is the Sigmoid function.

[0099] Step S3052: When the calibration identifiability is higher than the first threshold, perform a complete dynamic calibration and output the health test result.

[0100] Specifically, when the identifiability exceeds the first threshold of the higher level, it indicates that the metadata is complete, the skin area is sufficient, the physiological signals are clear, and the interference is controllable. Full dynamic calibration is allowed, enabling the use of all color unmixing matrices to complete imaging domain calibration, extracting standard rPPG waveforms, running a health indicator model to output complete health detection results such as heart rate, and simultaneously recording all calibration status information for this session for subsequent auditing.

[0101] Step S3053: When the calibration discernibility is higher than the second threshold and not higher than the first threshold, a restricted calibration is performed, where the second threshold is lower than the first threshold.

[0102] Specifically, a discernibility level in the medium range indicates some defects, such as missing metadata or slight lighting interference, but still some usable signals. In this case, a restricted calibration is performed, allowing only channel diagonal gain compensation and bias compensation, prohibiting cross-channel free color demixing, reducing calibration freedom, avoiding erroneous transformations caused by overfitting, and outputting health detection results under restricted conditions.

[0103] Step S3054: When the calibration discernibility is lower than the second threshold, stop outputting numerical results and guide the user to re-collect data.

[0104] Specifically, if the identifiability is below a lower second threshold, it indicates insufficient effective skin area, a low signal-to-noise ratio, and the inability to separate interference such as flicker and occlusion, thus lacking reliable measurement conditions. The system does not directly output quantitative health values ​​such as heart rate. Based on the identified main problems, it guides the user to adjust the lighting, shooting distance, and face angle, and initiates a re-acquisition. At the same time, it saves video quality summaries and device parameters for subsequent troubleshooting.

[0105] If the self-calibration iteration reaches the upper limit but still fails to meet the constraints, even if the identifiability is not lower than the second threshold, the process will proceed to the review process, retain the original session information, and conduct manual or contact-based device retesting.

[0106] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided in this embodiment calculates calibration discriminability across multiple dimensions, implements hierarchical and triage processing based on thresholds, executes complete calibration output results under high confidence, enables restricted calibration under medium confidence, and stops output and prompts for resampling if the quality is too low. This avoids false health data generated by low-quality videos, prevents over-correction of the calibration model, and ensures the reliability of facial recognition health detection output results.

[0107] In some optional implementations, if the same user completes the authorization measurement through different facial recognition terminals within a preset time interval, a cross-terminal consistency index is calculated, and the source of measurement deviation is located based on the cross-terminal consistency index.

[0108] Specifically, after gating, the system will divide the p of each valid areak The standard rPPG waveform p is obtained by quality weighting, and h is generated by a fixed version of the health indicator model. The health indicator model does not accept terminal numbers to prevent it from taking shortcuts by exploiting device distribution. If the same user voluntarily completes consecutive measurements on two terminals within a limited time, the system calculates cross-terminal consistency:

[0109] Where A and B represent two terminals; Δt(A,B) is the time interval between the two measurements; τ p The physiological timescale for comparison is defined by κ, which is the decay coefficient for the difference in health indicators. When C(A,B) is low, the system first compares the evidence from the imaging domains at both ends, the calibration mode, and the residuals to determine whether the problem is due to a mismatch in the data at a certain terminal, abnormal ambient light, or a change in the user's physiological state, without directly averaging the two health results.

[0110] The mobile terminal prioritizes reading actual exposure, gain, white balance, frame time, and encoding parameters via Camera2, AVFoundation, or similar interfaces; the self-service terminal reads these parameters from the device registry, driver interface, and ambient light sensor. The client sends parameter summaries and time alignment information along with the video, and the server performs device profile lookup, dynamic calibration, rPPG inference, and auditing. If the terminal supports locking exposure and white balance, the system selects parameters during the pre-capture phase that prevent skin overexposure and ensure observable pulse bands, maintaining these parameters during the actual capture phase; if locking is not supported, parameter changes are recorded window by window and managed by g. t and b t Dynamic compensation.

[0111] The imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided in this embodiment calculates cross-terminal consistency indicators for authorized measurements of the same user across multiple terminals within a short period of time. It objectively compares the measurement outputs of different terminals, quickly locates the source of deviations from device file mismatch, abnormal ambient light, or changes in the user's physiological state, avoids simply averaging the measurement results, facilitates the investigation of terminal problems, and further ensures that the facial recognition health detection results are auditable and traceable.

[0112] In one specific embodiment, the complete process is illustrated by taking the example of the same user continuously completing facial recognition detection on a mobile phone and a self-service health terminal. The example values ​​are only used to illustrate the calculation and traffic splitting logic and do not constitute experimental results data.

[0113] Table 1 Example Project Table

[0114] Step 1: The user first completes a 30-second face scan on mobile terminal A, and then completes a second scan three minutes later on self-service terminal B. Both sessions are authorized by the user for this cross-terminal consistency comparison.

[0115] Step 2: The mobile device reads the camera module, exposure, ISO, white balance, frame time, and encoding information; the self-service terminal reads the camera and driver version from the device registry, but ISO and white balance are missing. The imaging domain evidence builder generates ξ for each two-second window. t and in m t Mark missing fields in the middle.

[0116] Step 3: Ambient light analysis revealed a common periodic component in the background and skin highlight areas of terminal B. This component fell near the candidate physiological frequency band after video sampling. If only the maximum peak of the power spectrum is selected, terminal B will obtain a spurious peak of 91 times per minute.

[0117] Step 4: The system reads A0(d) for terminal A. Terminal B does not yet have its own file. The system selects the nearest device file based on the camera module, frame rate, exposure behavior, and visual domain characteristics, while increasing the calibration uncertainty U. Hθ outputs ΔA for each window. t b t and g t .

[0118] Step 5: After completing the initial standard domain mapping, the system extracts p from the forehead and left and right cheeks. k The n component was extracted from the background and highlight areas. The three regions of terminal A were synchronized around 78 times / minute and had low coherence with n; the 91 times / minute component of terminal B had high coherence with n, while the forehead and right cheek were synchronized around 79 times / minute.

[0119] Step 6: The session self-calibrator minimizes Lself. Terminal B's background coherence term suppresses spurious peaks of 91 Hz, while the regional synchronization term retains physiological peaks of 79 Hz. The left cheek receives a lower w due to shadows and low effective pixels. ij The calibration is not dominant. The calibration matrix residuals are constrained by ρ and Lbound.

[0120] Step 7: The gater calculates Qcal=0.91 for terminal A, which meets the full dynamic calibration conditions; terminal B has Qcal=0.73 due to missing metadata and flicker burden, which is between the two thresholds. The system only allows restricted diagonal calibration and exposure compensation, and does not allow free cross-channel demixing.

[0121] Step 8: The standard rPPG extractor outputs a peak of 78 times / minute for terminal A and 79 times / minute for terminal B, generating their respective health indicator vectors h(A) and h(B). Instead of averaging the two raw results, the system calculates C(A,B) using the gated p(A), p(B), h(A), and h(B).

[0122] Step 9: If C(A,B) reaches the institutional threshold, the health results of the two terminals are saved as consistent outputs within the same time period. Since terminal B only performs limited calibration, its session does not participate in the complete color matrix anchor point update, but it can be used as a device file missing statistics.

[0123] Step 10: When terminal B subsequently acquires complete driving parameters, Qcal reaches a high threshold, and cross-terminal consistency is passed multiple times, Iupd is set to one, and the system updates the camera and firmware version Ā in small steps ζ. d And generate new candidate versions of the archive.

[0124] Step 11: If the LED flickering of terminal B is inseparable from the physiological frequency band and the Qcal is lower than the low threshold, the system stops the numerical output, provides reasoned guidance such as "adjust to the continuous light source area or re-acquire", and retains the device parameters and quality summary for terminal maintenance.

[0125] This embodiment also provides an imaging domain adaptive calibration device for cross-terminal facial recognition health detection. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0126] This embodiment provides an imaging domain adaptive calibration device for cross-terminal facial recognition health detection, such as... Figure 5 As shown, it includes: The evidence vector construction module 501 is used to construct imaging domain evidence vectors based on the device profile, acquisition parameters and raw imaging data of the face scanning terminal.

[0127] The bounded residual calculation module 502 is used to extract the reference color unmixing matrix from the device file, combine it with the imaging domain evidence vector, use the calibration supernetwork to generate the bounded residual, and generate the initial calibration matrix based on the bounded residual.

[0128] The standard mapping module 503 is used to map the raw imaging data to the standard imaging domain using the initial calibration matrix to obtain the standardized face temporal signal.

[0129] The image self-calibration module 504 is used to extract candidate pulse signals corresponding to multiple preset skin regions from the standardized face time-series signal, and simultaneously extract noise signals from preset comparison regions. Based on the candidate pulse signals and noise signals, iterative self-calibration is performed to correct the initial calibration matrix and obtain the self-calibration matrix.

[0130] The detection result output module 505 is used to calculate the calibration discriminability based on the imaging domain evidence vector and the self-calibration matrix, and to perform stream splitting based on the relationship between the calibration discriminability and the discriminability threshold, and output the detection result.

[0131] The imaging domain adaptive calibration device for cross-terminal facial recognition health detection provided in this embodiment of the invention can execute the imaging domain adaptive calibration method for cross-terminal facial recognition health detection provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0132] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0133] The following is a detailed reference. Figure 6 This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0134] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0135] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the imaging domain adaptive calibration method for cross-terminal facial recognition health detection according to embodiments of the present invention.

[0136] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0137] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the imaging domain adaptive calibration method for cross-terminal facial scanning health detection shown in the above embodiments is implemented.

[0138] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. An imaging domain adaptive calibration method for cross-terminal facial recognition health detection, characterized in that, The method includes: An imaging domain evidence vector is constructed based on the device profile, acquisition parameters, and raw imaging data of the facial recognition terminal. The reference color unmixing matrix is ​​extracted from the device file, combined with the imaging domain evidence vector, and a bounded residual is generated using a calibration supernetwork. An initial calibration matrix is ​​then generated based on the bounded residual. Using the initial calibration matrix, the raw imaging data is mapped to the standard imaging domain to obtain a standardized face temporal signal; Candidate pulse signals corresponding to multiple preset skin regions are extracted from the standardized face time-series signal, and noise signals of preset comparison regions are extracted. Self-calibration is then performed iteratively based on the candidate pulse signals and the noise signals to correct the initial calibration matrix and obtain the self-calibration matrix. The calibration discriminability is calculated based on the imaging domain evidence vector and the self-calibration matrix, and the detection result is output based on the relationship between the calibration discriminability and the discriminability threshold.

2. The method according to claim 1, characterized in that, The training process of the calibration supernetwork includes: Obtain an original face video sample, apply a counterfactual imaging transformation to the original face video sample to obtain a counterfactual video sample, wherein the physiological timeline of the counterfactual video sample is consistent with that of the original face video sample; Based on the original face video samples and counterfactual video samples, a counterfactual consistency loss is constructed using calibrated photoplethysmography pulse waves, health indicator vectors, and standard imaging domain features. A total loss function is constructed based on the counterfactual consistency loss, contact reference loss, session self-calibration loss, and calibration boundary loss. The parameters of the calibration supernetwork are then optimized based on the total loss function to obtain the final calibration supernetwork.

3. The method according to claim 1 or 2, characterized in that, The method further includes: High-trust sessions are selected based on calibration identifiability, and consistency checks are performed on the high-trust sessions. The consistency checks include: cross-terminal consistency checks and / or single-device repeated measurement consistency checks. When the high-confidence session passes the consistency check and meets the calibration matrix constraint conditions, the session residual of the high-confidence session is extracted. Based on the session residuals, the device residual anchor points in the device file are iteratively updated according to a preset small step update rate.

4. The method according to claim 1, characterized in that, Using the initial calibration matrix, the raw imaging data is mapped to the standard imaging domain to obtain a standardized face temporal signal, including: Based on the camera response curve in the device file, inverse gamma or piecewise inverse response is performed on the raw imaging data, and the channel bias and exposure gain are compensated using the initial calibration matrix to obtain compensated imaging data. The compensated imaging data is mapped to the standard imaging domain to obtain a standardized face temporal signal.

5. The method according to claim 1, characterized in that, Self-calibration is performed iteratively based on the candidate pulse signal and the noise signal to correct the initial calibration matrix, resulting in a self-calibration matrix, including: Construct a session self-calibration loss function, including multi-skin region pulse synchronization loss, candidate pulse and noise signal coherence suppression loss, physiological frequency band out-of-band energy suppression loss, session residual amplitude constraint loss, and calibration parameter time smoothing loss. Based on the candidate pulse signal and the noise signal, the bounded residual is iteratively optimized using the self-calibration loss function to obtain the self-calibration matrix.

6. The method according to claim 1, characterized in that, The calibration discriminability is calculated based on the imaging domain evidence vector and the self-calibration matrix, and the detection result is output based on the relationship between the calibration discriminability and the discriminability threshold, including: The calibration identifiability is calculated based on the reliability of device metadata, the effective redundancy of multiple skin regions, the observability of physiological frequency bands, the uncertainty of calibration parameters, and the burden of imaging quality. When the calibration discernibility is higher than the first threshold, perform a complete dynamic calibration and output the health test result. When the calibration discernibility is higher than the second threshold but not higher than the first threshold, a restricted calibration is performed, where the second threshold is lower than the first threshold. When the calibration discernibility falls below the second threshold, stop outputting numerical results and guide the user to re-collect data.

7. The method according to claim 6, characterized in that, If the same user completes the authorization measurement through different facial recognition terminals within a preset time interval, the cross-terminal consistency index is calculated, and the source of measurement deviation is located based on the cross-terminal consistency index.

8. An imaging domain adaptive calibration device for cross-terminal facial recognition health detection, characterized in that, The device includes: The evidence vector construction module is used to construct imaging domain evidence vectors based on the device profile, acquisition parameters, and raw imaging data of the face scanning terminal. The bounded residual calculation module is used to extract the reference color unmixing matrix from the device file, combine it with the imaging domain evidence vector, use the calibration supernetwork to generate bounded residuals, and generate an initial calibration matrix based on the bounded residuals. The standard mapping module is used to map the original imaging data to the standard imaging domain using the initial calibration matrix to obtain a standardized face temporal signal. The image self-calibration module is used to extract candidate pulse signals corresponding to multiple preset skin regions from the standardized face time-series signal, and simultaneously extract noise signals from preset comparison regions. Based on the candidate pulse signals and the noise signals, iterative self-calibration is performed to correct the initial calibration matrix and obtain the self-calibration matrix. The detection result output module is used to calculate the calibration discriminability based on the imaging domain evidence vector and the self-calibration matrix, and to perform stream splitting based on the relationship between the calibration discriminability and the discriminability threshold, and output the detection result.

9. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images