Skin texture real-time monitoring method and system based on image recognition

Abstract

The application relates to the technical field of image processing, and discloses a skin texture real-time monitoring method and system based on image recognition, which comprises the following steps: acquiring a continuous skin video image stream and performing spatial frequency domain decomposition on a luminance channel of the continuous skin video image stream to obtain a multi-scale image sequence; synchronously generating a subsurface scattering mapping matrix, an environmental rigid motion tensor and a morphological prior mask; extracting a micro-strain phase tensor from the multi-scale sequence based on dynamic constraint of parameters of an Euler space-time micro-motion amplification mechanism of the subsurface scattering mapping matrix; taking local extreme points of the micro-strain phase tensor as graph nodes, and using the environmental rigid motion tensor and the morphological prior mask to jointly constrain an adjacency matrix between the nodes, so as to construct a dynamic graph topology structure; and finally, performing graph convolution aggregation processing on the dynamic graph topology structure to output multi-dimensional quantitative skin texture parameters. The application makes real physical mechanisms into network topology weights, breaks the limitation of a pure algorithm black box, and realizes skin state quantitative monitoring under a complex environment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for real-time monitoring of skin texture based on image recognition. Background Technology

[0002] Skin condition monitoring is in high demand in health management and care. Traditional methods often employ contact-based measuring instruments, which rely on professional operators and are limited by space and equipment. Currently, non-contact monitoring technology based on computer vision is widely used. The common approach involves acquiring images of the face or a specific area, converting the images to grayscale, performing conventional filtering and noise reduction, and then using traditional operators such as local binary mode and gray-level co-occurrence matrix, or conventional-scale convolutional neural networks, to extract texture features such as pores and fine lines on the epidermis and assess its condition.

[0003] However, existing non-contact image processing technologies still face significant bottlenecks in practical applications. Firstly, at the image preprocessing and feature extraction level, due to the presence of reflections, skin color differences, and keratin interference in human skin, current methods often employ static filtering enhancement techniques such as bilateral filtering or histogram equalization. These methods are prone to over-smoothing, leading to the loss of superficial fine lines. Even when using networks such as multi-scale dilated convolution to extract features, these conventional methods, which rely on static two-dimensional images, can only analyze the two-dimensional topological morphology of the skin surface, making it difficult to extract local dynamic features caused by physiological pulsations or subtle facial movements. These dynamic micro-motion features are usually closely related to the elasticity, hydration state, and deep structural information of the underlying skin. To obtain dynamic information, existing technologies utilize video micro-motion magnification algorithms to process continuous image sequences. However, conventional micro-motion magnification algorithms typically set globally fixed passbands and magnification factors, without considering the differences in physical properties such as subsurface scattering of light in different skin regions. This fixed-parameter magnification method indiscriminately amplifies background noise (such as fuzz and reflective noise) and causes feature distortion in skin regions with different physical properties.

[0004] On the other hand, to address the interference problem during dynamic monitoring, existing technologies often employ simple optical flow methods combined with affine transformation matrices to perform overall geometric translation or scaling correction on the image. This macroscopic geometric correction only reaches the pixel alignment level. When performing network extrapolation on the feature sequences extracted from the video, conventional deep learning models often lack physical environment constraints. In actual non-contact monitoring scenarios, the object being measured typically exhibits rigid displacements such as head shaking. Because existing network structures fail to effectively isolate interference caused by rigid motion and do not establish spatial topological relationships between local micro-motion signals within the model, feature information is easily affected by external factors during network forward propagation. This results in significant errors in the analysis results of existing image recognition methods under uncontrolled environments, making it difficult to balance real-time performance and detection accuracy, and unable to stably output quantitative monitoring indicators with physical meaning. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a real-time skin texture monitoring method and system based on image recognition. This solves the problems of existing non-contact skin monitoring methods based on image recognition, which cannot effectively overcome skin surface interference such as local reflections and skin color differences, lack an adaptive feature extraction mechanism for differences in skin physical properties, and have weak resistance to rigid motion interference in dynamic acquisition scenarios. This leads to the easy loss of epidermal texture details such as fine lines and pores, as well as distortion of dynamic micro-motion features, ultimately resulting in inaccurate quantitative evaluation results of skin texture parameters.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a real-time skin texture monitoring method based on image recognition, comprising the following steps: A continuous stream of skin video images is acquired, and the luminance channel of the continuous skin video image stream is spatially decomposed in the frequency domain to obtain a multi-scale image sequence. Based on the continuous skin video image stream, a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask are generated synchronously. Micro-strain phase tensors are extracted from the multi-scale image sequence based on the Euler spatiotemporal micro-amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. A dynamic graph topology is constructed using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology is jointly constrained by the environmental rigid motion tensor and the morphological prior mask. The dynamic graph topology is subjected to graph convolution aggregation processing to output quantized skin texture parameters.

[0007] Furthermore, the step of acquiring a continuous skin video image stream and performing spatial frequency domain decomposition on the luminance channel of the continuous skin video image stream to obtain a multi-scale image sequence specifically includes: converting the acquired continuous skin video image stream from a color space to an orthogonal color space, separating and extracting the luminance channel; using the Laplacian pyramid to perform iterative Gaussian smoothing downsampling and interpolation upsampling on the luminance channel to separate the global luminance trend and local high-frequency information of the image, and generating the multi-scale image sequence with different spatial resolutions.

[0008] Further, the step of simultaneously generating a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask based on the continuous skin video image stream specifically includes: extracting the local pixel gradient matrix of the luminance channel and performing local integration calculation on the local pixel gradient matrix using a Gaussian smoothing kernel of a preset scale to generate the subsurface scattering mapping matrix used to characterize the physical elasticity and hydration state of the skin; tracking the inter-frame displacement of feature points in non-skin target regions in the continuous skin video image stream and calculating global displacement features as the environmental rigidity motion tensor, specifically including: locating non-skin target regions using skin color segmentation or target detection and extracting static feature points in these regions; tracking the displacement vectors of the static feature points using an inter-frame feature matching algorithm, fitting the displacement vectors through an affine transformation model, and calculating the environmental rigidity motion tensor containing translation and rotation components; applying an adaptive threshold morphological top-hat transformation operation to the continuous skin video image stream to remove abrupt noise regions and generate the morphological prior mask composed of binary values ​​of 0 and 1.

[0009] Furthermore, the micro-strain phase tensor is extracted from the multi-scale image sequence based on the Euler spatiotemporal micro-amplification mechanism. The amplification parameters of the Euler spatiotemporal micro-amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. Specifically, this includes: applying a time-domain bandpass filter to each level of the multi-scale image sequence; dynamically adjusting the upper and lower frequency limits and phase amplification coefficient of the time-domain bandpass filter according to the numerical values ​​corresponding to each pixel in the subsurface scattering mapping matrix; performing filtering processing using the dynamically adjusted time-domain bandpass filter; and then superimposing the obtained filtering result after multi-scale reconstruction with the brightness channel to generate the corresponding micro-strain phase tensor.

[0010] Further, the construction of a dynamic graph topology using local extrema of the micro-strain phase tensor as graph nodes, wherein the adjacency matrix of the dynamic graph topology is jointly constrained by the environmental rigid motion tensor and the morphological prior mask, specifically includes: extracting feature vectors of each graph node in the spatiotemporal domain, wherein the feature vectors are composed of spatial coordinate information of the graph node in the image plane, micro-strain phase information, and temporal phase change rate; calculating the feature distance similarity between any two graph nodes; calculating the directional coincidence between the temporal displacement component of the feature vector and the environmental rigid motion tensor, and using this directional coincidence as a motion penalty term to numerically attenuate the feature distance similarity; multiplying the attenuated feature distance similarity with the mask values ​​corresponding to the two graph nodes in the morphological prior mask to obtain the matrix elements of the adjacency matrix, thereby forcibly cutting off the connection paths of graph nodes in noisy regions through the product operation. As a preferred specific calculation method, the adjacency matrix in the joint constraint mechanism... matrix elements The formula for construction is: In the formula, For nodes With nodes Feature distance similarity between them; This is a motion penalty term calculated based on directional coincidence. and This represents the binarized mask value of the corresponding node in the morphological prior mask. By introducing the subtraction operation in this formula, pseudo-motion displacements of the environment can be eliminated from the topological level; through the product operation, when any node is in a noisy region (i.e., the mask value is 0), it can be forcibly set... This cuts off the connection paths of graph nodes in noisy regions from the mathematical level.

[0011] Further, the step of performing graph convolutional aggregation processing on the dynamic graph topology to output quantized skin texture parameters specifically includes: extracting the feature vectors of the graph nodes in the spatiotemporal domain to construct a hidden node feature matrix; calculating the degree matrix based on the adjacency matrix; and performing symmetric normalization processing on the adjacency matrix; in the dynamic graph convolutional network, multi-level aggregation and nonlinear activation mapping are performed using the normalized adjacency matrix, the hidden node feature matrix, and the learnable weight matrix; specifically, this includes: constructing a dynamic graph network architecture containing an input layer, multiple hidden layers, and an output layer; in each hidden layer, performing graph convolution operations to aggregate neighbor node features, and sequentially performing dimensionality reduction and mapping updates of the features through batch normalization processing and nonlinear activation functions; and sequentially passing the hidden feature matrix output from the highest level through a global average pooling layer and a fully connected classification layer to map and output the quantized skin texture parameters.

[0012] A second aspect of the present invention provides a real-time skin texture monitoring system based on image recognition, comprising: The spatial frequency domain decoupling module is used to acquire a continuous skin video image stream and perform spatial frequency domain decomposition on the brightness channel of the continuous skin video image stream to obtain a multi-scale image sequence. The parallel physical feature generation module is used to synchronously generate a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask based on the continuous skin video image stream. The micro-motion phase controlled extraction module is used to extract the micro-strain phase tensor from the multi-scale image sequence based on the Euler spatiotemporal micro-motion amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-motion amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. A composite graph topology construction module is used to construct a dynamic graph topology structure using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology structure is jointly constrained by the environmental rigid motion tensor and the morphological prior mask. The graph network inference output module is used to perform graph convolution aggregation processing on the dynamic graph topology and output quantized skin texture parameters.

[0013] Furthermore, in the system, the parallel physical feature generation module is specifically used to: extract the local pixel gradient matrix of the brightness channel and perform local integration calculation using a Gaussian smoothing kernel of a preset scale to generate the subsurface scattering mapping matrix; track the inter-frame displacement of feature points in the non-skin target region in the continuous skin video image stream, calculate the global displacement features as the environmental rigid motion tensor; apply an adaptive threshold morphological top-hat transformation operation to the continuous skin video image stream to remove abrupt noise regions and generate the morphological prior mask composed of binary values ​​of 0 and 1.

[0014] Furthermore, in the system, the micro-motion phase controlled extraction module is specifically used to: dynamically adjust the upper and lower frequency limits and phase amplification coefficient of the time-domain bandpass filter according to the value of the subsurface scattering mapping matrix, and perform filtering processing using the dynamically adjusted time-domain bandpass filter, and then superimpose the obtained filtering result with the brightness channel after multi-scale reconstruction to generate the micro-strain phase tensor.

[0015] Furthermore, in the system, when the composite graph topology construction module calculates the adjacency matrix, it specifically executes the following logic: based on the feature distance similarity between any two graph nodes, it subtracts the motion penalty term calculated based on the directional coincidence between the temporal displacement component of the feature vector and the rigid motion tensor of the environment, and multiplies the attenuated result with the mask value corresponding to the two graph nodes in the morphological prior mask.

[0016] This invention provides a method and system for real-time monitoring of skin texture based on image recognition. It has the following beneficial effects: 1. To address the problem that traditional algorithms' global indiscriminate amplification easily leads to distortion of weak physiological signals, this invention utilizes a subsurface scattering mapping matrix to dynamically constrain the parameters of the Euler micro-motion amplification mechanism. This design enables the model to keenly perceive the local elasticity and hydration state of the skin, adaptively adjusting the filtering bandwidth and amplification coefficient, avoiding strong signal overload and weak signal annihilation, and significantly improving the accuracy and precision of skin micro-strain feature extraction from a physical perspective.

[0017] 2. To address the shortcomings of deep network inference links that are easily contaminated by environmental displacement and surface noise, this invention utilizes the environmental rigidity motion tensor and morphological prior masks to jointly constrain the adjacency matrix of the dynamic graph topology. This scheme introduces a directional blocking mechanism at the mathematical level, eliminating environmental jitter artifacts through displacement penalty terms and forcibly severing noise node connections using mask multiplication. This effectively blocks the cascading spread of invalid interference in the graph convolutional network, significantly improving the robustness and data stability of skin quantification monitoring in complex dynamic scenes. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a block diagram of the system module structure of the present invention; Figure 3 This is a schematic diagram of the hardware physical structure of the electronic device of the present invention; Figure 4 This is a schematic diagram illustrating the principle of the generation process of the three parallel physical prior features of the present invention; Figure 5 This is a controlled feedback calculation logic diagram of the dynamic micro-amplification mechanism of the present invention; Figure 6 This is a visual logic diagram of the dynamic graph topology construction and adjacency matrix calculation of the present invention; Figure 7 This is a diagram of the forward propagation and multidimensional parameter output architecture of the dynamic graph convolutional neural network of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see the appendix Figure 1 -Appendix Figure 7 This invention provides a method and system for real-time monitoring of skin texture based on image recognition. The method for real-time monitoring of skin texture based on image recognition may include the following steps: A continuous stream of skin video images is acquired, and the luminance channel of the continuous skin video image stream is spatially decomposed in the frequency domain to obtain a multi-scale image sequence. Based on the continuous skin video image stream, a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask are generated synchronously. Micro-strain phase tensors are extracted from the multi-scale image sequence based on the Euler spatiotemporal micro-amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. A dynamic graph topology is constructed using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology is jointly constrained by the environmental rigid motion tensor and the morphological prior mask. The dynamic graph topology is subjected to graph convolution aggregation processing to output quantized skin texture parameters.

[0021] In a specific implementation environment, the image acquisition device at the system's front end captures changes in optical reflection in the skin region, generating a continuous stream of skin video images. This original input video sequence is defined as... ,in, and Represents the spatial physical coordinates of a pixel on a two-dimensional image plane. Represents the time frame index of the video sequence.

[0022] To eliminate the interference of sudden changes in ambient light color on subtle underlying texture features, the system will input video sequences Mapped to an orthogonal color space, the brightness channel image matrix, which independently represents structural information, is separated. Spatial-frequency domain decomposition is performed on the brightness channel image matrix to construct an image pyramid architecture, outputting the multi-scale image sequence composed of different spatial frequency components. ,in This is a hierarchical index for the decomposition of the spatial frequency domain.

[0023] While decoupling the spatial frequency of the image, the system is based on the brightness channel image matrix. A parallel feature extraction mechanism is initiated, simultaneously outputting three prior physical constraints through a series of mathematical transformations. Specifically, the system extracts a subsurface scattering mapping matrix used to quantify the physical properties of the skin surface (such as tissue elasticity and hydration). Calculate the environmental rigid motion tensor used to characterize environmental jitter and the overall rigid body displacement of the target. It also generates a binarized morphological prior mask for isolating invalid image noise regions (such as hair-covered areas and sweat highlight areas). .

[0024] After extracting the physical prior features, the algorithm enters the nonlinear amplification stage of micro-strain features. The system processes multi-scale image sequences. Euler spatiotemporal micro-amplification is applied. This process overcomes the limitation of traditional micro-amplification mechanisms that use globally constant parameters. The system maps the subsurface scattering matrix... This amplification operator is introduced as a feedback adjustment variable. Based on this feedback adjustment variable, the system dynamically controls the upper and lower frequency limits and phase amplification coefficient of the time-domain bandpass filter at the pixel level, outputting a micro-strain phase tensor. This ensures that even minor deformations in skin areas with different physical properties can be reproduced with high fidelity.

[0025] Subsequently, the present invention transforms the extracted continuous micro-strain matrix data into a graph topological structure. The system extracts the micro-strain phase tensor. Construct a graph node set at local extrema in the spatial domain. The connection weights between any two graph nodes are calculated to construct an adjacency matrix. At that time, the system uses mathematical formulas to convert the rigid motion tensor of the environment. and morphological prior masks Strongly coupled to the underlying topology computation. Environmental rigidity motion tensor. Transformed into a numerical attenuation term to offset the environmental pseudo-displacement correlation between nodes; morphological prior mask This is then used as a Boolean multiplier to directly perform zero-value filtering on matrix elements. This joint constraint mechanism forcibly removes invalid and erroneous topological connections, establishing a dynamic graph topology with extremely high noise resistance.

[0026] Finally, the system inputs the completed dynamic graph topology into the graph convolutional network model. The dynamic graph convolutional network performs layer-by-layer aggregation and nonlinear mapping of node features along the physically constrained topological path, deeply integrates micro-strain features in the spatiotemporal dimensions, and finally outputs quantitative skin texture parameters that characterize the current skin physiological state through classification or regression fully connected layers, completing the monitoring closed loop from optical signals to physiological quantitative indicators.

[0027] In this embodiment, in order to effectively isolate the interference of ambient light fluctuations on the subtle physiological characteristics of the skin's underlying layers, the system first performs spatial frequency domain decoupling processing on the acquired continuous skin video image stream.

[0028] In this invention, the system converts the input color video frames from the original red-green-blue color space to an orthogonal color space. Through a preset orthogonal transformation matrix, the system separates the luminance channel matrix, which independently represents the image's structural details and illumination intensity. And the chroma channel matrix representing color attributes. All subsequent feature parsing operations are performed independently on this luminance channel matrix. This is done by expanding the chroma channel to eliminate the negative impact of redundant information on micro-strain calculations.

[0029] Furthermore, the system utilizes the Laplace pyramid algorithm to process the luminance channel matrix. Multi-scale spatial frequency domain decomposition is performed. In the specific calculation process, the system first uses a Gaussian smoothing filter to perform spatial low-pass filtering on the brightness channel matrix, and then combines it with downsampling operations to construct each level of the Gaussian pyramid sequentially.

[0030] Subsequently, the system performs interpolation upsampling on each level of the Gaussian pyramid image to restore its spatial resolution to the same scale as the previous level, and then subtracts the upsampling result from the original image of the previous level. Through this iterative Gaussian smoothing downsampling and interpolation upsampling difference operation, the system accurately extracts the global low-frequency brightness trend and local high-frequency detail information of the image, generating the multi-scale image sequence composed of different spatial resolution frequency bands. .in, This serves as the index for the pyramid's decomposition levels. To balance the accuracy of extracting subtle skin texture features with the system's real-time computational overhead, as a preferred embodiment, the total number of pyramid decomposition levels is set to 4 (i.e., ...). This allows the spatial resolution of the top-level image in the pyramid to be approximately 1 / 16 of the original input image. In other practical applications, the system can also adaptively determine the decomposition level depth based on the original resolution of the input video stream. For example, it can stop the decomposition iteration when the size of the decomposed and downsampled image is less than or equal to 32×32 pixels. Spatial texture structure corresponding to a specific frequency band.

[0031] In the technical solution provided in this embodiment, in addition to spatial frequency domain decoupling, the system synchronously starts a parallel physical feature generation mechanism based on the continuous skin video image stream, which aims to provide prior constraints for subsequent micro-motion amplification and topology deduction by extracting the physical and motion attributes of the underlying tissue.

[0032] In this invention, the system first calculates and generates the subsurface scattering mapping matrix to characterize the physical elasticity and hydration state of the skin. Considering that the physical elasticity of the skin is a relatively stable static property within a short monitoring window, the system preferably extracts the initial reference frame from the video sequence (or averages multiple frames within a preset time window), and extracts the local pixel gradient vector of the luminance channel matrix of the reference frame in two-dimensional physical space using a difference operator. To realistically simulate the multi-layer diffuse reflection physical behavior of light within complex skin tissue, the system uses a Gaussian smoothing kernel of a preset scale to perform local spatial integration calculation on the aforementioned local pixel gradient. This subsurface scattering mapping matrix... The mathematical model is defined as follows: in, Represents the coordinates in the neighborhood The local pixel gradient vector at that location. This indicates the calculation of the L2 norm magnitude of the gradient vector; It is a two-dimensional Gaussian smoothing kernel function. To control the variance parameter of the integral smoothing window scale. To ensure that the spatial smoothness of the subsurface scattering matrix conforms to the actual optical scattering characteristics of skin, as a basic implementation example, It can be set to a fixed constant, for example, preferably set to . Pixels. Furthermore, considering the variations in camera resolution and shooting distance in actual monitoring scenarios, The specific values ​​are preferably determined using an adaptive method: the system first obtains the image resolution of the current continuous video stream or the pixel width of the effective skin region. and will Dynamically set to a scale value related to that scale (e.g., set to 1% of the overall image width, or set to...). This adaptive adjustment ensures that subsequent micro-amplification parameters are not affected by differences in device resolution. Matrix The values ​​of each element directly map to the elastic physical state of the skin at the corresponding spatial coordinates.

[0033] Simultaneously, to accurately capture the systematic displacement deviation introduced in dynamic monitoring scenarios, the system further calculates and generates an environmental rigid motion tensor. The system extracts static feature points in non-skin target regions (such as rigid background references) of a continuous skin video image stream and tracks the physical displacement vectors of these feature points in a continuous time series using an inter-frame feature matching algorithm. As a preferred embodiment, the system uses a skin color segmentation model or facial keypoint detection algorithm to calibrate and eliminate effective skin regions to determine the non-skin target regions where the background or static references are located. Within this region, the system preferably uses the Shi-Tomasi algorithm to extract high-quality static feature points and uses the Lucas-Kanade optical flow method to track the displacement vectors of these feature points between consecutive frames. To eliminate local noise interference, the system further uses the RANSAC algorithm to eliminate mismatched optical flow points, and then uses the least squares method to fit a global affine transformation model to extract global displacement features, which are then used to construct the environmental rigid motion tensor. This tensor The horizontal displacement components caused by equipment position vibration or involuntary rigid translation of the human body were fully quantified. Vertical displacement components and rotational transformation components .

[0034] Furthermore, to address the common localized abrupt noise issues on the skin surface, such as hair occlusion, fine fuzz, and sweat highlights, this invention generates a morphological prior mask through mathematical morphological operations. The system defines structural elements of specific shapes and sizes. Considering that localized noise such as fine fuzz or sweat highlights typically exhibits isotropic characteristics as points or short lines, as a preferred embodiment, the structural elements are preferably circular structural elements with a radius of 5 pixels, or rectangular structural elements with a radius of 7×7 pixels. Using these defined structural elements, the system applies a morphological top-hat transformation operation to the brightness channel matrix. By calculating the difference between the original image matrix and the morphological opening operation result matrix, the system accurately highlights localized abrupt noise regions with a spatial area smaller than the given structural element.

[0035] Based on this, the system applies an adaptive threshold function to the output of the top-hat transform operation for binarized threshold segmentation. In the specific calculation implementation, the system preferably uses Otsu's method or a local adaptive Gaussian thresholding algorithm to dynamically solve for the optimal segmentation threshold. Further, as a preferred implementation, considering that the actual skin surface is easily affected by uneven local illumination, the system preferably uses the aforementioned local adaptive Gaussian thresholding algorithm, and its local calculation window size is preferably set to, for example, 11×11 pixels or 15×15 pixels, so as to effectively suppress background illumination interference while accurately stripping noise. Through the segmentation operation with the above parameter settings, abrupt noise regions are thoroughly stripped, ultimately generating the morphological prior mask composed of binary values. In this morphological prior mask, the matrix elements corresponding to the pixel coordinates that belong to the noise region are set to zero, while the elements representing the effective skin texture region are set to one, thus defining an absolute physical isolation boundary in geometric space.

[0036] In this embodiment, based on the multi-scale image sequence and physical prior features obtained in the aforementioned steps, the system initiates a controlled Euler spatiotemporal micro-motion amplification mechanism to extract micro-strain phase tensors characterizing latent physiological changes from the continuous image sequence.

[0037] The system applies a temporal bandpass filter along the time dimension for each spatial resolution level of the multi-scale image sequence. This temporal bandpass filter aims to isolate and extract specific frequency band periodic micro-deformation signals hidden in consecutive video frames, caused by underlying microcirculatory blood perfusion or subtle muscle contractions.

[0038] Traditional micro-amplification algorithms employ a globally constant frequency range and magnification factor across the entire image, which can easily lead to signal distortion in highly reflective or low-elasticity areas of the skin. In this invention, the subsurface scattering mapping matrix is ​​introduced as a spatial weight constraint term to dynamically adjust the core parameters of the temporal bandpass filter at the pixel level.

[0039] Specifically, the system adaptively resets the upper and lower frequency limits of the local filter based on the numerical values ​​of each pixel in the subsurface scattering mapping matrix. As a specific feasible implementation, the system first presets the reference frequency range of the filter; for example, for typical physiological micro-motion signals, a reference frequency lower limit is set. 0.5 Hz, upper limit of the reference frequency The frequency is 2.0 Hz. Based on this, the system establishes a linear mapping function between the upper and lower frequency limits and the values ​​of the subsurface scattering mapping matrix. The specific dynamic adjustment formula is expressed as follows: in, and These are the dynamically adjusted lower and upper limits of the local frequency, respectively. The subsurface scattering mapping values ​​are normalized to the [0,1] interval; The preset frequency shift control coefficient (e.g., preferably 1.0 Hz).

[0040] Based on the above mapping rules, an inverse mapping relationship was established between the translation direction of the upper and lower frequency limits and the value of the subsurface scattering mapping matrix: when the subsurface scattering mapping matrix indicates that a certain local area has high skin density and poor elasticity (i.e., When the frequency approaches 1), the system dynamically shifts the upper and lower limits of the filter corresponding to that pixel towards the higher frequency range; conversely, it shifts them towards the lower frequency range. This operation ensures that skin regions with different tissue densities and physical elasticities can be assigned to the optimal monitoring frequency band that matches their mechanical conduction response.

[0041] Simultaneously, the system establishes a nonlinear mapping relationship based on the subsurface scattering mapping matrix to dynamically adjust the phase amplification factor of the time-domain bandpass filter. When the scattering value indicates low skin hydration or poor elasticity in a local area, the system automatically increases the amplification weight of that area to compensate for the attenuated physiological and mechanical deformation. This phase amplification factor... The dynamic adjustment calculation formula is expressed as: in, Spatial coordinates The local phase amplification factor after dynamic adjustment; The preset reference amplification constant is used to set the baseline level of amplification (in a specific embodiment for monitoring human skin microcirculation, this parameter is preferably set in the range of 10 to 50). To adjust the nonlinear decay weight of the response sensitivity (preferably in the range of 0.1 to 1.0); The value of the subsurface scattering mapping matrix at this spatial coordinate is given.

[0042] To enable those skilled in the art to implement this solution without extensive trial and error, a set of specific and usable parameter configuration pairs and their applicable scenarios are provided here: As a typical embodiment, when the system is deployed in a normal indoor lighting environment and uses a conventional RGB camera with a frame rate of 30fps and a resolution of 1080P to capture continuous skin video image streams, the following settings are preferred: as well as This set of specific values ​​can effectively amplify microcirculatory deformation while suppressing motion artifacts to the greatest extent.

[0043] Furthermore, the above parameters support adaptive calibration adjustments based on actual hardware conditions and the monitored object: for example, when an increase in the frame rate of the input video stream (e.g., 60fps) causes a decrease in the absolute displacement between adjacent frames, the system needs to adjust accordingly. Increase the weight (e.g., raise it to 40) to ensure sufficient micro-motion signals are extracted; conversely, when the system identifies a target with dark skin or a thick stratum corneum, due to its lower surface transmittance, the system will appropriately reduce the nonlinear attenuation weight. (Adjusted to 0.3 below) to reduce the attenuation magnitude and ensure that the local area still has sufficient amplification weight.

[0044] After determining the local dynamic constraint parameters, the system uses the adjusted time-domain bandpass filter to perform temporal phase filtering and signal amplification processing on each level of the multi-scale image sequence. This process, through the intervention of physical prior features, precisely enhances the minute phase shifts strongly correlated with physiological features while suppressing the amplification of spatial structural noise.

[0045] After completing the filtering and amplification processing at each spatial level, the system performs multi-scale reconstruction on the filtered results. The system merges the amplified frequency band details layer by layer through interpolation upsampling and matrix addition operations, restoring its spatial resolution to the same scale as the input source.

[0046] Finally, the system performs a spatial superposition operation on the amplified micro-motion signal, which has undergone multi-scale reconstruction, and the original brightness channel matrix. This superposition operation seamlessly embeds the amplified micro-strain dynamic information into the original anatomical structure, generating the micro-strain phase tensor that fully characterizes the current dynamic micro-features of the skin.

[0047] In this embodiment, after obtaining the high-fidelity micro-strain phase tensor, the system uses graph theory and deep learning theory to construct a dynamic graph topology that integrates multi-dimensional physical prior constraints, and performs quantization texture parameter deduction through a graph convolutional network.

[0048] In this invention, the system first scans the micro-strain phase tensor and uses a sliding window with a preset spatial scale (e.g., 3×3 or 5×5 pixels) to perform pixel-by-pixel extremum determination within the spatial domain, extracting all local extrema points in the tensor space. These extrema points correspond to the spatial locations where skin microcirculation pulsation or subtle muscle contractions are most pronounced. The system uses all extracted local extrema points as graph nodes to construct a graph node set, thereby laying the discrete node foundation for the dynamic graph topology.

[0049] To establish the association weight relationships between graph nodes to construct an adjacency matrix, the system first extracts the feature vectors of each graph node in the spatiotemporal domain. Specifically, in this embodiment, the dimension of the feature vectors is preferably set to 5 dimensions, specifically consisting of the two-dimensional spatial coordinates of the graph nodes in the image plane. Micro-strain phase value of the current frame Inter-frame phase time derivative characterizing temporal dynamics And the numerical value of the subsurface scattering mapping matrix corresponding to that point. This multi-dimensional feature fusion, achieved through concatenation, comprehensively characterizes the physiological and physical states of nodes in this spatiotemporal context. After obtaining the feature vector, the system first calculates the Euclidean distance between the feature vectors of any two graph nodes using a distance metric algorithm. Subsequently, the system uses a Gaussian kernel function to transform this Euclidean distance into a normalized similarity score, which is then used as the initial feature distance similarity. Specifically, The complete calculation formula is: in, Represents graph nodes Graph Nodes 5-dimensional feature vector and The square of the Euclidean distance between them; The kernel function bandwidth parameter controls the similarity decay rate. As a specific embodiment, The preferred setting is a dynamic value, that is, 0.5 to 2.0 times the average feature distance between all node pairs in the graph network (for example, set to 1.0 times), in order to quantify the physiological rhythm synchronization between nodes in the original signal dimension.

[0050] The system uses this similarity score as the initial feature distance similarity. (This value is between 0 and 1, and the larger the value, the more similar the node features are), thereby quantifying the physiological rhythm synchronization between nodes in the original signal dimension.

[0051] Because involuntary slight swaying of the human head or camera position shift of the imaging equipment is unavoidable in dynamic monitoring environments, these environmental rigid displacements will generate strong spurious strain correlations in distance similarity calculations. To eliminate this systematic error, the system calls the aforementioned generated environmental rigid motion tensor. To eliminate dimensional differences, the system first converts the environmental rigid motion tensor (containing global horizontal translation components) into a single tensor. Vertical translation component and rotation angle components Mapped to the current spatial coordinates of each graph node At this point, generate the expected global displacement vector of the corresponding node affected by environmental jitter. .

[0052] Subsequently, the system extracts the two-dimensional time-domain displacement vector representing the actual motion of the node itself from the feature vector. And by calculating the vectors of the two mentioned above ( and The cosine similarity between the nodes is used to quantify the degree of coincidence between the temporal displacement of the node and the direction of environmental jitter. The calculation formula is as follows: .

[0053] After determining the directional coincidence of individual nodes, for any connected graph nodes in the topology... Graph Nodes (Their directional coincidence is respectively) and The system will include exercise penalty items. Strictly defined as the arithmetic mean of these two values, the calculation formula is: The calculation obtained from this It is directly used as a weight coefficient term, and its value is reduced from the basic feature distance similarity when constructing the core adjacency matrix in order to completely remove the artifact association caused by environmental jitter.

[0054] Besides the interference from macroscopic displacement, localized hair occlusion or specular noise can also disrupt the topological logic of the graph network. Therefore, the system further introduces the aforementioned generated morphological prior mask to implement zero-value masking filtering. The system multiplies the feature distance similarity after motion penalty attenuation with the binarized mask values ​​corresponding to each of the two graph nodes in the morphological prior mask.

[0055] Based on the aforementioned computational mechanism of joint physical priors, this invention defines the core adjacency matrix of the dynamic graph topology. In the middle, the first The graph node and the first Matrix elements between graph nodes The construction analytic expression is: In the formula, For nodes With nodes Initial feature distance similarity between them; This is the motion penalty term calculated based on the time-domain displacement components and the environmental rigidity tensor; The penalty adjustment coefficient, used to control the environmental disturbance immunity, is set to a value that depends on the system's maximum allowable tolerance threshold (in actual micro-motion signal monitoring projects, this penalty adjustment coefficient...). It is preferable to set it within the empirical range of 0.5 to 2.0 to balance noise resistance and the retention rate of true features. The function is used to truncate excessive negative penalties to maintain nonnegativity; and These are nodes respectively. With nodes The corresponding Boolean mask value in the morphological prior mask.

[0056] Through the subtraction mechanism in this core analytical expression, the system completely cancels out the pseudo-motion displacement of the environment at the underlying network topology. Simultaneously, using the multiplication mechanism, once a graph node is in a noisy region, its corresponding mask value becomes zero; this product operation forces the corresponding matrix elements to... Set to zero. This absolute physical cutoff operation completely blocks the node connection path containing invalid feature parameters, preventing low-level noise interference from propagating deeper into the network.

[0057] After constructing the rigorous physical constraints of the dynamic graph topology, the system performs deep graph network deduction on it. The system extracts the feature vectors of all graph nodes in the spatiotemporal domain and stacks them column-wise to form an initial hidden node feature matrix. Simultaneously, to prevent nodes from losing their own features during information aggregation, the system first constructs the adjacency matrix... An identity matrix is ​​added to construct an augmented adjacency matrix with self-loops. Subsequently, the system calculates the degree matrix based on this augmented adjacency matrix and performs symmetric normalization using the diagonal inverse square root operation of the degree matrix, ultimately generating a normalized adjacency matrix for network inference. .

[0058] Subsequently, the system performs multi-level forward propagation operations in the constructed dynamic graph convolutional network. As a specific network architecture design, the dynamic graph convolutional network preferably contains three hidden layers, and the learnable weight matrices of each layer are initialized using the Xavier uniform distribution method to ensure the stability of gradient propagation. In the specific feature inference process: the first graph convolutional layer maps the input graph node features to output 64-dimensional hidden features; the second graph convolutional layer further aggregates and reduces the dimensionality to 32-dimensional features; the third graph convolutional layer outputs 16-dimensional features. At each layer, the system uses a normalized adjacency matrix to aggregate information from neighboring nodes and performs linear multiplication with the hidden node feature matrix of the previous layer and the learnable weight matrix in the network. Then, the system sequentially processes this linear aggregation result through batch normalization layers and maps it to the ReLU nonlinear activation function, thereby completing the complex high-dimensional feature abstraction between network layers.

[0059] At the end of the network structure, the system sequentially feeds the hidden feature matrix output from the highest-level graph convolution operation into a global average pooling layer to extract a global topological abstract representation. Finally, this representation data undergoes dimensionality reduction and linear mapping through a fully connected classification layer, outputting comprehensive quantitative skin texture parameters representing the current skin elasticity, smoothness, and dynamic microtexture state. To ensure the output results have clear clinical and physiological interpretability, the output node dimension of the fully connected classification layer is specifically set to 3 dimensions, and a Sigmoid nonlinear activation function is applied at its end to strictly normalize and map the original predicted values ​​of the network output to the [0,1] interval. Subsequently, it is linearly scaled (multiplied by 100) to transform it into a standardized scoring interval of [0,100].

[0060] In practice, the quantified skin texture parameters are output as a one-dimensional, three-channel feature vector. Each channel of this feature vector independently corresponds to the aforementioned skin elasticity index, smoothness score, and micro-motion activity index. As a specific embodiment, the theoretical value ranges and corresponding physical state descriptions of each index are defined as follows: The first channel is the skin elasticity index, with a value range of 0-100. For example, a value range of 0-30 indicates that the skin is dry, aging, and has lost collagen; 31-70 indicates normal elasticity; and 71-100 indicates that the skin is hydrated, plump, and has excellent elasticity. The second channel is the smoothness score, ranging from 0 to 100. This score inversely maps the roughness and texture depth of the skin surface. The lower the score (e.g., 0-40), the more obvious the fine lines or large pores are on the skin, while the higher the score (e.g., 80-100), the smoother and finer the epidermis. The third channel is a dynamic microcirculation activity index, with a value range of 0-100. This index directly characterizes the blood perfusion intensity of the subcutaneous microcirculation. For example, 0-30 indicates slow local microcirculation or microvascular constriction; while 70-100 indicates abundant blood flow and active metabolism in the subcutaneous microvascular network.

[0061] Through the established dimensional mapping and numerical interpretation rules, this system achieves an integrated intelligent physiological monitoring function with high precision, high robustness, and clear physiological reference value.

[0062] Based on the above method embodiments, the present invention also provides a real-time skin texture monitoring system based on image recognition. In this embodiment, the system serves as the software and hardware execution carrier of the above monitoring method, and is deployed in an electronic device with a parallel computing architecture to realize dynamic environmental intelligent monitoring driven by physical priors.

[0063] In this invention, the system mainly includes a spatial frequency domain decoupling module, a parallel physical feature generation module, a micro-motion phase controlled extraction module, a composite graph topology construction module, and a graph network deduction output module. Each module achieves seamless flow and joint constraints of tensor data through a high-speed data bus.

[0064] The system acquires a continuous stream of skin video images from a front-end photoelectric sensor via a spatial-frequency domain decoupling module. This module converts the input image sequence to an orthogonal color space to separate the luminance channels and calls the underlying operator of the Laplacian pyramid algorithm to perform iterative Gaussian smoothing downsampling and interpolation upsampling. After decomposition operations by this module, the system outputs a multi-scale image sequence that removes global luminance trends and contains details at multiple spatial resolutions.

[0065] The parallel physical feature generation module and the spatial frequency domain decoupling module operate in strict synchronization and parallelism, generating three major physical constraint priors based on the continuous skin video image stream. This module extracts the local pixel gradient matrix of the brightness channel and performs Gaussian smoothing integration to generate a subsurface scattering mapping matrix that realistically represents the skin's elasticity and hydration state. Simultaneously, this module calculates and aggregates the environmental rigidity motion tensor by tracking the inter-frame optical flow displacement of feature points in non-skin target regions. For local noise, this module utilizes morphological top-hat transformation combined with adaptive thresholding segmentation techniques to remove highlight and hair-occluded regions, generating a binary morphological prior mask with absolute isolation properties.

[0066] The micro-motion phase-controlled extraction module takes over the aforementioned multi-scale image sequence and the generated subsurface scattering mapping matrix. This module incorporates a configurable temporal bandpass filter and establishes a nonlinear mapping path strictly based on the numerical values ​​of each spatial pixel in the subsurface scattering mapping matrix to dynamically adjust the filter's upper and lower frequency limits and phase amplification coefficient. After adjustment, the module performs micro-strain phase amplification on the images at each scale in the time dimension, and then performs multi-scale reconstruction of the filtering results, spatially superimposing them with the original brightness information to finally output a high-fidelity micro-strain phase tensor.

[0067] The composite graph topology construction module is responsible for converting data from the continuous image matrix domain into graph structure data in non-Euclidean space. This module extracts local extrema of the micro-strain phase tensor to establish a set of graph nodes and calculates the feature distance similarity between any two graph nodes. When constructing the core adjacency matrix, this module strictly adheres to the joint physical constraint mechanism of this invention: on one hand, it extracts the temporal displacement component from the graph node feature vectors, calculates the directional coincidence of this component with the environment's rigid motion tensor to construct a motion penalty term, and subtracts this penalty term from the feature distance similarity; on the other hand, it multiplies the attenuated similarity value with the Boolean value of any two graph nodes in the morphological prior mask to generate a dynamic graph topology with absolute resistance to environmental jitter and noise.

[0068] The graph network deduction output module receives the constructed dynamic graph topology and performs high-dimensional abstract deduction. This module utilizes the spatiotemporal characteristics of graph nodes to construct a hidden node feature matrix and performs symmetric normalization on the received adjacency matrix. Specifically, the module first adds an identity matrix to the received adjacency matrix to construct an augmented adjacency matrix with self-loops, and then performs symmetric normalization on this augmented matrix to generate the final normalized adjacency matrix. Within the dynamic graph convolutional network architecture, this module utilizes the normalized adjacency matrix, hidden node feature matrix, and learnable weight matrix within the network to perform multi-level node information aggregation. The calculation formula for the single-layer forward propagation feature map of this module is defined as: in, The adjacency matrix is ​​the one that has undergone symmetric normalization and contains self-loops; For the first The input hidden node feature matrix of the network layer (especially when , (The initial node feature matrix is ​​formed by stacking the spatiotemporal feature vectors of each graph node initially extracted by the system). For the first The network layers contain learnable weight matrices with optimizable parameters; This represents a non-linear activation function mapping operation (such as the ReLU activation function). This is the updated feature matrix output to the next network layer after aggregation and mapping.

[0069] After the aforementioned forward propagation through multiple layers (to avoid oversmoothing of node features during deep propagation in graph neural networks, the network depth is preferably 2 to 3 layers in this embodiment), the hidden feature matrix of the highest level is fed to a global average pooling layer for topological dimensionality reduction. Finally, this module outputs high-precision quantized skin texture parameters to the terminal interface through the linear mapping mechanism of the fully connected classification layer. In the specific system output format, these parameters are represented as a one-dimensional multi-channel feature vector. The system synchronously and independently outputs the skin elasticity index, smoothness score, and micro-motion activity index of the current tested object through different dimensional channels of this feature vector.

[0070] Furthermore, the present invention also provides a computer-readable storage medium and a corresponding electronic computing device. In this embodiment, the electronic computing device includes a high-performance communication bus, a memory, and a processor. The processor reads computer program instructions pre-programmed in the memory and mobilizes underlying computing resources to perform the aforementioned multidimensional tensor operations, feature map frequency domain decoupling, and deep graph neural network inference calculations. The computer-readable storage medium non-volatilely stores computer-executable instructions. When the instruction sequence is read and executed by the processor of the computing device, the complete technical logic of the above-mentioned image recognition-based real-time skin texture monitoring method is strictly triggered and executed in sequence.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time skin texture monitoring method based on image recognition, characterized in that, include: A continuous stream of skin video images is acquired, and the luminance channel of the continuous skin video image stream is spatially decomposed in the frequency domain to obtain a multi-scale image sequence. Based on the continuous skin video image stream, a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask are generated synchronously. Micro-strain phase tensors are extracted from the multi-scale image sequence based on the Euler spatiotemporal micro-amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. A dynamic graph topology is constructed using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology is jointly constrained by the environmental rigid motion tensor and the morphological prior mask. The dynamic graph topology is subjected to graph convolution aggregation processing to output quantized skin texture parameters.

2. The real-time skin texture monitoring method based on image recognition according to claim 1, characterized in that, The process of acquiring a continuous skin video image stream and performing spatial frequency domain decomposition on the luminance channel of the continuous skin video image stream to obtain a multi-scale image sequence specifically includes: The acquired continuous skin video image stream is converted from a color space to an orthogonal color space, and the luminance channel is extracted. By performing iterative Gaussian smoothing downsampling and interpolation upsampling on the brightness channel using the Laplacian pyramid, the global brightness trend and local high-frequency information of the image are separated, generating the multi-scale image sequence with different spatial resolutions.

3. The real-time skin texture monitoring method based on image recognition according to claim 1, characterized in that, The process of synchronously generating a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask based on the continuous skin video image stream specifically includes: The local pixel gradient matrix of the brightness channel is extracted, and the local pixel gradient matrix is ​​locally integrated using a Gaussian smoothing kernel of a preset scale to generate the subsurface scattering mapping matrix used to characterize the physical elasticity and hydration state of the skin. Non-skin target regions are located using skin color segmentation or target detection, and static feature points are extracted within these regions. The displacement vectors of the static feature points are tracked using an inter-frame feature matching algorithm, and the displacement vectors are fitted using an affine transformation model to calculate the environmental rigid motion tensor containing translation and rotation components. An adaptive threshold morphological top-hat transform operation is applied to the continuous skin video image stream to remove abrupt noise regions and generate the morphological prior mask composed of binary values.

4. The real-time skin texture monitoring method based on image recognition according to claim 1, characterized in that, The method for extracting micro-strain phase tensors from the multi-scale image sequence based on the Euler spatiotemporal micro-amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix, specifically including: A time-domain bandpass filter is applied to each level of the multi-scale image sequence; Based on the numerical values ​​of each pixel in the subsurface scattering mapping matrix, the upper and lower frequency limits and phase amplification coefficient of the time-domain bandpass filter are dynamically adjusted. The time-domain bandpass filter with dynamically adjusted parameters is used for filtering. The filtered result is then reconstructed at multiple scales and superimposed on the brightness channel to generate the corresponding micro-strain phase tensor.

5. The real-time skin texture monitoring method based on image recognition according to claim 1, characterized in that, The dynamic graph topology is constructed using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology is jointly constrained by the environmental rigidity motion tensor and the morphological prior mask, specifically including: Extract the feature vector of each graph node in the spatiotemporal domain, wherein the feature vector is composed of the spatial coordinate information of the graph node in the image plane, the micro-strain phase information, and the temporal phase change rate; calculate the feature distance similarity between any two graph nodes. Calculate the directional coincidence degree between the temporal displacement component of the feature vector and the rigid motion tensor of the environment, and use the directional coincidence degree as a motion penalty term to numerically decay the feature distance similarity; The decayed feature distance similarity is multiplied by the mask value corresponding to any two graph nodes in the morphological prior mask to obtain the matrix elements of the adjacency matrix, so as to forcibly cut off the connection path of graph nodes in the noise region through the product operation.

6. The real-time skin texture monitoring method based on image recognition according to claim 1, characterized in that, The step of performing graph convolution aggregation processing on the dynamic graph topology to output quantized skin texture parameters specifically includes: The feature vectors of the graph nodes in the spatiotemporal domain are extracted to construct the hidden node feature matrix. The degree matrix is ​​calculated based on the adjacency matrix, and the adjacency matrix is ​​subjected to symmetric normalization. In the dynamic graph convolutional network, multi-level aggregation and non-linear activation mapping are performed using the normalized adjacency matrix, the hidden node feature matrix and the learnable weight matrix. Specifically, this includes: performing graph convolution operations in each hidden layer to aggregate the features of neighboring nodes, and sequentially performing feature dimensionality reduction and mapping updates through batch normalization and non-linear activation functions. The hidden feature matrix output from the highest level is sequentially passed through a global average pooling layer and a fully connected classification layer to map and output the quantized skin texture parameters.

7. A real-time skin texture monitoring system based on image recognition, characterized in that, include: The spatial frequency domain decoupling module is used to acquire a continuous skin video image stream and perform spatial frequency domain decomposition on the brightness channel of the continuous skin video image stream to obtain a multi-scale image sequence. The parallel physical feature generation module is used to synchronously generate a subsurface scattering mapping matrix, an environmental rigidity motion tensor, and a morphological prior mask based on the continuous skin video image stream. The micro-motion phase controlled extraction module is used to extract the micro-strain phase tensor from the multi-scale image sequence based on the Euler spatiotemporal micro-motion amplification mechanism, wherein the amplification parameters of the Euler spatiotemporal micro-motion amplification mechanism are dynamically constrained by the subsurface scattering mapping matrix. A composite graph topology construction module is used to construct a dynamic graph topology structure using the local extrema of the micro-strain phase tensor as graph nodes. The adjacency matrix of the dynamic graph topology structure is jointly constrained by the environmental rigid motion tensor and the morphological prior mask. The graph network inference output module is used to perform graph convolution aggregation processing on the dynamic graph topology and output quantized skin texture parameters.

8. The real-time skin texture monitoring system based on image recognition according to claim 7, characterized in that, The parallel physical feature generation module is specifically used for: The local pixel gradient matrix of the brightness channel is extracted and local integration is performed using a Gaussian smoothing kernel of a preset scale to generate the subsurface scattering mapping matrix. Non-skin target regions are located using skin color segmentation or target detection, and static feature points are extracted within these regions. The displacement vectors of the static feature points are tracked using an inter-frame feature matching algorithm, and the displacement vectors are fitted using an affine transformation model to calculate the environmental rigid motion tensor containing translation and rotation components. An adaptive threshold morphological top-hat transform operation is applied to the continuous skin video image stream to remove abrupt noise regions and generate the morphological prior mask composed of binary values.

9. The real-time skin texture monitoring system based on image recognition according to claim 7, characterized in that, The micro-motion phase-controlled extraction module is specifically used for: The frequency upper and lower limits and phase amplification factor of the time-domain bandpass filter are dynamically adjusted according to the value of the subsurface scattering mapping matrix. The time-domain bandpass filter with dynamically adjusted parameters is used for filtering. The obtained filtering results are reconstructed at multiple scales and then superimposed on the brightness channel to generate the micro-strain phase tensor.

10. The real-time skin texture monitoring system based on image recognition according to claim 7, characterized in that, When calculating the adjacency matrix, the composite graph topology construction module performs the following logic: based on the feature distance similarity between any two graph nodes, subtracts the motion penalty term calculated based on the directional coincidence between the temporal displacement component of the feature vector and the rigid motion tensor of the environment, and multiplies the attenuated result by the mask value corresponding to the two graph nodes in the morphological prior mask.

Images