Three-dimensional face aesthetic evaluation method based on graph neural network
By constructing a 3D face aesthetic evaluation method based on graph neural networks and introducing a region-learnable weight mechanism, the problems of scoring stability and interpretability in 3D face aesthetic evaluation are solved, and stable evaluation and consistent output under different conditions are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies struggle to balance structural realism, scoring stability, and interpretability in 3D facial aesthetic evaluation. Two-dimensional methods are susceptible to lighting and pose variations, traditional 3D feature methods rely on manual features and lack robustness, and 3D deep learning methods do not adequately represent the structured expression of aesthetic knowledge.
A graph neural network-based approach is adopted to construct a graph structure through 3D facial key point recognition. By combining 3D coordinates, region labels and angular features, a region-learnable weight mechanism is introduced to perform message passing and aggregation, and output an aesthetic score.
It achieves stable output of aesthetic scores under different lighting and postures, improves the consistency and interpretability of the assessment, adapts to different groups of people and collection conditions, and has a clear process that is easy to implement in engineering.
Smart Images
Figure CN121921476A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical cosmetic surgery, specifically to a three-dimensional facial aesthetic evaluation method based on graph neural networks. Background Technology
[0002] With the development of 3D imaging and 3D reconstruction technologies, 3D facial data (such as triangular meshes and point clouds obtained from structured light / multi-view scanning) is widely used in medical plastic surgery evaluation, virtual avatar generation, digital humans, and image analysis. Compared with 2D images, 3D faces can more directly reflect the true geometric shape and spatial structure of the face. Therefore, aesthetic evaluation of 3D faces (such as quantitative evaluation of facial proportions, symmetry, and contour lines) has gradually become an important research direction in the fields of computer vision and intelligent evaluation. These algorithms are typically expected to output stable and reproducible aesthetic scores under different lighting, poses, expressions, and acquisition device variations, and possess a certain degree of interpretability, in order to serve professional aesthetic analysis and decision support.
[0003] The existing technical solutions most similar to this invention mainly include: First, aesthetic evaluation methods based on two-dimensional facial images, which typically use convolutional neural networks to extract features from pixel textures and perform regression scoring. However, these methods are easily affected by lighting, pose, makeup, etc., and are difficult to accurately reflect three-dimensional structural aesthetic factors (such as symmetry and proportional relationships). Second, evaluation methods based on three-dimensional geometric features, which manually calculate features such as distance, angle, and curvature from three-dimensional meshes / key points and perform regression or classification. However, these methods rely on feature engineering, have limited expressive power, and are sensitive to the accuracy of key point positioning and alignment. Third, deep learning methods (point clouds / mesh / graph networks, etc.) for three-dimensional irregular data can directly learn three-dimensional features, but some solutions require additional resampling or standardization processing, and lack structured expression of aesthetic knowledge (such as the three courts and five eyes, symmetry, facial proportions, arcs, etc.), making it difficult to simultaneously ensure scoring stability, generalization, and interpretability.
[0004] To address the challenges of balancing structural realism, scoring stability, and interpretability in 3D facial aesthetic evaluation, existing methods often rely on 2D image textures or a small number of artificial geometric features for modeling. While these methods can output aesthetic scores, they are susceptible to interference from factors such as lighting and pose, and fail to accurately depict key aesthetic elements, resulting in inconsistent evaluation results. Summary of the Invention
[0005] This invention aims to overcome the problems in the prior art, such as the inability of two-dimensional methods to reflect the true three-dimensional structure, the reliance of traditional three-dimensional feature methods on artificial features and insufficient robustness, and the insufficient structural expression of aesthetic knowledge and the complexity of the process in existing three-dimensional deep learning methods, which makes it difficult to balance the stability and interpretability of the scoring. It proposes a three-dimensional face aesthetic evaluation method based on graph neural networks.
[0006] The present invention provides a three-dimensional facial aesthetic evaluation method based on graph neural networks, comprising the following steps: S1. Read the original patch mesh data obtained from the 3D face scan, and perform 3D key point recognition on the original patch based on the improved face key point detection model to obtain a set of key points containing 3D coordinate information; S2. Construct a three-dimensional face image structure using the key points as nodes, and generate node features containing three-dimensional coordinates, region labels, and angle features; S3. Construct a graph neural network aesthetic evaluation model, perform message passing and aggregation on node features in the graph structure, and form a graph-level aesthetic representation through a region-learnable weight mechanism, and finally output a three-dimensional face aesthetic score; the model includes a node feature encoding module, a message passing and node update module, a region-learnable weight module, a graph-level aggregation module, and a regression scoring module; The node feature encoding module is used to encode node features. The initial implicit vector is obtained by performing a linear mapping or nonlinear transformation. The message passing and node update module uses a multi-layer graph neural network to iteratively update the latent vectors of nodes; the region learnable weight module assigns learnable weight parameters to nodes or edges in different regions based on the proportional features of region labels and / or angle codes; the graph-level aggregation module performs global aggregation of the high-level latent vectors of all nodes to obtain a graph-level representation g; the regression scoring module inputs the graph-level representation g into the regression head and outputs a continuous score s. S4. During the training phase, the model parameters are optimized using labeled aesthetic rating data so that the model can stably output rating results consistent with aesthetic standards. S5. Input the graph structure and node features into the graph neural network and output the 3D face aesthetic score.
[0007] In step S1, the process of first obtaining a 3D face image and then using the improved facemesh recognition to obtain facial key point information includes the following steps: S11: Obtain a 3D patch of a human face through scanning or other 3D generation methods; S12: Perform principal vector analysis on the 3D face image and adjust the 3D face image to a frontal state; S13: Project the 3D face image into a 2D model. Use facemesh to obtain the coordinates of the center point of the nose, the center point of the eyes, and the corners of the mouth in 2D coordinates. Use these coordinate points to rotate and align the face image so that the center point of the nose, the center point of the eyes, and the center point of the corners of the mouth are on the same vertical line, with the center point of the eyes at the top and the center point of the corners of the mouth at the bottom. S14: Repeat S13; S15: Perform parallel projection 2D rendering on the aligned face patch to eliminate the effect of perspective; re-identify the face key point coordinates of the 2D rendered image using facemesh to obtain the 2D coordinates of the key points. S16: Obtain the three-dimensional coordinates of the key points by projecting their two-dimensional coordinates onto a face patch. Where N is the total number of key points, preferably N=468, and each key point .
[0008] In step S2, a three-dimensional face image structure is constructed based on the key points, and proportional features are generated under the guidance of preset facial aesthetic knowledge, thereby forming a node feature matrix that can be used for graph neural network evaluation. This includes the following steps: S21: For each key point in the key point set P... As a graph node This yields the node set V; S22: To characterize the geometric adjacency and structural topological relationships between key points, construct the edge set E, thereby obtaining the 3D face image G=(V,E); S23: For each node Construct its node feature vector The node features include at least the following parts: S24: Convert the feature vectors of each node Combine the nodes sequentially to obtain the node feature matrix. It outputs the 3D face image structure G=(V,E) and the corresponding node feature matrix X.
[0009] Furthermore, the node features described in S23 include the following parts: (1) Three-dimensional coordinate features: the three-dimensional coordinates of key points The normalized coordinates serve as the basic features of the nodes; (2) Region label features: Assign region labels (forehead, eye area, nose, mouth area, cheekbone, jaw, etc.) to each key point and add node features in one-hot format; (3) Proportional features: Proportional features are constructed under the guidance of preset facial aesthetic knowledge, which includes the three courts and five eyes, symmetry, OJL curve and facial proportion, and the key point combination and the calculation rules of distance / angle / proportion of the proportional features are used to determine the proportional features.
[0010] The proportional features mentioned in step (3) include: Distance / proportion features: Calculate distances based on specific keypoint groups and form normalized proportions based on facial scale. Symmetry deviation features: Features are constructed based on the deviation and difference components of the left and right symmetrical key points to the facial midline / reference point; Angular encoding features: based on keypoint triples Calculated angles are used to characterize local angles, contour curvature, and geometric relationships; among which Corner key points and To and Angular encoding of connected or related key points on the same structure Defined as:
[0011] Furthermore, the angle encoding includes the following types: a) Local angle encoding: Construct triplets along the same local contour or adjacent key points in the same region to calculate angles, which are used to characterize local curvature and shape changes (e.g., zygomatic arch, mandibular border, inner and outer contours of the lip, eyelid contour, bridge of the nose / alar contour, etc.). b) Left and right symmetry angle encoding: Select the left and right paired key points and the central reference point to construct a triplet to calculate the angle or angle difference, which is used to quantify symmetry (the central reference point can be the tip of the nose, the chin point, or the key point of the midline of the face). c) Outer contour angle encoding: Select key point sequences along the outer boundary contour of the face to construct triplets and calculate turning angles, which are used to characterize the overall face shape and the coordination of the outer contour (e.g., jawline, cheek contour, forehead contour, etc.). The angle-coded features are used in subsequent evaluations in the following manner: Incorporated as part of node features : Will be with nodes Several related angle codes are concatenated to form the node proportional feature components. As input for edge features or edge weights: (This will be used in conjunction with the edge...) , The relevant angle encoding is mapped to edge features. Or, the input of attention weights can be used to adjust the intensity of neighbor information propagation; As input to the region-learnable weight module: the left-right symmetry angle encoding, outer contour angle encoding, etc. are associated with the region label to drive the adaptive learning of the region weight parameters, so that the weight learning corresponds to aesthetic factors.
[0012] In the message passing and node feature update module, the initial latent vector of the node is denoted as... For the l-th layer graph neural network, for each node From its set of adjacent nodes Aggregating neighbor information and updating node characteristics, the update process can be represented as follows:
[0013] in, For message functions, Aggregation functions (including but not limited to summation, averaging, maximization, or attention aggregation). For update functions; The edge features or edge weights are input, and the edge features are obtained by mapping the angle encoding features, distance / proportion features or their combination in step 2; through multi-layer message passing, the local structural information is gradually fused into a higher-level global structural representation.
[0014] The feature is that, in the region-learnable weight module, to enable the model to adaptively learn the contribution of different facial regions to the aesthetic score, a region-learnable weight mechanism is introduced to perform weighted modeling of node / edge information in different regions. This mechanism includes one or more of the following implementation methods: 1) Region Node Weights: Set learnable weight parameters for each facial region. Features of nodes belonging to region r Weighting; 2) Attention weights: Attention coefficients are calculated based on node features, region labels, and / or angular encoding features. Used to adjust the edge ( , The intensity of information dissemination; 3) Gated fusion weights: Gated weights are applied to the subgraph representations of different regions to obtain the final graph-level representation.
[0015] Through the above mechanism, the model can automatically learn the contribution of regions during training, so that the scoring results can correspond to aesthetic factors such as symmetry, contour curvature, and facial proportions.
[0016] In the graph-level aggregation module, after updating the L-level nodes, the high-level latent vectors of the nodes are obtained. The graph-level aesthetic representation g is obtained through the graph-level aggregation module, and a continuous score s is output. The process can be represented as follows:
[0017] in, This refers to global aggregation functions, including but not limited to global average pooling, global weighted pooling, or attention pooling. The node weights can be determined by a region-learnable weighting mechanism. In the regression scoring module, based on the graph-level aesthetic representation g, the regression process is expressed as follows:
[0018] in is the regression function (multilayer perceptron regression head), and s is the 3D facial aesthetic score.
[0019] The key point of this invention is: 1. Graph representation and overall geometric relationship modeling of 3D face structure: Unlike existing technologies that mostly rely on 2D textures or a small number of artificial geometric features, this invention uses 3D scanned patches as input and constructs a 3D face graph structure through 3D key point recognition. This allows the model to simultaneously depict local contours and global proportional relationships on the graph topology, thereby preserving and utilizing the overall structural information of the 3D face to complete aesthetic evaluation.
[0020] 2. Construction of proportional features guided by aesthetic knowledge: This invention uses aesthetic principles such as the three courts and five eyes, symmetry, OJL curve and facial proportions to guide the construction of proportional features. In particular, it explicitly represents the contour curvature, symmetry and facial structure through the angle encoding of key point triples, so that the model learning establishes a correspondence with specific aesthetic factors, improves the consistency of scoring and enhances interpretability.
[0021] 3. Graph Neural Network Evaluation Mechanism with Region-Learnable Weights: This invention introduces a region-learnable weight mechanism in the process of message passing and graph-level aggregation in graph neural networks. This allows the contribution of different facial regions to the score to be adaptively learned during the training process, reducing the subjectivity of manually setting weights and improving the adaptability and robustness to data distribution under different populations and collection conditions.
[0022] The advantages of this invention are: based on original 3D facial scan data, a facial image structure is constructed through 3D key point recognition, region labels and proportional features are integrated, and aesthetic knowledge is introduced. Structured feature learning and learnable region weight modeling are achieved in a graph neural network, thereby improving the stability and consistency of the evaluation while outputting continuous aesthetic scores. This invention can be used in applications such as medical aesthetics evaluation, facial analysis, and digital humans. The advantages of this invention include: 1. Directly utilizing 3D structural information leads to more stable evaluation. This invention uses raw face data obtained from 3D face scanning as input. Through 3D key point recognition and graph structure modeling, it avoids the strong dependence of 2D image methods on factors such as lighting, makeup, and shooting angle, and can more stably reflect the true geometric shape.
[0023] 2. Using graph structure modeling to fully express local-global relationships. By constructing facial key points as a graph structure and using graph neural networks for feature propagation and aggregation, this invention can simultaneously characterize local organ morphology and global proportional relationships, overcoming the limited expressive power of traditional artificial geometric features.
[0024] 3. Integrating aesthetic knowledge to improve evaluation consistency and interpretability. This invention incorporates aesthetic knowledge such as the three courts and five eyes principle, symmetry, OJL curve, and facial proportions into the feature construction and modeling process. This establishes a correspondence between the model learning process and aesthetic factors, thereby improving scoring consistency and enhancing the interpretability of the structural level.
[0025] 4. The regional weights are learnable and adaptable to different populations and data distributions. The contribution of different facial regions in this invention is obtained through adaptive learning during model training, reducing subjective bias caused by manually setting weights and better adapting to data differences of different genders, ages, face shapes, and under different collection conditions.
[0026] 5. The process loop is clear, facilitating engineering implementation and expansion. This invention forms a complete closed loop from 3D patch data input, 3D key point acquisition, graph construction, graph neural network inference to continuous scoring output, which facilitates deployment and application; it can also be expanded into a rating evaluation, sorting and filtering, or as an evaluation subsystem for subsequent aesthetic optimization modules. Attached Figure Description
[0027] Figure 1 This is a flowchart of the method of the present invention.
[0028] Figure 2 This is the face geometric adjacency topology graph of the present invention. Detailed Implementation
[0029] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0030] The process of the method of the present invention is as follows: Figure 1 As shown, the process includes three steps: 1. Reading the original patch mesh data obtained from 3D face scanning, and performing 3D key point recognition on the original patches based on the improved face key point detection model to obtain a set of key points containing 3D coordinate information; 2. Constructing a 3D face graph structure with the key points as nodes, and generating node features containing 3D coordinates, region labels, and angle features; 3. Inputting the graph structure and node features into a graph neural network to output a 3D face aesthetic score result.
[0031] The present invention proposes a three-dimensional face aesthetic evaluation method based on graph neural networks, the specific steps of which are as follows: Step 1: Acquisition of 3D Key Points To obtain information about the three key points, we first need to obtain a 3D face patch, and then use the improved Facemesh recognition to obtain the facial key point information. Facemesh is an existing facial key point detection model.
[0032] The steps for obtaining 3D facial key points are as follows: Step 1-1: Obtain a 3D face patch using scanning or other 3D generation methods; Steps 1-2: Perform principal vector analysis on the 3D face image and adjust the 3D face image to a frontal state.
[0033] Steps 1-3: Project the 3D face image into a 2D model. Use FaceMesh to obtain the coordinates of the center point of the nose, the center point of the eyes, and the corners of the mouth in 2D coordinates. Use these coordinate points to rotate and align the face image so that the center point of the nose, the center point of the eyes, and the center point of the corners of the mouth are on the same vertical line, with the center point of the eyes at the top and the center point of the corners of the mouth at the bottom.
[0034] Steps 1-4: Repeat steps 1-3 three times to ensure good alignment.
[0035] Steps 1-5: Perform parallel projection 2D rendering on the aligned face patch to eliminate the effects of perspective. Then, re-identify the face keypoint coordinates of this 2D rendered image using Facemesh to obtain the 2D coordinates of the keypoints.
[0036] Steps 1-6: Project the two-dimensional coordinates of the key points onto the face image to obtain the three-dimensional coordinates of the key points. Where N=468, each key point .
[0037] Step 2: Construction of 3D Face Graph Nodes and Their Features
[0038] After obtaining the set of three-dimensional key points P output in step 1, this step constructs a three-dimensional face image structure based on the key points and generates proportional features under the guidance of preset face aesthetic knowledge, thereby forming a node feature matrix that can be used for graph neural network evaluation.
[0039] The steps for constructing 3D face image nodes and their features are as follows: Step 2-1: Select each key point in the key point set P. As a graph node This yields the node set V; Step 2-2: To characterize the geometric adjacency and structural topological relationships between key points, an edge set E is constructed, thereby obtaining the 3D face image G=(V,E). The edge set E is constructed in a manner described by... Figure 2 As shown; Steps 2-3: For each node Construct its node feature vector The node features include at least the following parts: (1) Three-dimensional coordinate features: the three-dimensional coordinates of key points The normalized coordinates serve as the basic features of the nodes; (2) Region label features: Assign region labels (forehead, eye area, nose, mouth area, cheekbone, jaw, etc.) to each key point and add node features in one-hot format; (3) Proportional Features: Proportional features are constructed under the guidance of pre-defined facial aesthetic knowledge, including the three courts and five eyes principle, symmetry, OJL curve, and facial proportions, used to determine the key point combinations of proportional features and the calculation rules for their distances / angles / proportions. The proportional features include at least the following: Distance / proportion features: Calculate distances based on specific keypoint groups and form normalized proportions based on facial scale. Symmetry deviation features: Features are constructed based on the deviation and difference components of the left and right symmetrical key points to the facial midline / reference point; Angular encoding features: based on keypoint triples Calculated angles are used to characterize local angles, contour curvature, and geometric relationships. Corner key points and To and Angular encoding of connected or related key points on the same structure Defined as:
[0040] To cover different aesthetic factors, the angle encoding includes the following types: a) Local angle encoding: Construct triplets along the same local contour or adjacent key points in the same region to calculate angles, which are used to characterize local curvature and shape changes (e.g., zygomatic arch, mandibular border, inner and outer contours of the lip, eyelid contour, bridge of the nose / alar contour, etc.). b) Left and right symmetry angle encoding: Select the left and right paired key points and the central reference point to construct a triplet to calculate the angle or angle difference, which is used to quantify symmetry (the central reference point can be the tip of the nose, the chin point, or the key point of the midline of the face). c) Outer contour angle encoding: Select key point sequences along the outer boundary contour of the face to construct triplets and calculate turning angles, which are used to characterize the overall face shape and the coordination of the outer contour (e.g., jawline, cheek contour, forehead contour, etc.). The angle-coded features are used in subsequent evaluations in the following manner: Incorporated as part of node features : Will be with nodes Several related angle codes are concatenated to form the node proportional feature components. As input for edge features or edge weights: (This will be used in conjunction with the edge...) , The relevant angle encoding is mapped to edge features. Or, the input of attention weights can be used to adjust the intensity of neighbor information propagation; As input to the region-learnable weight module: the left and right symmetry angle encoding, outer contour angle encoding, etc. are associated with the region label to drive the adaptive learning of the region weight parameters, so that the weight learning corresponds to aesthetic factors. Steps 2-4: Convert the feature vectors of each node Combine the nodes sequentially to obtain the node feature matrix. It outputs the 3D face image structure G=(V,E) and the corresponding node feature matrix X.
[0041] Step 3: Graph Neural Network Construction, Training, and Scoring Output
[0042] Based on the output of the 3D face graph structure G=(V,E) and node feature matrix in step 2, this step constructs a graph neural network aesthetic evaluation model, performs message passing and aggregation on the node features in the graph structure, and forms a graph-level aesthetic representation through a region-learnable weight mechanism, finally outputting a 3D face aesthetic score; during the training phase, the model parameters are optimized using labeled aesthetic score data so that the model can stably output score results consistent with aesthetic standards.
[0043] Step 3-1: Construct a graph neural network model F, whose inputs are a 3D face image G=(V,E) and a node feature matrix X, and whose output is an aesthetic score s. The graph neural network includes the following components: 1) Node Feature Encoding Module: Encodes node features The initial implicit vector is obtained by performing a linear mapping or nonlinear transformation. ; 2) Message passing and node update module: A multi-layer graph neural network is used to iteratively update the hidden vectors of nodes; 3) Region Learnable Weight Module: Based on the proportional features of region labels and / or angle codes, learnable weight parameters are assigned to different region nodes or edges; 4) Graph-level aggregation module: Performs global aggregation of the high-level latent vectors of all nodes to obtain the graph-level representation g; 5) Regression Scoring Module: Input the graphical representation g into the regression header and output the continuous score s; Step 3-2: In the message passing and node feature update module, denote the initial latent vector of the node as... For the l-th layer graph neural network, for each node... From its set of adjacent nodes Aggregating neighbor information and updating node characteristics, the update process can be represented as follows:
[0044] in, For message functions, Aggregation functions (including but not limited to summation, averaging, maximization, or attention aggregation). For update functions; The edge features or edge weights are input, and these features can be obtained by mapping the angle-encoded features, distance / proportional features, or a combination thereof from step 2. Through multi-layer message passing, local structural information is gradually fused into a higher-level global structural representation. Step 3-3: In the region-learnable weight module, to enable the model to adaptively learn the contribution of different facial regions to the aesthetic score, this invention introduces a region-learnable weight mechanism to perform weighted modeling of node / edge information in different regions. This mechanism includes at least one or more of the following implementation methods: 1) Region Node Weights: Set learnable weight parameters for each facial region. Features of nodes belonging to region r Weighting; 2) Attention weights: Attention coefficients are calculated based on node features, region labels, and / or angular encoding features. Used to adjust the edge ( , The intensity of information dissemination; 3) Gated fusion weights: Gated weights are applied to the subgraph representations of different regions to obtain the final graph-level representation.
[0045] Through the above mechanism, the model can automatically learn the contribution of regions during training, so that the scoring results can correspond to aesthetic factors such as symmetry, contour curvature, and facial proportions. Steps 3-4: In the graph-level aggregation module, after completing the L-level node update, the high-level latent vectors of the nodes are obtained. The graph-level aesthetic representation g is obtained through the graph-level aggregation module, and a continuous score s is output. The process can be represented as follows:
[0046] in, This refers to global aggregation functions, including but not limited to global average pooling, global weighted pooling, or attention pooling. The node weights can be determined by a region-learnable weighting mechanism. Step 3-5: In the regression scoring module, based on the graph-level aesthetic representation g from Step 3-4, the regression process can be represented as follows:
[0047] in is the regression function (multilayer perceptron regression head), and s is the 3D facial aesthetic score.
[0048] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for evaluating the aesthetics of a 3D face based on graph neural networks, comprising the following steps: S1. Read the original patch mesh data obtained from the 3D face scan, and perform 3D key point recognition on the original patch based on the improved face key point detection model to obtain a set of key points containing 3D coordinate information; S2. Construct a three-dimensional face image structure using the key points as nodes, and generate node features containing three-dimensional coordinates, region labels, and angle features; S3. Construct a graph neural network aesthetic evaluation model, perform message passing and aggregation on node features in the graph structure, and form a graph-level aesthetic representation through a region-learnable weight mechanism, ultimately outputting a 3D face aesthetic score; the model includes a node feature encoding module, a message passing and node update module, a region-learnable weight module, a graph-level aggregation module, and a regression scoring module; The node feature encoding module is used to encode node features. The initial implicit vector is obtained by performing a linear mapping or nonlinear transformation. The message passing and node update module uses a multi-layer graph neural network to iteratively update the latent vectors of nodes; the region learnable weight module assigns learnable weight parameters to nodes or edges in different regions based on the proportional features of region labels and / or angle codes; the graph-level aggregation module is used to globally aggregate the high-level latent vectors of all nodes to obtain the graph-level representation g. The regression scoring module is used to input the graphical representation g into the regression header and output continuous scores s; S4. During the training phase, the model parameters are optimized using labeled aesthetic rating data so that the model can stably output rating results consistent with aesthetic standards. S5. Input the graph structure and node features into the graph neural network and output the 3D face aesthetic score.
2. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In step S1, a three-dimensional face image is first obtained, and then the improved facemesh recognition is used to obtain facial key point information, including the following steps: S11: Obtain a 3D patch of a human face through scanning or other 3D generation methods; S12: Perform principal vector analysis on the 3D face image and adjust the 3D face image to a frontal state; S13: Perform a 2D projection on the 3D face image, use facemesh to obtain the coordinates of the center point of the nose, the center point of the eyes, and the corners of the mouth in 2D coordinates, use these coordinate points to perform rotation and alignment operations on the face image, so that the center point of the nose, the center point of the eyes, and the center point of the corners of the mouth are on the same vertical line, with the center point of the eyes at the top and the center point of the corners of the mouth at the bottom. S14: Repeat S13; S15: Perform parallel projection 2D rendering on the aligned face patch to eliminate the effect of perspective; re-identify the face key point coordinates of the 2D rendered image using facemesh to obtain the 2D coordinates of the key points. S16: Obtain the three-dimensional coordinates of the key points by projecting their two-dimensional coordinates onto a face patch. N is the total number of key points, and each key point .
3. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In step S2, a three-dimensional face image structure is constructed based on the key points, and proportional features are generated under the guidance of preset facial aesthetic knowledge, thereby forming a node feature matrix that can be used for graph neural network evaluation, including the following steps: S21: For each key point in the key point set P... As a graph node This yields the node set V; S22: To characterize the geometric adjacency and structural topological relationships between key points, construct the edge set E, thereby obtaining the 3D face image G=(V,E); S23: For each node Construct its node feature vector The node features include at least the following parts: S24: Convert the feature vectors of each node Combine the nodes sequentially to obtain the node feature matrix. It outputs the 3D face image structure G=(V,E) and the corresponding node feature matrix X.
4. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 3, characterized in that, The node features described in S23 include the following: (1) Three-dimensional coordinate characteristics: 3D coordinates of key points The normalized coordinates serve as the basic features of the nodes; (2) Region label features: Assign region labels (forehead, eye area, nose, mouth area, cheekbone, jaw, etc.) to each key point and add node features in one-hot format; (3) Proportional features: Proportional features are constructed under the guidance of preset facial aesthetic knowledge, which includes the three courts and five eyes, symmetry, OJL curve and facial proportion, and the key point combination and the calculation rules of distance / angle / proportion of the proportional features are used to determine the proportional features.
5. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 4, characterized in that, The proportional features mentioned in step (3) include: Distance / proportion features: Calculate distances based on specific keypoint groups and form normalized proportions based on facial scale. Symmetry deviation features: Features are constructed based on the deviation and difference components of the left and right symmetrical key points to the facial midline / reference point; Angular encoding features: based on keypoint triples Calculated angles are used to characterize local angles, contour curvature, and geometric relationships; among which Corner key points and To and Angular encoding of connected or related key points on the same structure Defined as: 。 6. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 5, characterized in that, The angle encoding includes the following types: a) Local angle encoding: Construct triplets along the same local contour or adjacent key points in the same region to calculate angles, which are used to characterize local curvature and shape changes; b) Left-right symmetry angle encoding: Select the left and right paired key points and the central reference point to construct a triplet to calculate the angle or angle difference, which is used to quantify symmetry; c) Outer contour angle encoding: Select key point sequences along the outer boundary contour of the face to construct triplets and calculate turning angles, which are used to characterize the overall face shape and the coordination of the outer contour (e.g., jawline, cheek contour, forehead contour, etc.). The angle-coded features are used in subsequent evaluations in the following manner: Incorporated as part of node features : Will be with nodes Several related angle codes are concatenated to form the node proportional feature components. As input for edge features or edge weights: (This will be used in conjunction with the edge...) , The relevant angle encoding is mapped to edge features. Or, the input of attention weights can be used to adjust the intensity of neighbor information propagation; As input to the region-learnable weight module: the left-right symmetry angle encoding, outer contour angle encoding, etc. are associated with the region label to drive the adaptive learning of the region weight parameters, so that the weight learning corresponds to aesthetic factors.
7. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In the message passing and node feature update module, the initial latent vector of the node is denoted as... ; For the l-th layer graph neural network, for each node From its set of adjacent nodes Aggregating neighbor information and updating node characteristics, the update process can be represented as follows: in, For message functions, Aggregation functions (including but not limited to summation, averaging, maximization, or attention aggregation). For update functions; The edge features or edge weights are input, and the edge features are obtained by mapping the angle encoding features, distance / proportion features or their combination in step 2; through multi-layer message passing, the local structural information is gradually fused into a higher-level global structural representation.
8. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In the region-learnable weight module, to enable the model to adaptively learn the contribution of different facial regions to the aesthetic score, a region-learnable weight mechanism is introduced to perform weighted modeling of node / edge information in different regions. The mechanism includes one or more of the following implementation methods: 1) Region Node Weights: Set learnable weight parameters for each facial region. Features of nodes belonging to region r Weighting; 2) Attention weights: Attention coefficients are calculated based on node features, region labels, and / or angular encoding features. Used to adjust the edge ( , The intensity of information dissemination; 3) Gated fusion weights: Gated weights are applied to the subgraph representations of different regions to obtain the final graph-level representation.
9. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In the graph-level aggregation module, after completing the L-level node update, the high-level latent vectors of the nodes are obtained. The graph-level aesthetic representation g is obtained through the graph-level aggregation module, and a continuous score s is output. The process can be represented as follows: in, This refers to global aggregation functions, including but not limited to global average pooling, global weighted pooling, or attention pooling. The node weights are determined by a region-learnable weighting mechanism.
10. The three-dimensional face aesthetic evaluation method based on graph neural networks as described in claim 1, characterized in that, In the regression scoring module, based on the graph-level aesthetic representation g, the regression process is expressed as: in is the regression function (multilayer perceptron regression head), and s is the 3D facial aesthetic score.