Surface type index measurement model construction method, development condition evaluation method and system

By generating an elliptical Gaussian heat map and introducing a graph attention layer to optimize the detection of key points on the side face, the problem of large deviation in key point positioning in the measurement of side face facial indicators is solved, and higher-precision facial data measurement is achieved.

CN120726683AActive Publication Date: 2025-09-30HANGZHOU FEIYANG BIOTECHNOLOGY CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202511158056.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-19
Publication Date
2025-09-30
Estimated Expiration
2045-08-19

AI Technical Summary

Technical Problem

The existing technology has a problem in measuring profile facial indicators, such as large deviations in key point positioning, which leads to inaccurate angle calculations. It is particularly difficult to achieve high-precision facial data measurement in a home environment.

Method used

The facial index measurement model construction method is adopted. By generating an elliptical Gaussian heat map and introducing a graph attention layer (GAT layer), key point detection is optimized. The anatomical structure relationship of the side face image is used for global spatial optimization. The MSE and angular loss function are combined for supervised learning to improve the accuracy of key point detection.

Benefits of technology

It improves the accuracy of profile key point detection, enhances the ability to predict missing key points caused by factors such as head deviation and swing, and improves the accuracy and stability of facial angle calculation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120726683A_ABST
    Figure CN120726683A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of surface shape correction, in particular to a surface shape index measurement model construction method and a development condition evaluation method and system, and the method comprises the steps: carrying out the key point marking of a side face image, so as to make a standard side face image data set; generating a structural relationship among the key points corresponding to different standard angles based on the plurality of key points, and generating an optimized Gaussian heat map containing an elliptical heat map corresponding to each key point based on the structural relationship; introducing a GAT layer to take the key points as nodes, taking the structural relationship between the key points as edges, and calculating attention weights between the nodes and neighbor nodes to perform global space optimization; and performing supervised learning on the optimized Gaussian heat map and the real heat map based on an MSE loss function and an angle loss function to obtain a trained surface type index measurement model. The method has the effect of realizing accurate positioning of the side face key points so as to improve the calculation accuracy of the development angle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of facial correction, and in particular to a facial index measurement model construction method, a development assessment method and a system. Background Art

[0002] Measuring facial shape data through imaging and analyzing facial development has widespread medical applications. Visual facial shape data is primarily measured using three methods: facial skeletal imaging through X-rays, 3D facial imaging using multi-lens visible light or structured light, and multi-view imaging using monocular visible light. However, all of these methods have drawbacks.

[0003] Although X-rays can provide relatively accurate data, they rely on professional equipment and cannot meet the needs of users in daily home environment records. They also have radiation risks and are not suitable for frequent use.

[0004] Three-dimensional imaging methods such as multi-eye visible light or structured light usually require specialized customized equipment, which is very expensive and difficult to popularize in home scenarios.

[0005] Using monocular visible light for multi-view imaging is a feasible low-cost solution. Monocular multi-view imaging usually captures images of the human face from the front and side, and measures human facial data (such as pupil distance, lip distance, facial angle, nasolabial angle, chin-lip angle, etc.) through face positioning technology and facial key point extraction technology.

[0006] Specifically, it generally first uses face positioning technology to obtain the image area containing the face, and then uses facial key point extraction technology, such as deep learning methods (heat map regression, coordinate regression), traditional computer vision analysis methods (such as corner detection, SIFT, SURF) or commercial APIs to predict the coordinates of key points in the side face. After the key point positions are predicted, the lines and angles are connected based on the arrangement of each key point to calculate the required facial indicators.

[0007] Facial features like pupil distance and lip distance can be calculated by detecting key points in frontal facial images, while face angles and nasolabial angles can only be calculated by detecting key points in profile images. Currently, mainstream facial landmark detection methods, including various commercial APIs and community-based open-source algorithm models, are implemented for frontal faces. Directly applying these methods to profile images will result in significant deviations and is impractical.

[0008] One solution involves recreating a new dataset of profile landmarks based on existing frontal facial landmark algorithms, retraining the algorithm to create a model suitable for profile landmark detection. This model then calculates and measures the relevant facial shape metrics. In practice, keypoint positioning often deviates, which can be amplified in the angle calculation. For example, the chin-lip angle is typically large (over 110°), and the three landmarks corresponding to this angle are very close. A deviation of 1-2 pixels in this angle can result in an angular deviation of over 20°. During actual filming, subtle facial movements, changes in shooting angle, and even slight changes in head tilt can lead to unpredictable deviations in the independently predicted landmark positions. This deviation can be further amplified in environments with complex textures.

[0009] In summary, this solution divides the prediction of the position of key points of the side face and the calculation of the final angle into two independent technical steps, which has systematic defects and cannot be used for measuring side face facial shape indicators. Summary of the Invention

[0010] In order to accurately locate the key points of the side face and improve the calculation accuracy of the development angle, the present application provides a facial index measurement model construction method, a development status evaluation method and system.

[0011] In a first aspect, the present application provides a method for constructing a face index measurement model, which adopts the following technical solution:

[0012] A method for constructing a face index measurement model comprises the following steps:

[0013] Annotate key points of profile images to create a standard profile image dataset;

[0014] Generating a structural relationship between key points corresponding to different standard angles based on a plurality of key points, and generating an optimized Gaussian heat map including an elliptical heat map corresponding to each key point based on the structural relationship, wherein the elliptical heat map has different probability density distributions in a collinear direction and a perpendicular collinear direction between the key points matching the structural relationship;

[0015] A GAT layer is introduced to treat the key points as nodes, the structural relationships between the key points as edges, and calculate the attention weights between each node and its neighboring nodes to perform global spatial optimization;

[0016] Based on the MSE loss function and the angle loss function, supervised learning is performed on the optimized Gaussian heat map and the real heat map to obtain a trained side face key point detection model.

[0017] In some embodiments, generating a structural relationship between key points corresponding to different standard angles based on a plurality of key points includes the following steps:

[0018] Obtain each standard angle to define the angle relationship, obtain the key points required for each standard angle and obtain the marked coordinates to define the point coordinate relationship;

[0019] Obtaining participating edges corresponding to the standard angles to define a collinear relationship between the key points located between the same participating edges;

[0020] The structural relationship is generated based on the angular relationship, the collinear relationship and the point coordinate relationship.

[0021] In some embodiments, the optimized Gaussian heat map further includes a circular heat map, and before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are further included:

[0022] Determine the participation and distance of each key point to define each key point as an ellipse constraint point and a circle constraint point, specifically,

[0023] The participation degree is characterized by the number of angular relationships associated with each key point. If it is greater than 1, the participation degree corresponds to the circular constraint point; if it is not greater than 1, the participation degree corresponds to the elliptical constraint point;

[0024] The distance is characterized by the corner point distance between each key point and its associated angular relationship. If it is greater than the preset distance, the distance corresponds to the circular constraint point. If it is not greater than the preset distance, the distance corresponds to the elliptical constraint point.

[0025] In some embodiments, before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are also included:

[0026] Determining the stability of each of the ellipse constraint points to determine the collinear extension direction;

[0027] If the stability is not lower than the preset requirement, confirming the collinear extension direction of the ellipse constraint point as being toward a corner point in the associated angular relationship;

[0028] If the stability is lower than a preset requirement, the co-occurrence extension direction of the ellipse constraint point is determined to be toward other adjacent ellipse constraint points in the associated angular relationship.

[0029] In some embodiments, generating an optimized Gaussian heat map including an elliptical heat map corresponding to each key point based on the structural relationship includes the following steps:

[0030] Generate the circular heat map centered on the circular constraint point based on the real coordinates of the key point and a preset basic diffusion coefficient;

[0031] Determining a direction vector corresponding to the collinear direction based on each of the collinear extension directions, obtaining a deflection angle corresponding to the direction vector using a four-quadrant inverse tangent function, and aligning each original coordinate in the heat map toward the collinear direction based on the deflection angle to obtain an optimized coordinate;

[0032] The elliptical heat map corresponding to the elliptical constraint point is generated based on the optimized coordinates and the preset major axis diffusion coefficient and minor axis diffusion coefficient, wherein the major axis diffusion coefficient corresponds to the diffusion ability of the probability density distribution in the collinear direction, and the minor axis diffusion coefficient corresponds to the constraint ability of the probability density distribution in the perpendicular collinear direction.

[0033] In some embodiments, when the stability is lower than the preset requirement, the following steps are further included:

[0034] The probability density distribution between the ellipse constraint point whose stability is lower than the preset requirement and other adjacent ellipse constraint points is lower the closer it is to the midpoint of the two points.

[0035] In some embodiments, the following steps are also included:

[0036] Acquire specific group marker information, and match corresponding processing tags based on the specific group marker information;

[0037] A key point optimization coefficient is generated based on the processing tag, the standard angle is adjusted according to the key point optimization coefficient, and the major axis diffusion coefficient and the minor axis diffusion coefficient are adjusted based on the specific feature corresponding to the processing tag.

[0038] In some embodiments, calculating the attention weights between each node and its neighboring nodes to perform global spatial optimization on each structural relationship includes the following steps:

[0039] Obtaining and concatenating the feature vectors of the node and its neighboring nodes, calculating the relevance score corresponding to each edge through linear transformation and activation function, and normalizing the relevance scores corresponding to the node to calculate the attention weight;

[0040] Global spatial optimization is performed on the structural relationship and the nodes with abnormal feature vectors based on the size of the attention weight.

[0041] In a second aspect, the present application provides a method for evaluating developmental status, which employs the following technical solutions:

[0042] A developmental assessment method, based on the above method, comprises the following steps:

[0043] Obtain a two-dimensional profile face image and detect the ROI region containing the profile face, obtain the coordinates of several key points of the profile face through the face shape index measurement model and calculate the development angle;

[0044] Obtain a two-dimensional frontal face image and obtain the coordinates of key points of the front face and the head posture angle based on the recognition API, and calculate the facial soft tissue ratio based on the head posture angle and the coordinates of the key points of the front face;

[0045] A developmental assessment result is generated based on the developmental angle and the facial soft tissue proportion.

[0046] In a third aspect, the present application provides a developmental assessment method system, which adopts the following technical solutions:

[0047] A developmental status assessment method system is used to implement the above-mentioned developmental status assessment method.

[0048] The technical solutions provided by the embodiments of this application have the following technical effects:

[0049] First, for those specific key points that contribute to the formation of important facial angles, an isotropic circular Gaussian heat map is no longer generated. Instead, an elliptical Gaussian heat map region is generated whose direction adapts to its anatomical structure relationship. The major axis of this ellipse extends along the collinear direction between the two key points (that is, the direction least sensitive to the angle change), while the minor axis shrinks accordingly. This explicitly encodes the prior knowledge that "positioning deviations along specific axis directions have little effect on angle calculation" in the learning objective of the model. Secondly, the key point prediction is coordinated and corrected not only based on the response of local point features, but also through the GAT layer under the constraints of the global angle structure, so that the missing key points caused by head offset, swing, and hair occlusion can achieve neighbor perception and feature aggregation through graph attention, and enhance the prediction ability of missing key points through overall structured prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic diagram of the steps of a method for constructing a facial index measurement model provided in this embodiment.

[0051] Figure 2 It is a side face schematic diagram including various key points in the embodiment of the present application.

[0052] Figure 3 It is a schematic diagram of the relationship between various side face angles in the embodiment of the present application.

[0053] Figure 4Schematic diagram comparing the optimized Gaussian heat map and the conventional Gaussian heat map in the embodiment of the present application.

[0054] Figure 5 This is a comparison chart of the test data after the structural prior and the introduction of GAT in the embodiment of the present application and the test data of the conventional solution.

[0055] Figure 6 Schematic diagram of the steps of a developmental status assessment method provided in this embodiment.

[0056] Figure 7 It is a logical diagram of the developmental status assessment method provided in this embodiment. DETAILED DESCRIPTION

[0057] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. However, it should be understood by those skilled in the art that the present application can be implemented without these details. In some cases, in order to avoid unnecessary descriptions that make various aspects of the present application obscure, the well-known methods, processes, systems, components and / or circuits that have been described at a higher level will not be described in detail. It is obvious to those skilled in the art that various changes can be made to the embodiments disclosed in the present application, and the general principles defined in the present application can be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the embodiments shown, but conforms to the broadest scope consistent with the scope claimed for protection in the present application.

[0058] It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0059] In the description of this application, "several" means one or more, "many" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, while "above," "below," and "within" are understood to include the number itself. The use of "first" and "second" in the description is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0060] In the description of this application, reference to the terms "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any combination in one or more embodiments or examples.

[0061] An embodiment of the present application discloses a method for generating a profile key point detection model.

[0062] like Figure 1 and Figure 2 As shown, a method for constructing a face index measurement model includes the following steps:

[0063] S100 , annotating key points of the side face image to create a standard side face image dataset.

[0064] Collect pictures of people's profile faces, and use Labelme software to annotate key points in the profile face area according to anatomical definitions to obtain several key points required for facial development angles, such as the glabella point (G'), the tip of the columella (Cm'), the subnasal point (Sn'), the midpoint of the upper lip (Ls'), the midpoint of the lower lip (Li'), the chin-labial groove point (Sl'), the tip of the chin (Pog'), etc. The number and requirements of key points can be adaptively adjusted based on different detection scenarios and detection requirements.

[0065] For the purpose of clear description, this application mainly explains the above seven key points.

[0066] Collect several annotated side face pictures and integrate them to obtain a standard side face image dataset for training the model.

[0067] Furthermore, to improve the model's adaptability and applicable population, when creating a standard profile image dataset, the dataset can be further categorized by age, gender, and race. Because profile angles and contours vary among people of different ages, genders, and races, categorizing the dataset can improve the model's applicability.

[0068] S200, based on a number of key points, a structural relationship between key points corresponding to different standard angles is generated, and based on the structural relationship, an optimized Gaussian heat map including an elliptical heat map corresponding to each key point is generated, wherein the elliptical heat map has different probability density distributions in the collinear direction and the vertical collinear direction between the key points matching the structural relationship.

[0069] like Figure 3 As shown, when monitoring and evaluating the development of the face based on the side face, the main angles include the facial angle (defined by G, Sn, and Pog points), the nasolabial angle (defined by Ls, Sn, and Cm), and the chin-lip angle (defined by Li, Sl, and Pog). Based on different detection scenarios and detection requirements, the angles can also include the nasofacial angle, the upper lip concave angle, the lower lip inclination angle, the facial protrusion angle, the upper triangle, the lower triangle, the upper lip-chin protrusion angle, etc.

[0070] For the purpose of clear description, this application is mainly explained from the above three perspectives.

[0071] Therefore, when predicting angles, its accuracy does not only depend on the absolute positioning accuracy of each key point in a single facial tissue, but more importantly, the spatial configuration between the several points that constitute the angle. That is to say, the structural relationship of an angle should include the exact point coordinate position of each key point, and also the accurate angle relationship formed by multiple points.

[0072] Based on the structural relationship corresponding to each angle, prior knowledge is formed to perform structured optimization on the generation of key point heat maps. The optimized Gaussian heat map obtained after structured optimization is different from the conventional heat map.

[0073] like Figure 4 As shown in the figure, the conventional heat map only focuses on the accurate distribution probability between the predicted position and the true position of each key point independently, so it generates a circular heat map. The circular heat map assumes that the key point positioning error is isotropic and uniform in all directions, and its probability density decays symmetrically from the true center to the surrounding area to consider the accuracy of key point positioning.

[0074] The optimized Gaussian heat map also includes an elliptical heat map. The biggest difference between an ellipse and a circle is that the ellipse can reflect a certain directionality. It can adapt to the anatomical structure relationship and extend the probability density along the angular structure relationship formed by a certain key point and other key points, so that the probability density in the long axis direction is longer, while the probability density in the short axis direction is relatively short.

[0075] Because the key points are all on the contour of the human face, when a key point is extended to other key points that form the same angle relationship based on the structural relationship, the probability distribution of its elliptical heat map should be extended along the edge direction of the human face, that is, the collinear direction between the two key points involved in an angle. Then, how to predict a certain point has a certain offset in its true position, but when this offset happens to occur in its collinear direction, its final impact on the angle is small, and the angle values ​​calculated by these points will not have significant errors due to the positioning deviation of this key point.

[0076] On the contrary, if the offset between the key point and the true position occurs in the vertical collinear direction, that is, toward the inside or outside of the face, the offset of the point will cause it to deviate from the collinearity with other points, and its impact on the angle will be more significant. Therefore, the probability distribution density corresponding to the short axis of the vertical collinear direction in the elliptical heat map needs to be contracted.

[0077] S300, introduces the GAT layer to treat key points as nodes, the structural relationships between key points as edges, and calculate the attention weights between each node and its neighboring nodes for global spatial optimization.

[0078] Given the distinct and stable structural characteristics of a human face's profile, key points do not exist in isolation but rather have strict relative positional relationships and spatial constraints. To effectively capture and utilize this inherent anatomical knowledge within the model, this application introduces a graph attention network layer into the head prediction network. Key points on the profile are treated as nodes in a graph structure, and edge connections between nodes are defined based on their anatomical locations.

[0079] To address situations such as key points being blocked, missing, or blurred due to head shaking, changes in shooting angles, hair occlusion, and complex background, the edge relationship between multiple points in the graph structure is used to treat several key points as a whole, which together satisfy the inherent spatial relationship of the facial contour, thereby providing a high and consistent spatial basis for subsequent predicted facial angles. Missing key points are predicted as a whole through the complete spatial triangulation relationship that can be formed with other key points.

[0080] S400, based on the MSE loss function and angle loss function, supervised learning is performed on the optimized Gaussian heat map and the real heat map to obtain a trained profile face key point detection model.

[0081] The output heat map of the model and the true heat map are supervised by MSE loss function and angle loss function, the difference between the two is calculated, the deviation between the predicted heat map and the true heat map is minimized, and the model is guided to learn the positioning rules of key points (especially the angle-related spatial configuration). At the same time, the loss value is backpropagated to each layer of the model, and the convolution layer parameters of the head prediction network, the attention weight of the GAT layer and other parameters are updated, so that the model gradually optimizes the accuracy of the predicted heat map to obtain a trained side face key point detection model.

[0082] Specifically, the formula corresponding to the joint loss function supervision of heat map MSE and angle is:

[0083] .

[0084] in, Characterized as the mean square error loss of the keypoint heatmap, Characterized as the loss function of angle, 、 It is represented as a dynamic weight coefficient, which satisfies the requirement that the sum is equal to 1.

[0085] Among them, the formula of the MSE loss function is:

[0086] ;

[0087] Among them, N represents the batch size, K represents the number of key points, H and W are the height and width of the heat map, The representation is the predicted heat map output by the model, and Y is the optimized heat map generated by the real generation.

[0088] The formula of the angle cosine similarity loss function is:

[0089] ;

[0090] Where N represents the batch size and M represents the number of angles.

[0091] Through the above method, first, for those specific key points that are involved in the formation of important facial angles, an isotropic circular Gaussian heat map is no longer generated. Instead, an elliptical Gaussian heat map area with a directionality adapted to its anatomical structure relationship is generated. The major axis of this ellipse will extend along the collinear direction between the two key points (that is, the direction least sensitive to the angle change), while the minor axis will shrink accordingly, thereby explicitly encoding the prior knowledge that "positioning deviation along a specific axis direction has little effect on angle calculation" in the learning objective of the model; secondly, the key point prediction is coordinated and corrected not only based on the response of local point features, but also through the GAT layer under the global angle structure constraint, so that the missing key points caused by head offset, swing, and hair occlusion can achieve neighbor perception and feature aggregation through graph attention, and the prediction ability of missing key points is enhanced through overall structured prediction.

[0092] In some other embodiments, generating a structural relationship between key points corresponding to different standard angles based on a plurality of key points includes the following steps:

[0093] S210 , acquiring each standard angle to define an angle relationship, acquiring key points required for each standard angle, and acquiring standard coordinates to define a point coordinate relationship.

[0094] Taking the nasolabial angle as an example, it is defined by the subnasal point (Sn'), the upper lip convex point (Ls'), and the columellar point (Cm'), and is generally around 95 degrees. Therefore, the angle relationship introduced by the nasolabial angle during subsequent model training is set based on its standard angle. At the same time, the point coordinate relationship during training is defined based on the point coordinates marked in the side face image and the standard angle corresponding to the angle.

[0095] When analyzing and predicting key points, the key points themselves must not only satisfy the corresponding point coordinate relationship, but also the angle relationship formed by them and other key points.

[0096] S220: Acquire participating edges corresponding to the standard angles to define collinear relationships between key points located between the same participating edges.

[0097] Each angle is composed of three points and two edges. For example, in the nasolabial angle, it is composed of an edge composed of Cm and Sn and an edge composed of Sn and Ls. Then, the two key points corresponding to each participating edge have a collinear relationship. This collinear relationship is an important feature for the subsequent generation of elliptical heat maps and the prediction of point positions based on the GAT layer.

[0098] S230: Generate a structural relationship based on the angle relationship, the collinear relationship, and the point coordinate relationship.

[0099] The structural relationship is composed of angular relationship, collinear relationship and point coordinate relationship, so when introducing structural prior and GAT layer in this application, we not only focus on the real coordinate prediction of each key point, but also pay attention to the angle formed by the spatial relationship between different key points and the identity of the key points involved in the angular relationship.

[0100] like Figure 4 As shown, in some other embodiments, the optimized Gaussian heat map further includes a circular heat map, and before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are also included:

[0101] S240 , determining the participation degree and the distance degree of each key point to define each key point as an ellipse constraint point and a circle constraint point, respectively.

[0102] The optimized Gaussian heat map in this application is different from the existing conventional Gaussian heat map in that each key point will match and generate an elliptical heat map or a circular heat map based on its own characteristics during heat map generation.

[0103] The characteristic of the elliptical heat map is that it has a directional constraint on the probability distribution, and its probability density in the long axis direction is higher than that in the short axis direction. The characteristic of the circular heat map is that it mainly focuses on the accuracy of key point positioning, and its probability density is uniform in all directions.

[0104] In the embodiment of the present application, whether a key point should correspond to an elliptical heat map or a circular heat map is determined mainly based on two characteristics of the key point: participation and proximity.

[0105] The degree of participation is characterized by the angular relationship of the key point in which it actually participates, that is, how many angles require the key point to participate.

[0106] The distance is characterized by the distance between each key point and the corner point in the corner where it is located. When a key point itself is a corner point, its corresponding distance is zero.

[0107] S241, the participation degree is characterized by the number of angular relationships associated with each key point. If it is greater than 1, the participation degree corresponds to a circular constraint point. If it is not greater than 1, the participation degree corresponds to an elliptical constraint point.

[0108] For participation, if a keypoint corresponds to more than one angular relationship, that is, it needs to participate in the definition of two angles, the point corresponds to a circular constraint point.

[0109] Taking the chin tip point Pog as an example, it defines the facial angle together with G and Sn, and also defines the chin-lip angle together with Li and SI. Therefore, its collinear direction is not fixed, and there are two collinear directions. Then, when the long axis direction of the elliptical heat map is subsequently defined based on the collinear direction, the structural prior will cause the model to use both collinear directions as prior knowledge, resulting in the inability to generate an accurate elliptical heat map.

[0110] Therefore, in this case, no elliptical heat map is generated for this type of key point, and only the circular heat map is used to constrain it to find the accurate point coordinates.

[0111] S242, the distance is characterized by the corner point distance between each key point and its associated angular relationship. If it is greater than the preset distance, the distance corresponds to a circular constraint point. If it is not greater than the preset distance, the distance corresponds to an elliptical constraint point.

[0112] Regarding the distance, such as the glabella point G, which is located in the upper half of the face, it is far away from the corner point Sn of the facial angle it participates in defining. If the G point generates an elliptical heat map, its corresponding long axis distance Sn will be far away from Pog, and its structural relationship will be unstable, which will reduce the confidence of the area. Therefore, for points with a distance greater than the preset distance, they are defined as circular constraint points in the embodiment of the present application.

[0113] For other points whose distance is less than the preset distance, they can all correspond to elliptical constraint points. Since the distances between these key points and other key points of the corners they participate in defining are relatively close, the overall structural relationship is stable.

[0114] It should be noted that when determining key points by distance and participation, only if a point corresponds to an elliptical constraint point in both determinations will an elliptical heat map be generated for that point. Conversely, as long as there is a point that corresponds to a circular constraint point in one determination, only a circular heat map will be generated for that point.

[0115] In some other embodiments, before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are also included:

[0116] S250: Determine the stability of each ellipse constraint point to determine the collinear extension direction.

[0117] For the ellipse constraint points, it is necessary to clarify the co-occurrence relationship between them and other key points to determine the extension direction of the major axis of the ellipse.

[0118] At this point, the stability of each elliptical constraint point must be determined. Stability is characterized by whether a key point is clearly defined within the facial structure and whether slight changes in facial movement during side-on face photography will significantly disrupt the final predicted position.

[0119] The high or low stability determines the method used to determine the collinear extension direction of the ellipse constraint point.

[0120] S251: If the stability is not lower than the preset requirement, the collinear extension direction of the ellipse constraint point is confirmed to be toward the corner point in the associated angular relationship.

[0121] When stability is at least as high as the preset requirement, the collinear extension direction of the elliptical constraint points is oriented toward the corner point of the defined angle relationship. Because a corner point is the intersection of two sides of an angle, its overall structural characteristics are more distinct. Furthermore, the line connecting a key point and a corner point is an edge of the angle relationship, and its structural properties are also more obvious.

[0122] Therefore, the collinear extension direction of the ellipse constraint points with higher stability is determined to be generated toward the corner point in the angular relationship.

[0123] S252: If the stability is lower than the preset requirement, the co-occurrence extension direction of the ellipse constraint point is determined to be toward other adjacent ellipse constraint points in the associated angular relationship.

[0124] For example, the chin-labial groove point SI is located between the chin and the lips. Its structural feature information is not obvious. At the same time, during the evaluation process, the slightest movement of the lips will greatly interfere with the predicted positioning of this point, so the stability of this point is lower than the preset requirement.

[0125] For this type of key points, it is necessary to improve their ability to resist interference from complex backgrounds and other factors. Therefore, in the embodiment of the present application, the collinear extension direction is determined based on other adjacent elliptical constraint points in the defined angle relationship.

[0126] When determining the collinear relationship between two points, if the features of each point are stable and obvious, then any point only needs to be connected to another point. When the features of another key point are not obvious, its point position and collinear position need to be determined by other stable point positions and line positions. Therefore, for the SI point, on the one hand, its features are unstable, and on the other hand, as a corner point in an angle, its collinear relationship needs to refer to the direction information of other stable key points to assist in determination. Therefore, in this application, the collinear extension direction of the SI point needs to be determined with reference to the adjacent Li point. That is to say, the long axis of the ellipse heat map of the two points Sl and Li is determined by Li-Sl, so that the two points can fit the edge of the face as much as possible. SI-Pog is not used for determination because the participation of Pog is greater than 1, so it is a circular constraint point, which cannot provide accurate collinear direction information for the SI point.

[0127] It should be noted that the collinear extension direction and the collinear direction are two different concepts. The collinear extension direction is characterized by the direction in which the major axis of the elliptical constraint point is to be extended, which is determined by the structural prior before the contribution direction is determined. It has a single directionality, such as towards the corner point or towards other key points. Once the collinear extension direction is determined, the collinear direction between the two points can be obtained. The collinear direction is bidirectional, which is specifically considered depending on which key point the object is. For example, in Sl-Li, when the reference object is SI, its collinear direction is from SI to Li, and when the reference object is Li, its collinear direction is from Li to SI; and for Cm, its collinear extension direction is towards the corner point Sn, that is, the reference object of Cm when looking for the collinear relationship is along Sn. When the collinear extension direction is determined, the collinear relationship of Cm-Sn can be obtained, and the collinear direction of Cm itself is along the connecting line and extension line of Cm-Sn.

[0128] In some other embodiments, generating an optimized Gaussian heat map including an elliptical heat map corresponding to each key point based on the structural relationship includes the following steps:

[0129] S260: Generate a circular heat map centered on the circular constraint point based on the real coordinates of the key point and the preset basic diffusion coefficient.

[0130] For circular constraint points, the method for generating a circular heat map is the same as the existing Gaussian heat map regression method. The specific formula is as follows:

[0131] .

[0132] Among them, (x, y) represents the integer coordinates of any point on the heat map. , It is represented by the coordinates of the center of the key point on the heat map, that is, the actual position of the key point after scaling, and σ is represented by the basic diffusion coefficient (standard deviation) in pixels, which controls the diffusion range of the Gaussian distribution.

[0133] S261, determine the direction vector corresponding to the collinear direction based on each collinear extension direction, use the four-quadrant inverse tangent function to obtain the deflection angle corresponding to the direction vector, and align each original coordinate in the heat map to the collinear direction based on the deflection angle to obtain the optimized coordinate.

[0134] For the ellipse constraint point, we must first determine the direction of the ellipse's major axis, so here we first obtain the direction vector between the coordinates of the two key points, specifically:

[0135] .

[0136] in, 、 It is represented by the coordinates of the key point in the topological relationship as the starting point on the heat map (such as Ls in Ls-Sn), 、 It is represented by the coordinates of the key point as the end point in the topological relationship on the heat map (such as in Ls-Sn, the end point is Sn), and v is represented by the direction vector pointing from the starting point to the end point, indicating the directionality of the key point.

[0137] Then, the four-quadrant inverse tangent function is used to obtain the deflection angle. The deflection angle is represented by the angle value corresponding to the co-occurrence direction. Specifically:

[0138] .

[0139] Align the original coordinates to the new coordinate system by the deflection angle, that is, rotate the original coordinates on the heat map , making the long axis of the elliptical heat map the new coordinate system of the reference, eliminating the direction error. For example, for the midpoint of the upper lip Ls', its long axis is along the line Ls'-Sn'. After rotation, the x' axis coincides with the direction of the line, and the y' axis is perpendicular to the direction (that is, the direction sensitive to angle calculation). Specifically:

[0140] .

[0141] Finally, the coordinates of each key point in the rotated coordinate system are used to generate the elliptical Gaussian. The specific formula is:

[0142] .

[0143] in, Characterized as the long axis diffusion coefficient, it is used to control the diffusion capacity of the probability density distribution in the long axis direction, that is, the distance the heat map extends along the collinear direction. Characterized as the short-axis diffusion coefficient, it is used to control the constraint ability of the probability density distribution in the short-axis direction, that is, the distance length compressed by the heat map along the vertical collinear direction.

[0144] S262, generating an elliptical heat map corresponding to the elliptical constraint point based on the optimized coordinates and the preset major axis diffusion coefficient and minor axis diffusion coefficient, wherein the major axis diffusion coefficient corresponds to the diffusion ability of the probability density distribution in the collinear direction, and the minor axis diffusion coefficient corresponds to the constraint ability of the probability density distribution in the perpendicular collinear direction.

[0145] For elliptical constraint points, when the model learns to locate key points, a greater tolerance is allowed for positioning errors in specific directions (collinear directions), and its directionality adapts to its anatomical structure relationship.

[0146] For example, in the nasolabial angle, point Sn is considered a relatively stable anchor point. For points Cm and Ls, the heat map we generate will no longer be a simple circle. Instead, we will design an elliptical distribution area based on the collinear direction indicated by the line connecting point Sn and points Cm and Ls (that is, the direction least sensitive to the angle change). The long axis of this ellipse will extend along the direction of the line connecting Sn-Cm or Sn-Ls, while the short axis will shrink accordingly, thereby explicitly encoding the prior knowledge that "positioning deviation along a specific axis direction has little effect on angle calculation" in the learning objective of the model.

[0147] In this way, during the optimization process, the model is more tolerant of positional variations of points Cm and Ls along the collinear direction of the nasolabial contour, while deviations perpendicular to this direction are subject to stricter constraints. Furthermore, these key points that constitute the angle (such as Cm, Ls, and Sl) are typically located along the edge contour of the profile. Background areas outside the contour often contain complex and disruptive texture information such as texture variations, shadows, and even clothing or other irrelevant objects. Traditional circular heatmaps symmetrically spread their high-response areas inward and outward of the contour. This means that circular heatmaps can unnecessarily, or even harmfully, direct the model's learning attention to complex background areas outside the contour, as the model attempts to match the "false" high-response signals generated by the circular geometry truth in these areas.

[0148] In contrast, elliptical heatmaps offer significant advantages: by precisely extending the long axis of the ellipse along the contour direction (i.e., the potential collinear direction between key points) while shrinking the short axis perpendicular to the contour direction, the high-response regions can be effectively restricted to the contour line itself and its adjacent, anatomically consistent regions. This targeted focusing strategy avoids ineffective diffusion into complex background texture areas, allowing the model to focus its learning resources more on capturing true anatomical edge features rather than being misled by background noise. The ultimate goal is to significantly enhance the prediction accuracy of facial angles calculated from these points, at the expense of the absolute accuracy of individual key points, and to reduce the model's sensitivity to complex background interference.

[0149] In other embodiments, when the stability is lower than the preset requirement, the following steps are further included:

[0150] S2521, the probability density distribution between the ellipse constraint point whose stability is lower than the preset requirement and other adjacent ellipse constraint points is lower the closer it is to the midpoint of the two points.

[0151] In the above embodiment, it is disclosed that the SI point feature is unstable, so its collinear extension direction needs to refer to the adjacent Li point. The SI point is a corner point, and the collinear extension direction of the Li point needs to be determined by the SI point. Then the collinear directions of the Li point and the SI point are consistent, and both need to be extended along the long axis through the elliptical heat map. If the extension condition of the long axis is not constrained, the probability density distribution between SI and Li will be relatively high. Regardless of whether it is the Li point or the SI point, when the predicted point coordinates are generated by the predicted heat map, if the probability density between SI and Li is high, resulting in the prediction being generated at a position between the two points, although it does not affect the angle of the chin-lip angle, the actual point coordinates will have a large error from the true point coordinates.

[0152] Therefore, in the embodiment of the present application, it is necessary to constrain the probability density distribution assigned to the long axis of the two points in the above situation, so that the probability density distribution between the two points is lower the closer it is to the midpoint. Then, when the model predicts the point coordinates, it will not be generated at a position with large errors such as the midpoint of the two points.

[0153] In some other embodiments, the following steps are further included:

[0154] S270: Acquire specific group identification information, and match corresponding processing tags based on the specific group identification information.

[0155] When users take photos of their own faces to obtain developmental change data, they often use mobile devices to take photos of their side faces, or have a doctor take the photos. After the photos are taken, they can manually enter relevant specific population marker information, such as gender, age, race, etc., based on the population characteristics of the actual person being photographed.

[0156] Based on the information entered by the user, the system will automatically select a matching processing label. The model will select parameters trained by training data sets of different categories under different processing labels to analyze and predict the key points of the side face.

[0157] There are certain differences in the positions of facial feature points among different genders, age groups, and races. For example, younger children are not fully developed, so their nasolabial angles are relatively large; Europeans have relatively high brow bones, and their actual facial angles are smaller.

[0158] Especially for people of different ages and races, there is no uniformity between the corresponding standard angles, and there are large angle differences. In order to improve the prediction accuracy for people with different differences, it is necessary to adjust the various parameters of the model by processing labels.

[0159] S271 , generating a key point optimization coefficient based on the processing tag, adjusting the standard angle according to the key point optimization coefficient, and adjusting the long-axis diffusion coefficient and the short-axis diffusion coefficient based on the specific features corresponding to the processing tag.

[0160] By processing labels to generate key point optimization coefficients, the standard angles trained in the model are adjusted through the key point optimization coefficients to adapt to the analysis basis of different populations. At the same time, specific features corresponding to different populations are generated based on different processing labels, such as high cheekbone features for Mongolians, high nose bridge features for Caucasians, and low brow bone features for young children. The long-axis diffusion coefficient and short-axis diffusion coefficient are adaptively adjusted according to different specific features.

[0161] Different features lead to large differences in the positions of key points in different populations. Therefore, the angular relationships, collinear relationships, and major and minor axis distances of elliptical heat maps of different populations are all different. Therefore, in order to improve the prediction accuracy under different scenarios, this application needs to configure different model parameters for different populations through different specific features.

[0162] Through the above steps, relevant priors such as age, gender, and race are added when generating the elliptical heat map to improve the model's learning of the angular characteristics of different groups of people.

[0163] This application introduces the GAT layer. The core function of the GAT layer is that it gives the model the ability to dynamically learn the interdependence between key points. In this layer, each key point (node) no longer relies solely on its own local image features when making the final heat map prediction, but can adaptively and selectively aggregate feature information from its topological neighbor key points. This aggregation does not treat all neighbors equally, but learns a weight distribution through the attention mechanism: the model calculates the "correlation" or "importance" score (i.e., attention weight) between the current node and each of its neighbor nodes. The feature information of neighbor nodes with high attention weights will have a greater impact on the prediction of the current node. This information interaction mechanism based on graph attention directly encodes the strict relative position and spatial constraint relationship between key points.

[0164] For example, when predicting the G-point, which constitutes facial angles, the GAT layer can focus on the characteristic states of the Sn and Pog points, which have strong anatomical connections with it. Through learned attention weights, the model understands the anatomical principle that the reasonable position of the G-point should be strongly constrained by the positions of the Sn and Pog points. Therefore, the predicted G-point position is not only a response based on local image evidence but also the result of coordination and correction under global structural constraints. It must form a reasonable spatial triangular relationship with the Sn and Pog points.

[0165] In actual side face pictures, the offset and swing of the head may cause occlusion of facial soft tissue points. This neighbor perception and feature aggregation achieved through graph attention greatly enhances the model's ability to predict occluded key points.

[0166] Essentially, the GAT layer establishes a powerful structured reasoning module within the model, enabling it to go beyond local feature matching and perform global spatial optimization based on anatomical relationships, which is key to improving the robustness and accuracy of angle prediction.

[0167] Therefore, in some other embodiments, the attention weights between each node and its neighboring nodes are calculated to perform global spatial optimization on each structural relationship, including the following steps:

[0168] S310, obtain the feature vectors of the node and its neighboring nodes and concatenate them, calculate the correlation score corresponding to each edge through linear transformation and activation function, and normalize the several correlation scores corresponding to the node to calculate the attention weight.

[0169] Obtain feature vectors of the target node and neighboring nodes, which include local image features, multi-scale information, context association, etc.

[0170] The original correlation scores of the target node and each neighbor node are calculated through a learnable parameter vector. For example, when calculating the correlation between G and Sn, the feature vectors of the two are concatenated and the scores are obtained through linear transformation and activation function (such as LeakyReLU). ; Similarly calculate the scores of G and Pog .

[0171] The raw scores of all neighbor nodes are softmax normalized to obtain the attention weights. After normalization, the sum of the weights is equal to 1, and the larger the value, the stronger the anatomical correlation between the neighbor node and the target node.

[0172] S320, performing global spatial optimization on nodes with abnormal structural relationships and feature vectors based on the size of the attention weight.

[0173] When optimizing through the GAT layer, two situations are mainly optimized:

[0174] The first is to optimize the relative position and spatial constraint relationship between different key points, that is, to strengthen the strong anatomical relationship between one point and the other two points, and the reasonable position of this point should be strongly constrained by the positions of the other two points.

[0175] The above-mentioned enhancement process is achieved through feature aggregation of attention weights. The final features of the target node (such as G) are obtained by the weighted sum of its own features and the features of neighboring nodes (such as Sn and Pog). If Sn and Pog have a stronger anatomical correlation with point G, the model will learn a higher attention weight, so that the features of Sn and Pog account for a higher proportion in the final features of point G.

[0176] The second method is to optimize the prediction process of a point when it is occluded or missing. Specifically, the prediction process is strengthened by the correlation score between the occluded point and its topological neighboring nodes. For example, the correlation score between Sl and Li is calculated. Since the two are direct components of the chin-lip angle (with a strong anatomical connection), the score is high.

[0177] The correlation score between Sl and Pog was calculated. The two were indirectly related through the chin-lip angle (Pog is the tip of the chin and has a stable position), and the score was second only to that of Pog.

[0178] However, the SI point is occluded, resulting in low reliability of its local features;

[0179] After Softmax normalization, the attention weight of Sl and Li is high, the attention weight of Sl and Pog is medium, and the weight of Sl's own features is the lowest.

[0180] Then, the optimized features of the occluded point SI can be obtained by aggregating the features between neighboring nodes. After aggregation, the optimized features of Sl no longer rely on the fuzzy local occlusion features, but are mainly based on the position information of Li and Pog, following the spatial constraint that "Sl is located on the anatomical contour line between Li and Pog".

[0181] like Figure 5 As shown, in the test phase, different face images in different time periods are selected for key point prediction based on the present application scheme and conventional Gaussian heat maps. Multiple images in the same day are of the same person with different backgrounds. The shooting distance and the tilt angle of the head are kept as consistent as possible to simulate the slight tilt in normal shooting conditions.

[0182] After the improvement of this scheme, the difference between the maximum and minimum values ​​of the facial angle measured based on partial data was 1.48, while the difference without optimization was 4.09; the difference between the maximum and minimum values ​​of the chin-lip angle was 7.78 with the improvement, while it was 17.64 without improvement; the difference between the maximum and minimum values ​​of the nasolabial angle was 5.7 with the improvement, while it was 15.57 without improvement.

[0183] After the improvement of this scheme, the variation of the measurement results at different angles is smaller, which is much smaller than the variation of the front and back angles in the conventional scheme.

[0184] like Figure 6 、 7 As shown, the present application also discloses a developmental status assessment method, which is implemented based on the above method and includes the following steps:

[0185] S500: Obtain a two-dimensional side face image and detect the ROI region containing the side face, obtain the coordinates of several key points of the side face through a face shape index measurement model, and calculate the development angle.

[0186] During the actual developmental data evaluation process, users take and upload side face pictures using mobile devices such as mobile phones and tablets. Based on the acquired two-dimensional side face pictures, the system detects the ROI area containing the side face through mobilenet or YOLOV3. At the same time, the ROI area is input into the side face key point detection model trained by the above method to obtain the coordinates of each key point and calculate each developmental angle.

[0187] S600: Obtain a two-dimensional frontal face image and obtain the coordinates of key points of the front face and the head posture angle based on the recognition API, and calculate the facial soft tissue ratio based on the head posture angle and the coordinates of key points of the front face.

[0188] After obtaining the frontal face picture of the user taken by mobile devices such as mobile phones and tablets, and obtaining the key points on the front face and the head posture angle through commercial API, the facial soft tissue proportion is compensated and calculated based on the head posture angle and the coordinates of the key points on the front face.

[0189] S700, generating a developmental assessment result based on developmental angles and facial soft tissue proportions.

[0190] For example, when evaluating the recovery of the soft tissues of the upper lip and chin, the relative proportions of the two Euclidean distances between the key points of the frontal face and the pupils of the two eyes are calculated through the coordinates of the key points of the frontal face, which assists the system in determining the accuracy of the side face angle assessment.

[0191] At the same time, the distance and relative ratio will change with the pitch of the head. Therefore, after the pitch angle data is returned by the API, for images with excessively large angles, the calculated distance is multiplied by the cosine of the pitch angle to correct the vertical distance.

[0192] The present application also discloses a developmental status assessment method system for implementing the above-mentioned developmental status assessment method.

[0193] The implementation principle is:

[0194] First, for specific key points involved in forming important facial angles, an isotropic circular Gaussian heat map is no longer generated. Instead, an elliptical Gaussian heat map region is generated whose directionality adapts to its anatomical structure. The major axis of this ellipse extends along the collinear direction between the two key points (i.e., the direction least sensitive to angle changes), while the minor axis shrinks accordingly. This explicitly encodes the prior knowledge that "positioning deviations along specific axis directions have little impact on angle calculation" in the model's learning objectives.

[0195] Secondly, it is not only based on the response of local point features, but also coordinates and corrects the key point prediction under the global angle structure constraint through the GAT layer, so that the missing key points caused by head offset, swing, and hair occlusion can achieve neighbor perception and feature aggregation through graph attention, and enhance the prediction ability of missing key points through overall structured prediction.

[0196] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps and they may be performed in other orders.

[0197] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A method for constructing a face index measurement model, characterized in that: The following steps are involved: Annotate key points of profile images to create a standard profile image dataset; Generating a structural relationship between key points corresponding to different standard angles based on a plurality of key points, and generating an optimized Gaussian heat map including an elliptical heat map corresponding to each key point based on the structural relationship, wherein the elliptical heat map has different probability density distributions in a collinear direction and a perpendicular collinear direction between the key points matching the structural relationship; A GAT layer is introduced to treat the key points as nodes, the structural relationships between the key points as edges, and calculate the attention weights between each node and its neighboring nodes to perform global spatial optimization; Based on the MSE loss function and the angle loss function, supervised learning is performed on the optimized Gaussian heat map and the real heat map to obtain a trained side face key point detection model.

2. The method for constructing a face index measurement model according to claim 1, wherein: Generating structural relationships between key points corresponding to different standard angles based on a number of key points includes the following steps: Obtain each standard angle to define the angle relationship, obtain the key points required for each standard angle and obtain the marked coordinates to define the point coordinate relationship; Obtaining participating edges corresponding to the standard angles to define a collinear relationship between the key points located between the same participating edges; The structural relationship is generated based on the angular relationship, the collinear relationship and the point coordinate relationship.

3. The method for constructing a face index measurement model according to claim 2, wherein: The optimized Gaussian heat map also includes a circular heat map. Before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are also included: Determine the participation and distance of each key point to define each key point as an ellipse constraint point and a circle constraint point, specifically, The participation degree is characterized by the number of angular relationships associated with each key point. If it is greater than 1, the participation degree corresponds to the circular constraint point; if it is not greater than 1, the participation degree corresponds to the elliptical constraint point; The distance is characterized by the corner point distance between each key point and its associated angular relationship. If it is greater than the preset distance, the distance corresponds to the circular constraint point. If it is not greater than the preset distance, the distance corresponds to the elliptical constraint point.

4. The method for constructing a face index measurement model according to claim 3, wherein: Before generating the optimized Gaussian heat map corresponding to each key point based on the structural relationship, the following steps are also included: Determining the stability of each of the ellipse constraint points to determine the collinear extension direction; If the stability is not lower than the preset requirement, confirming the collinear extension direction of the ellipse constraint point as being toward a corner point in the associated angular relationship; If the stability is lower than a preset requirement, the collinear extension direction of the ellipse constraint points is determined to be toward other adjacent ellipse constraint points in the associated angular relationship.

5. The method for constructing a face index measurement model according to claim 3 or 4, characterized in that: Generating an optimized Gaussian heat map including an elliptical heat map corresponding to each of the key points based on the structural relationship includes the following steps: Generate the circular heat map centered on the circular constraint point based on the real coordinates of the key point and a preset basic diffusion coefficient; Determining a direction vector corresponding to the collinear direction based on each of the collinear extension directions, obtaining a deflection angle corresponding to the direction vector using a four-quadrant inverse tangent function, and aligning each original coordinate in the heat map toward the collinear direction based on the deflection angle to obtain an optimized coordinate; The elliptical heat map corresponding to the elliptical constraint point is generated based on the optimized coordinates and the preset major axis diffusion coefficient and minor axis diffusion coefficient, wherein the major axis diffusion coefficient corresponds to the diffusion ability of the probability density distribution in the collinear direction, and the minor axis diffusion coefficient corresponds to the constraint ability of the probability density distribution in the perpendicular collinear direction.

6. The method for constructing a face index measurement model according to claim 4, wherein: When the stability is lower than the preset requirement, the following steps are further included: The probability density distribution between the ellipse constraint point whose stability is lower than the preset requirement and other adjacent ellipse constraint points is lower the closer it is to the midpoint of the two points.

7. The method for constructing a face index measurement model according to claim 5, wherein: The following steps are also included: Acquire specific group marker information, and match corresponding processing tags based on the specific group marker information; A key point optimization coefficient is generated based on the processing tag, the standard angle is adjusted according to the key point optimization coefficient, and the major axis diffusion coefficient and the minor axis diffusion coefficient are adjusted based on the specific feature corresponding to the processing tag.

8. The method for constructing a facial index measurement model according to claim 1, wherein: Calculating the attention weights between each of the nodes and neighboring nodes to perform global spatial optimization on each of the structural relationships comprises the following steps: Obtaining and concatenating the feature vectors of the node and its neighboring nodes, calculating the relevance score corresponding to each edge through linear transformation and activation function, and normalizing the relevance scores corresponding to the node to calculate the attention weight; Global spatial optimization is performed on the structural relationship and the nodes with abnormal feature vectors based on the size of the attention weight.

9. A method for evaluating developmental status, characterized in that: The method according to claims 1 to 8 is implemented, comprising the following steps: Obtain a two-dimensional profile face image and detect the ROI region containing the profile face, obtain the coordinates of several key points of the profile face through the face shape index measurement model and calculate the development angle; Obtain a two-dimensional frontal face image and obtain the coordinates of key points of the front face and the head posture angle based on the recognition API, and calculate the facial soft tissue ratio based on the head posture angle and the coordinates of the key points of the front face; A developmental assessment result is generated based on the developmental angle and the facial soft tissue proportion.

10. A developmental assessment method system, characterized in that: Used to implement the developmental status assessment method as described in claim 9.

Citation Information

Patent Citations

  • Data processing, training and recognition method and device and storage medium

    CN111881705A

  • Face aesthetics evaluation and enhancement method based on structural reasoning

    CN112699816A

  • Facial development analysis and evaluation method and system

    CN113705444A

  • Key point detection model training method and device and storage medium

    CN116311422A

  • Unsupervised key point detection method based on graph attention network

    CN118229991A