Computer vision-based facial image intelligent recognition analysis method and system

Abstract

The application discloses a computer vision-based facial image intelligent recognition analysis method and system, belongs to the technical field of traditional Chinese medicine facial recognition, and comprises the following steps: applying physical stimulation with predetermined parameters to the face of a subject, and acquiring a facial hyperspectral image sequence in a stimulation response period; based on the facial hyperspectral image sequence, constructing a time-series intensity curve of each pixel point at a preset wave band, and extracting a relaxation characteristic parameter from the time-series intensity curve; deducing the static spectral reflection characteristics of each pixel point according to the relaxation characteristic parameter and the end data of the time-series intensity curve, and forming a facial static optical base map; and respectively calculating the facial static optical base map and a baseline frame in the facial hyperspectral image sequence; the computer vision technology is combined with the hyperspectral image analysis, the objective extraction and quantification of facial features are realized, the technical pain points that the traditional Chinese medicine facial diagnosis relies on artificial experience and is highly subjective are solved, and the objectivity and reliability of the diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology in traditional Chinese medicine, and more specifically, to a method and system for intelligent facial image recognition and analysis based on computer vision. Background Technology

[0002] With the rapid development of artificial intelligence and computer vision technology, applying computer vision technology to intelligent facial recognition and analysis in traditional Chinese medicine (TCM) to achieve standardization and objectification of facial diagnosis has become an important direction for promoting the modernization and informatization of TCM. This approach can effectively address industry pain points such as reliance on physicians' personal experience, strong subjectivity, and poor diagnostic consistency in traditional TCM facial diagnosis. In the actual process of acquiring facial images, human facial skin is easily affected by various physiological factors, resulting in temporary changes in appearance, such as facial reflection due to oily skin, localized darkening of skin color due to sweating, and acquired non-pathological changes. Pigmentation, acne scars, and skin redness caused by emotional fluctuations and environmental stimuli are all physiological skin manifestations that are highly similar in appearance to pathological colors used in traditional Chinese medicine (TCM) facial diagnosis to determine pathological conditions, such as sallow, dull, or flushed complexion, as well as the characteristics of pathological moles. This makes it difficult for computer vision-based facial recognition systems to effectively distinguish between temporary skin conditions and the long-term, fundamental facial color and texture features relied upon in TCM diagnosis. Consequently, facial feature extraction becomes mixed, effective features are interfered with, and ultimately, TCM diagnostic conclusions are misjudged, seriously affecting the reliability of intelligent facial recognition analysis.

[0003] In-depth analysis reveals that existing methods generally treat facial images as a homogeneous, static appearance signal field, attempting to simultaneously extract two fundamentally different types of optical signals from single-frame or finite-frame RGB image and texture information: transient physiological optical noise, which corresponds to temporary physiological disturbances such as skin oiling and sweating, and a stable facial optical substrate, which corresponds to the long-term, fundamental facial color and texture features required for traditional Chinese medicine diagnosis. In reality, at the physical level, physiological optical noise primarily originates from specular reflection on the skin surface, such as reflections caused by oil and light scattering effects from sweat, while the facial optical substrate originates from the skin... The color of the underlying tissue and the texture of the dermis are completely different in terms of light propagation paths and imaging mechanisms. Statistically, physiological optical noise is transient, random, and has weak temporal correlation, while surface optical substrates are long-term, stable, and have strong temporal correlation. These two fundamentally different optical signals highly overlap in image pixel representation and share the same pixel information. However, existing technologies lack a physically driven mechanism that can effectively decouple the two at the imaging source or signal representation level, making it impossible to accurately separate physiological optical noise from surface optical substrates. This is the deep-seated technical root cause of mixed feature extraction and misjudgment.

[0004] It is known that current computer vision-based intelligent facial image recognition and analysis methods suffer from serious misjudgment problems in the context of skin physiological interference. The superficial reason is the inability to distinguish between temporary skin conditions and long-term facial features. The deeper root cause is the lack of an effective decoupling mechanism between physiological optical noise and facial optical substrate. Furthermore, in single-frame and short-time sequence diagnosis scenarios, it is unable to balance the contradiction between imaging signal intensity, texture clarity, and spectral aliasing. These technical defects make it difficult for existing methods to meet the standardization and precision requirements of TCM facial diagnosis and prevent their widespread application in clinical diagnosis, health screening, and other practical scenarios. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a computer vision-based intelligent facial image recognition and analysis method and system. This method can achieve the objective extraction and quantification of facial features by combining computer vision technology with hyperspectral image analysis. It solves the technical pain points of traditional Chinese medicine facial diagnosis, which relies on human experience, is highly subjective, and has poor consistency, thereby improving the objectivity and reliability of diagnosis.

[0006] To solve the above problems, the present invention adopts the following technical solution: The first aspect is a computer vision-based intelligent face image recognition and analysis method, including: Physical stimulation with predetermined parameters is applied to the subject's face, and a sequence of hyperspectral images of the face is acquired during the stimulation response period; Based on facial hyperspectral image sequences, a temporal intensity curve of each pixel in a preset band is constructed, and relaxation feature parameters are extracted from the temporal intensity curve. Based on the relaxation feature parameters and the final data of the time series intensity curve, the static spectral reflectance characteristics of each pixel are deduced to form a static optical base map of the face. Pixel-level spectral differences between the static optical base map of the face and the baseline frames and quasi-steady-state frames in the facial hyperspectral image sequence are calculated respectively; based on the spectral differences and relaxation feature parameters, a reliability heatmap of the static optical base map of the face is generated. The static spectral reflectance characteristics of each pixel in the static optical base image of the face are matched with the predefined baseline facial color spectral reflectance curve, and combined with the confidence heatmap, facial pixels are classified and labeled to generate a facial color classification symbol map. Based on the preset facial organ partition map, the pixels of each partition in the facial color classification symbol map are weighted and statistically analyzed according to their facial color classification and the corresponding credibility in the credibility heat map to generate a dialectical vector.

[0007] Furthermore, for each subject's face, an individual thermodynamic portrait was created, including: Multiple subthreshold thermal pulses with increasing energy are applied to a selected skin region, and time-series data of the region's radiance in a preset wavelength band are collected to obtain a cluster of transient thermodynamic curves of the skin micro-region. Based on the cluster of transient thermodynamic curves in skin microregions, an individualized heat transfer function is established, and the personalized master stimulus waveform for inducing standard physiological responses is obtained by inversely solving this function.

[0008] Furthermore, relaxor feature parameters are extracted, including: The temporal intensity curve is subjected to Laplace domain transformation and mode decomposition to obtain the intrinsic relaxation mode set corresponding to each pixel; Based on the intrinsic relaxation mode set, the dominant mode dominance and intermodal antagonism index of each pixel are calculated.

[0009] Furthermore, the relaxor feature parameters are relaxor fingerprint vectors, and the dominant mode dominance and intermodal antagonism index are calculated, including: Calculate the first and second derivatives of the time series intensity curve, and identify the zero-crossing points of the first and second derivatives as feature point sequences; The temporal intensity curve is divided into multiple stages based on the sequence of feature points, and its dynamic type is determined based on the local geometric features of each stage. For each stage, the corresponding intrinsic dynamic model is called to extract the stage intrinsic parameters according to its calibrated dynamic type. Based on the time order of the feature point sequence, the extracted stage intrinsic parameters are combined to form a relaxor fingerprint vector.

[0010] Further, the generation of a static optical base map of the face includes: A sliding window statistical analysis was performed on the end of the time series intensity curve to determine the start time of the quasi-steady-state segment, and the noise floor level within the quasi-steady-state segment was calculated. By utilizing the relaxation fingerprint vector, the quasi-steady-state start time, and the noise floor level, a residual response function is constructed and inverse compensation is performed to obtain the initial value of the instantaneous spectral intensity. The instantaneous spectral intensity of each pixel is used as the initial estimation field. Iterative optimization is performed based on the relaxed fingerprint vector to generate a static optical base map of the face.

[0011] Furthermore, the generation of the credibility heatmap includes: Calculate the pixel-level spectral angular distance between the facial static optical base map and the baseline frame and quasi-steady-state frame, generate a total difference map, and perform linear trend fitting and residual analysis on the total difference map to separate the linear difference components and nonlinear difference components. Based on the relaxed fingerprint vector and nonlinear difference components, a dynamic consistency index map is constructed. Using the dynamic consistency index map as the initial seed, and combining the spatial gradient of nonlinear difference components and the relaxation fingerprint similarity between adjacent pixels, anisotropic spatial propagation and conflict resolution are performed to generate a credibility heatmap.

[0012] Furthermore, the generation of the facial color classification symbol map includes: Calculate the fuzzy membership degree of the static spectral reflectance properties to the spectral reflectance curve of each predefined reference surface color, and generate an initial surface color probability distribution map; Based on the relaxed fingerprint vector and the spatial neighborhood information of pixels, the initial face color probability distribution map is collaboratively corrected to obtain the corrected face color probability distribution map.

[0013] Furthermore, the generation of facial color classification symbol maps also includes: High-conflict regions in the corrected fading probability distribution map are identified and reclassified in the low-dimensional spectral subspace by combining a confidence heatmap to generate an updated fading probability distribution map. Based on the confidence heatmap, a decision is made on the updated face color probability distribution map, and a face color classification symbol map is generated.

[0014] Furthermore, the dialectical vector is generated, including: For each organ region, based on the facial color classification symbol map and the credibility heatmap, the credibility weighted area ratio of each facial color, the spatial clustering degree of each facial color, and the spatial cross-correlation between different facial colors are calculated and combined into a multi-dimensional facial color topological feature vector. The multidimensional facial color topological feature vector is input into a pre-set TCM organ-organ mutual generation and restraint and five-color association rule library for parsing, generating a structured diagnostic vector.

[0015] Secondly, a computer vision-based intelligent face image recognition and analysis system includes: The stimulation acquisition module is used to apply physical stimulation with predetermined parameters to the subject's face and acquire a sequence of hyperspectral images of the face during the stimulation response period; The feature extraction module is used to construct the temporal intensity curve of each pixel in a preset band based on the facial hyperspectral image sequence, and extract relaxation feature parameters from the temporal intensity curve. The substrate generation module is used to deduce the static spectral reflectance characteristics of each pixel based on the relaxation feature parameters and the final data of the time-series intensity curve, and to form a static optical substrate map of the face. The credibility assessment module is used to calculate the pixel-level spectral differences between the baseline frame and the quasi-steady-state frame in the facial static optical base map and the facial hyperspectral image sequence, respectively; and to generate a credibility heatmap of the facial static optical base map based on the spectral differences and relaxation feature parameters. The facial color classification module is used to match the static spectral reflectance characteristics of each pixel in the static optical base image of the face with the predefined benchmark facial color spectral reflectance curve, and combine it with the confidence heatmap to classify and label the facial pixels and generate a facial color classification symbol map. The dialectical generation module is used to generate a dialectical vector by weighting the pixels of each region in the facial color classification symbol map according to their corresponding credibility in the facial color classification and credibility heat map, based on the preset facial organ partition map.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This solution uses computer vision technology combined with hyperspectral image analysis to achieve objective extraction and quantification of facial features, which solves the technical pain points of traditional Chinese medicine facial diagnosis that relies on human experience, is highly subjective, and has poor consistency. It improves the objectivity and reliability of diagnosis and allows standardized diagnosis to be completed without the need for technical personnel to rely on subjective judgment.

[0017] (2) By constructing a static optical base map of the face, this scheme effectively removes temporary physiological optical noise interference such as skin oiliness and sweating, accurately captures the long-term stable facial optical base features that reflect the state of the internal organs, provides more realistic and accurate facial color data support for TCM diagnosis, and avoids interference of temporary skin condition on the diagnosis conclusion.

[0018] (3) This scheme generates a confidence heatmap and performs multiple rounds of correction and conflict resolution on face color classification by combining relaxation feature parameters. This reduces the interference of low reliability pixels on the diagnosis results, improves the accuracy of face color classification and diagnosis conclusions, meets the practical needs of clinical diagnosis, and ensures that the diagnosis results are referable.

[0019] (4) Based on the theory of visceral division and five-color association in traditional Chinese medicine, this scheme transforms the quantitative facial color topological features into structured diagnostic vectors, realizes the standardized and normalized output of TCM facial diagnosis, facilitates clinical promotion and application, and provides a feasible technical path for TCM digital and intelligent diagnosis, promoting the integration of traditional Chinese medicine facial diagnosis with modern computer technology. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0021] Figure 1 This is a flowchart of the intelligent face image recognition and analysis method based on computer vision according to the present invention; Figure 2 This is a data flow diagram between various modules in the computer vision-based intelligent face image recognition and analysis system of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] Example 1: Please see Figure 1 A computer vision-based intelligent recognition and analysis method for face images, which includes: Step 1: Apply physical stimulation with predetermined parameters to the subject's face and acquire a sequence of hyperspectral images of the face within the stimulation response period. The specific operation is as follows: In practice, the first step is to determine the predetermined parameters of the physical stimulation that are suitable for the needs of TCM facial diagnosis. The selection of parameters must take into account both safety and response effectiveness. Thermal stimulation is usually selected as the stimulation type. The predetermined parameters include stimulation intensity, stimulation duration, and stimulation range. The stimulation intensity is controlled within the comfort threshold of human skin to avoid skin damage or strong physiological stress response. The stimulation range covers the entire facial area to ensure that the corresponding areas of facial organs required for TCM diagnosis are uniformly stimulated. During the stimulation process, a stimulation device that matches the predetermined parameters is used to act on the subject's face in a stable and uniform manner. At the same time, a hyperspectral image acquisition device is activated to synchronously acquire a sequence of hyperspectral images of the face within the stimulation response period. The stimulation response period is defined as the complete time period from the start of physical stimulation to the recovery of the skin's physiological state to the baseline level before stimulation. The acquisition frequency must meet the requirements of capturing the temporal changes in the skin's optical properties. It is usually set to an interval of no more than 100ms per frame to ensure that the optical response relaxation process of the skin under stimulation can be completely recorded. The acquisition bands of hyperspectral image sequences need to cover the key spectral range that reflects the color of subcutaneous tissue and the texture of the dermal layer of the skin. Typically, the 400 to 1000 nm band is selected. This band can effectively capture the spectral characteristics of components such as hemoglobin and melanin that are related to the diagnosis of facial color in traditional Chinese medicine. Each frame of the image needs to ensure sufficient spatial and spectral resolution, with a spatial resolution of no less than 1080P and a spectral resolution of no less than 10nm, to ensure that the temporal spectral information of individual pixels can be extracted subsequently.

[0024] This step allows transient physiological optical noise and stable surface optical substrate to exhibit different response patterns in the temporal dimension. Physiological optical noise changes rapidly with stimulation, while surface optical substrate maintains relatively stable relaxation characteristics, thus providing a distinguishable temporal characteristic basis for separating the two.

[0025] Creating an individual thermodynamic portrait for each subject's face includes the following steps: Step 11: Apply multiple subthreshold thermal pulses with increasing energy to the selected skin area and collect time-series data of the area's radiance in a preset wavelength band to obtain a cluster of transient thermodynamic curves for the skin micro-region. The specific operation is as follows: First, representative skin areas of the face need to be selected as the selected skin areas. The selection principle should take into account both the typicality of the area and the availability of data. Usually, four areas are selected: the forehead, the left cheek, the right cheek, and the chin. These areas correspond to different organ divisions in traditional Chinese medicine, and the skin thickness and subcutaneous tissue distribution are representative, which can reflect the thermodynamic characteristics of the skin of the entire face. The area of ​​each selected skin area is controlled within 1cm×1cm to ensure the uniformity of the collected micro-area data, while avoiding heat transfer interference caused by the area being too large. Next, multiple subthreshold thermal pulses with increasing energy were designed. The core of subthreshold thermal pulses is defined as thermal pulses with energy below the skin damage threshold, ensuring that they only induce reversible thermal responses in the skin surface and subcutaneous tissues after application, without causing irreversible skin damage. The energy of the thermal pulses is set in a gradient manner, starting from the lowest effective stimulation energy, with each increment having a fixed energy gradient, typically 5 mJ, for a total of 5 to 8 energy levels, covering the skin response range under different intensities of thermal stimulation. The duration of each energy level thermal pulse is fixed at 100 ms, and the pulse interval is set to 30 s, ensuring that the skin thermal response induced by the previous thermal pulse has fully recovered before applying the next energy level thermal pulse, avoiding the superposition and interference of responses from adjacent thermal pulses.

[0026] During the application of each energy level thermal pulse, a radiance acquisition device is simultaneously activated to collect time-series data of radiance in selected skin areas within a preset wavelength band. The preset wavelength band is selected as a key band reflecting skin temperature changes and the optical properties of subcutaneous tissue, typically 800 to 900 nm. Radiance in this band is linearly correlated with skin temperature and can indirectly reflect the skin's heat transfer process. The acquisition of radiance time-series data begins the instant the thermal pulse is applied and continues until the skin radiance returns to the baseline level before the thermal pulse was applied. The acquisition frequency is consistent with the hyperspectral image acquisition frequency in step 1 to ensure... To ensure the synchronicity and comparability of time-series data, a complete set of radiance time-series data was collected for each energy level of thermal pulse. With time as the horizontal axis and radiance as the vertical axis, each set of time-series data was plotted as a curve. Multiple sets of curves corresponding to different energy levels together constitute a cluster of transient thermodynamic curves in the skin microregion. This cluster of curves can clearly reflect the response pattern of the individual subject's skin to thermal stimulation of different intensities. There are differences in the peak value, relaxation time, and baseline recovery speed of the curves corresponding to different energy levels of thermal pulses. These differences are directly related to individual characteristics such as skin thickness, water content, and subcutaneous capillary distribution.

[0027] Step 12: Based on the transient thermodynamic curves of skin microregions, establish an individualized heat transfer function, and then use this function to inversely solve for the personalized master stimulus waveform used to induce a standard physiological response. The specific steps are as follows: Based on the physical principles of heat conduction, the skin, as a multi-layered medium, undergoes heat stimulation transmission according to the heat conduction equation. The heat transfer function quantifies the relationship between the input heat stimulation energy and the skin's radiant output. The transient thermodynamic curve cluster of skin micro-regions obtained in step 11 is essentially the time-series response of skin radiant output under different heat stimulation energy inputs. Therefore, based on this curve cluster data and combined with the heat conduction physical model, an individualized heat transfer function can be derived. In the actual derivation process, the heat conduction equation is first simplified. Considering that the thickness of the skin micro-region is much smaller than the thermal diffusion length, the skin micro-region can be approximated as a one-dimensional heat conduction medium. The simplified heat conduction calculation is: the partial derivative of temperature T with respect to time t is equal to the thermal diffusivity α multiplied by the second partial derivative of temperature T with respect to depth h. The derivation logic of this calculation is: heat transfer in the skin micro-region is mainly by heat conduction, ignoring the effects of heat convection and heat radiation. According to Fourier's law of heat conduction and the law of conservation of energy, the differential equation of skin temperature changing with time and space can be derived, namely the aforementioned one-dimensional heat conduction equation, where α represents the thermal diffusivity of the skin, in meters. 2 / s is a parameter reflecting the efficiency of heat transfer through the skin, and it varies significantly from person to person; h represents the skin depth, measured in meters.

[0028] Based on the simplified heat conduction equation described above, and combined with the transient thermodynamic curve cluster data of the skin micro-regions obtained in step 11, the parameters of the individualized heat transfer function are determined through data fitting methods. The thermal stimulation energy corresponding to each energy level of thermal pulse is used as the input parameter, and the corresponding radiance time-series data is converted into temperature time-series data. Based on the linear correspondence between radiance and temperature in a preset band, the temperature time-series data is then substituted into the simplified heat conduction equation, and the thermal diffusivity of the subject's skin is obtained through least squares fitting. In addition to other relevant parameters, an individualized heat transfer function is constructed. This function can accurately describe the temporal response of the subject's skin to any thermal stimulus energy, reflecting the influence of individualized characteristics such as the subject's skin thickness, water content, and subcutaneous tissue distribution on heat transfer, and realizing the quantitative mapping between thermal stimulus input and skin temperature response output. Subsequently, based on the individualized heat transfer function, the personalized master stimulus waveform used to induce the standard physiological response is solved in reverse. The standard physiological response is defined as the skin optical response that can clearly distinguish between physiological optical noise and the facial optical substrate. Specifically, during the stimulus response period, physiological optical noise exhibits rapid transient characteristics, while the facial optical substrate exhibits stable relaxation characteristics, ensuring that the two can be effectively separated through temporal characteristics. In the reverse solution process, the skin temperature temporal response curve corresponding to the standard physiological response is first preset. This curve needs to meet the requirements of moderate peak temperature, complete relaxation process, and stable baseline recovery. Then, the preset temperature temporal response curve is substituted into the individualized heat transfer function, and the waveform of the corresponding thermal stimulation energy changing with time is obtained through reverse derivation, which is the personalized master stimulus waveform. This waveform is tailored to the subject's skin thermodynamic characteristics, ensuring that the physical stimulus applied in step 1 can induce a standard physiological response of sufficient intensity without causing damage to the skin, while maximizing the amplification of the temporal characteristics of the facial optical substrate and suppressing the interference of physiological optical noise.

[0029] In a preferred embodiment of the present invention, step 2 is further included: based on the facial hyperspectral image sequence, a temporal intensity curve of each pixel in a preset band is constructed, and relaxation feature parameters are extracted from the temporal intensity curve. The specific operation is as follows: In practice, the facial hyperspectral image sequence acquired in step 1 first needs to be preprocessed. The preprocessing process mainly includes image alignment and noise removal. Image alignment is used to eliminate pixel position shifts caused by slight facial movements of the subject. The facial feature point matching method is used, and fixed feature points such as the corners of the eyes and the wings of the nose are selected as references. Each frame of the sequence is aligned with the baseline frame, that is, the facial hyperspectral image acquired before the stimulus is applied, to ensure that the same pixel always corresponds to the same position on the face. Noise removal is used to eliminate random electronic noise generated during the acquisition of hyperspectral images. The Gaussian filtering method is used, and the filter kernel size is set to 3×3, which effectively removes noise without destroying the temporal characteristics of the spectral intensity of pixels in the image. After preprocessing, for each pixel in the facial image, its spectral intensity temporal data in the preset band is extracted. The preset band adopts the key spectral range of 400 to 1000 nm selected in step 1. This band can effectively capture the spectral signals corresponding to the color of subcutaneous tissue and the texture of the dermal layer of the skin, and eliminate the interference of irrelevant bands.

[0030] For each pixel, a curve is plotted with time on the horizontal axis and the average spectral intensity of that pixel in a preset band in each frame of the hyperspectral image on the vertical axis. This curve represents the temporal intensity curve of that pixel in the preset band. The temporal intensity curve contains two fundamentally different signal components: one is the signal corresponding to transient physiological optical noise, which appears as a rapidly fluctuating, short-duration component in the curve, corresponding to temporary physiological disturbances such as skin oiling and sweating; the other is the signal corresponding to a stable facial optical substrate, which appears as a slowly changing, long-duration component with a stable relaxation pattern, corresponding to the long-term facial color and texture characteristics required for TCM diagnosis. The relaxation feature parameter is the core parameter used to quantify the relaxation process in the temporal intensity curve, and it can accurately characterize the difference between the two signal components. Therefore, it is necessary to extract the relaxation feature parameter from the temporal intensity curve.

[0031] The extraction of relaxed feature parameters also includes the following steps: Step 21: Perform Laplace domain transformation and mode decomposition on the time-series intensity curve to obtain the intrinsic relaxation mode set corresponding to each pixel. The specific operations are as follows: As a complex non-stationary temporal signal, the temporal intensity curve contains physiological optical noise signals that differ significantly from the surface optical substrate signals in frequency and relaxation rate. Directly extracting feature parameters makes it difficult to accurately distinguish between the two types of signals. Therefore, a two-step operation of Laplace domain transformation and mode decomposition is required to achieve signal separation and purification. The purpose of Laplace domain transformation is to convert the temporal intensity curve in the time domain to the Laplace domain, highlighting the relaxation characteristics of the signal. Because the relaxation process in the time domain is manifested as the signal decaying or rising over time, after transformation to the Laplace domain, signals with different relaxation rates will exhibit different eigenvalues, which facilitates subsequent mode decomposition. The specific operation of Laplace domain transformation is to perform a Laplace transformation on the temporal intensity curve f(t) of each pixel to obtain F(s), where s is the Laplace variable. This transformation converts the differential relationship in the time domain into an algebraic relationship in the Laplace domain, simplifying signal analysis and processing.

[0032] After completing the Laplace domain transform, the transformed signal undergoes mode decomposition. This mode decomposition employs a method based on the inherent characteristics of the signal, requiring no additional model assumptions. It decomposes the signal within the Laplace domain into multiple independent intrinsic relaxation modes based solely on the signal's frequency and amplitude characteristics. Each intrinsic relaxation mode corresponds to a signal component with specific relaxation speed and frequency characteristics. Some modes correspond to transient signals of physiological optical noise (i.e., fast relaxation speed and high frequency), while others correspond to stable signals of a planar optical substrate (i.e., slow relaxation speed and low frequency). All the decomposed intrinsic relaxation modes together form the intrinsic relaxation mode set corresponding to the pixel. Each mode contains complete temporal intensity information and relaxation characteristics, ensuring accurate differentiation between the two types of signal components through subsequent analysis of the characteristics of each mode. This also provides data support for quantifying the intensity and interrelationships of each mode.

[0033] Step 22: Based on the intrinsic relaxation mode set, calculate the dominant mode dominance and intermodal antagonism index for each pixel. The specific operations are as follows: The dominant mode dominance is used to characterize the contribution of a particular mode in the intrinsic relaxation mode set to the overall time-series intensity curve. The mode with the highest contribution is the dominant mode. The larger the dominance value, the higher the proportion of the signal component corresponding to that mode in the entire time-series signal. Since the signal corresponding to the surface optical substrate has stability and long-term characteristics, its corresponding intrinsic relaxation mode usually has the highest contribution to the time-series signal. Therefore, the dominant mode dominance can be used to determine the dominance of the surface optical substrate signal in the time-series signal of that pixel. The specific calculation method of the dominant mode dominance is as follows: First, calculate the sum of squares of the amplitude of each mode in the intrinsic relaxation mode set. The sum of squares of the amplitude is used to characterize the energy of that mode. The higher the energy, the higher the contribution to the entire time-series signal. Then, calculate the proportion of the sum of squares of the amplitude of each mode to the total sum of squares of the amplitudes of all modes. This proportion is the dominance of that mode. Select the proportion with the highest dominance as the dominant mode dominance of that pixel, and at the same time determine the corresponding dominant mode.

[0034] The intermodal antagonism index is used to quantify the concentration of intrinsic relaxation modes and the degree of mutual inhibition or independence between different modes. It mainly reflects the antagonistic relationship between the dominant mode and other non-dominant modes. The higher the antagonism index, the stronger the independence of the dominant mode from other modes and the better the distinction between the two types of signal components. Conversely, it indicates that the two types of signal components are more severely mixed. The dominant mode usually corresponds to the surface optical substrate, and other non-dominant modes usually correspond to physiological optical noise. The specific calculation method of the intermodal antagonism index is as follows: First, calculate the correlation coefficient between the dominant mode and each non-dominant mode. The correlation coefficient is used to characterize the degree of linear correlation between the two modes. The closer the correlation coefficient is to 0, the more independent the two modes are and the higher the degree of antagonism. Then, calculate the average value of the correlation coefficients between all dominant modes and non-dominant modes. Subtracting the average value from 1 gives the intermodal antagonism index. By calculating the dominant mode dominance and the intermodal antagonism index, the proportion and distinction between the two types of signal components in the temporal signal of each pixel can be accurately quantified.

[0035] The relaxation feature parameters are relaxation fingerprint vectors, and the calculation of dominant mode dominance and intermodal antagonism index also includes the following steps: Step 23: Calculate the first and second derivatives of the time series intensity curve, and identify the zero-crossing points of the first and second derivatives as feature point sequences. The specific operations are as follows: The relaxation process of a time-series intensity curve includes multiple different stages of change, each corresponding to a different pattern of signal change. The transitions between these stages are reflected in the inflection points and extreme points of the curve. The derivative is the core tool for characterizing the rate of change and inflection points of the curve. Therefore, by calculating the first and second derivatives, these key feature points can be accurately identified. Before calculating the derivatives, the time-series intensity curve constructed in step 2 needs to be pre-processed with a smoothing method using moving average smoothing. Five consecutive time points are selected as a smoothing window, and the average intensity within the window is calculated as the smoothed intensity value at the middle time point of the window. This process is repeated to complete the smoothing of the entire curve. The purpose is to eliminate random noise in the time-series curve, avoid errors in derivative calculation due to noise, and prevent misjudgment of feature points. Simultaneously, it ensures that the smoothed curve retains its original relaxation characteristics and key inflection points. After smoothing, the first derivative of the time-series intensity curve is calculated. The first derivative is used to characterize the instantaneous rate of change of the time-series intensity curve at each time point. f'(t) equals f(t+Δt) minus f(t) and then divided by Δt. The derivation logic of this calculation is based on the definition of the derivative. The instantaneous rate of change cannot be directly calculated. Therefore, the ratio of the intensity difference between two adjacent time points in the discrete time series to the time interval is used to approximate the instantaneous rate of change at that time point. This is consistent with the characteristic that the time series intensity curve is a discrete time series. Here, f'(t) represents the first derivative of the time series intensity curve at time t, which represents the instantaneous rate of change of intensity at time t; f(t+Δt) represents the time series intensity value at time t+Δt; f(t) represents the time series intensity value at time t; Δt represents the time interval between two adjacent acquisition time points, which is consistent with the acquisition interval of the hyperspectral image sequence in step 1, usually 100ms.

[0036] After calculating the first derivative, the second derivative is further calculated. The second derivative is used to characterize the rate of change of the first derivative, i.e., the curvature change of the time series intensity curve. It is calculated as follows: f''(t) equals f′(t+Δt) minus f'(t) and then divided by Δt. This derivation logic is consistent with that of the first derivative. The ratio of the difference in the first derivative values ​​between two adjacent time points to the time interval is used to approximate the second derivative at time t, which is used to identify the inflection point of the curve. In the formula, f''(t) represents the second derivative of the time series intensity curve at time t, characterizing the curvature change of the curve at time t; f′(t+Δt) represents the first derivative value at time t+Δt; f'(t) represents the first derivative value at time t; Δt is consistent with the time interval in the first derivative formula. After the derivative calculation is complete... After completion, the zero-crossing points of the first and second derivatives are identified as feature points. The zero-crossing point of the first derivative refers to the time point when the value of the first derivative changes from positive to negative or from negative to positive, corresponding to the extreme points of the time-series intensity curve, i.e., peaks or valleys. These points are key nodes of intensity change during relaxation, such as the time point when the intensity reaches its peak after thermal stimulation. The zero-crossing point of the second derivative refers to the time point when the value of the second derivative changes from positive to negative or from negative to positive, corresponding to the inflection point of the time-series intensity curve. These points are nodes where the rate of change of intensity changes during relaxation, marking the transition of the relaxation stage. All identified zero-crossing points of the first and second derivatives are arranged in chronological order to form a feature point sequence. Each feature point contains corresponding time coordinates and intensity coordinates.

[0037] Step 24: Divide the temporal intensity curve into multiple stages based on the feature point sequence, and determine its dynamic type based on the local geometric features of each stage. The specific operations are as follows: Each feature point in the feature point sequence, as a key node in the relaxation process of the temporal intensity curve, can divide the entire curve into multiple continuous line segments with a single variation pattern. Each line segment is an independent relaxation stage, and each stage corresponds to a specific physical process in the signal relaxation process, such as the intensity increase process under stimulus, the rapid relaxation process after stimulus ends, the slow relaxation process tending to stabilize, and the steady-state process of returning to the baseline. The specific operation of stage division is to divide the temporal intensity curve segment between two adjacent feature points into an independent stage according to the chronological order of the feature point sequence. Starting from the starting point of the temporal curve, i.e., the stimulus application time, and then to the first feature point as the first stage, and so on, until the end point of the temporal curve, i.e., the time of recovery to the baseline, ensuring that the temporal intensity curve in each stage has a single variation trend without additional inflections. To avoid inaccurate dynamic type calibration due to multiple variation patterns within a stage, local geometric feature analysis is performed on the time-series intensity curve segments of each stage after stage division. Local geometric features mainly include the slope, curvature, and intensity variation range of the curve within the stage. These features accurately reflect the signal variation pattern of that stage, thus corresponding to different dynamic types. The slope characterizes the speed and direction of intensity change within the stage; a positive slope indicates that the intensity increases with time, while a negative slope indicates that the intensity decreases with time. The larger the absolute value of the slope, the faster the change. The curvature characterizes the direction and degree of curvature within the stage; a positive curvature indicates that the curve bends upwards, while a negative curvature indicates that the curve bends downwards. The larger the absolute value of the curvature, the more pronounced the bend. The intensity variation range characterizes the maximum difference in intensity within the stage, reflecting the amplitude of signal change in that stage.

[0038] Based on the aforementioned local geometric features, the dynamic type of each stage is calibrated. The calibration logic follows the physical laws of signal relaxation. Specifically, the calibration rules are as follows: a stage with a positive slope, negative curvature, and a large absolute value of the slope, and a large range of intensity variation, corresponds to the rapid rise process under stimulation and is calibrated as the rapid rise dynamic type. This stage mainly corresponds to the transient signal of physiological optical noise. A stage with a negative slope, positive curvature, and a large absolute value of the slope, and a large range of intensity variation, corresponds to the rapid relaxation process after stimulation ends and is calibrated as the rapid relaxation dynamic type. This stage mainly corresponds to the attenuation signal of physiological optical noise. A stage with a negative slope, near-zero curvature, and a small absolute value of the slope, and a small range of intensity variation, corresponds to the slow relaxation process tending towards stability and is calibrated as the slow relaxation dynamic type. This stage mainly corresponds to the stable signal of the surface optical substrate. A stage with a slope close to zero, near-zero curvature, and a very small range of intensity variation corresponds to the steady-state process of returning to the baseline and is calibrated as the steady-state dynamic type. This stage mainly corresponds to the stable state of the surface optical substrate. Each stage is calibrated with a unique dynamic type.

[0039] Step 25: For each stage, extract the stage intrinsic parameters by calling the corresponding intrinsic dynamic model according to its calibrated dynamic type, and combine the extracted stage intrinsic parameters according to the time order of the feature point sequence to form a relaxor fingerprint vector. The specific operation is as follows: Different dynamic types correspond to different physical processes of signal relaxation, and their intensity changes follow different dynamic characteristics. Therefore, for each dynamic type, the corresponding intrinsic dynamic model is invoked. The intrinsic dynamic models used here are all classic basic models established based on physical relaxation laws and verified through long-term practice in the field of optical signal relaxation analysis. They are not models additionally designed by this solution. Each model has clear physical theoretical support and mature application scenarios, ensuring that those skilled in the art can directly and accurately invoke the corresponding model and complete the parameter extraction process according to the dynamic type, fully meeting the reproducibility requirements of the technical solution. The selection of intrinsic dynamic models corresponding to each dynamic type is strictly matched to the corresponding stage. Physical relaxation mechanism: For rapid rise dynamics and rapid relaxation dynamics, the classical exponential model is used. This model is the most commonly used basic model in optical transient signal analysis. Its core is based on the physical law that the signal intensity changes exponentially with time, which is just right for the signal characteristics of the rapid rise and rapid relaxation stages. These stages correspond to the transient process of physiological optical noise on the skin surface. The changes in optical signals caused by physical effects such as specular reflection and light scattering by sweat on the skin surface are fast and conform to the law of exponential growth or decay. The classical exponential model can accurately quantify its rise or decay rate. At the same time, the model has few parameters and simple fitting logic. Those skilled in the art can quickly extract feature parameters through conventional fitting methods. For slow relaxation dynamics, the classic logarithmic decay model is used. This model is the fundamental model for the relaxation analysis of stable optical signals in subcutaneous tissue. Based on the optical substrate signal corresponding to the color of subcutaneous tissue and the texture of the dermis, its relaxation process is regulated by physical processes such as subcutaneous blood circulation and tissue metabolism. The rate of change is slow and the decay trend follows a logarithmic law, which is highly consistent with the core characteristics of the logarithmic decay model. It can accurately characterize the stable relaxation characteristics of the surface optical substrate and accurately capture the intensity change law in the slow relaxation stage. For steady-state dynamics, the classic constant model is used. This model is the fundamental model for the characterization of steady-state optical signals. It is suitable for physical scenarios where the signal intensity tends to be stable and there is no significant time change. This stage corresponds to the steady-state process of the skin optical signal recovering to the baseline. The intensity fluctuation is minimal. The constant model can directly and accurately characterize its steady-state intensity value without complex parameter fitting. Moreover, the application of this model is a conventional technique in this field.

[0040] After the model is invoked, the temporal intensity data for each stage is fitted, and the intrinsic parameters of the stage are extracted. The fitting process uses the least squares method. By adjusting the model parameters, the sum of squared errors between the theoretical intensity value output by the model and the actual temporal intensity value collected within the stage is minimized, ensuring that the extracted intrinsic parameters of the stage can accurately reflect the signal relaxation characteristics of the stage. For the exponential rise model, the extracted intrinsic parameters of the stage include the rise rate and the peak intensity. The rise rate represents how fast the intensity rises, and the peak intensity represents the maximum intensity value of the stage. The model formula is f(t) equal to I0ekt. The derivation logic of this formula is based on the rise law of physiological optical noise transient signals. The signal intensity increases exponentially with time, which is consistent with the physical characteristics of the rapid rise stage. In the formula, f(t) represents the intensity value at time t, I0 represents the intensity value at the beginning of the stage, k represents the rise rate, and t represents the time within the stage.

[0041] For the exponential decay model, the extracted intrinsic parameters of the stage include the decay coefficient and the relaxation time. The decay coefficient characterizes the rate of intensity decay, and the relaxation time characterizes the time required for the intensity to decay to the initial value 1 / e. The model formula is f(t) equal to Ip - t / τ. The derivation logic is based on the decay law of physiological optical noise transient signals. The signal intensity decays exponentially with time, which is consistent with the physical characteristics of the rapid relaxation stage. In the formula, f(t) represents the intensity value at time t, Ip represents the peak intensity value at the beginning of the stage, τ represents the relaxation time, and t represents the time within the stage.

[0042] For the logarithmic decay model, the extracted stage intrinsic parameters include decay rate and steady-state intensity. The decay rate characterizes the speed of the slow relaxation process, and the steady-state intensity characterizes the intensity value when the stage tends to stabilize, which can accurately reflect the stability characteristics of the surface optical substrate. For the constant model, the extracted stage intrinsic parameters are the steady-state intensity values, which characterize the stability level of the intensity in that stage and correspond to the baseline characteristics of the surface optical substrate. After extracting the corresponding stage intrinsic parameters for each stage, the intrinsic parameters of all stages are arranged sequentially according to the time order of the stage division in step 24 to form a multi-dimensional vector. This vector is the relaxation fingerprint vector of the pixel. The relaxation fingerprint vector contains the core quantitative features of all relaxation stages of the temporal intensity curve of the pixel, which can uniquely characterize the relaxation characteristics of the skin region corresponding to the pixel. The parameters of the rapid rise and rapid relaxation stages correspond to the physiological optical noise signal, and the parameters of the slow relaxation and steady-state stages correspond to the surface optical substrate signal. The two types of signal components can be accurately distinguished through this vector.

[0043] In a preferred embodiment of the present invention, step 3 is further included: based on the relaxation feature parameters and the final data of the time-series intensity curve, the static spectral reflectance characteristics of each pixel are deduced to form a static optical base map of the face. The specific operation is as follows: Static spectral reflectance characteristics refer to the long-term stable spectral reflectance patterns of the skin after the removal of temporary physiological interference, determined by the color of subcutaneous tissue and the texture structure of the dermis. This is also a core characteristic required for TCM facial diagnosis. The data at the end of the time-series intensity curve is close to the skin's steady-state state, and physiological optical noise interference has significantly decreased. Combining this with the quantification indicators in the relaxation characteristic parameters that can distinguish between the two types of signals, effective separation of physiological optical noise from the facial optical base can be achieved. In practice, firstly, a sliding window statistical analysis is performed on the end of the time-series intensity curve to screen out the stable quasi-steady-state segment and calculate the noise base level, thus clarifying the residual noise in the end-stage data. The noise interference intensity is determined; then, by utilizing relaxation feature parameters, namely the relaxation fingerprint vector, the quasi-steady-state segment start time, and the noise floor level, a residual response function is constructed and inverse compensation is performed to remove the signal components corresponding to physiological optical noise, obtaining an initial value of the instantaneous spectral intensity that closely approximates the real facial optical floor. Finally, using the initial value of the instantaneous spectral intensity as the initial estimation field, iterative optimization is performed under the guidance of the relaxation fingerprint vector to correct the residual error in the initial value, so that the spectral reflectance characteristics of each pixel accurately correspond to the facial optical floor. Then, the static spectral reflectance characteristics of all pixels are combined according to their facial coordinate positions to form a complete facial static optical floor map. This floor map can completely preserve the long-term, basic facial color and texture features required for TCM diagnosis, and completely eliminate the appearance changes caused by temporary physiological interferences such as skin oiliness and sweating.

[0044] The generation of the facial static optical base map also includes the following steps: Step 31: Perform sliding window statistical analysis on the end of the time series intensity curve to determine the start time of the quasi-steady-state segment, and calculate the noise floor level within the quasi-steady-state segment. The specific operations are as follows: The end of the temporal intensity curve corresponds to the later stage of the stimulus-response cycle. At this time, the physical stimulus to the skin has been greatly reduced, the transient signal of physiological optical noise has been basically attenuated, and the overall signal tends to be stable, but there is still weak random noise. The quasi-steady-state segment refers to the continuous period of time in the end of the cycle where the signal intensity fluctuation is minimal and tends to be stable. The signal in this period mainly consists of the surface optical substrate signal and weak random noise, which is the core data source for inferring static spectral reflectance characteristics. The definition of the end of the temporal intensity curve needs to be combined with the total duration of the stimulus-response cycle. Usually, the last 1 / 3 of the stimulus-response cycle is selected as the end. For example, if the stimulus-response cycle is 30s, the acquisition frequency is 10 frames / s, and there are a total of 300 frames of data, then the last 100 frames, usually 20s to 30s, are selected as the end data. This definition method takes into account both stability and data volume, ensuring that the end data contains a complete quasi-steady-state process, while avoiding the selection of too much unstable data in the early stage that may affect the analysis results.

[0045] The specific operation of sliding window statistical analysis must follow clear parameter settings and judgment logic to ensure reproducibility. The size of the sliding window is set according to the acquisition frequency of the hyperspectral image sequence, and 5 to 8 consecutive frames of data are selected as a sliding window. The choice of window size must balance the noise suppression effect and the preservation of signal details. If the window is too small, random noise cannot be effectively suppressed, and if the window is too large, the starting node of the quasi-steady-state segment may be lost. The sliding step size is set to 1 frame, that is, the window starts from the first frame of the final stage and slides forward 1 frame at a time until the last frame of the final stage, ensuring that statistical analysis is performed on the data of each frame of the final stage. The statistical analysis of each window is to calculate the standard deviation of the time series intensity data within the window. The standard deviation is used to quantify the degree of fluctuation of the signal intensity within the window. The smaller the standard deviation, the more stable the signal within the window is, and the closer it is to the quasi-steady-state characteristics. The logic for determining the start time of the quasi-steady-state segment is as follows: starting from the first frame of the final stage, the standard deviation of each sliding window is judged sequentially. When the standard deviation of three or more consecutive sliding windows is less than the preset threshold, the start frame of the first window that meets the condition is taken as the start time of the quasi-steady-state segment. The preset threshold is set based on the signal characteristics after noise removal in the early stage. It is usually set to 1.2 times the standard deviation of the baseline segment of the time-series intensity curve. The baseline segment is the signal collected before the stimulus is applied. Its standard deviation can directly reflect the inherent noise level of the system. This threshold setting can ensure that the signal fluctuation in the quasi-steady-state segment is only due to weak random noise, rather than transient interference from physiological optical noise.

[0046] After the quasi-steady-state segment is determined, the noise floor level within this segment is calculated. The noise floor level is used to quantify the intensity of weak random noise within the quasi-steady-state segment. The calculation method is as follows: select the temporal intensity data of all frames within the quasi-steady-state segment, calculate its average intensity, then calculate the deviation of each frame's intensity value from the average value, and average the absolute values ​​of all deviations. The result obtained is the noise floor level. This calculation method can accurately reflect the average intensity of random noise within the quasi-steady-state segment and avoid the influence of outliers in a single frame on the accuracy of noise quantization. The smaller the noise floor level, the weaker the noise interference within the quasi-steady-state segment, and the closer the subsequently inferred static spectral reflectance characteristics are to the real surface optical substrate. If the noise floor level is too high, the previous image preprocessing and relaxation feature parameter extraction process needs to be re-examined to eliminate abnormal interference and ensure the reliability of subsequent steps. The parameter settings and calculation logic of the entire process are all conventional techniques in the field of optical signal analysis. Those skilled in the art can directly follow the above operations to determine the start time of the quasi-steady-state segment and calculate the noise floor level.

[0047] Step 32: Using the relaxation fingerprint vector, the quasi-steady-state start time, and the noise floor level, construct the residual response function and perform inverse compensation to obtain the initial value of the instantaneous spectral intensity. The specific operations are as follows: The relaxation fingerprint vector contains quantized parameters for each relaxation stage of the temporal intensity curve. The parameters for the rapid rise and rapid relaxation stages correspond to the transient signal characteristics of physiological optical noise, accurately characterizing its relaxation pattern. The quasi-steady-state start time clarifies the time node when the physiological optical noise has basically decayed, while the noise floor level quantifies the intensity of weak random noise within the quasi-steady-state segment. The combination of these three parameters enables accurate modeling and removal of residual physiological optical noise signals. The residual response function is constructed based on the relaxation pattern of physiological optical noise. Using the parameters corresponding to transient signals in the relaxation fingerprint vector, it simulates the complete process of physiological optical noise from generation to decay, especially the weak transient signals that may remain within the quasi-steady-state segment. This function is a simple fitting function constructed based on the previously extracted relaxation feature parameters and the physical laws of signal decay.

[0048] The construction process of the residual response function is as follows: First, parameters such as the rise rate and peak intensity of the rapid rise phase, as well as the attenuation coefficient and relaxation time of the rapid relaxation phase, are extracted from the relaxation fingerprint vector. These parameters can accurately characterize the intensity change law of the transient signal of physiological optical noise. Then, taking the start time of the quasi-steady-state segment as the dividing point, the physiological optical noise signal before this time point is fitted to simulate its attenuation trend, thereby deducing the possible residual physiological optical noise intensity at each time point within the quasi-steady-state segment, forming the residual response function. Its functional expression can be simplified to R(t) equal to Ae__t / τ. The derivation logic of this formula is based on the transient signal of physiological optical noise. The exponential decay law of the signal has been clearly defined in step 25, which shows that the signal in the rapid relaxation phase follows an exponential decay model. Therefore, this law is used here, and only the residual signal in the quasi-steady-state segment is simplified and fitted without complex derivation. In the formula: R(t) represents the residual response value at time t, that is, the residual physiological optical noise intensity at that time; A represents the residual noise amplitude at the beginning of the quasi-steady-state segment, which is derived from the parameters of the rapid relaxation phase in the relaxation fingerprint vector; τ represents the relaxation time of the physiological optical noise, which is directly adopted from the relaxation time parameter extracted from the rapid relaxation phase in the relaxation fingerprint vector; t represents the time in the quasi-steady-state segment, which is calculated with the beginning of the quasi-steady-state segment as t=0.

[0049] After the residual response function is constructed, an inverse compensation operation is performed. The logic of the inverse compensation is to subtract the corresponding residual response function value and noise floor level from the actual intensity value at each time point within the quasi-steady-state segment of the time-series intensity curve, thereby removing the dual interference of physiological optical noise residual signal and weak random noise, to obtain the spectral intensity value after removing all interference, i.e., the initial value of instantaneous spectral intensity. In specific operation, for each frame of data within the quasi-steady-state segment, its corresponding initial value of instantaneous spectral intensity is equal to the actual time-series intensity value of that frame, minus the corresponding R(t) value of that frame, and then minus the noise. The substrate level ensures that interference has been removed from each frame of data. Since the signal in the quasi-steady state is basically stable and the residual response function and noise substrate level have been accurately quantized, the initial value of the instantaneous spectral intensity obtained through this inverse compensation operation is very close to the true static spectral reflectance characteristics of the pixel. It can initially reflect the characteristics of the surface optical substrate. Each pixel completes the inverse compensation according to the above process to obtain its own initial value of instantaneous spectral intensity. The initial values ​​of all pixels together constitute the initial data basis for subsequent iterative optimization, ensuring that the iterative optimization can quickly converge to the true value.

[0050] Step 33: Using the initial instantaneous spectral intensity of each pixel as the initial estimation field, iterative optimization is performed based on the relaxed fingerprint vector to generate a static optical base map of the face. The specific operations are as follows: The initial instantaneous spectral intensity values ​​of each pixel obtained in step 32 are used as the initial estimation field. Guided by the relaxation fingerprint vector, iterative optimization is performed to correct residual errors in the initial values, ensuring that the spectral reflectance characteristics of each pixel accurately correspond to the facial optical substrate. Finally, the optimized results of all pixels are combined to generate a complete facial static optical substrate map. The initial estimation field is the set of initial instantaneous spectral intensity values ​​of all pixels. Although most interference has been removed through inverse compensation, there may still be a small amount of residual error caused by residual response function fitting deviation and noise substrate level calculation error. These errors will affect the accuracy of the static optical substrate map, so it is necessary to further correct them through iterative optimization to ensure the accuracy of the static spectral reflectance characteristics. The iterative optimization is guided by the relaxation fingerprint vector, where the parameters of the slow relaxation stage and steady-state stage correspond to the stable relaxation characteristics of the facial optical substrate. These parameters have long-term stability and can provide a clear convergence direction for iterative optimization, avoiding deviations in the optimization process. At the same time, there is no need to introduce an additional optimization model; it is only based on the stable characteristics of the signal for gradual correction, ensuring reproducibility.

[0051] The iterative optimization follows a clear process and termination condition. First, the initial instantaneous spectral intensity of each pixel is used as the initial value for optimization to construct an initial estimation field. The value of each pixel in the initial estimation field corresponds to its preliminary static spectral reflectance characteristics. The pixels are arranged according to their facial coordinates to form an initial substrate map prototype. Each round of iterative optimization is guided by the relaxation fingerprint vector. Specifically, the decay rate, steady-state intensity, and steady-state intensity parameters of the slow relaxation phase are extracted from the relaxation fingerprint vector. These parameters characterize the stability of the surface optical substrate of the pixel. The current spectral intensity value of each pixel is compared with the corresponding steady-state intensity parameter in the relaxation fingerprint vector, and the difference between the two is calculated. If the difference is greater than a preset error threshold, the current spectral intensity value is corrected. The correction direction is to move closer to the steady-state intensity parameter. The correction magnitude is determined according to the difference and the decay rate of the slow relaxation phase. The larger the difference and the smaller the decay rate, the larger the correction magnitude, and vice versa, ensuring that the correction process conforms to the stable relaxation law of the surface optical substrate.

[0052] Two termination conditions are set for iterative optimization. The iteration will terminate when either condition is met, ensuring the efficiency and accuracy of the optimization process. The first termination condition is that after two consecutive iterations, the change in the spectral intensity value of all pixels is less than the preset change threshold. The change threshold is usually set to 0.5% of the average intensity of all pixels in the initial estimation field. This threshold ensures that the iterative optimization has converged and further iterations cannot significantly improve the accuracy. The second termination condition is that the number of iterations reaches the preset maximum number of iterations. The maximum number of iterations is usually set to 20 to avoid infinite loops caused by errors in individual pixels. At the same time, 20 iterations are sufficient to achieve convergence, balancing efficiency and accuracy. After the iteration terminates, the spectral intensity value corresponding to each pixel is the final static spectral reflectance characteristic of that pixel. This characteristic has completely eliminated the interference of physiological optical noise and random noise, accurately reflecting the long-term stable characteristics determined by the color of subcutaneous tissue and the texture structure of the dermis, which fully meets the needs of traditional Chinese medicine facial diagnosis.

[0053] Finally, the static spectral reflectance characteristics of all pixels are arranged and combined one by one according to their coordinate positions in the facial image to form a complete static optical base map of the face. This base map can clearly and accurately present the long-term, basic facial color and texture characteristics of the entire face. The differences in static spectral reflectance characteristics of different facial regions correspond to the pathological characteristics of different organ regions in traditional Chinese medicine facial diagnosis, solving the problem of temporary skin condition interfering with the diagnosis conclusion in the existing technology.

[0054] In a preferred embodiment of the present invention, step 4 is further included: calculating the pixel-level spectral differences between the facial static optical base map and the baseline frames and quasi-steady-state frames in the facial hyperspectral image sequence; and generating a reliability heatmap of the facial static optical base map based on the spectral differences and relaxation feature parameters. The specific operations are as follows: The baseline frame is a hyperspectral image of the face acquired before the physical stimulus is applied in step 1. Its pixel spectral features correspond to the initial steady-state state of the skin without any stimulation or physiological optical noise interference, and it is highly consistent with the theoretical features of the facial static optical base map. The quasi-steady-state frame is a representative frame within the quasi-steady-state segment determined in step 31. The middle frame of the quasi-steady-state segment is selected as the quasi-steady-state frame. Its pixel spectral features have been largely stripped of transient physiological optical noise, containing only facial optical base signals and weak random noise. It has a strong correlation with the actual derivation results of the facial static optical base map. By calculating the pixel-level spectral differences between the facial static optical base map and these two frames, the errors that may exist in the base map derivation process can be accurately captured. The smaller the difference, the closer the base map is to the real facial optical base, and the higher the reliability.

[0055] The calculation of pixel-level spectral differences needs to be performed independently for each pixel, focusing on the complete spectral features within a preset band, rather than the intensity difference of a single wavelength. This ensures the comprehensiveness and accuracy of the difference assessment, because traditional Chinese medicine facial diagnosis relies on the overall spectral reflectance characteristics of the face. The intensity difference of a single wavelength cannot reflect the complete features of facial color. After the calculation is completed, the rationality of the difference needs to be further analyzed in conjunction with relaxation feature parameters. The relaxation fingerprint vector in the relaxation feature parameters contains the relaxation pattern of each pixel, which can distinguish whether the difference is due to the error of the base map inference or the normal physiological difference of the skin itself, providing quantitative support for the credibility assessment. First, the spectral differences are separated into components, then a dynamic consistency index is constructed, and finally, through spatial propagation and conflict resolution, the correlation between spectral differences and relaxation feature parameters is transformed into the credibility value of each pixel, which is arranged by pixel coordinates to form a credibility heatmap.

[0056] The generation of the credibility heatmap also includes the following steps: Step 41: Calculate the pixel-level spectral angular distance between the facial static optical base map and the baseline frame and quasi-steady-state frame, generate the total difference map, and perform linear trend fitting and residual analysis on the total difference map to separate the linear difference components and nonlinear difference components. The specific operations are as follows: Spectral angular distance is a classic method for quantifying the similarity between two spectral vectors. Compared to simple spectral intensity difference, it more accurately reflects the overall similarity of spectral features and is unaffected by global interference factors such as illumination intensity. It is suitable for evaluating the spectral differences between a static optical base map and a baseline frame and a quasi-steady-state frame, because there may be slight global illumination differences among the three. Simple intensity difference can exaggerate or reduce the actual differences in spectral features, while spectral angular distance can effectively avoid this problem. The calculation of pixel-level spectral angular distance needs to be performed independently for the spectral vector of each pixel. The specific operation is as follows: For each pixel, extract the spectral reflectance data of the facial static optical base map, baseline frame, and quasi-steady-state frame in a preset wavelength band to form three spectral vectors with the same dimension. The preset wavelength band adopts the previously selected key spectral range of 400 to 1000 nm. The dimension of each spectral vector is equal to the preset wavelength. The number of wavelengths within a segment ensures the integrity of spectral features. The formula for calculating the spectral angular distance is: the spectral angular distance θ equals the dot product of vectors a and b divided by the product of the magnitudes of the two vectors, plus the inverse cosine function value. The derivation logic of this formula is: consider the two spectral vectors as vectors in a high-dimensional space, and the spectral angular distance is the angle between the two vectors. The smaller the angle, the more similar the two spectral vectors are, and the smaller the difference. According to the definition of the vector dot product, the vector dot product equals the product of the magnitudes of the two vectors and the cosine of the angle. By rearranging terms and performing inverse cosine operations, the formula for calculating the spectral angular distance can be derived. This derivation process conforms to the basic mathematical laws of high-dimensional vector analysis. Here, θ represents the spectral angular distance between the two spectral vectors, in radians; vector a represents the spectral vector of a pixel in the static optical base image of the face; vector b represents the spectral vector of the corresponding pixel in the baseline frame or quasi-steady frame.

[0057] The spectral angular distance between the static optical base map of the face and the baseline frame, and the spectral angular distance between the static optical base map of the face and the quasi-steady-state frame are calculated for each pixel. The arithmetic mean of the two spectral angular distances is taken to obtain the comprehensive spectral difference value of the pixel. The comprehensive spectral difference values ​​of all pixels are arranged according to their facial coordinate positions to form a total difference map. The total difference map visually presents the spectral difference distribution between the static optical base map and the two reference frames within the entire facial area. After generating the total difference map, linear trend fitting and residual analysis are performed on it to separate two essentially different difference components. This is because the differences in the total difference map may originate from different factors such as systematic errors and physiological differences, and different components have different meanings for credibility assessment. The linear trend fitting adopts the least squares method, with the coordinates of the pixel as the independent variable and the corresponding comprehensive spectral difference value as the dependent variable, to obtain the linear trend surface of the total difference map. This linear trend surface reflects the component of the total difference that changes linearly with the coordinates, i.e., the linear difference component. This type of difference usually originates from global systematic interference such as system installation errors and illumination gradients, and is unrelated to the base features of the pixel itself, and does not belong to the base map inference error.

[0058] Residual analysis refers to calculating the difference between the actual comprehensive spectral difference value of each pixel and the fitted value of the linear trend surface. This difference is the residual. The residuals of all pixels are arranged according to their coordinates to form a residual map. The difference components corresponding to the residual map are the nonlinear difference components. The nonlinear difference components originate from the feature differences of the pixels themselves, including the extrapolation error of the static optical base map, the local physiological differences of the skin itself, and a small amount of unexfoliated physiological optical noise residue. These differences are directly related to the reliability of the static optical base map and are the focus of subsequent credibility assessment. The separation of linear difference components and nonlinear difference components can effectively eliminate the influence of global systematic interference on credibility assessment, ensuring that subsequent analysis only focuses on nonlinear differences related to the reliability of the base map, thereby improving the accuracy of the credibility heatmap.

[0059] Step 42: Based on the relaxed fingerprint vector and nonlinear difference components, construct a dynamic consistency index map. The specific operations are as follows: The relaxed fingerprint vector contains the complete relaxed feature parameters of each pixel. The parameters in the slow relaxation and steady-state phases correspond to the stable features of the facial optical substrate, exhibiting strong spatial consistency. That is, if adjacent pixels belong to the same facial region, the corresponding stable parameters in their relaxed fingerprint vectors should have high similarity. In contrast, the parameters in the rapid rise and rapid relaxation phases correspond to the transient features of physiological optical noise, exhibiting weaker spatial consistency. The nonlinear difference component is related to the pixel's own substrate feature differences and inference errors. The correlation between these two factors directly reflects the rationality of the nonlinear difference, thereby quantifying the reliability of the substrate map. The construction of the dynamic consistency index requires independent calculation for each pixel. The core is to quantify the consistency in two dimensions: first, the consistency between the pixel's own relaxation pattern and the nonlinear difference; second, the consistency between the pixel's own relaxation pattern and the nonlinear difference. The consistency of the relaxation pattern and nonlinear differences between this pixel and its neighboring pixels determines the dynamic consistency index value of this pixel. The higher the index value, the more reasonable the nonlinear difference of this pixel is, and the higher the reliability of the static optical substrate map. The calculation logic of self-consistency is as follows: extract the decay rate and steady-state intensity parameters of the slow relaxation stage in the relaxation fingerprint vector, and calculate the correlation between the nonlinear difference value of this pixel and these stable parameters. If the correlation is low, it means that the nonlinear difference does not originate from the change of the substrate features, and is likely caused by the inference error, resulting in poor self-consistency. If the correlation is high, it means that the nonlinear difference originates from the reasonable difference of the substrate features themselves, resulting in good self-consistency. The correlation calculation uses the Pearson correlation coefficient, which is simple and reproducible, and does not require the introduction of an additional model.

[0060] The calculation logic for neighbor consistency is as follows: For each pixel, select its 8 neighboring pixels as the neighbor set. The 8 neighbors are the 8 pixels adjacent to the pixel in the vertical, horizontal, and diagonal directions, ensuring coverage of the main area surrounding the pixel while avoiding excessive selection that would increase computational complexity. For each pair of adjacent pixels, calculate the similarity of stable parameters and nonlinear difference values ​​in their relaxed fingerprint vectors. Then, take the arithmetic mean of the two similarity values ​​as the neighbor consistency value for that pair. Finally, calculate the average of the neighbor consistency values ​​of the pixel and all 8 neighboring pixels as the neighbor consistency index for that pixel. The relaxation fingerprint similarity is calculated using the cosine similarity method, similar to the calculation logic of the spectral angular distance in step 41. Consistency is quantified by assessing the similarity of stable parameters in two relaxed fingerprint vectors. Nonlinear difference similarity is calculated using the normalized absolute value of the difference; the smaller the difference, the higher the similarity. The dynamic consistency index value of each pixel is obtained by taking the arithmetic mean of its own consistency index and the consistency index of its neighbors with a 1:1 weight. The weighting is based on the equal importance of both for reliability assessment, requiring no additional optimization to ensure reproducibility. The dynamic consistency index values ​​of all pixels are arranged according to their facial coordinates to form a dynamic consistency index map. The value of each pixel in this map quantifies the reasonableness of its own differences and those between it and its neighbors; high index values ​​correspond to high reasonableness and high reliability, while low index values ​​correspond to low reasonableness and low reliability.

[0061] Step 43: Using the dynamic consistency index map as the initial seed, and combining the spatial gradient of the nonlinear difference components and the relaxation fingerprint similarity between adjacent pixels, anisotropic spatial propagation and conflict resolution are performed to generate a confidence heatmap. The specific operations are as follows: Using the dynamic consistency index map constructed in step 42 as the initial seed, combined with the spatial gradient of the nonlinear difference components separated in step 41 and the relaxation fingerprint similarity between adjacent pixels, anisotropic spatial propagation and conflict resolution are performed to correct the initial dynamic consistency index value, ultimately generating a credibility heatmap that accurately reflects the reliability of each pixel. Although the initial dynamic consistency index map has quantified the rationality of pixel differences, there may be isolated high or low index values. These isolated points usually originate from local calculation errors rather than the true reliability of the pixel itself. Spatial propagation can utilize the correlation between adjacent pixels to correct such errors and improve the spatial continuity of the index map. Conflict resolution is used to solve the problem of excessive differences in the dynamic consistency index of adjacent pixels, ensuring that the corrected index value conforms to the spatial distribution law of facial skin features. Because the relaxation features and basal features of facial skin have strong spatial continuity, the reliability of adjacent pixels should not have significant differences.

[0062] The specific operation of anisotropic spatial propagation follows clear logic and parameter settings to ensure reproducibility. The propagation process uses a dynamic consistency index map as the initial seed, meaning that the initial propagation value of each pixel is its dynamic consistency index value. The propagation weight is determined by two factors: first, the spatial gradient of the nonlinear difference component. The spatial gradient quantifies the rate of change of nonlinear difference values ​​between adjacent pixels. The smaller the gradient, the smoother the change of nonlinear difference between adjacent pixels, the stronger the spatial correlation, and the larger the propagation weight; conversely, the larger the gradient, the smaller the propagation weight. Second, the relaxation fingerprint similarity between adjacent pixels. The higher the similarity, the closer the basis features of adjacent pixels, and the stronger the correlation of the reliability assessment results; the larger the propagation weight, and vice versa. The propagation weight is calculated by multiplying the two factors and then normalizing them to ensure that the weight value is between 0 and 1. The normalization process uses a linear normalization method to map the weight value to a preset range, avoiding excessively large or small weights that could affect the propagation effect.

[0063] The propagation process employs an iterative approach, with 10 iterations to balance propagation effectiveness and computational efficiency. Ten iterations are sufficient to correct isolated point errors and improve spatial continuity. Too many iterations would lead to uniformity of index values ​​and loss of local reliability differences, while too few iterations would fail to effectively correct errors. During each iteration, the dynamic consistency index value of each pixel is updated to the weighted sum of its initial value and the index values ​​of its neighboring pixels. The weights are the anisotropic propagation weights calculated above, ensuring that the propagation process conforms to the spatial distribution of facial skin and only propagates effectively between pixels with strong correlation. After spatial propagation is complete, conflict resolution is required to resolve conflicts where the index values ​​of adjacent pixels differ too much. The conflict judgment threshold is set to 0.8 times the initial average value of the dynamic consistency index map. When the difference between the index values ​​of two adjacent pixels exceeds this threshold, a conflict is determined to exist.

[0064] The specific logic of conflict resolution is as follows: extract the relaxed fingerprint vector of the conflicting pixel pair, calculate the similarity of their stable parameters. If the similarity is high, it means that their basic features are similar and their reliability should be consistent. In this case, the arithmetic mean of the two pixel index values ​​is taken as the corrected value. If the similarity is low, it means that there are real differences in their basic features and their reliability may have reasonable differences. In this case, their respective propagated index values ​​are retained without forced correction to avoid destroying the real reliability differences due to conflict resolution. After conflict resolution, a corrected dynamic consistency index map is obtained. Each index value in the index map is normalized and mapped to a confidence value between 0 and 1. The higher the index value, the higher the confidence value. The confidence values ​​are then arranged according to pixel coordinates and combined with color gradient mapping to generate a confidence heatmap. The color changes from blue (low confidence) to red (high confidence) to visually present the reliability of the static optical basis map of each pixel.

[0065] In a preferred embodiment of the present invention, step 5 is further included: matching the static spectral reflectance characteristics of each pixel in the facial static optical base image with a predefined reference facial color spectral reflectance curve, and combining this with a confidence heatmap to classify and label the facial pixels to generate a facial color classification symbol map. The specific operations are as follows: The static spectral reflectance characteristics of the facial static optical base map have been completely freed from physiological optical noise interference, accurately reflecting the long-term stable characteristics of subcutaneous tissue color, which is highly consistent with the facial color characteristics required for TCM face diagnosis. The predefined benchmark facial color spectral reflectance curve is a typical spectral curve obtained by collecting a large number of standard facial color samples based on the TCM five-color theory, namely green, red, yellow, white and black, and after statistical averaging and standardization. Each benchmark facial color corresponds to a unique spectral reflectance curve, covering the complete spectral characteristics within the preset band, ensuring the professionalism and accuracy of facial color matching.

[0066] The face color matching process requires independent processing of each pixel. It quantifies the similarity between the static spectral reflectance characteristics of the pixel and the spectral reflectance curves of each reference face color. Then, it combines the reliability of the pixel in the confidence heatmap to select reasonable matching results and avoid the matching errors of unreliable pixels from interfering with the classification accuracy. Since facial colors have transitional and ambiguous characteristics and are not absolute single reference colors, directly determining a single category will lead to classification bias. First, fuzzy membership degrees are calculated to generate an initial probability distribution map. Then, local biases are corrected through collaborative correction, high-conflict areas are identified and processed, and finally, confidence is combined to make decisions, gradually optimizing the classification results to ensure the accuracy and rationality of face color classification.

[0067] The generation of the facial color classification symbol map also includes the following steps: Step 51: Calculate the fuzzy membership degree of the static spectral reflectance characteristics to the spectral reflectance curve of each predefined reference surface color, and generate an initial surface color probability distribution map. The specific operations are as follows: Fuzzy membership degree is used to characterize the degree to which the static spectral reflectance characteristics of a pixel belong to a certain reference surface color. The value ranges from 0 to 1. The closer the membership degree value is to 1, the higher the similarity between the pixel's spectral features and the corresponding reference surface color, and the greater the probability that it belongs to that reference surface color. The closer the membership degree value is to 0, the lower the similarity and the smaller the probability. The sum of the fuzzy membership degrees of each pixel to all reference surface colors is 1, which conforms to the basic logic of fuzzy classification, accurately characterizing the transition and fuzziness of surface colors, and avoiding the bias caused by absolute classification. The calculation of fuzzy membership degree uses the Gaussian fuzzy membership degree function, which can well fit the fuzzy distribution law of surface color similarity and is a classic method in the field of fuzzy classification. Its calculation is as follows: The membership degree μi(x) of pixel x to the i-th reference surface chromatogram is equal to the value of an exponential function with the natural constant e as the base. The square of the Euclidean distance d(x,ci) between pixel x and reference surface chromatogram ci with a negative exponent is divided by twice the square of the variance σi. The derivation logic of this calculation method is that, based on the distribution characteristics of spectral similarity, the smaller the difference between the pixel spectrum and the reference surface chromatogram, the larger the membership degree should be, and the difference and membership degree have a negative exponential relationship. The Gaussian function has a smooth bell-shaped distribution characteristic, which can accurately simulate this fuzzy relationship where the smaller the difference, the larger the membership degree. The influence of the difference is strengthened by the squared difference, and the rate of change of the membership degree is adjusted by the standard deviation. Finally, the Gaussian fuzzy membership degree function is derived. Where μi(x) represents the fuzzy membership degree of the static spectral reflectance characteristics of pixel x to the i-th reference surface color; x represents the pixel to be calculated; ci represents the spectral reflectance curve of the i-th reference surface color, i.e., in vector form; d(x,ci) represents the Euclidean distance between the static spectral vector of pixel x and the spectral vector of the i-th reference surface color, used to quantify the spectral difference between the two; the smaller the distance, the higher the similarity; σ The variance of the membership function for the i-th reference face color is used to adjust the rate of change of membership. It is obtained by statistical analysis of the spectral data of the predefined reference face color, specifically the average variance of the spectral vector of the standard sample of the reference face color. This ensures that the variance setting fits the spectral distribution characteristics of the corresponding reference face color. Each reference face color corresponds to a unique variance value, which does not require additional optimization or adjustment.

[0068] The calculation process is as follows: For each pixel, its static spectral reflectance characteristics are extracted to form a spectral vector. Then, the Euclidean distance between this vector and the spectral vector of each reference face color is calculated. The distance value is substituted into the Gaussian fuzzy membership function mentioned above to calculate the fuzzy membership degree of the pixel to each reference face color. After the calculation is completed, an initial face color probability distribution map is generated. This distribution map contains the same number of sub-maps as the reference face colors. Each sub-map corresponds to one reference face color. The value of each pixel in the sub-map is the fuzzy membership degree of that pixel to the corresponding reference face color. They are arranged according to facial coordinate positions to visually present the initial probability distribution of each pixel belonging to each reference face color.

[0069] Step 52: Based on the relaxed fingerprint vector and the spatial neighborhood information of pixels, the initial face color probability distribution map is collaboratively corrected to obtain the corrected face color probability distribution map. The specific operation is as follows: The initial facial color probability distribution map calculates membership degrees based solely on the spectral features of individual pixels, neglecting the spatial continuity of facial skin features. This can lead to situations where the membership degrees of isolated pixels differ significantly from those of neighboring pixels. Such deviations typically originate from local noise or computational errors in the pixel's spectral data, rather than representing true facial color features, and require correction through collaborative calibration. Collaborative calibration is based on two factors: the relaxation fingerprint vector and the spatial neighborhood information of the pixel. These two factors work together to ensure that the calibration results both conform to the pixel's own spectral features and the spatial distribution patterns of facial skin. In the relaxation fingerprint vector, stable parameters such as the decay rate and steady-state intensity during the slow relaxation phase are highly correlated with the static spectral reflectance characteristics of the face, and thus have a clear association with facial color features. That is, pixels with similar relaxation fingerprint vectors should also have high similarity in their facial color features, and their membership degree distributions should tend to be consistent. The spatial neighborhood information is based on the physiological characteristics of facial skin. Adjacent pixels have similar skin structures and subcutaneous tissue distributions, resulting in strong spatial continuity of facial color features. Therefore, their membership degree distributions should not show significant differences. This is the logic behind correcting local isolation biases.

[0070] The specific correction operation follows a clear process and parameter settings to ensure reproducibility. First, the eight neighboring pixels of each pixel are selected as the set of adjacent pixels. The eight neighbors are the eight adjacent pixels in the vertical, horizontal, and diagonal directions of the pixel. This not only fully covers the skin area around the pixel, reflecting spatial continuity, but also avoids increasing computational complexity due to selecting too many neighboring pixels. Then, for each pixel, its own and the relaxation fingerprint vectors of its eight neighboring pixels are extracted. The similarity between the pixel and the relaxation fingerprints of each neighboring pixel is calculated. The similarity calculation uses the cosine similarity method to quantify the similarity of their stable parameters. The higher the similarity value, the stronger the correlation between their facial color features, and the greater the weight during correction. Subsequently, the co-correction weight of the pixel is calculated, and the weights are determined by... The confidence level of the self-relaxed fingerprint and the similarity of adjacent pixels are jointly determined. The self-weight is set to 0.6, and the total weight of adjacent pixels is set to 0.4. The weights are normalized according to the similarity of the relaxed fingerprints of adjacent pixels. The higher the similarity, the greater the weight of the adjacent pixels. This ensures that the correction process prioritizes the reasonableness of the self-spectral features while fully considering the reasonable distribution of adjacent pixels. Finally, the fuzzy membership degree of the pixel after correction is calculated according to the weight, which is the weighted sum of the initial membership degree of the self and the initial membership degree of each adjacent pixel. The membership degree of each reference face color is corrected independently according to this logic. After all pixels are corrected, the corrected face color probability distribution map is obtained. This map has corrected the deviation of local isolated pixels, significantly improved spatial continuity, and the membership degree distribution is more in line with the real face color features.

[0071] Step 53: Identify high-conflict regions in the corrected fading probability distribution map and reclassify them in the low-dimensional spectral subspace using a confidence heatmap to generate an updated fading probability distribution map. The specific operations are as follows: High-conflict regions refer to areas in the corrected color probability distribution map where the fuzzy membership degree of a pixel to all reference colors is relatively small, and the difference between the maximum and second-largest membership degree is less than a preset conflict threshold, making it impossible to clearly determine its color category. These regions typically appear in transition areas between two or more reference colors, or in areas with concentrated pixels of low confidence. Directly retaining the initial correction result will lead to subsequent classification errors. The identification of high-conflict regions needs to be performed independently for each pixel. The conflict threshold is set based on the global membership degree difference statistics of the corrected color probability distribution map, specifically the average of the difference between the maximum and second-largest membership degree of all pixels. This ensures that the threshold setting fits the overall membership degree distribution characteristics, neither missing real conflict regions nor misclassifying non-conflict regions as high-conflict regions. During the identification process, the difference between the maximum and second-largest fuzzy membership degree of each pixel is calculated one by one. If the difference is less than the conflict threshold, the pixel is determined to be a high-conflict pixel. All high-conflict pixels are arranged by coordinates to form high-conflict regions. High-conflict regions may be isolated pixels or continuous regions, requiring unified reclassification processing.

[0072] Reclassification is performed within a low-dimensional spectral subspace to reduce redundancy in high-dimensional spectral data, highlight core differences in spectral features, and improve the accuracy of similarity calculations. Simultaneously, a confidence heatmap is used to guide reclassification of high-confidence regions by leveraging the reasonableness of high-confidence pixel classifications. The low-dimensional spectral subspace is constructed using principal component analysis. Specifically, the static spectral vectors of all pixels within the high-conflict region and the spectral vectors of each reference surface color are subjected to principal component analysis. The first two principal components are retained to construct a two-dimensional low-dimensional spectral subspace. These first two principal components retain over 95% of the core features of the original spectral data, balancing dimensionality reduction with feature integrity.

[0073] During reclassification, the spectral vector of each pixel in a high-conflict region is projected onto a low-dimensional spectral subspace, and the spectral vectors of each reference surface color are also projected onto this subspace. The Euclidean distance between the projected vector of the pixel and the projected vectors of each reference surface color is calculated to quantify the similarity. Simultaneously, the confidence values ​​of the pixel and its eight adjacent pixels in the confidence heatmap are extracted. High-confidence pixels with a confidence value ≥ 0.8 are selected, and their membership distribution in the low-dimensional subspace is referenced to perform weighted correction on the similarity calculation results of high-conflict pixels. High-confidence pixels have higher weights, guiding them towards a reasonable category. After correction, the fuzzy membership of high-conflict pixels to each reference surface color is recalculated, and their membership values ​​are updated. After all high-conflict regions are reclassified, an updated color probability distribution map is obtained. This map has resolved the classification bias in high-conflict regions, the membership distribution is more reasonable, and the classification accuracy is further improved.

[0074] Step 54: Based on the confidence heatmap, make a decision on the updated face color probability distribution map and generate a face color classification symbol map. The specific operations are as follows: The updated color probability distribution map has corrected local biases and classification issues in high-conflict areas. The fuzzy membership degree of each pixel to each baseline color is now more reasonable. However, it is still necessary to combine it with the confidence heatmap to screen reliable classification results and avoid the classification errors of low-confidence pixels affecting the overall classification quality, ensuring the accuracy and practicality of the color classification symbol map. The decision-making process is based on the confidence value of each pixel in the confidence heatmap, formulating clear decision rules to ensure that the decision logic is reproducible and unambiguous. First, a confidence threshold is set, which is determined based on the global confidence statistics of the confidence heatmap and is usually set to 0.7. This threshold can effectively distinguish between high-confidence and low-confidence pixels. Pixels with a confidence value ≥ 0.7 have higher reliability in their static optical base map and a reasonable membership degree distribution, and the category can be directly decided based on the membership degree. Pixels with a confidence value < 0.7 have lower reliability, and their membership degree distribution may have errors. The category of adjacent high-confidence pixels needs to be referenced for decision-making to avoid bias caused by direct decision-making.

[0075] The specific decision-making rules are divided into two cases. The first case is when the pixel's confidence value is ≥0.7. In this case, the reference face color with the highest fuzzy membership degree in the updated face color probability distribution map is directly selected as the final face color category of the pixel. If there are two or more reference face colors with the same membership degree and both are the highest value, the reference face color with the highest similarity to the relaxed fingerprint of the pixel is selected as the final category to ensure the rationality of the decision result. The second case is when the pixel's confidence value is <0.7. In this case, the decision is not made directly based on its own membership degree. Instead, all high confidence pixels with a confidence value ≥0.7 in the 8-neighborhood of the pixel are selected, and the final category of these high confidence pixels is counted. The category with the highest frequency is selected as the final category of the pixel. If there are no high confidence pixels in the neighborhood, it is temporarily marked as an undetermined category.

[0076] After each pixel completes its category decision, it is labeled with simple and easy-to-understand symbols. Each baseline facial color corresponds to a unique symbol, which can be Arabic numerals, such as 1 to 5, corresponding to the five baseline facial colors. The labeling process must ensure that the symbol of each pixel corresponds one-to-one with the final facial color category without confusion. After all pixels are labeled, the labels are arranged according to their coordinate positions in the facial image to form a facial color classification symbol map. This map clearly presents the distribution of facial color categories across the entire face, with each symbol representing the facial color category at the corresponding position. This preserves the facial color characteristics required for traditional Chinese medicine facial diagnosis while also achieving standardization and objectification of facial color classification.

[0077] In a preferred embodiment of the present invention, step 6 is further included: based on a preset facial organ partitioning map, the pixels of each partition in the facial color classification symbol map are weighted and statistically analyzed according to their facial color classification and the corresponding credibility in the credibility heatmap to generate a diagnostic vector. The specific operation is as follows: The pre-defined facial organ partitioning map is a standardized partitioning template based on the theory of organ correspondence in traditional Chinese medicine. It clearly delineates the specific areas on the face that correspond to the five internal organs, such as the forehead corresponding to the heart, the cheeks on both sides corresponding to the liver, the tip of the nose corresponding to the spleen, and the chin corresponding to the kidneys. The boundaries of each partition are standardized and precisely defined using pixel coordinates to ensure that the partitioning positions of different subjects are consistent and highly repeatable.

[0078] The logic of weighted statistics is to combine the confidence value of each pixel to highlight the contribution of high-confidence pixels to the facial color classification statistics, reduce the interference of low-confidence pixels, and ensure that the statistical results can truly reflect the facial color characteristics of each organ region, which is in line with the requirements of TCM diagnosis for the authenticity of facial color. The facial color classification symbol map has clearly defined the final facial color category of each pixel, while the confidence heatmap quantifies the reliability of each pixel. The two work together to make the weighted statistical results have both classification accuracy and reliability discrimination. Subsequently, by extracting the core facial color topological features of each organ region, the scattered pixel-level classification information is transformed into a region-level quantitative feature vector. Then, by combining the pre-set TCM organ mutual generation and restraint and five-color matching rule library, the feature vector is analyzed dialectically, transforming the technical facial color features into TCM theoretical diagnostic conclusions, forming a structured diagnostic vector. This realizes a complete closed loop from facial image acquisition to TCM diagnostic conclusion output, effectively solving the pain points of strong subjectivity and poor consistency in traditional TCM facial diagnosis.

[0079] The process of generating dialectical vectors also includes the following steps: Step 61: For each organ region, based on the facial color classification symbol map and the confidence heatmap, calculate the confidence-weighted area ratio of each facial color, the spatial clustering degree of each facial color, and the spatial cross-correlation between different facial colors, and combine them into a multi-dimensional facial color topological feature vector. The specific operation is as follows: Based on the facial color classification symbol map generated in step 5 and the credibility heatmap generated in step 4, the credibility weighted area ratio of each facial color, the spatial clustering degree of each facial color, and the spatial cross-correlation between different facial colors are calculated one by one. These three types of quantitative features are combined in a fixed order to form a multi-dimensional facial color topological feature vector exclusive to each organ region. The three types of features comprehensively depict the facial color features of each organ region from the three dimensions of area distribution, spatial clustering, and interrelation, covering the information required for TCM facial diagnosis, and ensuring that the feature vector can accurately reflect the facial color manifestations related to the physiological and pathological state of the corresponding organ in that region.

[0080] The calculation of the confidence-weighted area ratio for each face color is achieved by quantifying the proportion of different face color categories within each organ region using a confidence-weighted method, highlighting the contribution of high-confidence pixels. The specific calculation process is as follows: For a single organ region, firstly, all pixels within that region are selected, excluding irrelevant pixels outside the region to ensure the accuracy of the statistical range. For each preset baseline face color, the sum of the confidence values ​​of all pixels belonging to that face color category within that region is calculated; this sum is the total confidence-weighted pixel count for that face color within that region. Then, the sum of the confidence values ​​of all pixels within that region is calculated as the total confidence-weighted pixel count for that region. Finally, the total confidence-weighted pixel count for each face color is divided by the total confidence-weighted pixel count for that region, resulting in the confidence-weighted area ratio for that face color within that region. The calculation is as follows: the weighted area ratio ri is equal to the sum of the confidence values ​​w(x) of all pixels x in the pixel set Si belonging to the baseline face color i, divided by... The formula is derived by summing the confidence values ​​w(x) of all pixels x in the set S of all pixels within the visceral region. The logic behind this formula is that traditional area proportions do not consider differences in pixel reliability, and low-confidence pixels may interfere with the statistical results. Therefore, confidence is introduced as a weight, giving high-confidence pixels a higher weight in the area proportion calculation. By calculating the ratio of the weighted total number of pixels of a certain face color to the total weighted number of pixels in the region, the actual proportion of that face color in the region can be accurately quantified, which aligns with the focus on the true characteristics of face color in traditional Chinese medicine diagnosis. This leads to the derivation of the weighted area proportion formula, where ri represents the confidence-weighted area proportion of the i-th reference face color in the current visceral region; Si represents the set of all pixels belonging to the i-th reference face color in the current visceral region; w(x) represents the confidence value of pixel x, taken from the confidence heatmap; S represents the set of all pixels in the current visceral region; the numerator is the total weighted number of pixels of that face color, and the denominator is the total weighted number of pixels in the region. Each reference face color is independently weighted for its area ratio, and the sum of the ratios of all reference face colors is 1, ensuring the rationality and completeness of the calculation results.

[0081] The calculation of spatial clustering degree for each facial color is used to quantify the spatial concentration of pixels of the same facial color within each organ region. In traditional Chinese medicine facial diagnosis, the degree of clustering of the same facial color is often related to the pathological state of the organ. The higher the clustering degree, the more it reflects the specific manifestation of that organ. For each type of facial color within each organ region, connected component analysis is used to identify all interconnected pixel regions of that facial color. The criterion for determining a connected region is that there are pixels of the same facial color in the eight neighboring regions. That is, if any one of the eight pixels above, below, to the left, right, and diagonally of a certain pixel has the same facial color category as that pixel, then the two belong to the same connected region. The number of connected regions for that facial color is counted, and the number of pixels in each connected region is calculated, which is the sum of the weighted confidence values. The average of the weighted pixel counts of all connected regions is taken, and the number of connected regions is divided by the average value. The result is the spatial clustering degree of that facial color within that region. The smaller the value, the higher the spatial clustering degree of that facial color, and the more concentrated the pixels are; the larger the value, the lower the clustering degree, and the more dispersed the pixels are.

[0082] Spatial cross-correlation between different facial colors is used to quantify the degree of spatial adjacency between pixels of two different facial colors within each organ region, reflecting the distributional correlation of different facial colors and providing a quantitative basis for multi-facial color correlation analysis in TCM diagnosis. For any two different baseline facial colors within an organ region, the number of pixels belonging to the other facial color within the 8-neighborhood of one facial color pixel is counted (i.e., the sum of weighted confidence values), and then divided by the total weighted pixel count of that facial color to obtain the one-way spatial cross-correlation between the two facial colors. The one-way correlation between the two facial colors is then calculated. The average value is used as the spatial cross-correlation of these two types of facial colors. The correlation is calculated independently for all different combinations of facial colors. If the number of baseline facial colors is n, then n(n-1) / 2 sets of cross-correlation need to be calculated to ensure comprehensive coverage of the association relationships of various facial colors. The weighted area ratio of all facial colors in each organ region, the spatial clustering degree of all facial colors, and the spatial cross-correlation between all different facial colors are arranged in a fixed order to form a multidimensional facial color topological feature vector for that organ region. The vector dimension is fixed to ensure that the feature vectors of different organ regions and different subjects are comparable.

[0083] Step 62: Input the multidimensional facial color topological feature vector into the pre-set TCM organ-branch mutual generation and restraint and five-color association rule library for parsing to generate a structured diagnostic vector. The specific operation is as follows: The multidimensional facial color topological feature vectors of each organ region generated in step 61 are input into a pre-set TCM organ-organ mutual generation and restraint and five-color association rule library for analysis. Combining TCM diagnostic theory, the quantitative feature vectors at the technical level are transformed into structured diagnostic vectors at the TCM theoretical level, achieving a standardized mapping from facial features to TCM diagnostic conclusions. This ensures that the diagnostic conclusions conform to the TCM clinical diagnostic logic and can directly provide a reference for clinical diagnosis. The pre-set TCM organ-organ mutual generation and restraint and five-color association rule library is a set of rules constructed after standardization, quantification, and calibration based on traditional TCM facial diagnostic theory, organ-organ mutual generation and restraint theory, and five-color association theory. The pre-set TCM organ-organ mutual generation and restraint and five-color association rule library mainly contains two core rules: five-color association rules and organ-organ mutual generation and restraint rules. These two types of rules work together to form a complete diagnostic analysis logic. The five-color association rules... These are the fundamental rules for TCM facial diagnosis, clearly defining the relationship between each baseline facial color and its corresponding organ. For example, green corresponds to the liver, red to the heart, yellow to the spleen, white to the lungs, and black to the kidneys. They also specify the weighted area percentage threshold for each facial color within its corresponding organ region. When the weighted area percentage of a certain facial color within an organ region exceeds the corresponding threshold, it is preliminarily determined that the organ may have a physiological or pathological abnormality related to that facial color. The threshold settings are based on a large number of TCM clinical facial color samples, closely aligning with clinical diagnosis and requiring no additional optimization or adjustment. The organ-related generating and restraining rules are based on the TCM Five Elements theory, clearly defining the generating and restraining relationships between various organs. For example, liver wood generates heart fire, heart fire generates spleen earth, liver wood restrains spleen earth, and kidney water restrains heart fire. This is used to analyze the mutual influence of facial color characteristics in different organ regions, avoiding biases caused by isolated diagnosis of a single organ and ensuring the comprehensiveness of the diagnostic conclusions.

[0084] The specific operations of the analysis process follow clear logical steps to ensure reproducibility. First, the multidimensional facial color topological feature vector of each organ region is input into the rule base. The weighted area ratio of each facial color in the vector is extracted one by one and compared with the threshold of the facial color ratio corresponding to the organ in the rule base. If the weighted area ratio of a certain facial color exceeds the corresponding threshold, the organ is initially marked as having an abnormal tendency related to that facial color in combination with the five-color association rule, and the degree of abnormal tendency is recorded. The degree of abnormality is divided according to the proportion exceeding the threshold. The larger the proportion, the higher the degree of abnormality. Subsequently, the spatial clustering degree of facial color in the multidimensional feature vector is combined to correct the judgment of abnormal tendency. If the spatial clustering degree of abnormal facial color is high, it indicates that the abnormality is specific, and the judgment of abnormal tendency is strengthened; if the clustering degree is low, it indicates that the abnormality may be diffuse, and the weight of abnormal tendency is appropriately reduced, which is in line with the logic of TCM syndrome differentiation that clustering indicates disease and dispersion indicates mildness.

[0085] Next, combining the spatial cross-correlation between different complexion colors and the organ-related mutual generation and restraint rules in the rule base, the mutual influence of abnormal tendencies of different organs is analyzed. For example, if the proportion of blue in a certain organ region exceeds the standard and has a high spatial cross-correlation with yellow in the region corresponding to the spleen, combined with the rule of liver wood restraining spleen earth, it is determined that liver abnormalities may affect spleen function, and relevant diagnostic hints are added. After all organ regions are analyzed, the abnormal tendencies, abnormalities, and mutual influences of each organ are integrated. According to a fixed organ order, such as heart, liver, spleen, lung, kidney, etc., the analysis results are quantified into structured numerical vectors, i.e., structured diagnostic vectors. Each dimension in the vector corresponds to an organ, and the dimension value quantifies the degree of abnormality of that organ. A value of 0 indicates no obvious abnormality, and a larger value indicates a more obvious abnormality. At the same time, the mutual generation and restraint relationships between organs are marked by specific numerical ranges to ensure that the diagnostic vector is both quantitative and has clear TCM diagnostic meaning. It can be directly used for clinical diagnostic reference or health screening. The entire analysis process strictly follows the preset logic of the rule base, without subjective judgment, ensuring that different technicians obtain consistent diagnostic results.

[0086] Example 2: Please see Figure 2 Based on Example 1, this embodiment provides a computer vision-based intelligent face image recognition and analysis system, including: The stimulation acquisition module is used to apply physical stimulation with predetermined parameters to the subject's face and acquire a sequence of hyperspectral images of the face during the stimulation response period; The feature extraction module is used to construct the temporal intensity curve of each pixel in a preset band based on the facial hyperspectral image sequence, and extract relaxation feature parameters from the temporal intensity curve. The substrate generation module is used to deduce the static spectral reflectance characteristics of each pixel based on the relaxation feature parameters and the final data of the time-series intensity curve, and to form a static optical substrate map of the face. The credibility assessment module is used to calculate the pixel-level spectral differences between the baseline frame and the quasi-steady-state frame in the facial static optical base map and the facial hyperspectral image sequence, respectively; and to generate a credibility heatmap of the facial static optical base map based on the spectral differences and relaxation feature parameters. The facial color classification module is used to match the static spectral reflectance characteristics of each pixel in the static optical base image of the face with the predefined benchmark facial color spectral reflectance curve, and combine it with the confidence heatmap to classify and label the facial pixels and generate a facial color classification symbol map. The dialectical generation module is used to generate a dialectical vector by weighting the pixels of each region in the facial color classification symbol map according to their corresponding credibility in the facial color classification and credibility heat map, based on the preset facial organ partition map.

[0087] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A computer vision-based facial image intelligent recognition analysis method, characterized in that, include: Physical stimulation with predetermined parameters is applied to the subject's face, and a sequence of hyperspectral images of the face is acquired during the stimulation response period; Based on facial hyperspectral image sequences, a temporal intensity curve of each pixel in a preset band is constructed, and relaxation feature parameters are extracted from the temporal intensity curve. Based on the relaxation feature parameters and the final data of the time series intensity curve, the static spectral reflectance characteristics of each pixel are deduced to form a static optical base map of the face. Calculate the pixel-level spectral differences between the baseline frame and the quasi-steady frame in the facial static optical base map and the facial hyperspectral image sequence, respectively. A reliability heatmap of the facial static optical substrate image is generated based on spectral differences and relaxation characteristic parameters. The static spectral reflectance characteristics of each pixel in the static optical base image of the face are matched with the predefined baseline facial color spectral reflectance curve, and combined with the confidence heatmap, facial pixels are classified and labeled to generate a facial color classification symbol map. Based on the preset facial organ partition map, the pixels of each partition in the facial color classification symbol map are weighted and statistically analyzed according to their facial color classification and the corresponding credibility in the credibility heat map to generate a dialectical vector.

2. The computer vision-based facial image intelligent recognition analysis method according to claim 1, wherein, For each subject's face, an individual thermodynamic portrait was created, including: Multiple subthreshold thermal pulses with increasing energy are applied to a selected skin region, and time-series data of the region's radiance in a preset wavelength band are collected to obtain a cluster of transient thermodynamic curves of the skin micro-region. Based on the cluster of transient thermodynamic curves in skin microregions, an individualized heat transfer function is established, and the personalized master stimulus waveform for inducing standard physiological responses is obtained by inversely solving this function.

3. The intelligent face image recognition and analysis method based on computer vision according to claim 1 or 2, characterized in that, Extract relaxor feature parameters, including: The temporal intensity curve is subjected to Laplace domain transformation and mode decomposition to obtain the intrinsic relaxation mode set corresponding to each pixel; Based on the intrinsic relaxation mode set, the dominant mode dominance and intermodal antagonism index of each pixel are calculated.

4. The intelligent face image recognition and analysis method based on computer vision according to claim 3, characterized in that, The relaxor feature parameters are relaxor fingerprint vectors, and the dominant mode dominance and intermodal antagonism index are calculated, including: Calculate the first and second derivatives of the time series intensity curve, and identify the zero-crossing points of the first and second derivatives as feature point sequences; The temporal intensity curve is divided into multiple stages based on the sequence of feature points, and its dynamic type is determined based on the local geometric features of each stage. For each stage, the corresponding intrinsic dynamic model is called to extract the stage intrinsic parameters according to its calibrated dynamic type. Based on the time order of the feature point sequence, the extracted stage intrinsic parameters are combined to form a relaxor fingerprint vector.

5. The intelligent face image recognition and analysis method based on computer vision according to claim 4, characterized in that, The generation of a static optical base map of the face includes: A sliding window statistical analysis was performed on the end of the time series intensity curve to determine the start time of the quasi-steady-state segment, and the noise floor level within the quasi-steady-state segment was calculated. By utilizing the relaxation fingerprint vector, the quasi-steady-state start time, and the noise floor level, a residual response function is constructed and inverse compensation is performed to obtain the initial value of the instantaneous spectral intensity. The instantaneous spectral intensity of each pixel is used as the initial estimation field. Iterative optimization is performed based on the relaxed fingerprint vector to generate a static optical base map of the face.

6. The intelligent face image recognition and analysis method based on computer vision according to claim 5, characterized in that, The generation of the credibility heatmap includes: Calculate the pixel-level spectral angular distance between the facial static optical base map and the baseline frame and quasi-steady-state frame, generate the total difference map, and perform linear trend fitting and residual analysis on the total difference map to separate the linear difference components and nonlinear difference components. Based on the relaxed fingerprint vector and nonlinear difference components, a dynamic consistency index map is constructed. Using the dynamic consistency index map as the initial seed, and combining the spatial gradient of nonlinear difference components and the relaxation fingerprint similarity between adjacent pixels, anisotropic spatial propagation and conflict resolution are performed to generate a credibility heatmap.

7. The intelligent face image recognition and analysis method based on computer vision according to claim 6, characterized in that, The generation of facial color classification symbol maps includes: Calculate the fuzzy membership degree of the static spectral reflectance properties to the spectral reflectance curve of each predefined reference surface color, and generate an initial surface color probability distribution map; Based on the relaxed fingerprint vector and the spatial neighborhood information of pixels, the initial face color probability distribution map is collaboratively corrected to obtain the corrected face color probability distribution map.

8. The intelligent face image recognition and analysis method based on computer vision according to claim 7, characterized in that, The generation of facial color classification symbol maps also includes: High-conflict regions in the corrected fading probability distribution map are identified and reclassified in the low-dimensional spectral subspace by combining a confidence heatmap to generate an updated fading probability distribution map. Based on the confidence heatmap, a decision is made on the updated face color probability distribution map, and a face color classification symbol map is generated.

9. The intelligent face image recognition and analysis method based on computer vision according to claim 8, characterized in that, Generate dialectical vectors, including: For each organ region, based on the facial color classification symbol map and the credibility heatmap, the credibility weighted area ratio of each facial color, the spatial clustering degree of each facial color, and the spatial cross-correlation between different facial colors are calculated and combined into a multi-dimensional facial color topological feature vector. The multidimensional facial color topological feature vector is input into a pre-set TCM organ-organ mutual generation and restraint and five-color association rule library for parsing, generating a structured diagnostic vector.

10. A computer vision-based intelligent face image recognition and analysis system, applied to any one of the computer vision-based intelligent face image recognition and analysis methods according to claims 1-9, characterized in that, include: The stimulation acquisition module is used to apply physical stimulation with predetermined parameters to the subject's face and acquire a sequence of hyperspectral images of the face during the stimulation response period; The feature extraction module is used to construct the temporal intensity curve of each pixel in a preset band based on the facial hyperspectral image sequence, and extract relaxation feature parameters from the temporal intensity curve. The substrate generation module is used to deduce the static spectral reflectance characteristics of each pixel based on the relaxation feature parameters and the final data of the time-series intensity curve, and to form a static optical substrate map of the face. The reliability assessment module is used to calculate the pixel-level spectral differences between the baseline frame and the quasi-steady frame in the facial static optical base map and the facial hyperspectral image sequence, respectively. A reliability heatmap of the facial static optical substrate image is generated based on spectral differences and relaxation characteristic parameters. The facial color classification module is used to match the static spectral reflectance characteristics of each pixel in the static optical base image of the face with the predefined benchmark facial color spectral reflectance curve, and combine it with the confidence heatmap to classify and label the facial pixels and generate a facial color classification symbol map. The dialectical generation module is used to generate a dialectical vector by weighting the pixels of each region in the facial color classification symbol map according to their corresponding credibility in the facial color classification and credibility heat map, based on the preset facial organ partition map.

Images