Facial feature index measurement model construction method, developmental assessment method and system

Abstract

This application relates to the technical field of facial contour correction, and in particular to a method for constructing a facial contour index measurement model, a method for assessing developmental status, and a system. The method includes: annotating key points in profile images to create a standard profile image dataset; generating structural relationships between key points corresponding to different standard angles based on several key points; generating an optimized Gaussian heatmap containing elliptical heatmaps for each key point based on the structural relationships; introducing a GAT layer to treat key points as nodes and the structural relationships between key points as edges; calculating attention weights between each node and its neighboring nodes for global spatial optimization; and performing supervised learning on the optimized Gaussian heatmap and the real heatmap based on the MSE loss function and the angle loss function to obtain a trained facial contour index measurement model. This application has the effect of accurately locating key points of the profile to improve the accuracy of calculating developmental angles.

Description

Technical Field

[0001] This application relates to the technical field of facial correction, and in particular to a method for constructing a facial index measurement model, a method for assessing development, and a system. Background Technology

[0002] Measuring facial contour data through imaging and analyzing facial development has wide applications in the medical field. Visual facial contour data measurement mainly uses three methods: facial skeletal imaging via X-rays, three-dimensional facial imaging using multi-view visible light or structured light, and multi-view imaging using monocular visible light. However, all of these methods have certain limitations.

[0003] While X-rays can provide relatively accurate data, they require specialized equipment and cannot meet the needs of users recording their daily home environment. In addition, they pose a radiation risk and are not suitable for frequent use.

[0004] Multi-view visible light or structured light 3D imaging methods usually require specialized custom equipment, which is very expensive and difficult to popularize in home settings.

[0005] Using monocular visible light for multi-view imaging is a feasible low-cost solution. Monocular multi-view imaging typically acquires images of the front and side of the human face, and uses face localization technology and facial key point extraction technology to measure human facial data (such as pupil distance, lip distance, facial angle, nasolabial angle, chin-lip angle, etc.).

[0006] Specifically, it generally involves first obtaining an image region containing the face through face localization technology. Based on this, facial key point extraction techniques, such as deep learning methods (heatmap regression, coordinate regression), traditional computer vision analysis methods (such as corner detection, SIFT, SURF), or commercial APIs, are used to predict the coordinates of key points in the profile. After the key point positions are predicted, lines are connected and angles are measured based on the arrangement of each key point to calculate the required facial shape indicators.

[0007] Facial landmarks such as interpupillary distance and lip distance can be calculated by detecting key points in frontal facial images, while facial 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 open-source algorithms, are all designed for frontal faces. Directly applying these methods to profile images will result in significant inaccuracies and render them impractical.

[0008] One solution is to create a new dataset of side-face keypoints based on existing frontal facial landmark detection algorithms, retrain the algorithm to obtain a model suitable for side-face keypoint detection, and then calculate and measure relevant facial features. In practical applications, keypoint localization often has errors, which are amplified in the final angle calculation. For example, the chin-lip angle is generally large (over 110°), and the three keypoints corresponding to this angle are very close. If the corner point has a deviation of 1-2 pixels, it may result in an angle deviation of over 20° in the final angle. During actual shooting, minute facial movements, slight changes in shooting angle, and slight changes in head rotation angle can all cause unpredictable deviations in the independently predicted keypoint positions. If shooting in a complex textured environment, these deviations may be further amplified.

[0009] In conclusion, this approach treats the prediction of key points on the side of the face and the calculation of the final angle as two independent and separate technical steps, which has systemic flaws and cannot be used for measuring side facial features. Summary of the Invention

[0010] In order to achieve accurate positioning of key points on the side of the face and improve the accuracy of development angle calculation, this application provides a method for constructing a facial shape index measurement model, a method for evaluating development status, and a system.

[0011] Firstly, this application provides a method for constructing a surface shape index measurement model, which adopts the following technical solution:

[0012] A method for constructing a surface shape index measurement model includes the following steps:

[0013] Keypoint annotation was performed on profile images to create a standard profile image dataset;

[0014] Based on several key points, structural relationships between key points corresponding to different standard angles are generated. Based on the structural relationships, optimized Gaussian heatmaps containing elliptical heatmaps are generated for each key point. The elliptical heatmaps have different probability density distributions in the collinear and perpendicular collinear directions between key points matching the structural relationships.

[0015] A GAT layer is introduced to treat the key points as nodes and the structural relationships between the key points as edges. Attention weights between each node and its neighboring nodes are calculated to perform global space optimization.

[0016] Supervised learning is performed on the optimized Gaussian heatmap and the real heatmap based on the MSE loss function and the angle loss function to obtain a trained side profile key point detection model.

[0017] In some embodiments, the structural relationships between key points corresponding to different standard angles are generated based on several key points, including the following steps:

[0018] Obtain standard angles to define angular relationships, obtain key points required for each standard angle, and obtain labeled coordinates to define point coordinate relationships;

[0019] Obtain the participating edge corresponding to the standard angle to define the collinear relationship of the key points located between the same participating edge;

[0020] The structural relationships are generated based on the aforementioned angular relationships, collinear relationships, and point coordinate relationships.

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

[0022] The engagement degree and distance degree of each key point are determined to define each key point as an elliptical constraint point and a circular constraint point, respectively.

[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 phase distance is characterized as the angular distance between each key point and its associated angle. If it is greater than a preset distance, the phase distance corresponds to the circular constraint point; if it is not greater than the preset distance, the phase distance corresponds to the elliptical constraint point.

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

[0026] Determine the stability of each elliptical constraint point to determine the collinear extension direction;

[0027] If the stability is not lower than the preset requirement, then the collinear extension direction of the elliptical constraint point is confirmed as the direction toward the corner point in the associated angular relationship;

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

[0029] In some embodiments, generating an optimized Gaussian heatmap containing an elliptical heatmap for each of the key points based on the structural relationship includes the following steps:

[0030] A circular heat map centered on the circular constraint point is generated based on the real coordinates of the key points and a preset basic diffusion coefficient.

[0031] Based on the collinear extension directions, the direction vector corresponding to the collinear direction is determined, and the bias angle corresponding to the direction vector is obtained using the four-quadrant arctangent function. Based on the bias angle, the original coordinates in the heat map are aligned to the collinear direction to obtain optimized coordinates.

[0032] Based on the optimized coordinates and the preset major axis diffusion coefficient and minor axis diffusion coefficient, an elliptical heatmap corresponding to the elliptical constraint point is generated, wherein the major axis diffusion coefficient corresponds to the diffusion capability of the probability density distribution in the collinear direction, and the minor axis diffusion coefficient corresponds to the constraint capability 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 elliptical constraint point whose stability is lower than the preset requirement and other adjacent elliptical constraint points is lower the closer it is to the midpoint between the two points.

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

[0036] Obtain specific group identifier information and match corresponding processing tags based on the specific group identifier information;

[0037] Keypoint optimization coefficients are generated based on the processed labels, the standard angle is adjusted according to the keypoint optimization coefficients, and the major axis diffusion coefficient and the minor axis diffusion coefficient are adjusted based on the specific features corresponding to the processed labels.

[0038] In some embodiments, the attention weights between each node and its neighboring nodes are calculated to perform global spatial optimization of each structural relationship, including the following steps:

[0039] The feature vectors of the node and its neighboring nodes are obtained and concatenated. The relevance scores corresponding to each edge are calculated through linear transformation and activation function. The relevance scores corresponding to the node are normalized to calculate the attention weight.

[0040] Global space optimization is performed on the structural relationships and nodes with abnormal feature vectors based on the magnitude of the attention weights.

[0041] Secondly, this application provides a method for assessing developmental status, employing the following technical solution:

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

[0043] Two-dimensional side profile images are acquired and ROI regions containing the side profile are detected. The coordinates of several key points of the side profile are obtained through the facial shape index measurement model and the development angle is calculated.

[0044] A two-dimensional frontal face image is obtained, and the coordinates of key points on the frontal face and the head posture angle are obtained based on the recognition API. The proportion of facial soft tissue is calculated based on the head posture angle and the coordinates of key points on the frontal face.

[0045] Developmental assessment results are generated based on the developmental angle and the proportion of facial soft tissue.

[0046] Thirdly, this application provides a developmental assessment method system, which adopts the following technical solution:

[0047] A developmental assessment method system for implementing the aforementioned developmental assessment method.

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

[0049] First, for specific key points that contribute to important facial angles, instead of generating isotropic circular Gaussian heatmaps, an elliptical Gaussian heatmap region is generated that adapts to the 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 contracts accordingly. This explicitly encodes the prior knowledge that "positioning deviations along specific axes have little impact on angle calculations" in the model's learning objective. Second, instead of relying solely on local point feature responses, the GAT layer coordinates and corrects key point predictions under global angle structure constraints. This allows missing key points caused by head offset, swaying, or hair occlusion to achieve neighbor perception and feature aggregation through graph attention, enhancing the predictive ability for missing key points through overall structured prediction. Attached Figure Description

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

[0051] Figure 2 This is a side view diagram including key points in the embodiments of this application.

[0052] Figure 3 This is a schematic diagram showing the relationship between the various side profile angles in the embodiments of this application.

[0053] Figure 4This is a schematic diagram comparing the optimized Gaussian heatmap and the conventional Gaussian heatmap in the embodiments of this application.

[0054] Figure 5 This is a comparison chart of the test data of the structural prior and the test data after introducing GAT in the embodiments of this application with the test data of the conventional scheme.

[0055] Figure 6 This is a schematic diagram of the steps of a developmental assessment method provided in this embodiment.

[0056] Figure 7 This is a logical schematic diagram of the developmental assessment method provided in this embodiment. Detailed Implementation

[0057] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0058] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, 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, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0060] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0061] This 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 surface shape index measurement model includes the following steps:

[0063] S100, key point annotation is performed on side profile images to create a standard side profile image dataset.

[0064] Images of human profiles are collected, and Labelme software is used to annotate key points within the profile 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 mentolabial sulcus point (Sl'), and the tip of the chin (Pog'). The number and requirements of key points can be adaptively adjusted based on different detection scenarios and requirements.

[0065] For the purpose of clarity, this application mainly focuses on the seven key points mentioned above.

[0066] We collected several labeled profile images and integrated them to obtain a standard profile image dataset for training the model.

[0067] Furthermore, to improve the model's adaptability and applicability to different groups, the standard profile image dataset can be further categorized by age group, gender, and ethnicity. Since the profile angles and contours of people of different ages, genders, and ethnicities vary, categorizing the dataset can enhance the model's applicability.

[0068] S200 generates structural relationships between key points at different standard angles based on several key points, and generates optimized Gaussian heatmaps containing elliptical heatmaps for each key point based on the structural relationships. The elliptical heatmaps have different probability density distributions in the collinear and perpendicular collinear directions between key points matching structural relationships.

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

[0070] For the purpose of clarity, this application mainly focuses on the three aspects mentioned above.

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

[0072] Prior knowledge is formed based on the structural relationships corresponding to each angle to perform structured optimization of the generation of key point heatmaps. The optimized Gaussian heatmap obtained after structured optimization is different from the conventional heatmap.

[0073] like Figure 4 As shown, a conventional heatmap only focuses on the accurate distribution probability between the predicted and actual locations of each key point, thus generating a circular heatmap. The circular heatmap assumes that the key point positioning error is isotropic and uniform in all directions, and its probability density decays symmetrically from the actual center outwards to assess the accuracy of key point positioning.

[0074] Optimized Gaussian heatmaps also include elliptical heatmaps. The biggest difference between elliptical and circular heatmaps is that elliptical heatmaps can reflect a certain directionality. They can adapt to anatomical structural relationships and extend the probability density along the angular structural relationship between a key point and other key points, making the probability density in the major axis direction longer, while the probability density in the minor axis direction is relatively shorter.

[0075] Since the key points are all located on the contour of the human face, when a key point extends to other key points that constitute the same angular relationship based on structural relationships, the probability distribution of its elliptical heatmap should extend along the edge direction of the human face, that is, the collinear direction between two key points involved in the same angle. So, if the predicted position of a certain point has a certain offset, but this offset happens to occur in its collinear direction, its final impact on the angle is small. The angle value calculated by these points will not have a significant error due to the positioning deviation of this key point.

[0076] Conversely, if the offset of a key point from its actual position occurs in the vertical collinear direction, that is, towards the inside or outside of the face, the offset of that point will cause it to deviate from the collinearity of other points. In this case, its impact on the angle is more significant. Therefore, the probability distribution density corresponding to the minor axis in the vertical collinear direction in the elliptical heatmap needs to be reduced.

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

[0078] Considering the significant and stable structural characteristics of the human profile, key points are not isolated but have strict relative positional relationships and spatial constraints. To significantly capture and utilize this inherent anatomical knowledge within the model, this application introduces a graph attention network layer into the head prediction network. Each key point on the profile is treated as a node in a graph structure, and the edge connections between nodes are defined based on their anatomical location.

[0079] To address situations where key points are obscured, missing, or blurred due to head movement, changes in shooting angle, hair obscuring the view, or complex backgrounds, this method treats several key points as a whole by utilizing the edge relationships between multiple points in the graph structure. This collectively satisfies the inherent spatial relationships of the facial contour, thus providing a highly consistent spatial basis for subsequent facial angle prediction. Missing key points are predicted as a whole by using the complete spatial triangular relationships that can be formed with other key points.

[0080] S400 uses the MSE loss function and the angle loss function to perform supervised learning on the optimized Gaussian heatmap and the real heatmap to obtain a well-trained side profile key point detection model.

[0081] Supervised learning is performed on the model's output heatmap and the real heatmap using the MSE loss function and the angle loss function. The difference between the two is calculated, and the deviation between the predicted heatmap and the real heatmap is minimized. This guides the model to learn the localization rules of key points (especially angle-related spatial configurations). At the same time, the loss value is backpropagated to each layer of the model to update the parameters of the convolutional layer of the head prediction network, the attention weights of the GAT layer, and other parameters, so that the model can gradually optimize the accuracy of the predicted heatmap to obtain a well-trained profile key point detection model.

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

[0083] .

[0084] in, This is represented by the mean square error loss of the keypoint heatmap. The loss function is represented by the angle. , It is represented as a dynamic weighting coefficient, which satisfies the condition that the sum of the coefficients equals 1.

[0085] The formula for the MSE loss function is as follows:

[0086] ;

[0087] Where N represents the batch size, K represents the number of key points, and H and W represent the height and width of the heatmap. Y represents the predicted heatmap output by the model, while Y represents the optimized heatmap generated in reality.

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

[0089] ;

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

[0091] The above method firstly, for specific key points that contribute to important facial angles, instead of generating isotropic circular Gaussian heatmaps, it generates elliptical Gaussian heatmap regions that are directionally adapted to their anatomical relationships. The major axis of this ellipse extends along the collinear direction between the two key points (i.e., the direction in which the angle change is least sensitive), while the minor axis shrinks accordingly. This explicitly encodes the prior knowledge that "positioning deviations along specific axes have little impact on angle calculations" in the model's learning objective. Secondly, it not only relies on the response of local point features but also coordinates and corrects key point predictions under global angle structure constraints through the GAT layer. This allows missing key points caused by head offset, swaying, or hair occlusion to achieve neighbor perception and feature aggregation through graph attention, enhancing the predictive ability for missing key points through overall structured prediction.

[0092] In other embodiments, the structural relationships between key points corresponding to different standard angles are generated based on several key points, including the following steps:

[0093] S210: Obtain each standard angle to define the angle relationship, obtain the key points required for each standard angle and obtain the standard coordinates to define the point coordinate relationship.

[0094] Taking the nasolabial angle as an example, it is defined by the subnasal point (Sn'), the midpoint of the upper lip (Ls'), and the columellar point (Cm'), and is generally around 95 degrees. The nasolabial angle is set based on its standard angle to introduce the angle relationship during subsequent model training. At the same time, the point coordinate relationship during training is defined based on the coordinates of the points marked on the side face image combined with the standard angle corresponding to this 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, obtain the participating edges corresponding to the standard angle to define the collinear relationship of key points between the same participating edges.

[0097] Each corner consists of three points and two sides. For example, in the nasolabial angle, it is formed by the side composed of Cm and Sn and the side composed of Sn and Ls. Then, the two key points corresponding to each participating side are collinear. This collinear relationship is an important feature for the subsequent generation of elliptical heatmaps and the prediction of point positions based on GAT layers.

[0098] S230 generates structural relationships based on angular relationships, collinear relationships, and point coordinate relationships.

[0099] The structural relationships are composed of angular relationships, collinear relationships, and point coordinate relationships. This allows the application to focus not only on the prediction of the true coordinates of each key point when introducing structural priors and GAT layers, but also on the angles formed by different key points based on spatial relationships and the identities of key points participating in these angular relationships.

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

[0101] S240, determine the participation degree and distance degree of each key point to define each key point as an elliptical constraint point and a circular constraint point respectively.

[0102] The optimized Gaussian heatmap in this application differs from the existing conventional Gaussian heatmap in that each key point, based on its own characteristics, will be matched to generate an elliptical or circular heatmap during heatmap generation.

[0103] Elliptical heatmaps are characterized by directional constraints on probability distribution, with higher probability density along the major axis than along the minor axis. Circular heatmaps, on the other hand, primarily focus on the accuracy of key point localization, and their probability density is uniform in all directions.

[0104] In the embodiments of this application, the determination of whether a key point should correspond to an elliptical heatmap or a circular heatmap is mainly based on two characteristics of the key point: participation and distance.

[0105] The participation level is characterized by how many angle relationships the key point actually participates in, that is, how many angles require the participation of the key point.

[0106] The distance characteristic represents the distance between each keypoint and the corner point within the corner it is located in. When a keypoint is itself a corner point, its distance characteristic is zero.

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

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

[0109] Taking the chin tip Pog as an example, it defines the face shape angle together with G and Sn, and the chin-lip angle together with Li and SI. Therefore, its collinear direction is not fixed, and there are two collinear directions. So when defining the major axis direction of the elliptical heatmap based on the collinear direction, it will cause the model to take both collinear directions as prior knowledge when the structure is known, resulting in the inability to generate an accurate elliptical heatmap.

[0110] Therefore, in this case, elliptical heatmaps are not generated for such key points; instead, circular heatmaps are used to constrain the accurate point coordinates.

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

[0112] For the distance, such as the eyebrow point G, which is located in the upper half of the face, the distance between it and the corner point Sn of the face shape angle it participates in defining is relatively far. If an elliptical heat map is generated for point G, the corresponding major axis distance Sn will be far from Pog, and its structural relationship will be unstable, which will reduce the confidence of the region. Therefore, for points with a distance greater than the preset distance, in this embodiment of the application, they are defined as circular constraint points.

[0113] For other points with a distance less than the preset distance, they can all correspond to elliptical constraint points, because these key points are close to other key points between the angles they participate in defining, and the overall structural relationship is stable.

[0114] It is important to note that when determining key points using distance and participation, only if a point corresponds to an elliptical constraint point in both determinations will an elliptical heatmap be generated for that point. Conversely, if a point corresponds to a circular constraint point in one determination, then only a circular heatmap will be generated for that point.

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

[0116] S250, determine the stability of each elliptical constraint point to determine the collinear extension direction.

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

[0118] At this point, the stability of each elliptical constraint point needs to be determined. Stability is characterized by whether the features of a key point are obvious in the structural position of the face, and whether slight changes in movement during the side-view shooting of the face will cause significant interference to the final predicted localization.

[0119] The level of stability determines the method used to select the collinear extension direction of the elliptical constraint point.

[0120] S251, if the stability is not lower than the preset requirement, then the collinear extension direction of the elliptical constraint points is confirmed as the corner point in the associated angular relationship.

[0121] When the stability is not lower than the preset requirement, the collinear extension direction of the elliptical constraint point is towards the corner point in its defined angular relationship. Since the corner point is the intersection point of two sides in an angle, its overall structural features are relatively obvious. At the same time, the line connecting a key point and a corner point is a side in an angular relationship, and its structural properties are also relatively obvious.

[0122] Therefore, the collinear extension direction of the elliptical constraint points with higher stability is determined by generating the corner points in the angular relationship.

[0123] S252, if the stability is lower than the preset requirement, the co-occurrence extension direction of the elliptical constraint point is confirmed as the direction towards other adjacent elliptical constraint points in the associated angular relationship.

[0124] For example, the mental-labial sulcus point SI is located in the middle of the chin and lips. Its structural features are not obvious. At the same time, during the evaluation process, even slight movements of the lips can greatly interfere with the predicted location of this point. Therefore, the stability of this point is lower than the preset requirements.

[0125] For such key points, it is necessary to improve their ability to resist interference from complex backgrounds and other human factors. Therefore, in this embodiment, the collinear extension direction is determined based on other adjacent elliptical constraint points in the defined angular 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 the other point. However, when the features of another key point are not obvious, its position and collinear position need to be determined by the positions of other stable points and lines. Therefore, for point S1, its features are unstable, and as a corner point in an angle, its collinear relationship needs to be determined with reference to the directional information of other stable key points. Thus, in this application, the collinear extension direction of point S1 needs to be determined with reference to the adjacent point Li. That is, the major axis of the elliptical heatmap of points S1 and Li is determined by Li-S1, which allows these two points to fit as close as possible to the edge of the face. SI-Pog is not used because Pog's participation degree is greater than 1, making it a circular constraint point, which cannot provide accurate collinear direction information for point S1.

[0127] It should be noted that collinear extension direction and collinear direction are two different concepts. Collinear extension direction represents the direction in which the major axis of an elliptical constraint point will extend, determined by structural priors before the contribution direction is determined. It has a single directionality, such as towards a corner point or towards other key points. Once the collinear extension direction is determined, the collinear direction between two points can be obtained. Collinear direction has a bidirectional nature, 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; when the reference object is Li, its collinear direction is from Li to SI. For Cm, its collinear extension direction is towards the corner point Sn. That is to say, when Cm is looking for collinear relationships, the reference object is along Sn. Once the collinear extension direction is determined, the collinear relationship between Cm and Sn can be obtained, and the collinear direction of Cm itself is along the line connecting Cm and Sn and its extension.

[0128] In other embodiments, generating an optimized Gaussian heatmap containing an elliptical heatmap for each key point based on structural relationships includes the following steps:

[0129] S260 generates a circular heatmap centered on a circular constraint point based on the real coordinates of key points and a preset basic diffusion coefficient.

[0130] For circular constraint points, the method for generating circular heatmaps is the same as the existing Gaussian heatmap regression method, and the specific formula is as follows:

[0131] .

[0132] Where (x, y) represents the integer coordinates of any point on the heatmap. , The coordinates of the center of the key point on the heatmap are represented, which is the actual location of the key point after scaling. σ represents the basic diffusion coefficient (standard deviation), which is in pixels and 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 arctangent function to obtain the bias angle corresponding to the direction vector, and align each original coordinate in the heat map to the collinear direction based on the bias angle to obtain optimized coordinates.

[0134] For elliptical constraint points, the direction of the major axis of the ellipse must first be determined. Therefore, 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 that serves as the starting point in the topological relationship on the heatmap (e.g., in Ls-Sn, the starting point is Ls). , The key is represented by the coordinates of the endpoint of the key point on the heatmap (e.g., in Ls-Sn, the endpoint is Sn), and v is represented by the direction vector from the starting point to the endpoint, indicating the directionality of the key point.

[0137] Next, the bias angle is obtained using the four-quadrant arctangent function. The bias angle represents the magnitude of the angle value corresponding to the co-occurrence direction. Specifically:

[0138] .

[0139] Aligning the original coordinates to the new coordinate system by tilting the original coordinates, which means rotating the original coordinates on the heatmap. This creates a new coordinate system with the major axis of the elliptical heatmap as the reference, eliminating direction errors. For example, for the midpoint Ls' of the upper lip, its major axis is along the line connecting Ls' and Sn'. After rotation, the x' axis coincides with this line, and the y' axis is perpendicular to this direction (i.e., the direction sensitive to angle calculations). Specifically:

[0140] .

[0141] Finally, an elliptical Gaussian is generated using the coordinates of each key point in the rotated coordinate system. The specific formula is as follows:

[0142] .

[0143] in, Characterized by the major axis diffusion coefficient, it controls the diffusion capability of the probability density distribution along the major axis, which is the distance the heatmap extends along the collinear direction. Characterized as the short-axis diffusion coefficient, it is used to control the constraint of the probability density distribution along the short axis, which is the distance length compressed by the heat map along the vertical collinear direction.

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

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

[0146] In the nasolabial angle, point Sn is considered a relatively stable anchor point. For points Cm and Ls, the generated heatmap will no longer be a simple circle. Instead, we design an elliptical distribution region based on the collinear direction indicated by the line connecting Sn with Cm and Ls (i.e., the direction least sensitive to angle changes). The major axis of this ellipse extends along the line connecting Sn-Cm or Sn-Ls, while the minor axis contracts accordingly, thus explicitly encoding the prior knowledge that "positioning deviations along a specific axis have little impact on angle calculation" into the model's learning objective.

[0147] In this way, the model is more tolerant of positional variations of points Cm and Ls along the nasolabial contour (collinear direction) during optimization, while offsets perpendicular to this direction are subject to stricter constraints. Furthermore, these key points constituting the angle (such as Cm, Ls, Sl) are typically located on the edge contour line of the profile. The background region outside the contour line often includes complex and distracting texture information such as texture variations, shadows, and even clothing or other irrelevant objects. Traditional circular heatmaps symmetrically diffuse their high-response regions both inside and outside the contour line. This means that circular heatmaps unnecessarily, and even harmfully, direct the model's attention to the complex background regions outside the contour line, as the model attempts to match the "pseudo" high-response signals generated by the circular Ground Truth in these regions.

[0148] In contrast, elliptical heatmaps offer significant advantages: by precisely extending the major axis of the ellipse along the contour direction (i.e., the potential collinearity between keypoints) while contracting the minor axis perpendicular to the contour direction, they effectively limit high-response regions to primarily the contour line itself and its adjacent, anatomically consistent areas. This targeted focusing strategy avoids ineffective diffusion into complex background textures, allowing the model to concentrate its learning resources more effectively on capturing genuine anatomical edge features, rather than being misled by background noise. The ultimate goal is to significantly enhance the accuracy of facial angle predictions calculated from individual keypoints, at the cost of sacrificing absolute accuracy of individual keypoints, and to reduce the model's sensitivity to complex background interference.

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

[0150] S2521, the probability density distribution between elliptical constraint points with stability lower than the preset requirement and other adjacent elliptical constraint points is lower the closer to the midpoint between the two points.

[0151] In the above embodiments, it is disclosed that the SI point is unstable, so its collinear extension direction needs to be referenced to the adjacent Li point. Since the SI point is a corner point, the collinear extension direction of the Li point also needs to be determined by the SI point. Therefore, the collinear directions of the Li point and the SI point are consistent, and they both need to be extended along the major axis through the elliptical heatmap. If the extension conditions of the major axis are not constrained, the probability density distribution between SI and Li will be high. Whether it is the Li point or the SI point, when generating the predicted point coordinates in the predicted heatmap, if the probability density between SI and Li is high, the predicted point will be generated in the position between the two points. Although this does not affect the angle of the chin-lip angle, the actual point coordinates will have a large error with the true point coordinates.

[0152] Therefore, in this embodiment of the application, it is necessary to constrain the probability density distribution of the major axis of the two points in the above case so that the probability density distribution between the two points is lower as it gets closer to the midpoint. In this way, the model will not generate at positions with large errors, such as the midpoint between the two points, when predicting the coordinates of the points.

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

[0154] S270: Obtain the identification information of a specific population group, and match the corresponding processing label based on the identification information of the specific population group.

[0155] When users take photos of their own faces to obtain data on developmental changes, they often use mobile devices to take photos of their profiles, or have a doctor take the photos. After the photos are taken, users can manually input relevant specific group information, such as gender, age, and race, based on the characteristics of the people being photographed.

[0156] Based on the information input by the user, the system will automatically select the matching processing label. Under different processing labels, the model will select parameters trained from training datasets of different categories to analyze and predict key points of the profile.

[0157] The positions of facial feature points vary among different genders, age groups, and ethnicities. For example, younger children, whose development is not yet complete, have relatively larger nasolabial angles; Europeans have higher brow bones and smaller actual facial angles.

[0158] In particular, there is no uniformity in the standard angles corresponding to people of different ages and ethnicities, and there are significant differences in angles. In order to improve the prediction accuracy for different groups of people, it is necessary to adjust the various parameters of the model by processing the labels.

[0159] S271, generate key point optimization coefficients based on the processed labels, adjust the standard angle according to the key point optimization coefficients, and adjust the major axis diffusion coefficient and minor axis diffusion coefficient based on the specific features corresponding to the processed labels.

[0160] By processing labels to generate keypoint optimization coefficients, the standard angles trained in the model are adjusted using these keypoint optimization coefficients to adapt to the analysis basis of different populations. At the same time, based on different processed labels, specific features corresponding to different populations are generated, such as high cheekbone features for some populations, high nose bridge features for some populations, 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 significant differences in the location of key points among different populations. Consequently, the angular relationships, collinear relationships, and major and minor axis distances of elliptical heatmaps also differ among different populations. Therefore, in order to improve the prediction accuracy under different scenarios, this application needs to configure different model parameters for different populations using different specific features.

[0162] By incorporating prior information such as age, gender, and race into the generation of elliptical heatmaps, the model's ability to learn angular features from different population groups is improved.

[0163] This application introduces the GAT layer, whose core function is to endow the model with the ability to dynamically learn the interdependencies between keypoints. In this layer, each keypoint (node) no longer relies solely on its own local image features when making the final heatmap prediction. Instead, it adaptively and selectively aggregates feature information from its topological neighbors. This aggregation does not treat all neighbors equally; instead, it learns a weight distribution through an attention mechanism: the model calculates the "relevance" or "importance" score (i.e., attention weight) between the current node and each of its neighboring nodes. Neighboring nodes with higher attention weights have a greater impact on the prediction of the current node. This graph attention-based information interaction mechanism directly encodes the strict relative positional and spatial constraints between keypoints.

[0164] For example, when predicting the G-point that constitutes a facial angle, the GAT layer can make it pay more attention to the feature states of the Sn and Pog points, which have strong anatomical connections with it. Through the learned attention weights, the model can understand the anatomical rule that "the reasonable location of the G-point should be strongly constrained by the locations of the Sn and Pog points." Therefore, the predicted location of the G-point 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 profile photos, head shifts and movements can occlude soft tissue points on the face. 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. This is key to improving the robustness and accuracy of angular prediction.

[0167] Therefore, in some other embodiments, the attention weights between each node and its neighboring nodes are calculated to perform global spatial optimization of 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 relevance scores corresponding to each edge through linear transformation and activation function. Normalize the relevance scores corresponding to the node to calculate the attention weight.

[0169] The feature vectors of the target node and its neighboring nodes are obtained, which include local image features, multi-scale information, contextual relationships, etc.

[0170] The original relevance scores between the target node and its neighbors are calculated using learnable parameter vectors. For example, when calculating the relevance between G and Sn, their feature vectors are concatenated, and the scores are obtained through linear transformation and activation functions (such as LeakyReLU). Similarly, calculate the fractions of G and Pog. .

[0171] The original scores of all neighboring nodes are normalized using softmax to obtain attention weights. After normalization, the sum of the weights equals 1, and the larger the value, the stronger the anatomical correlation between the neighboring node and the target node.

[0172] S320 performs global space optimization on nodes with structural relationships and abnormal feature vectors based on the magnitude of attention weights.

[0173] When optimizing through the GAT layer, two main scenarios are optimized:

[0174] The first approach is to optimize the relative positions and spatial constraints between different key points, which means strengthening the strong anatomical relationship between a point and two other points, and ensuring that the reasonable position of the point is 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 feature of the target node (such as G) is obtained by weighted summation of its own features and the features of its neighboring nodes (such as Sn and Pog). If Sn and Pog have a stronger anatomical relationship with point G, the model will learn higher attention weights, so that the features of Sn and Pog account for a higher proportion in the final features of point G.

[0176] The second approach is to optimize the prediction process for a point when it is occluded or missing. Specifically, this is done by strengthening the correlation score between the occluded point and its neighboring nodes in the topological relationship. For example, the correlation score between Sl and Li is calculated because they are the direct constituent points of the chin-lip angle (strong anatomical correlation) and have a high score.

[0177] Calculate the correlation score between Sl and Pog. The two are indirectly related through the mentolabial angle (Pog is the tip of the chin, which is in a stable position), and the score is the next highest.

[0178] The occlusion of SI points leads to lower reliability of their local features;

[0179] After Softmax normalization, the attention weights of Sl and Li are high, the attention weights of Sl and Pog are 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 feature aggregation between neighboring nodes. After aggregation, the optimized features of SI no longer depend on the fuzzy local occlusion features, but are mainly based on the position information of Li and Pog, following the spatial constraint that "SI is located on the anatomical contour line between Li and Pog".

[0181] like Figure 5 As shown, in the testing phase, different face images from different time periods were selected for key point prediction based on the proposed solution and conventional Gaussian heatmaps. Multiple images from the same day were taken of the same person, and the shooting distance and head tilt angle were kept as consistent as possible under different backgrounds to simulate slight tilting under normal shooting conditions.

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

[0183] The improved scheme results in smaller variations in measurement results at different angles, far less than the changes in front and back angles in the conventional scheme.

[0184] like Figure 6 , 7 As shown, this application also discloses a developmental assessment method, implemented based on the above method, comprising the following steps:

[0185] The S500 acquires a 2D profile image and detects the region of interest (ROI) containing the profile. It then uses a facial shape index measurement model to obtain the coordinates of several key points on the profile and calculates the development angle.

[0186] In the actual developmental data evaluation process, users take and upload side profile photos using mobile devices such as mobile phones and tablets. Based on the acquired two-dimensional side profile photos, the system detects the Region of Interest (ROI) containing the side profile using methods such as MobileNet or YOLOv3. At the same time, the ROI is input into the side profile key point detection model trained by the above method to obtain the coordinates of each key point and calculate each developmental angle.

[0187] S600 acquires a two-dimensional frontal face image and obtains the coordinates of key points on the frontal face and the head posture angle based on the recognition API. It then calculates the proportion of facial soft tissue based on the head posture angle and the coordinates of key points on the frontal face.

[0188] The system acquires frontal images of users taken by mobile devices such as smartphones and tablets, and obtains key points and head posture angles on the frontal face through commercial APIs. It then calculates compensation for the proportions of facial soft tissue based on the head posture angles and coordinates of the key points on the frontal face.

[0189] S700 generates developmental assessment results based on developmental angles and facial soft tissue proportions.

[0190] For example, when assessing the recovery of the soft tissues of the upper lip and chin, the relative proportions of these two points to the two Euclidean distances from the pupils of the two eyes are calculated using the coordinates of key points on the front of the face, thus assisting the system in judging the accuracy of the side profile angle assessment.

[0191] Meanwhile, the distance and relative proportion will change with the tilt of the head. Therefore, after returning the tilt angle data through the API, for images with excessively large angles, the calculated distance is multiplied by the cosine of the tilt angle to correct the vertical distance.

[0192] This application also discloses a developmental assessment method system for implementing the above-described developmental assessment method.

[0193] The implementation principle is as follows:

[0194] First, for specific key points that contribute to important facial angles, instead of generating isotropic circular Gaussian heatmaps, an elliptical Gaussian heatmap region is generated that is directionally adapted to the anatomical structure. The major axis of this ellipse extends along the collinear direction between the two key points (i.e., the direction in which the angle change is least sensitive), while the minor axis shrinks accordingly. This explicitly encodes the prior knowledge that "positioning deviations along a specific axis have little impact on angle calculation" in the model's learning objective.

[0195] Secondly, it is not only based on the response of local point features, but also coordinates and corrects the prediction of key points under the global angular structure constraint through the GAT layer. This allows missing key points caused by head offset, swinging, or hair occlusion to achieve neighbor perception and feature aggregation through graph attention, and enhances the prediction ability of missing key points through overall structured prediction.

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

[0197] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for constructing a surface shape index measurement model, characterized in that, Includes the following steps: Keypoint annotation was performed on profile images to create a standard profile image dataset; Based on several key points, generate the structural relationship between key points corresponding to different standard angles. Specifically, obtain each standard angle to define the angle relationship, obtain the key points required for each standard angle and obtain the labeled coordinates to define the point coordinate relationship. Obtain the participating edge corresponding to the standard angle to define the collinear relationship of the key points located between the same participating edge; The structural relationships are generated based on the aforementioned angular relationships, collinear relationships, and point coordinate relationships. Based on the structural relationship, an optimized Gaussian heatmap containing an elliptical heatmap is generated for each of the key points, wherein the elliptical heatmap has different probability density distributions in the collinear and perpendicular collinear directions between the key points matching the structural relationship; A GAT layer is introduced to treat the key points as nodes and the structural relationships between the key points as edges. Attention weights between each node and its neighboring nodes are calculated to perform global space optimization. Supervised learning is performed on the optimized Gaussian heatmap and the real heatmap based on the MSE loss function and the angle loss function to obtain a trained side profile key point detection model.

2. The method for constructing a surface shape index measurement model according to claim 1, characterized in that, The optimized Gaussian heatmap also includes a circular heatmap. Before generating the optimized Gaussian heatmap corresponding to each key point based on the structural relationship, the following steps are also included: The engagement degree and distance degree of each key point are determined to define each key point as an elliptical constraint point and a circular constraint point, respectively. 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 phase distance is characterized as the angular distance between each key point and its associated angle. If it is greater than a preset distance, the phase distance corresponds to the circular constraint point; if it is not greater than the preset distance, the phase distance corresponds to the elliptical constraint point.

3. The method for constructing a surface shape index measurement model according to claim 2, characterized in that, Before generating the optimized Gaussian heatmaps corresponding to each key point based on the structural relationship, the following steps are also included: Determine the stability of each elliptical constraint point to determine the collinear extension direction; If the stability is not lower than the preset requirement, then the collinear extension direction of the elliptical constraint point is confirmed as the direction toward the corner point in the associated angular relationship; If the stability is lower than the preset requirement, then the collinear extension direction of the elliptical constraint point is determined to be toward other adjacent elliptical constraint points in the associated angular relationship.

4. The method for constructing a surface shape index measurement model according to claim 2 or 3, characterized in that, The process of generating an optimized Gaussian heatmap containing an elliptical heatmap for each key point based on the structural relationship includes the following steps: A circular heat map centered on the circular constraint point is generated based on the real coordinates of the key points and a preset basic diffusion coefficient. Based on the collinear extension directions, the direction vector corresponding to the collinear direction is determined, and the bias angle corresponding to the direction vector is obtained using the four-quadrant arctangent function. Based on the bias angle, the original coordinates in the heat map are aligned to the collinear direction to obtain optimized coordinates. Based on the optimized coordinates and the preset major axis diffusion coefficient and minor axis diffusion coefficient, an elliptical heatmap corresponding to the elliptical constraint point is generated, wherein the major axis diffusion coefficient corresponds to the diffusion capability of the probability density distribution in the collinear direction, and the minor axis diffusion coefficient corresponds to the constraint capability of the probability density distribution in the perpendicular collinear direction.

5. The method for constructing a surface shape index measurement model according to claim 3, characterized in that, When the stability is lower than the preset requirement, the following steps are also included: The probability density distribution between the elliptical constraint point whose stability is lower than the preset requirement and other adjacent elliptical constraint points is lower the closer it is to the midpoint between the two points.

6. The method for constructing a surface shape index measurement model according to claim 4, characterized in that, It also includes the following steps: Obtain specific group identifier information and match corresponding processing tags based on the specific group identifier information; Keypoint optimization coefficients are generated based on the processed labels, the standard angle is adjusted according to the keypoint optimization coefficients, and the major axis diffusion coefficient and the minor axis diffusion coefficient are adjusted based on the specific features corresponding to the processed labels.

7. The method for constructing a surface shape index measurement model according to claim 1, characterized in that, Calculating the attention weights between each node and its neighboring nodes to perform global spatial optimization of each structural relationship includes the following steps: The feature vectors of the node and its neighboring nodes are obtained and concatenated. The relevance scores corresponding to each edge are calculated through linear transformation and activation function. The relevance scores corresponding to the node are normalized to calculate the attention weight. Global space optimization is performed on the structural relationships and nodes with abnormal feature vectors based on the magnitude of the attention weights.

8. A method for assessing developmental status, characterized in that, Based on the method described in any one of claims 1-7, the method includes the following steps: Two-dimensional profile images are acquired and ROI regions containing the profile are detected. The coordinates of several key points of the profile are obtained through the facial shape index measurement model and the development angle is calculated. A two-dimensional frontal face image is obtained and the coordinates of key points on the frontal face and the head posture angle are obtained based on the recognition API. The proportion of facial soft tissue is calculated based on the head posture angle and the coordinates of key points on the frontal face. Developmental assessment results are generated based on the developmental angle and the proportion of facial soft tissue.

9. A developmental assessment system, characterized in that, Used to implement the developmental assessment method as described in claim 8.

Images