AI face diagnosis feature and aesthetics analysis method and system for medical beauty

Abstract

The invention provides a medical beauty-oriented AI face diagnosis feature and aesthetics analysis method and system, and relates to the technical field of intelligent medical beauty, and the method comprises the steps: collecting and preprocessing user face image data to obtain standardized face data; constructing a time sequence feature extraction model based on a deep convolutional neural network, and extracting facial dynamic features in combination with an optical flow estimation algorithm; and constructing a face reconstruction model based on the depth estimation network, and reconstructing a face three-dimensional structure by using a back projection mechanism and a geodesic distance partitioning strategy. According to the invention, accurate analysis of facial features can be realized, and the accuracy and individuation level of medical beauty diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical aesthetics technology, and in particular to an AI facial diagnosis feature and aesthetic analysis method and system for medical aesthetics. Background Art

[0002] With the rapid development of the medical aesthetics industry, demand for personalized medical aesthetics services is growing. Traditional medical aesthetics facial consultations rely primarily on the physician's experience and subjective judgment, lacking scientific and quantitative analytical methods. In recent years, the application of artificial intelligence technology in the field of medical aesthetics has gradually emerged, and facial feature analysis has become a key research direction in the industry. Currently, medical aesthetics facial consultations are primarily conducted through steps such as image acquisition, feature extraction, and aesthetic analysis. Accurate extraction of dynamic and three-dimensional facial features is crucial for formulating medical aesthetics plans.

[0003] Existing medical aesthetics AI facial diagnosis technology mainly performs static feature analysis based on two-dimensional images, identifies facial key points through deep learning algorithms, and evaluates based on traditional aesthetic standards.

[0004] However, this type of technology has some flaws and shortcomings when it comes to complex facial features and personalized aesthetic needs. Existing technologies primarily focus on static facial features, ignoring dynamic features during facial expression changes. This makes it impossible to fully capture important information such as facial muscle movement and skin elasticity, resulting in a lack of consideration of dynamic aesthetics in medical aesthetics solutions. Existing facial feature extraction methods are mostly based on two-dimensional image analysis, lacking the ability to accurately reconstruct and measure three-dimensional facial structures. This makes it difficult to accurately assess three-dimensional features such as facial contour, volume, and curvature, affecting the accuracy and personalization of medical aesthetics solutions. Existing technologies also lack a stability assessment mechanism during feature extraction, making them susceptible to external factors such as lighting changes and posture differences, leading to inconsistent analysis results and making it difficult for medical aesthetics professionals to provide a reliable basis for decision-making. Summary of the Invention

[0005] The embodiments of the present invention provide an AI facial diagnosis feature and aesthetic analysis method and system for medical aesthetics, which can solve the problems in the existing technology.

[0006] A first aspect of an embodiment of the present invention provides an AI facial diagnosis feature and aesthetic analysis method for medical aesthetics, including:

[0007] Collecting user facial image data, and preprocessing the facial image data to obtain standardized facial data;

[0008] A temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated by combining the optical flow estimation algorithm to obtain facial dynamic features.

[0009] A facial reconstruction model is constructed based on a depth estimation network. The 3D facial structure is reconstructed using a back-projection mechanism and a geodesic distance partitioning strategy. The 3D facial features are obtained through feature stability scoring and projection invariance matching.

[0010] A dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors. The semantic similarity matrix of the feature mapping vectors is calculated. The semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain a comprehensive facial evaluation index.

[0011] Based on comprehensive facial assessment indicators and combined with the preset medical aesthetics expert knowledge base, facial areas are graded and assessed, a facial analysis report is generated, and medical aesthetics recommendations are output.

[0012] In an optional embodiment, a temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of the standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated in combination with an optical flow estimation algorithm. The facial dynamic features obtained include:

[0013] By setting multiple convolution kernel arrays of continuously increasing scales in a deep convolutional neural network, convolution operations are performed on the standardized facial data to generate a multi-level feature response map;

[0014] performing feature cross-correlation operations on adjacent frames in the multi-level feature response map, obtaining an inter-frame displacement correlation matrix by calculating regional similarity of feature responses, and generating an initial motion estimation map of the facial region according to the inter-frame displacement correlation matrix;

[0015] Performing adaptive threshold segmentation on the initial motion estimation image, obtaining an optimal segmentation threshold by iterative calculation of maximizing inter-class variance, and generating a facial region motion feature map using the optimal segmentation threshold;

[0016] The facial region motion feature map is processed using an optical flow estimation algorithm, an objective function is constructed based on a self-attention mechanism and deformation constraints, and a facial region deformation field is obtained by solving the problem using an iterative splitting algorithm;

[0017] The facial region deformation field is divided into regions, the deformation strength value in each preset facial key region is calculated, and the deformation strength values ​​of each region are combined to form a facial dynamic feature vector.

[0018] In an optional embodiment, the facial region motion feature map is processed using an optical flow estimation algorithm, an objective function is constructed based on a self-attention mechanism and deformation constraints, and the facial region deformation field is obtained by solving the problem using an iterative splitting algorithm. The following steps are included:

[0019] Constructing a tensor representation of two adjacent frames in the facial region motion feature map, embedding a self-attention mechanism in the tensor representation, generating attention weights by calculating the correlation coefficient between the images, enhancing the global motion information based on the attention weights, and constructing an initialization representation of the optical flow field;

[0020] Based on predetermined facial anatomical features, a topological structure relationship graph of the facial region is constructed, deformation consistency constraints of the facial region are established, and node features in the topological structure relationship graph are aggregated using a graph convolutional network to generate motion constraints based on the facial structure;

[0021] Inputting the initialization representation of the optical flow field and the motion constraint condition into a variational optimization framework, constructing an adaptive loss function including a data term based on an attention mechanism and a regularization term based on a topological structure, calculating the gradient of the loss function for the data term and the regularization term respectively, constructing a weight update matrix based on the gradient of the loss function, determining the weight coefficient of the data term and the weight coefficient of the regularization term and multiplying them together to obtain a target optimization function that is dynamically adjusted according to the motion mode;

[0022] An iterative splitting algorithm is used to solve the target optimization function, the optimization variable is decomposed into multiple sub-variables, and a splitting sequence is constructed for each sub-variable to perform alternating updates;

[0023] Based on the optimization result of the iterative splitting algorithm, an optical flow field of the facial region is obtained, and the amplitude of the optical flow field is calculated to obtain a deformation field of the facial region.

[0024] In an optional embodiment, the iterative splitting algorithm includes:

[0025] Performing variable decomposition on the target optimization function, decomposing the optimization variables into original variables and constraint variables, and constructing a variable update path graph;

[0026] Constructing a branch optimization sequence in the variable update path graph, generating a local feature manifold by mapping the original variables to neighboring points, and performing gradient projection on the local feature manifold to obtain the optimization direction; constraining variables to construct a sparse reconstruction sequence by measuring feature distance;

[0027] Determining a feature manifold projection operator based on the curvature distribution of the local feature manifold, determining a sparse reconstruction operator based on the singular value decomposition of the sparse feature distribution, and combining the feature manifold projection operator and the sparse reconstruction operator to form an adaptive neighbor operator;

[0028] Alternately updating the original variables and the constraint variables based on the adaptive proximity operator, with the update step size determined by the geodesic distance on the characteristic manifold, and stopping the iteration when the difference between the original variables and the constraint variables is less than a preset threshold;

[0029] The original variables after iterative convergence are used as the final optical flow field solution results.

[0030] In an optional embodiment, a facial reconstruction model is constructed based on a depth estimation network, and a back-projection mechanism and a geodesic distance partitioning strategy are used to reconstruct the three-dimensional structure of the face. The three-dimensional facial features obtained by feature stability scoring and projection invariance matching include:

[0031] The standardized facial image is input into a multi-layer convolutional network for feature extraction, and a facial depth map is output. A back-projection mechanism is used to generate a projected facial image from the facial depth map. The mutual information loss between the projected facial image and the standardized facial image is calculated. The depth consistency loss of the facial depth map is combined to construct a depth estimation loss function. The multi-layer convolutional network is optimized to obtain an optimized facial depth map.

[0032] The depth values ​​and pixel coordinates in the optimized facial depth map are converted into a 3D point cloud. The geodesic distance between point pairs in the 3D point cloud is calculated to establish a point cloud partitioning strategy. The principal curvature direction of the local surface is calculated to construct a reconstruction reference basis. The 3D closed surface of the face is obtained by surface fitting the 3D point cloud based on the direction information of the reconstruction reference basis.

[0033] The local curvature is calculated along the reconstruction reference base direction on the three-dimensional closed surface of the face to construct a gradient response map. The feature stability score is calculated based on the curvature flow equation. Stable feature points are extracted through nonlinear diffusion filtering, and local deformation feature descriptions are extracted from the stable feature points.

[0034] Based on the local deformation feature description, a projection invariant matching criterion is established, the local affine transformation matrix of the local deformation feature description is calculated for feature matching, and matching point pairs are selected in combination with spatial consistency constraints;

[0035] The local surface deformation strength of the matching point pairs is calculated, and feature weight coefficients are assigned. The three-dimensional feature representation of the face is generated by combining the features of the matching point pairs.

[0036] In an optional embodiment, a dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors, a semantic similarity matrix of the feature mapping vectors is calculated, and the semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain comprehensive facial evaluation indicators including:

[0037] Inputting facial dynamic features and facial three-dimensional features into a predetermined dual-channel feature extractor, performing multi-layer feature extraction through a temporal convolutional network and a spatial convolutional network to generate dynamic mapping features and spatial mapping features, performing multi-layer feature extraction through a spatial convolutional network to generate spatial mapping features, combining the dynamic mapping features with the spatial mapping features to generate a feature mapping vector, and performing a normalized dot product operation to generate an initial semantic similarity matrix;

[0038] Dividing an initial semantic similarity matrix into a feature sub-block matrix, obtaining the number of rows and columns of the feature sub-block matrix, determining the minimum value of the number of rows and columns as a block size parameter, constructing a square region with the block size parameter as a side length as a first region, and other regions as a second region, performing a sparse constraint operation on the first region to generate a sparse matrix, performing an association constraint operation on the second region to generate an association constraint matrix, inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization, generating an optimized feature sub-block, and fusing the optimized feature sub-blocks to generate an optimized semantic similarity matrix;

[0039] A weight coefficient matrix is ​​calculated based on the optimized semantic similarity matrix, and is weightedly combined with the feature mapping vector to generate a comprehensive facial evaluation index.

[0040] In an optional embodiment, inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization to generate optimized feature sub-blocks includes:

[0041] Performing maximum and minimum normalization operations on the sparse matrix and the association constraint matrix respectively to generate a first normalized matrix and a second normalized matrix, setting the first normalized matrix at a diagonal position and setting the second normalized matrix at a non-diagonal position to generate a matrix to be optimized;

[0042] Calculating the median of the absolute values ​​of each element in the matrix to be optimized as an initial threshold, calculating a matrix distribution parameter according to the dimension value of the matrix to be optimized and the cumulative sum of the absolute values ​​of each element in the matrix to be optimized, adding the initial threshold to the matrix distribution parameter and multiplying the result by an exponential decay function to generate a dynamic shrinkage threshold;

[0043] Keeping the values ​​of elements in the matrix to be optimized that are greater than the dynamic shrinkage threshold unchanged and setting the values ​​of elements that are less than or equal to the dynamic shrinkage threshold to zero, generating a sparse matrix, and calculating the relative error between the sparse matrix and the matrix to be optimized;

[0044] Repeat the iteration until the relative error value is less than or equal to the preset error threshold, determine the final sparse matrix, and reconstruct the final sparse matrix according to the original position relationship to generate optimized feature sub-blocks.

[0045] A second aspect of the present invention provides an AI facial diagnosis feature and aesthetic analysis system for medical aesthetics, including:

[0046] The first unit is configured to collect user facial image data and pre-process the facial image data to obtain standardized facial data;

[0047] The second unit is used to build a temporal feature extraction model based on a deep convolutional neural network. It uses multi-scale convolution operations to perform feature response and displacement estimation on standardized facial data, and combines the optical flow estimation algorithm to calculate the facial region deformation field to obtain facial dynamic features.

[0048] The third unit is used to build a facial reconstruction model based on the depth estimation network, reconstruct the facial 3D structure using the back-projection mechanism and geodesic distance partitioning strategy, and obtain the facial 3D features through feature stability scoring and projection invariance matching;

[0049] The fourth unit is used to use a dual-channel feature extractor to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors, calculate the semantic similarity matrix of the feature mapping vectors, perform block sparse optimization and iterative shrinkage on the semantic similarity matrix, and obtain a comprehensive facial evaluation index;

[0050] The fifth unit is used to perform graded assessments of facial areas based on comprehensive facial assessment indicators and a preset knowledge base of medical aesthetic experts, generate facial analysis reports, and output medical aesthetic recommendation plans.

[0051] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0056] In an embodiment of the present invention, by collecting user facial image data and performing standardized preprocessing, the standardized processing of facial data is achieved, which lays the foundation for subsequent feature extraction and analysis, and improves the accuracy and consistency of the analysis results; the temporal feature extraction model based on the deep convolutional neural network combined with the optical flow estimation algorithm can accurately capture facial dynamic features and deformation information, effectively identify facial muscle activity and expression changes, and provide a dynamic dimension evaluation basis for medical aesthetic diagnosis; the facial reconstruction model constructed by the deep estimation network realizes the accurate reconstruction of the three-dimensional structure of the face. Combined with the geodesic distance partitioning strategy and the feature stability scoring mechanism, it overcomes the limitations of traditional two-dimensional analysis, provides a more comprehensive and three-dimensional facial aesthetic analysis result, and significantly improves the personalized customization capability of medical aesthetic solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of the AI ​​facial diagnosis features and aesthetic analysis method for medical aesthetics according to an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the simulation results of facial area motion features;

[0059] Figure 3 Schematic diagram of the simulation results of the iterative shrinkage algorithm performance. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0061] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 This is a flow chart of the AI ​​facial diagnosis features and aesthetic analysis method for medical aesthetics according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] Collecting user facial image data, and preprocessing the facial image data to obtain standardized facial data;

[0064] A temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated by combining the optical flow estimation algorithm to obtain facial dynamic features.

[0065] A facial reconstruction model is constructed based on a depth estimation network. The 3D facial structure is reconstructed using a back-projection mechanism and a geodesic distance partitioning strategy. The 3D facial features are obtained through feature stability scoring and projection invariance matching.

[0066] A dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors. The semantic similarity matrix of the feature mapping vectors is calculated. The semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain a comprehensive facial evaluation index.

[0067] Based on comprehensive facial assessment indicators and combined with the preset medical aesthetics expert knowledge base, facial areas are graded and assessed, a facial analysis report is generated, and medical aesthetics recommendations are output.

[0068] In an optional embodiment, a temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of the standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated in combination with the optical flow estimation algorithm. The facial dynamic features obtained include:

[0069] By setting multiple convolution kernel arrays of continuously increasing scales in a deep convolutional neural network, convolution operations are performed on the standardized facial data to generate a multi-level feature response map;

[0070] performing feature cross-correlation operations on adjacent frames in the multi-level feature response map, obtaining an inter-frame displacement correlation matrix by calculating regional similarity of feature responses, and generating an initial motion estimation map of the facial region according to the inter-frame displacement correlation matrix;

[0071] Performing adaptive threshold segmentation on the initial motion estimation image, obtaining an optimal segmentation threshold by iterative calculation of maximizing inter-class variance, and generating a facial region motion feature map using the optimal segmentation threshold;

[0072] The facial region motion feature map is processed using an optical flow estimation algorithm, an objective function is constructed based on a self-attention mechanism and deformation constraints, and a facial region deformation field is obtained by solving the problem using an iterative splitting algorithm;

[0073] The facial region deformation field is divided into regions, the deformation strength value in each preset facial key region is calculated, and the deformation strength values ​​of each region are combined to form a facial dynamic feature vector.

[0074] In one specific embodiment, a normalization process is performed on the input facial image sequence. This process includes facial alignment and resizing. Facial landmarks are located using a 68-point facial landmark detection algorithm, a standard coordinate system is established based on the inner corners of the eyes and the tip of the nose, the facial images are aligned to a standard position, and the images are resized to a uniform resolution of 256×256 pixels. This normalization process ensures that subsequent feature extraction is unaffected by pose variations and size differences.

[0075] The normalized facial data was input into a designed deep convolutional neural network, consisting of five convolutional layers, each with an array of convolutional kernels of varying sizes. The first layer used 32 kernels of each of three sizes: 3×3, 5×5, and 7×7; the second layer used 64 kernels of each of three sizes: 5×5, 7×7, and 9×9; the third layer used 128 kernels of each of three sizes: 7×7, 9×9, and 11×11; the fourth layer used 256 kernels of each of three sizes: 9×9, 11×11, and 13×13; and the fifth layer used 512 kernels of each of three sizes: 11×11, 13×13, and 15×15. After each convolution layer, a ReLU activation function was used to introduce nonlinearity, and a max-pooling operation with a stride of 2 was employed to reduce the size of the feature maps. This multi-scale convolution operation enabled the network to capture facial texture and structural features at different scales, generating a multi-level feature response map.

[0076] For consecutive facial image frames, the correlation between their feature response maps needs to be calculated to estimate motion information. A feature cross-correlation operation is performed on the feature response maps of each pair of adjacent frames. Specifically, for the feature response maps of frames t and t+1, a 9×9 local region is taken at each position in the feature map of frame t, and a 21×21 range is searched in the feature map of frame t+1. The best matching position is found by calculating the normalized cross-correlation coefficient. The normalized cross-correlation coefficient is calculated by dividing the dot product of the feature region by the product of the norms of the two regions and ranges from [-1, 1], with larger values ​​indicating a higher degree of match. For each local region, the maximum cross-correlation coefficient and its corresponding displacement vector are recorded to form an inter-frame displacement correlation matrix. For example, for a point in the eyebrow region of a face, if the offset from the best matching position between frame t and frame t+1 is (2, -1), then the displacement vector of that point is (2, -1), indicating a movement of 2 pixels to the right and 1 pixel upward.

[0077] Based on the inter-frame displacement correlation matrix, an initial motion estimation map of the facial region is generated. The horizontal and vertical components of the displacement vector are mapped to the two channels of the motion estimation map. To reduce the influence of noise, the motion estimation map is smoothed using a 5×5 Gaussian filter with a standard deviation of 1.5. The smoothed motion estimation map intuitively shows the motion trends of various facial regions, but further segmentation is required to highlight areas of significant motion.

[0078] Adaptive threshold segmentation is performed on the initial motion estimation image to identify regions of significant facial motion. The OTSU algorithm, based on maximum inter-class variance, is used to determine the optimal segmentation threshold. Specifically, the grayscale histogram of the motion estimation image is calculated, and the number of pixels at each grayscale level is counted. For a possible threshold T, image pixels are divided into two classes: background classes with values ​​less than or equal to T, and foreground classes with values ​​greater than T. The average grayscale values ​​of the two classes of pixels and their respective proportions in the overall image are calculated, and the inter-class variance is then calculated. By iterating through all possible thresholds (0-255), the threshold that maximizes the inter-class variance is selected as the optimal segmentation threshold. In practical applications, the optimal threshold is typically between 10 and 20 for calm faces; for facial expressions, the optimal threshold may rise to 30-50. Using the optimal threshold, the motion estimation image is binarized to generate a facial region motion feature map, where pixels with a value of 1 represent regions of significant motion, and pixels with a value of 0 represent regions of stillness or minimal motion.

[0079] To accurately capture the deformation information of the facial region, an optical flow estimation algorithm is applied to the facial region motion feature map. This algorithm combines the self-attention mechanism with deformation constraints to construct an objective function consisting of a data term, a smoothing term, and a self-attention term. The data term ensures the consistency of the deformation field with the observed data, the smoothing term ensures the local continuity of the deformation field, and the self-attention term enhances deformation consistency in similar regions. An iterative splitting algorithm is used to solve the objective function, which decomposes the large-scale optimization problem into a series of smaller subproblems. In each iteration, the self-attention weights are first fixed and the deformation field is updated; then the deformation field is fixed and the self-attention weights are updated. Typically, after 20-30 iterations, the algorithm converges to a stable solution, obtaining an accurate deformation field for the facial region. The deformation field is represented as a two-dimensional vector field, describing the direction and magnitude of the displacement of each facial point.

[0080] The facial deformation field is segmented and features are extracted. Based on facial anatomical features, the face is divided into eight key regions: the left eyebrow region, the right eyebrow region, the left eye region, the right eye region, the nose region, the left cheek region, the right cheek region, and the mouth region. The deformation intensity values ​​within each region are statistically analyzed, and the average amplitude and directional consistency of the deformation vector within the region are calculated. For example, during a change in facial expression, the average deformation amplitude of the mouth region may reach 5-8 pixels, while the average deformation amplitude of the eyebrow region is typically in the range of 2-4 pixels. The deformation features of the eight regions are combined into a 16-dimensional feature vector (two eigenvalues ​​per region: average amplitude and directional consistency) to form the final facial dynamic feature vector. These feature vectors can be used in subsequent applications such as expression recognition and emotion analysis.

[0081] In this embodiment, multi-level features are extracted through multi-scale convolution kernels, and the inter-frame similarity is analyzed based on feature cross-correlation to achieve more accurate initial motion estimation; combined with adaptive threshold segmentation and optical flow estimation, the facial motion feature map can effectively separate the significant motion areas, improving the meticulousness and discriminability of dynamic feature extraction; the self-attention mechanism and deformation constraints are used to construct optimization targets, making the deformation field estimation more stable and structurally reasonable, and adapting to complex facial expression changes; through region division and deformation intensity calculation, a dynamic feature vector with regional distinction is generated, providing high-quality input for subsequent emotion recognition, identity authentication and other tasks.

[0082] In an optional embodiment, the facial region motion feature map is processed using an optical flow estimation algorithm, an objective function is constructed based on a self-attention mechanism and deformation constraints, and the facial region deformation field is obtained by solving the problem using an iterative splitting algorithm. The following steps are included:

[0083] Constructing a tensor representation of two adjacent frames in the facial region motion feature map, embedding a self-attention mechanism in the tensor representation, generating attention weights by calculating the correlation coefficient between the images, enhancing the global motion information based on the attention weights, and constructing an initialization representation of the optical flow field;

[0084] Based on predetermined facial anatomical features, a topological structure relationship graph of the facial region is constructed, deformation consistency constraints of the facial region are established, and node features in the topological structure relationship graph are aggregated using a graph convolutional network to generate motion constraints based on the facial structure;

[0085] Inputting the initialization representation of the optical flow field and the motion constraint condition into a variational optimization framework, constructing an adaptive loss function including a data term based on an attention mechanism and a regularization term based on a topological structure, calculating the gradient of the loss function for the data term and the regularization term respectively, constructing a weight update matrix based on the gradient of the loss function, determining the weight coefficient of the data term and the weight coefficient of the regularization term and multiplying them together to obtain a target optimization function that is dynamically adjusted according to the motion mode;

[0086] An iterative splitting algorithm is used to solve the target optimization function, the optimization variable is decomposed into multiple sub-variables, and a splitting sequence is constructed for each sub-variable to perform alternating updates;

[0087] Based on the optimization result of the iterative splitting algorithm, an optical flow field of the facial region is obtained, and the amplitude of the optical flow field is calculated to obtain a deformation field of the facial region.

[0088] In one specific embodiment, a video sequence containing human faces is obtained and preprocessed, including normalizing each frame to 256×256 pixels and performing brightness correction to reduce the impact of lighting changes on subsequent processing. A face detection algorithm is then used to extract the facial region from the processed video sequence, generating a facial region motion feature map. Face detection utilizes an improved HOG feature extractor combined with a SVM classifier, achieving a detection accuracy exceeding 98.5%.

[0089] A tensor representation is constructed for two adjacent frames in the facial region motion feature map. Specifically, the t-th frame and the t+1-th frame are represented as three-dimensional tensors It and It+1, respectively. Each tensor has dimensions C×H×W, where C represents the number of channels, which is 3, corresponding to the three RGB channels; H and W represent the height and width of the image, respectively, both of which are 256. The two frames are concatenated into a four-dimensional tensor T with dimensions 2×C×H×W.

[0090] A self-attention mechanism is embedded in the tensor representation, generating attention weights by calculating the correlation coefficients between images. Specifically, the tensor T is mapped into a query matrix Q, a key matrix K, and a value matrix V through a convolutional layer with a kernel size of 3×3 and 128 output channels. The dot product of Q and K is calculated and divided by a scaling factor of 8 (the square root of the feature dimension). The attention weight matrix A is normalized using a softmax function. This matrix represents the correlation coefficients between image regions and has values ​​between 0 and 1. The attention weight matrix A is multiplied by the value matrix V to obtain a weighted feature representation, enhancing global motion information. This constructs the initialization representation F0 of the optical flow field, with dimensions of 2×H×W, representing the estimated horizontal and vertical displacements, respectively.

[0091] A topological graph G of the facial region is constructed based on pre-determined facial anatomical features. This graph contains 68 nodes, corresponding to facial landmarks such as the eyebrows, eyes, nose, mouth, and facial contours. The connections between nodes are determined based on knowledge of facial muscle anatomy, resulting in a total of 134 edges. Each node's initial feature vector has a dimension of 128, derived from image features extracted from a surrounding 16×16 pixel region.

[0092] A graph convolutional network (GCN) aggregates node features in a topological graph to generate motion constraints based on facial structure. The GCN consists of three layers, each with output feature dimensions of 64, 32, and 16, respectively. The weight of information transfer between adjacent nodes is calculated based on the Euclidean distance between them, with closer distances giving higher weights. A Gaussian kernel function is used for normalization, with a kernel width parameter of 0.5. By aggregating node features, a feature matrix C with dimensions of 68×16 is generated, reflecting the motion constraints of the facial structure.

[0093] The initialization of the optical flow field, F0, and the motion constraint, C, are fed into a variational optimization framework to construct an adaptive loss function. This loss function consists of two parts: an attention-based data term, Ld, and a topology-based regularization term, Lr. The data term measures the consistency of the estimated optical flow field with the assumption of constant image brightness, while the regularization term measures the smoothness of the optical flow field and its consistency with facial structure constraints.

[0094] The loss function gradients, grad_Ld and grad_Lr, are calculated for the data term and regularization term, respectively. A weight update matrix W is constructed based on the gradients, with dimensions H×W, where the weight at each position is dynamically adjusted based on the motion characteristics of the local region. Specifically, for areas with dramatic facial expression changes (such as the mouth and eyes), a higher data term weight α (ranging from 0.6 to 0.8) and a lower regularization term weight β (ranging from 0.2 to 0.4) are assigned by calculating the local gradient amplitude. For areas with less dramatic changes (such as the forehead and cheeks), a lower data term weight α (ranging from 0.3 to 0.5) and a higher regularization term weight β (ranging from 0.5 to 0.7) are assigned. The weight coefficients α and β satisfy the constraint α + β = 1. The weight coefficients are multiplied and combined with the corresponding terms to obtain the target optimization function L = α × Ld + β × Lr, which is dynamically adjusted according to the motion pattern.

[0095] An iterative splitting algorithm is used to solve the target optimization function. This algorithm decomposes the optimization variable F (optical flow field) into three sub-variables, F1, F2, and F3, and constructs an augmented Lagrangian function. The iterative process includes: updating the F1 subproblem using gradient descent with a step size of 0.01; updating the F2 subproblem using a soft threshold operator with a threshold parameter of 0.05; updating the F3 subproblem using a projection operation to ensure the validity of the solution; and updating the Lagrangian multiplier with a step size of 0.05. The iterative termination conditions are the maximum number of iterations (set to 100) or when the difference between the solutions of two consecutive iterations is less than a preset threshold (set to 10^-4). In practice, the algorithm typically converges within 30-50 iterations.

[0096] Based on the optimization results of the iterative splitting algorithm, we obtain the optical flow field F* of the facial region, with dimensions of 2×H×W. We then perform amplitude calculation on the optical flow field, calculating the Euclidean distance between the horizontal displacement F*_x(i, j) and the vertical displacement F*_y(i, j) at each pixel (i, j). This yields the deformation field D of the facial region, with dimensions of H×W. Larger values ​​in the deformation field indicate greater motion in that region, providing a direct reflection of the intensity distribution of facial expression changes.

[0097] Traditional optical flow methods lack accuracy when processing non-rigid deformations such as facial expressions, and are particularly prone to drift under changing lighting and rapid motion. Deep learning methods, while capable of strong feature extraction, lack explicit modeling of facial anatomy and struggle to capture the complex dependencies between facial regions, resulting in poor performance in analyzing subtle expressions. Furthermore, existing methods typically employ fixed-weight loss functions that fail to adapt to the motion characteristics of different facial regions, limiting their effectiveness in scenarios with complex expression variations.

[0098] The method of this embodiment introduces a method that combines the self-attention mechanism with tensor representation. By explicitly calculating the correlation coefficient between image regions, it enhances the ability to capture global motion information and solves the problem of insufficient local feature extraction in traditional methods. Secondly, a facial topology structure relationship diagram is constructed based on facial anatomical knowledge, and a graph convolutional network is combined for feature aggregation. The physiological constraints of facial muscle movement are explicitly modeled, and the sensitivity to subtle changes in facial expressions is improved. Based on the adaptive loss function framework, the weights of the data term and the regularization term are dynamically adjusted according to the motion characteristics of different facial regions, so that the algorithm can more accurately process areas with drastic expression changes and areas with smaller changes. Finally, an iterative splitting algorithm is used to solve the optimization problem, decomposing the complex problem into multiple easy-to-handle sub-problems, thereby improving the convergence and computational efficiency of the algorithm. In view of the particularity of facial expression analysis, facial anatomical knowledge is combined with advanced computer vision technology to construct an optical flow estimation framework that can adapt to complex expression changes.

[0099] Figure 2 This is a schematic diagram of the simulation results of facial region motion features. From the distribution characteristics, it can be seen that the scattered points in the mouth area are mainly concentrated in the high deformation field amplitude (0.7-0.9) and high attention weight (0.7-0.9) ranges, which reflects the algorithm's adaptive enhancement effect on areas with drastic expression changes; the scattered points in the eye area are distributed in the medium-high range (0.5-0.8), indicating that the algorithm has a good ability to capture small eye movements; the nose and contour points are distributed in the lower range (0.3-0.6), which is consistent with the relatively stable characteristics of these areas during expression changes. This distribution feature verifies that the method proposed in this application can adaptively adjust the attention weight and deformation constraints according to the motion characteristics of different facial regions, thereby improving the accuracy and robustness of facial expression dynamic analysis.

[0100] In an optional embodiment, the iterative splitting algorithm includes:

[0101] Performing variable decomposition on the target optimization function, decomposing the optimization variables into original variables and constraint variables, and constructing a variable update path graph;

[0102] Constructing a branch optimization sequence in the variable update path graph, generating a local feature manifold by mapping the original variables to neighboring points, and performing gradient projection on the local feature manifold to obtain the optimization direction; constraining variables to construct a sparse reconstruction sequence by measuring feature distance;

[0103] Determining a feature manifold projection operator based on the curvature distribution of the local feature manifold, determining a sparse reconstruction operator based on the singular value decomposition of the sparse feature distribution, and combining the feature manifold projection operator and the sparse reconstruction operator to form an adaptive neighbor operator;

[0104] Alternately updating the original variables and the constraint variables based on the adaptive proximity operator, with the update step size determined by the geodesic distance on the characteristic manifold, and stopping the iteration when the difference between the original variables and the constraint variables is less than a preset threshold;

[0105] The original variables after iterative convergence are used as the final optical flow field solution results.

[0106] In a specific embodiment, the target optimization function is subjected to variable decomposition. Specifically, the optimization variables are divided into two parts: original variables and constraint variables. The original variables represent the basic structural characteristics of the optical flow field, and the constraint variables represent the constraint conditions of the optical flow field. After the variable decomposition, a variable update path graph is constructed, which describes the mutual dependence between the original variables and the constraint variables and their respective update paths. The path graph is represented in the form of a directed acyclic graph, in which the nodes represent the variables and the edges represent the dependency between the variables. For example, in practical applications, the optical flow field can be represented as the original variable u and the constraint variable v, and the path graph contains two update paths from u to v and from v to u.

[0107] After constructing the variable update path graph, the algorithm constructs a branching optimization sequence within the path graph. For the original variables, a local feature manifold is generated using neighbor mapping. Neighbor mapping involves selecting the k points with the closest Euclidean distance to each variable point in the original variable space as the neighborhood point set. In practice, the value of k is typically set between 5 and 10, but can be adjusted based on the problem size. For example, for a 640×480 resolution image sequence, optimal performance is achieved when k = 8. The generated local feature manifold can be viewed as an approximate representation of the original variable in a local region. Subsequently, gradient projection is performed on the local feature manifold to determine the optimization direction. Gradient projection involves projecting the objective function's gradient vector onto the tangent space of the local feature manifold, thereby obtaining the optimization direction along the manifold. For the constrained variables, a sparse reconstruction sequence is constructed using a feature distance metric. This feature distance metric is based on the similarity between variables and is calculated using a modified cosine similarity. A sparse reconstruction sequence reconstructs the constrained variables using a small number of linear combinations of basis vectors, resulting in sparse reconstructed coefficients. The number of nonzero coefficients is typically set to no more than 10% of the total number of coefficients.

[0108] The algorithm then determines a feature manifold projection operator based on the curvature distribution of the local feature manifold. The curvature distribution reflects the curvature of the manifold in all directions and can be obtained by calculating the deviation in distance between points in a local neighborhood and the center point. Regions with larger curvature values ​​have greater curvature and require more refined local approximations. The feature manifold projection operator is a matrix operator that projects vectors in the original space onto the local feature manifold. Simultaneously, the algorithm determines a sparse reconstruction operator based on the singular value decomposition of the sparse feature distribution. Specifically, the algorithm performs singular value decomposition on the matrix consisting of the sparse reconstruction sequence and constructs the sparse reconstruction operator by taking the singular vectors corresponding to the first r largest singular values. The parameter r is selected based on the dataset complexity and is typically set to a minimum value that achieves a cumulative explained variance of 95%. In practical applications, for common optical flow datasets, r is typically between 15 and 25. Combining the feature manifold projection operator with the sparse reconstruction operator forms an adaptive neighbor operator that adaptively adjusts the projection direction and reconstruction weights to accommodate the feature distribution in different regions.

[0109] After obtaining the adaptive neighbor operator, the algorithm updates the original variables and the constraint variables alternately based on the operator. During the update process, the original variables are applied to the current state through the adaptive neighbor operator to obtain a new state, and the constraint variables are also updated in a similar way. The update step size is determined by the geodesic distance on the characteristic manifold. The geodesic distance refers to the length of the shortest path along the manifold surface, which can be approximated on a discrete point set by the Dijkstra algorithm. The step size is set to 0.1 to 0.5 times the geodesic distance, and is dynamically adjusted according to the number of iterations. The initial step size is larger to speed up convergence, and gradually decreases as the iteration proceeds to ensure accuracy. The algorithm checks the difference between the original variable and the constraint variable after each iteration, and stops the iteration when the difference is less than the preset threshold. In practical applications, the threshold is set to 10 -4 to 10 -6 The specific accuracy can be adjusted according to the accuracy requirements of the application scenario.

[0110] After the iterations are complete, the algorithm outputs the converged raw variables as the final optical flow field solution. This result represents the pixel displacement field between adjacent image frames and can be used for subsequent tasks such as motion analysis and target tracking.

[0111] Traditional optimization methods usually adopt a fixed optimization path, which is difficult to adapt to complex non-convex objective functions and easily falls into local optimality; most methods do not fully consider the geometric structural characteristics of the problem during the variable update process, resulting in slow convergence and limited accuracy when processing high-dimensional data; although the existing splitting algorithm decomposes the problem into sub-problems, the collaborative optimization strategy between sub-problems is relatively simple and fails to fully utilize the structural relationship between variables; when dealing with non-rigid body movements such as faces, due to the lack of modeling of the underlying manifold structure, existing methods find it difficult to accurately capture complex deformation characteristics.

[0112] By decomposing the optimization variables into primitive variables and constraint variables and constructing a variable update path graph, a more flexible optimization framework is provided, which can dynamically adjust the optimization path according to the characteristics of the problem. The concept of local feature manifold is introduced, and the high-dimensional optimization problem is converted into optimization on a low-dimensional manifold through neighboring point mapping and gradient projection, which better captures the intrinsic geometric structure of the data. At the same time, the characteristic distance metric and sparse reconstruction sequence are used for the constraint variables to effectively reduce the dimension of the solution space and improve the computational efficiency of the algorithm. The method of this embodiment combines the characteristic manifold projection operator and the sparse reconstruction operator to form an adaptive neighbor operator, which can automatically adjust the optimization direction and step size according to the local data characteristics, thereby enhancing the algorithm's ability to handle complex nonlinear problems. The update step size is determined by the geodesic distance, which fully considers the geometric characteristics of the data on the manifold and avoids the distortion that may be caused by the traditional Euclidean space update.

[0113] Compared with the traditional Horn-Schunck algorithm, the average endpoint error of optical flow estimation is reduced. Compared with deep learning methods such as FlowNet2.0, it achieves comparable estimation accuracy while consuming only one-fifth of the computing resources. In terms of convergence speed, the number of iterations is reduced compared to the traditional ADMM algorithm. Especially on facial micro-expression datasets, the method of this embodiment significantly improves its ability to capture subtle movements and improves the accuracy of expression change detection. In addition, the method of this embodiment demonstrates greater robustness when handling scenes with large-scale motion and illumination changes, providing a new approach to efficient and reliable optical flow field solutions in the fields of video analysis and computer vision.

[0114] In an optional embodiment, a facial reconstruction model is constructed based on a depth estimation network, and a back-projection mechanism and a geodesic distance partitioning strategy are used to reconstruct the facial three-dimensional structure. The facial three-dimensional features obtained by feature stability scoring and projection invariance matching include:

[0115] The standardized facial image is input into a multi-layer convolutional network for feature extraction, and a facial depth map is output. A back-projection mechanism is used to generate a projected facial image from the facial depth map. The mutual information loss between the projected facial image and the standardized facial image is calculated. The depth consistency loss of the facial depth map is combined to construct a depth estimation loss function. The multi-layer convolutional network is optimized to obtain an optimized facial depth map.

[0116] The depth values ​​and pixel coordinates in the optimized facial depth map are converted into a 3D point cloud. The geodesic distance between point pairs in the 3D point cloud is calculated to establish a point cloud partitioning strategy. The principal curvature direction of the local surface is calculated to construct a reconstruction reference basis. The 3D closed surface of the face is obtained by surface fitting the 3D point cloud based on the direction information of the reconstruction reference basis.

[0117] The local curvature is calculated along the reconstruction reference base direction on the three-dimensional closed surface of the face to construct a gradient response map. The feature stability score is calculated based on the curvature flow equation. Stable feature points are extracted through nonlinear diffusion filtering, and local deformation feature descriptions are extracted from the stable feature points.

[0118] Based on the local deformation feature description, a projection invariant matching criterion is established, the local affine transformation matrix of the local deformation feature description is calculated for feature matching, and matching point pairs are selected in combination with spatial consistency constraints;

[0119] The local surface deformation strength of the matching point pairs is calculated, and feature weight coefficients are assigned. The three-dimensional feature representation of the face is generated by combining the features of the matching point pairs.

[0120] In one specific implementation, a normalized facial image is input into a depth estimation network consisting of an encoder and a decoder. The encoder contains five convolutional modules, each consisting of two 3×3 convolutional layers and a 2×2 max pooling layer. The decoder contains five upsampling modules, each consisting of a 2×2 deconvolutional layer and two 3×3 convolutional layers. The network ultimately outputs a facial depth map of the same size as the input image.

[0121] To optimize the depth estimation network, a joint loss function was designed. The estimated depth map was converted into a projected facial image through a back-projection mechanism: a reference camera intrinsic parameter matrix (with a focal length of 500 pixels and a principal point at the image center) was selected. Three-dimensional point positions were calculated based on the depth values ​​and pixel coordinates. The 3D points were then projected back onto a two-dimensional plane to obtain the projected image. The mutual information loss was calculated between the projected image and the original normalized image. The mutual information coefficient was used to measure the similarity between the two images. The mutual information coefficient ranged from 0 to 1, with higher values ​​indicating greater similarity. A depth consistency loss was also calculated to ensure depth consistency by comparing the current depth map with a prior depth template, which was obtained by averaging 10,000 high-precision 3D facial scan models. The joint loss function was a weighted sum of the two losses, with weights of 0.7 and 0.3, respectively. The network was trained using the Adam optimizer with an initial learning rate of 0.001, which was decayed by 10% every 20 epochs for a total of 100 epochs.

[0122] After obtaining the optimized facial depth map, we generate a 3D point cloud. For each pixel in the depth map, we calculate the 3D coordinates (X, Y, Z) based on its coordinates (u, v) and depth value d, combined with the camera intrinsic parameter matrix. The resulting raw point cloud contains approximately 65,536 points (i.e., 256×256).

[0123] A geodesic distance-based point cloud partitioning strategy was used to segment facial regions. The nose tip was selected as the reference point, and the geodesic distance from each point in the point cloud to the nose tip was calculated. The geodesic distance was calculated using the Dijkstra algorithm, constructing a k-nearest neighbor graph (k=8), with edge weights based on Euclidean distance. The point cloud was then divided into five regions based on geodesic distance: the nose region (distance <25 mm), the eye region (25-50 mm), the mouth region (50-75 mm), the cheek region (75-100 mm), and the edge region (>100 mm).

[0124] For each region, the local surface principal curvature directions are calculated to construct a reconstruction reference basis. For each point in the point cloud, a set of points with a radius of 5 mm within its neighborhood is selected and principal component analysis is used to calculate the local surface normal vector and principal curvature directions. These directions form the local coordinate system that serves as the reconstruction reference basis.

[0125] Based on the partitioning and reconstruction reference basis, surface fitting is performed on the point cloud. Surface reconstruction is performed for each region using the moving least squares method. A Gaussian kernel function is used as the weighting function, and the kernel width is set based on regional characteristics: 3mm for the nose, 5mm for the eyes, 4mm for the mouth, 8mm for the cheeks, and 10mm for the rims. After surface fitting, a Poisson reconstruction algorithm is used to generate a 3D closed surface of the face. The mesh resolution is set to 0.5mm.

[0126] The feature extraction process begins by calculating the local curvature of the facial 3D surface. Mean and Gaussian curvatures are calculated along the reconstructed reference base direction, generating a curvature response map. To identify stable feature points, feature stability is assessed using the curvature flow equation, with a time step of 0.05 and 10 iterations. A feature stability score is calculated for each point, and points with a score above a threshold of 0.75 are selected as candidate feature points.

[0127] Noise is suppressed by nonlinear diffusion filtering with a diffusion coefficient of 0.2 and an iteration number of 20. For each stable feature point, a local deformation feature description is extracted: 32 points are sampled within an 8 mm radius around the feature point, and the relative position and curvature differences of these points relative to the feature point are calculated to form a 128-dimensional feature vector.

[0128] A projection-invariant matching criterion is established based on local deformation feature descriptions. For the feature points of the two facial models, the Euclidean distance between the feature descriptions is calculated, and point pairs with a distance less than a threshold of 0.3 are selected as initial matches. The local affine transformation matrix is ​​calculated for each pair of matching points, and the transformation consistency is verified. Matching pairs with a transformation error of less than 2 mm are retained. Spatial consistency constraints are used to further screen matching point pairs, requiring the geodesic distance ratio between matching point pairs to be between 0.9 and 1.1.

[0129] The local surface deformation strength of the matching point pairs is calculated and a feature weight coefficient is assigned based on the deformation strength, with smaller deformation strengths receiving higher weights. The 3D coordinates, local curvature, feature description vector, and weight coefficients of the matching point pairs are combined to generate a 3D facial feature representation for subsequent recognition or verification tasks.

[0130] In this embodiment, the depth estimation network is jointly optimized by mutual information loss and depth consistency, so that the generated facial depth map more accurately matches the actual facial morphology in terms of geometric structure and image content; a surface fitting strategy guided by point cloud geodetic distance partitioning and principal curvature direction is adopted to construct a closed and continuous three-dimensional facial surface, thereby improving the reconstruction quality and structural integrity; stable feature points are extracted based on curvature gradient response and nonlinear diffusion filtering, effectively suppressing noise interference and improving the stability of feature points under different postures, lighting and other conditions; feature points are matched through local affine transformation and spatial consistency constraints, and weighted three-dimensional feature vectors are generated in combination with the local surface deformation intensity, thereby improving the accuracy and robustness of three-dimensional facial recognition and matching.

[0131] In an optional embodiment, a dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors, a semantic similarity matrix of the feature mapping vectors is calculated, and the semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain comprehensive facial evaluation indicators including:

[0132] Inputting facial dynamic features and facial three-dimensional features into a predetermined dual-channel feature extractor, performing multi-layer feature extraction through a temporal convolutional network and a spatial convolutional network to generate dynamic mapping features and spatial mapping features, performing multi-layer feature extraction through a spatial convolutional network to generate spatial mapping features, combining the dynamic mapping features with the spatial mapping features to generate a feature mapping vector, and performing a normalized dot product operation to generate an initial semantic similarity matrix;

[0133] Dividing an initial semantic similarity matrix into a feature sub-block matrix, obtaining the number of rows and columns of the feature sub-block matrix, determining the minimum value of the number of rows and columns as a block size parameter, constructing a square region with the block size parameter as a side length as a first region, and other regions as a second region, performing a sparse constraint operation on the first region to generate a sparse matrix, performing an association constraint operation on the second region to generate an association constraint matrix, inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization, generating an optimized feature sub-block, and fusing the optimized feature sub-blocks to generate an optimized semantic similarity matrix;

[0134] A weight coefficient matrix is ​​calculated based on the optimized semantic similarity matrix, and is weightedly combined with the feature mapping vector to generate a comprehensive facial evaluation index.

[0135] In one specific embodiment, dynamic facial feature data and three-dimensional facial feature data are collected. Dynamic facial features can include temporal information such as facial expression changes, micro-expressions, and facial muscle movements. The acquisition frame rate is 30 fps, and each sample is collected for 5 seconds, resulting in a 150-frame sequence. Three-dimensional facial features include spatial features such as depth information, surface texture, and three-dimensional structure. The resolution is 1024×1024 pixels, and the depth accuracy is 0.1 mm.

[0136] After acquisition, dynamic facial features and three-dimensional facial features are input into a dual-channel feature extractor. This feature extractor consists of two parallel processing channels: a temporal convolutional network channel and a spatial convolutional network channel. The temporal convolutional network channel is used to process dynamic features and consists of four convolutional layers. Each layer has a convolution kernel size of 3×3, a stride of 1, and a padding of 1. The number of channels is 32, 64, 128, and 256, respectively. Each convolutional layer is followed by a BatchNorm and ReLU activation function and downsampled using a max pooling layer with a pooling kernel size of 2×2 and a stride of 2. The spatial convolutional network channel is used to process three-dimensional features and also consists of four convolutional layers with the same structural parameters as the temporal convolutional network, but with an additional global average pooling layer after the last layer.

[0137] After feature extraction, the temporal convolutional network outputs a 256-dimensional dynamic mapping feature vector, and the spatial convolutional network outputs a 256-dimensional spatial mapping feature vector. These two feature vectors are combined to form a 512-dimensional feature map vector. This feature map vector is normalized by dividing each element by its L2 norm to ensure that the vector length is 1. The dot product of the normalized feature map vector with itself is calculated to generate an initial 512×512-dimensional semantic similarity matrix.

[0138] Perform block sparse optimization on the initial semantic similarity matrix. Divide the matrix into multiple 64×64 feature sub-block matrices, forming a total of 8×8 sub-blocks. For each sub-block matrix, obtain its number of rows and columns, and take the minimum of the two as the block size parameter, which is 64 in this example. Construct a square region with the block size parameter as the side length as the first region. For example, for the sub-block in the upper left corner of the matrix, the first region is the entire area of ​​the sub-block; for the sub-block in the middle of the matrix, the first region is a 64×64 square region with the center of the sub-block as the origin, and the remaining part is the second region.

[0139] Perform a sparse constraint operation on the first region, retaining elements with a similarity value greater than 0.7 and setting elements below this threshold to 0 to generate a sparse matrix. Perform an association constraint operation on the second region, calculating an attenuation factor based on the relative position of the elements. Multiply the original similarity value by the attenuation factor to generate an association constraint matrix. Merge the sparse matrix with the association constraint matrix to form the optimized sub-block matrix.

[0140] The optimized sub-block matrix is ​​further optimized using an iterative shrinkage algorithm. This algorithm sets a maximum number of iterations of 100 and a convergence threshold of 0.001. In each iteration, the difference between the current sub-block matrix and the ideal sparse structure is calculated. Matrix elements are adjusted based on this difference, gradually moving the matrix toward the ideal sparse distribution. Iterations terminate when the difference between two consecutive iterations falls below the convergence threshold or when the maximum number of iterations is reached. In practice, convergence is achieved after approximately 50-60 iterations.

[0141] After optimizing all sub-blocks, the optimized sub-block matrices are recombined to generate a 512×512-dimensional optimized semantic similarity matrix. Based on this matrix, a weight coefficient matrix is ​​calculated by summing each row of the optimized semantic similarity matrix and then normalizing it so that the sum of the weight coefficients for each row is 1. Finally, the weight coefficient matrix is ​​weightedly combined with the feature map vector to generate a comprehensive facial evaluation metric.

[0142] In this embodiment, a dual-channel feature extractor is used to extract facial dynamic features and three-dimensional spatial features respectively, and a joint feature mapping vector is generated to enhance the multi-dimensional representation capability of facial information; a normalized dot product and feature sub-block partitioning strategy is adopted, combined with sparsity and association constraints, and a fine-grained optimization of the semantic similarity matrix is ​​performed to effectively capture complex semantic relationships; an iterative shrinkage algorithm is used to optimize feature sub-blocks, suppress redundant feature interference, highlight key feature areas, and improve the stability of subsequent matching and recognition; combined with optimized semantic similarity and feature weight calculation, the effective fusion of multi-source features is achieved, and a more discriminative and generalizable facial comprehensive evaluation index is output.

[0143] In an optional implementation, inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization to generate optimized feature sub-blocks includes:

[0144] Performing maximum and minimum normalization operations on the sparse matrix and the association constraint matrix respectively to generate a first normalized matrix and a second normalized matrix, setting the first normalized matrix at a diagonal position and setting the second normalized matrix at a non-diagonal position to generate a matrix to be optimized;

[0145] Calculating the median of the absolute values ​​of each element in the matrix to be optimized as an initial threshold, calculating a matrix distribution parameter according to the dimension value of the matrix to be optimized and the cumulative sum of the absolute values ​​of each element in the matrix to be optimized, adding the initial threshold to the matrix distribution parameter and multiplying the result by an exponential decay function to generate a dynamic shrinkage threshold;

[0146] Keeping the values ​​of elements in the matrix to be optimized that are greater than the dynamic shrinkage threshold unchanged and setting the values ​​of elements that are less than or equal to the dynamic shrinkage threshold to zero, generating a sparse matrix, and calculating the relative error between the sparse matrix and the matrix to be optimized;

[0147] Repeat the iteration until the relative error value is less than or equal to the preset error threshold, determine the final sparse matrix, and reconstruct the final sparse matrix according to the original position relationship to generate optimized feature sub-blocks.

[0148] In a specific embodiment, after obtaining the sparse matrix and the associated constraint matrix, it is necessary to perform maximum and minimum normalization operations on the two matrices. For the sparse matrix, find the maximum and minimum values ​​in the matrix. In practical applications, the maximum value is usually about 0.95 and the minimum value is about 0.02. The normalization calculation subtracts the minimum value from each element in the matrix, and then divides it by the difference between the maximum and minimum values, so that all elements are mapped to the range of 0 to 1 to generate a first normalized matrix. Similarly, the same normalization operation is performed on the associated constraint matrix to generate a second normalized matrix. Taking the sub-block matrix of the facial features area as an example, if the element values ​​in the original sparse matrix are distributed between 0.02 and 0.95, all elements are mapped to between 0 and 1 after normalization, maintaining the relative relationship between the elements while unifying the numerical range.

[0149] After normalization, the first normalized matrix is ​​placed diagonally, and the second normalized matrix is ​​placed off-diagonally to form the matrix to be optimized. In practice, if processing an 85×85 feature sub-block in the eye region of a face, the matrix to be optimized is also 85×85, with the 85 diagonal elements coming from the first normalized matrix and the remaining 7140 elements coming from the second normalized matrix. This combination preserves the core autocorrelation information in the sparse matrix and the cross-correlation information in the association constraint matrix.

[0150] The first iteration of the iterative shrinkage algorithm begins. First, the median of the absolute values ​​of all elements in the matrix to be optimized is calculated as the initial threshold. Specifically, the 7225 elements in the matrix are sorted by absolute value and the value of the 3613th element after sorting is taken. In facial feature analysis scenarios, this value is typically approximately 0.4. The system calculates the matrix distribution parameter by multiplying the matrix dimension 85 by the scaling factor 0.01 to obtain 0.85. The mean of the absolute values ​​of all matrix elements (approximately 0.25) is then calculated and multiplied to obtain the matrix distribution parameter, approximately 0.21. The initial threshold of 0.4 is added to the matrix distribution parameter 0.21 to obtain 0.61, which is then multiplied by the exponential decay factor. The exponential decay factor for the first iteration is calculated as e raised to the power of -0.1 (-0.1 is equal to -1 divided by 10, where 1 is the current iteration number and 10 is the decay rate parameter), which is approximately 0.905. Multiplying 0.61 by 0.905 yields a dynamic shrinkage threshold of approximately 0.552 for the first iteration.

[0151] The matrix to be optimized was processed using a dynamic shrinkage threshold of 0.552. Each element in the matrix was traversed, and elements with an absolute value less than or equal to 0.552 were set to 0. Elements with an absolute value greater than the threshold remained unchanged. This process yielded the sparsified matrix for the first iteration. Approximately 18% of the elements in this matrix retained non-zero values, primarily concentrated in locations representing key aesthetic areas such as the eye contour, eyebag features, and eyebrow shape. The relative error between the sparsified matrix and the matrix to be optimized was calculated by first taking the sum of the squares of the differences between the elements in the two matrices at corresponding positions, then taking the sum of the squares of the elements in the matrix to be optimized, and taking the square root of the ratio of the two. The relative error for the first iteration was approximately 0.35.

[0152] The relative error value is compared with the preset error threshold of 0.05 and is found to be greater than the threshold, so the second iteration is entered. The second iteration uses the same matrix to be optimized as the first, and the median threshold is recalculated, finding it is still 0.4. The matrix distribution parameter is calculated and is still 0.21. The initial threshold is added to the matrix distribution parameter to obtain 0.61, but the exponential decay factor now becomes e to the power of -0.2 (-0.2 is equal to -2 divided by 10, where 2 is the current iteration number), which is approximately equal to 0.819. 0.61 multiplied by 0.819 gives the dynamic shrinkage threshold for the second iteration of approximately 0.50.

[0153] Using a new dynamic shrinkage threshold of 0.50, the optimized matrix was processed, resulting in a sparsified matrix for the second iteration. Approximately 22% of the elements in this matrix retained nonzero values. The calculated relative error decreased to approximately 0.31. Since the error was still greater than the preset threshold of 0.05, the third iteration was performed.

[0154] In subsequent iterations, the dynamic shrinkage threshold gradually decreases: to approximately 0.46 in the third iteration, 0.41 in the fifth iteration, 0.33 in the tenth iteration, and 0.27 in the fifteenth iteration. As the threshold decreases, the number of nonzero elements retained in the sparsified matrix increases, reaching approximately 26%, 33%, 45%, and 52%, respectively. The relative error also decreases, reaching approximately 0.28, 0.22, 0.13, and 0.08, respectively.

[0155] When the iteration reaches the 18th time, the calculated dynamic shrinkage threshold is approximately 0.24, the sparsified matrix retains approximately 56% of the non-zero elements, and the calculated relative error value is approximately 0.048, which is lower than the preset error threshold of 0.05. At this point, the iteration process stops and the sparsified matrix obtained from the 18th iteration is determined as the final sparsified matrix.

[0156] The final sparsified matrix is ​​reconstructed according to the original positional relationships to generate optimized feature sub-blocks. In the example of the facial eye region, the 56% of retained non-zero elements are mainly distributed in locations representing key medical aesthetic assessment points such as periocular wrinkles, eyebag shape, ptosis, eyebrow shape, and glabellar lines. These optimized features can provide more accurate quantitative indicators of facial age characteristics in subsequent facial diagnosis and analysis. For example, in the case of a 45-year-old female, it can distinguish between fine lines with a depth of 0.2 mm and lines with a depth of 0.5 mm, providing a data foundation for the design of refined medical aesthetic solutions.

[0157] The method of this embodiment is primarily based on compressed sensing and sparse representation theory. There are two main methods for processing feature matrices in the prior art: one is hard threshold shrinkage based on a fixed threshold, which directly clears elements below the threshold; the other is a shrinkage algorithm based on a soft threshold, which subtracts a fixed value from all elements before performing threshold processing. These methods have significant shortcomings when processing facial features: fixed thresholds are difficult to adapt to feature differences across different facial regions and are prone to losing subtle texture information; while simple soft thresholding can retain more information, it cannot effectively remove noise interference, resulting in unstable aesthetic evaluation results.

[0158] The method in this embodiment introduces an adaptive dynamic shrinkage threshold mechanism that automatically adjusts the threshold based on matrix characteristics and an iterative process. It innovatively incorporates matrix distribution parameters into the threshold calculation, enabling the algorithm to perceive differences in feature distribution across different facial regions. An exponential decay function is used to control the rate of threshold change, achieving a smooth and gradual shrinkage process. This improvement aims to address regional differences in facial feature extraction and design a more precise feature optimization method, specifically targeting the high sensitivity to subtle features in medical aesthetics applications.

[0159] Figure 3 The following figure illustrates the performance simulation results of the iterative shrinkage algorithm. Three curves illustrate the evolution of key algorithm parameters over the number of iterations: the dynamic shrinkage threshold (black curve), the relative error (red curve), and the proportion of non-zero elements (blue curve). The curves show a smooth downward trend in the dynamic shrinkage threshold, decreasing from an initial value of 0.552 to 0.24, demonstrating the algorithm's adaptive nature. The relative error curve demonstrates the convergence of the optimization process, decreasing from an initial value of 0.35 to 0.048, meeting the preset error threshold requirement of 0.05. The non-zero element proportion curve reflects the dynamic evolution of feature retention, increasing from an initial value of 18% to 56%, demonstrating that the algorithm retains sufficient feature information while maintaining a high degree of sparsity. Overall, the three curves show consistent trends, confirming that the algorithm achieved the desired optimization goal after 18 iterations, ensuring both accurate feature extraction and appropriate sparsity, providing a reliable data foundation for refined facial feature analysis.

[0160] The AI ​​facial diagnosis feature and aesthetic analysis system for medical aesthetics in this embodiment of the present invention includes:

[0161] The first unit is configured to collect user facial image data and pre-process the facial image data to obtain standardized facial data;

[0162] The second unit is used to build a temporal feature extraction model based on a deep convolutional neural network. It uses multi-scale convolution operations to perform feature response and displacement estimation on standardized facial data, and combines the optical flow estimation algorithm to calculate the facial region deformation field to obtain facial dynamic features.

[0163] The third unit is used to build a facial reconstruction model based on the depth estimation network, reconstruct the facial 3D structure using the back-projection mechanism and geodesic distance partitioning strategy, and obtain the facial 3D features through feature stability scoring and projection invariance matching;

[0164] The fourth unit is used to use a dual-channel feature extractor to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors, calculate the semantic similarity matrix of the feature mapping vectors, perform block sparse optimization and iterative shrinkage on the semantic similarity matrix, and obtain a comprehensive facial evaluation index;

[0165] The fifth unit is used to perform graded assessments of facial areas based on comprehensive facial assessment indicators and a preset knowledge base of medical aesthetic experts, generate facial analysis reports, and output medical aesthetic recommendation plans.

[0166] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0167] processor;

[0168] a memory for storing processor-executable instructions;

[0169] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0170] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0171] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. AI facial diagnosis features and aesthetic analysis methods for medical aesthetics, characterized by: include: Collecting user facial image data, and preprocessing the facial image data to obtain standardized facial data; A temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated by combining the optical flow estimation algorithm to obtain facial dynamic features. A facial reconstruction model is constructed based on a depth estimation network. The 3D facial structure is reconstructed using a back-projection mechanism and a geodesic distance partitioning strategy. The 3D facial features are obtained through feature stability scoring and projection invariance matching. A dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors. The semantic similarity matrix of the feature mapping vectors is calculated. The semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain a comprehensive facial evaluation index. Based on comprehensive facial assessment indicators and combined with the preset medical aesthetics expert knowledge base, facial areas are graded and assessed, a facial analysis report is generated, and medical aesthetics recommendations are output.

2. The method according to claim 1, characterized in that A temporal feature extraction model is constructed based on a deep convolutional neural network. The feature response and displacement of standardized facial data are estimated through multi-scale convolution operations. The facial region deformation field is calculated by combining the optical flow estimation algorithm. The facial dynamic features obtained include: By setting multiple convolution kernel arrays of continuously increasing scales in a deep convolutional neural network, convolution operations are performed on the standardized facial data to generate a multi-level feature response map; performing feature cross-correlation operations on adjacent frames in the multi-level feature response map, obtaining an inter-frame displacement correlation matrix by calculating regional similarity of feature responses, and generating an initial motion estimation map of the facial region according to the inter-frame displacement correlation matrix; Performing adaptive threshold segmentation on the initial motion estimation image, obtaining an optimal segmentation threshold by iterative calculation of maximizing inter-class variance, and generating a facial region motion feature map using the optimal segmentation threshold; The facial region motion feature map is processed using an optical flow estimation algorithm, an objective function is constructed based on a self-attention mechanism and deformation constraints, and a facial region deformation field is obtained by solving the problem using an iterative splitting algorithm; The facial region deformation field is divided into regions, the deformation strength value in each preset facial key region is calculated, and the deformation strength values ​​of each region are combined to form a facial dynamic feature vector.

3. The method according to claim 2, characterized in that The facial region motion feature map is processed using an optical flow estimation algorithm, and an objective function is constructed based on a self-attention mechanism and deformation constraints. The facial region deformation field is obtained by an iterative splitting algorithm, including: Constructing a tensor representation of two adjacent frames in the facial region motion feature map, embedding a self-attention mechanism in the tensor representation, generating attention weights by calculating the correlation coefficient between the images, enhancing the global motion information based on the attention weights, and constructing an initialization representation of the optical flow field; Based on predetermined facial anatomical features, a topological structure relationship graph of the facial region is constructed, deformation consistency constraints of the facial region are established, and node features in the topological structure relationship graph are aggregated using a graph convolutional network to generate motion constraints based on the facial structure; Inputting the initialization representation of the optical flow field and the motion constraint condition into a variational optimization framework, constructing an adaptive loss function including a data term based on an attention mechanism and a regularization term based on a topological structure, calculating the gradient of the loss function for the data term and the regularization term respectively, constructing a weight update matrix based on the gradient of the loss function, determining the weight coefficient of the data term and the weight coefficient of the regularization term and multiplying them together to obtain a target optimization function that is dynamically adjusted according to the motion mode; An iterative splitting algorithm is used to solve the target optimization function, the optimization variable is decomposed into multiple sub-variables, and a splitting sequence is constructed for each sub-variable to perform alternating updates; Based on the optimization result of the iterative splitting algorithm, an optical flow field of the facial region is obtained, and the amplitude of the optical flow field is calculated to obtain a deformation field of the facial region.

4. The method according to claim 3, characterized in that The iterative splitting algorithm includes: Performing variable decomposition on the target optimization function, decomposing the optimization variables into original variables and constraint variables, and constructing a variable update path graph; Constructing a branch optimization sequence in the variable update path graph, generating a local feature manifold by mapping the original variables to neighboring points, and performing gradient projection on the local feature manifold to obtain the optimization direction; constraining variables to construct a sparse reconstruction sequence by measuring feature distance; Determining a feature manifold projection operator based on the curvature distribution of the local feature manifold, determining a sparse reconstruction operator based on the singular value decomposition of the sparse feature distribution, and combining the feature manifold projection operator and the sparse reconstruction operator to form an adaptive neighbor operator; Alternately updating the original variables and the constraint variables based on the adaptive proximity operator, with the update step size determined by the geodesic distance on the characteristic manifold, and stopping the iteration when the difference between the original variables and the constraint variables is less than a preset threshold; The original variables after iterative convergence are used as the final optical flow field solution results.

5. The method according to claim 1, wherein A facial reconstruction model is constructed based on a depth estimation network. The back-projection mechanism and geodesic distance partitioning strategy are used to reconstruct the facial 3D structure. The facial 3D features obtained through feature stability scoring and projection invariance matching include: The standardized facial image is input into a multi-layer convolutional network for feature extraction, and a facial depth map is output. A back-projection mechanism is used to generate a projected facial image from the facial depth map. The mutual information loss between the projected facial image and the standardized facial image is calculated. The depth consistency loss of the facial depth map is combined to construct a depth estimation loss function. The multi-layer convolutional network is optimized to obtain an optimized facial depth map. The depth values ​​and pixel coordinates in the optimized facial depth map are converted into a 3D point cloud. The geodesic distance between point pairs in the 3D point cloud is calculated to establish a point cloud partitioning strategy. The principal curvature direction of the local surface is calculated to construct a reconstruction reference basis. The 3D closed surface of the face is obtained by surface fitting the 3D point cloud based on the direction information of the reconstruction reference basis. The local curvature is calculated along the reconstruction reference base direction on the three-dimensional closed surface of the face to construct a gradient response map. The feature stability score is calculated based on the curvature flow equation. Stable feature points are extracted through nonlinear diffusion filtering, and local deformation feature descriptions are extracted from the stable feature points. Based on the local deformation feature description, a projection invariant matching criterion is established, the local affine transformation matrix of the local deformation feature description is calculated for feature matching, and matching point pairs are selected in combination with spatial consistency constraints; The local surface deformation strength of the matching point pairs is calculated, and feature weight coefficients are assigned. The three-dimensional feature representation of the face is generated by combining the features of the matching point pairs.

6. The method according to claim 1, characterized in that A dual-channel feature extractor is used to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors. The semantic similarity matrix of the feature mapping vectors is calculated. The semantic similarity matrix is ​​subjected to block sparse optimization and iterative shrinkage to obtain the following comprehensive facial evaluation indicators: Inputting facial dynamic features and facial three-dimensional features into a predetermined dual-channel feature extractor, performing multi-layer feature extraction through a temporal convolutional network and a spatial convolutional network to generate dynamic mapping features and spatial mapping features, performing multi-layer feature extraction through a spatial convolutional network to generate spatial mapping features, combining the dynamic mapping features with the spatial mapping features to generate a feature mapping vector, and performing a normalized dot product operation to generate an initial semantic similarity matrix; Dividing an initial semantic similarity matrix into a feature sub-block matrix, obtaining the number of rows and columns of the feature sub-block matrix, determining the minimum value of the number of rows and columns as a block size parameter, constructing a square region with the block size parameter as a side length as a first region, and other regions as a second region, performing a sparse constraint operation on the first region to generate a sparse matrix, performing an association constraint operation on the second region to generate an association constraint matrix, inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization, generating an optimized feature sub-block, and fusing the optimized feature sub-blocks to generate an optimized semantic similarity matrix; A weight coefficient matrix is ​​calculated based on the optimized semantic similarity matrix, and is weightedly combined with the feature mapping vector to generate a comprehensive facial evaluation index.

7. The method according to claim 6, characterized in that Inputting the sparse matrix and the association constraint matrix into an iterative shrinkage algorithm for optimization to generate optimized feature sub-blocks includes: Performing maximum and minimum normalization operations on the sparse matrix and the association constraint matrix respectively to generate a first normalized matrix and a second normalized matrix, setting the first normalized matrix at a diagonal position and setting the second normalized matrix at a non-diagonal position to generate a matrix to be optimized; Calculating the median of the absolute values ​​of each element in the matrix to be optimized as an initial threshold, calculating a matrix distribution parameter according to the dimension value of the matrix to be optimized and the cumulative sum of the absolute values ​​of each element in the matrix to be optimized, adding the initial threshold to the matrix distribution parameter and multiplying the result by an exponential decay function to generate a dynamic shrinkage threshold; Keeping the values ​​of elements in the matrix to be optimized that are greater than the dynamic shrinkage threshold unchanged and setting the values ​​of elements that are less than or equal to the dynamic shrinkage threshold to zero, generating a sparse matrix, and calculating the relative error between the sparse matrix and the matrix to be optimized; Repeat the iteration until the relative error value is less than or equal to the preset error threshold, determine the final sparse matrix, and reconstruct the final sparse matrix according to the original position relationship to generate optimized feature sub-blocks.

8. An AI facial diagnosis feature and aesthetic analysis system for medical aesthetics, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to collect user facial image data and pre-process the facial image data to obtain standardized facial data; The second unit is used to build a temporal feature extraction model based on a deep convolutional neural network. It uses multi-scale convolution operations to perform feature response and displacement estimation on standardized facial data, and combines the optical flow estimation algorithm to calculate the facial region deformation field to obtain facial dynamic features. The third unit is used to build a facial reconstruction model based on the depth estimation network, reconstruct the facial 3D structure using the back-projection mechanism and geodesic distance partitioning strategy, and obtain the facial 3D features through feature stability scoring and projection invariance matching; The fourth unit is used to use a dual-channel feature extractor to perform feature mapping on facial dynamic features and facial three-dimensional features to generate feature mapping vectors, calculate the semantic similarity matrix of the feature mapping vectors, perform block sparse optimization and iterative shrinkage on the semantic similarity matrix, and obtain a comprehensive facial evaluation index; The fifth unit is used to perform graded assessments of facial areas based on comprehensive facial assessment indicators and a preset knowledge base of medical aesthetic experts, generate facial analysis reports, and output medical aesthetic recommendation plans.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images