A Visualization Method and Device for Facial Four-Region Segmentation Based on 3D Models

Abstract

This invention discloses a method and apparatus for visualizing four-region facial segmentation based on a 3D model, relating to the field of image processing technology. The method includes: calculating the unit normal of each vertex; extracting specular transition rings and generating a ring buffer set; performing weighted statistical analysis on mesh vertices using suppression weights to establish the left-right and top-bottom axes of the 3D face mesh; projecting the vertex set onto the left-right and top-bottom axes, and determining the left-right segmentation boundary and top-bottom segmentation threshold based on the ring buffer set to generate a four-region mask; performing ring attenuation backfilling processing on the scalar values ​​of the vertices to be visualized based on geodesic distance to obtain the visualization scalar values, and calculating the statistical indicators of each region based on the suppression weights and the four-region mask. This invention can significantly improve the stability and accuracy of 3D facial analysis.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and apparatus for visualizing facial four-region division based on a three-dimensional model. Background Technology

[0002] Regional visualization of 3D face models is commonly used in scenarios such as skin assessment, aesthetic analysis, and virtual human rendering and quality inspection. It typically requires dividing the face into regions on a 3D mesh model to map deformation amplitude, error, or scoring heatmaps onto the mesh surface and summarize statistics by region. However, under realistic acquisition conditions, the human face surface has a thin layer of sebum, and the surface normals change rapidly in areas such as the forehead to the bridge of the nose and the bridge of the nose to the tip of the nose. This easily produces significant specular reflection components under grazing incidence conditions, which not only leads to brightness saturation in the texture image but also causes geometric defects in 3D reconstruction geometry such as local holes, steps, or depth bias.

[0003] Existing technical solutions typically establish a unified coordinate framework on a 3D mesh, then generate region masks based on geometric segmentation lines, and directly map scalar data for display and statistics. However, this conventional approach struggles to effectively address the dual interference of texture and geometry caused by specular reflection. Since specular reflection areas often traverse regions near the facial center segmentation line, traditional coordinate axis estimation is easily influenced by geometric noise, leading to deviations. This forces region boundaries to cross high-risk distortion areas, resulting in cross-regional leakage and jagged edges in the visualized heatmap. Furthermore, existing statistical summarization methods often lack effective suppression of anomalous reflection regions, causing localized overexposure of highlights or reconstruction errors to be mistakenly included as valid signals in regional statistics. Consequently, the final regional statistical indicators are dominated by pseudo-anomaly values, severely impacting the accuracy of facial four-region visualization. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies, such as facial four-zone division being susceptible to mirror reflection interference leading to boundary distortion, cross-zone leakage in visualized heat maps, and regional statistical indicators being dominated by pseudo-anomalies. Therefore, this invention proposes a facial four-zone division visualization method and device based on a three-dimensional model.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: A visualization method for facial four-region segmentation based on a 3D model includes: S1. Obtain a 3D face mesh containing a set of vertices, and calculate the unit normal of each vertex in the vertex set; S2. Calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex. Extract specular transition rings based on the degree of transition in the neighborhood of the ring and generate a ring buffer set. S3. Based on the geodesic distance from the vertex to the specular transition ring, construct the suppression weight, and perform weighted statistical analysis on the grid vertices through the suppression weight to establish the left and right axes and the top and bottom axes of the three-dimensional face grid. S4. Project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds, and generate a four-region mask based on the left and right segmentation boundaries and the top and bottom segmentation thresholds. S5. Based on geodesic distance, perform ring attenuation backfilling on the vertex scalar to be visualized to obtain the visualized scalar, and calculate the statistical index of each region according to the suppression weight and the four-region mask.

[0006] Preferably, calculating the specular reflection proxy and texture brightness includes: The incident cosine is calculated by the dot product of the unit normal of the vertex and the incident direction of the light source, where the incident direction of the light source is the direction vector obtained by subtracting the position vector of the corresponding vertex from the position vector of the light source and normalizing it. Set the effective refractive index of the sebum layer, and calculate the normal incident reflection base value based on the effective refractive index of the sebum layer; Based on the normal incident reflection basis value and the incident cosine, the specular reflection surrogate quantity of each vertex is calculated using the Schlick approximation formula. Obtain the texture color data of the vertex, and perform a weighted summation of the red, green, and blue components of the texture color data to obtain the texture brightness.

[0007] Preferably, the specular transition band is extracted based on the degree of transition between specular reflection proxy and texture brightness in a ring neighborhood, including: Construct a ring neighborhood for each vertex; The first transition quantity is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the target vertex and the adjacent vertices in the ring neighborhood of the mirror reflection surrogate quantity. The second transition amount is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the texture brightness and the target vertex in the neighborhood of the ring. The product of the first transition and the second transition is used as the ring score; The median and absolute median difference of the ring zone scores of all grid vertices are statistically calculated, and the ring zone threshold is determined based on the median and the absolute median difference. Vertices with ring scores greater than or equal to the ring threshold are used as constituent vertices of specular transition rings.

[0008] Preferably, generating the annular buffer set includes: Calculate the geodesic distance from each vertex in the vertex set to the specular transition ring; Define a buffer radius, and define the set of vertices whose geodesic distance is less than or equal to the buffer radius as the annular buffer set.

[0009] Preferably, establishing the left and right axes and the top and bottom axes of the three-dimensional face mesh includes: The geodesic distance from the vertex to the specular transition ring is mapped to a suppression weight using an exponential decay function. Using the suppression weight as a coefficient, the coordinates of the vertex set are weighted and averaged to obtain the weighted centroid; Using the suppression weights as coefficients, calculate the weighted covariance matrix of the vertex set relative to the weighted centroid; Perform eigenvalue decomposition on the weighted covariance matrix, and select the unit eigenvectors corresponding to the largest and second largest eigenvalues ​​as the left and right axes and the top and bottom axes, respectively.

[0010] Preferably, the vertex set is projected onto the left and right axes and the top and bottom axes of the 3D face mesh, and the left and right segmentation boundaries and the top and bottom segmentation thresholds are determined by combining the annular buffer set, including: Using the weighted centroid as the base point, construct left and right dividing planes with the normal to the left and right axes, and use the left and right dividing planes as the left and right dividing lines; The optimal upper and lower segmentation thresholds are searched along the projection directions of the upper and lower axes, and an objective function containing a weighted balance term and a cross-penetration penalty term is constructed. The weighted balance term is used to measure the difference in the sum of the suppression weights of the regions on both sides of the upper and lower segmentation thresholds. The cross-penetration penalty term is used to measure the degree to which the segmentation line corresponding to the upper and lower segmentation thresholds crosses the annular buffer set. The value that minimizes the objective function is selected as the upper and lower segmentation threshold.

[0011] Preferably, the vertex scalar to be visualized is subjected to ring attenuation backfilling based on geodesic distance, including: Attenuation coefficients are constructed based on the geodesic distance from the vertex to the high-light transition ring. Within the ring neighborhood of each vertex, the difference between the attenuation coefficients of the vertex and its neighboring vertex is used as a weight to perform a weighted average of the scalar values ​​of the vertices to be visualized in the neighborhood, thus obtaining the non-ring reference value. Based on the attenuation coefficient, linear interpolation is performed between the vertex scalar to be visualized and the non-annular reference value to obtain the visualized scalar.

[0012] Preferably, the statistical indicators of each region are calculated based on the suppression weight and the four-region mask, including: The visualization scalar is divided into four regions based on the four-region mask; Within each region, the visualized scalar is weighted according to the suppression weight to obtain the region mean as a statistical indicator.

[0013] To address the aforementioned problems, the present invention also provides a facial four-region segmentation visualization device based on a three-dimensional model, the device comprising: The mesh normal module is used to obtain a 3D face mesh containing a set of vertices and calculate the unit normal of each vertex in the vertex set. The ring extraction module is used to calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex, extract specular transition rings based on the degree of transition of specular reflection proxy and texture brightness in a ring neighborhood, and generate a ring buffer set. The coordinate construction module is used to construct suppression weights based on the geodesic distance from the vertex to the specular transition ring. The suppression weights are used to perform weighted statistical analysis on the mesh vertices to establish the left and right axes and the top and bottom axes of the 3D face mesh. The four-region segmentation module is used to project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, and combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds. Based on the left and right segmentation boundaries and the top and bottom segmentation thresholds, a four-region mask is generated. The visualization statistics module is used to perform ring attenuation backfilling on the vertex scalars to be visualized based on geodesic distance to obtain the visualization scalars, and calculate the statistical indicators of each region according to the suppression weight and the four-region mask.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention extracts the hyperspectral transition ring and constructs a suppression weight based on geodesic distance. When establishing the coordinate frame and segmentation threshold, it actively reduces the weight and avoids high-risk distortion areas. By calculating the weighted centroid and weighted covariance matrix through the suppression weight, it ensures that the left and right axes and the up and down axes are not pulled by local geometric bias. At the same time, when searching the segmentation threshold, a band-crossing penalty term is introduced, which forces the four-zone segmentation boundary to automatically bypass the hyperspectral transition ring and its buffer set. This ensures the spatial geometric stability of the four facial regions and effectively prevents the region boundaries from becoming jagged or crossing the distortion area due to specular reflection interference.

[0015] 2. This invention utilizes a geodesic distance-driven annular attenuation backfill mechanism and a weighted statistical strategy to achieve the repair of the visualized scalar distribution and the noise-resistant aggregation of district-level indicators. By performing smooth backfilling of vertex scalars near the annular zone based on neighborhood reference values, holes and overexposure noise in the heatmap are eliminated, avoiding visual misleading caused by outliers. At the same time, the application of suppression weights when calculating regional statistical indicators reduces the contribution of pseudo-outlier data to the final results, ensuring that the output district-level statistics can truly reflect the surface state of stable facial regions, significantly improving the accuracy and interpretability of the three-dimensional facial analysis results. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart illustrating a facial four-region division visualization method based on a three-dimensional model, provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a facial four-zone division visualization device based on a three-dimensional model, provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example: This example provides a visualization method for facial four-region division based on a 3D model. See [link to example]. Figure 1 Specifically, including: S1. Obtain a 3D face mesh containing a set of vertices, and calculate the unit normal of each vertex in the vertex set; In an embodiment of the present invention, obtaining a three-dimensional face mesh containing a set of vertices and calculating the unit normal of each vertex in the vertex set includes: Obtain a 3D face mesh containing a set of vertices; Calculate the unit normal of each vertex in the vertex set; Specifically, obtaining a 3D face mesh containing a vertex set refers to acquiring a set of 3D geometric data to describe the shape of the outer surface of a face. The 3D face mesh is composed of a large number of triangular facets connected end to end, which can approximate the continuous surface of real facial skin in a discrete manner. The vertex set is the set of all geometric sampling points in the mesh. Each vertex corresponds to a position coordinate in 3D space and serves as a common connection point for adjacent triangular facets, used to carry local surface shape information and subsequent calculations. Calculating the unit normal of each vertex in the vertex set means that for each vertex, a direction vector representing the overall orientation of the surface near that point is formed based on the orientation and area contribution of multiple triangular facets connected to that vertex, and the length of this direction vector is normalized to one, thus obtaining the unit normal. The direction of the unit normal is used to characterize the outward direction of the facial surface at that vertex, reflecting the degree of inclination of the surface and the geometric relationship of light incidence.

[0019] In detail, when acquiring a 3D face mesh containing a set of vertices, the mesh data output by the 3D face acquisition device is used as input, and the vertex set and triangle set are read. The vertex set consists of several vertices, each containing 3D coordinate components. The triangle set consists of several triangles, each determined by three vertex indices. To ensure the stability of subsequent unit normal calculations, the 3D face mesh is preprocessed. The preprocessing includes removing duplicate vertices by vertex index and merging nearest neighbor vertices whose distance is less than a preset merging threshold. The merging threshold is preferably one ten-thousandth to one thousandth of the face scale to avoid damaging facial details and suppressing acquisition noise. The face scale is preferably the diagonal length of the bounding box of the 3D face mesh. Furthermore, degenerate triangles with an area less than a preset minimum triangle area threshold are removed. The minimum triangle area threshold is preferably one ten-thousandth to one thousandth of the median area of ​​all triangles in the mesh to avoid introducing abnormal normals by noisy triangles.

[0020] After obtaining the preprocessed 3D face mesh, when calculating the unit normal of each vertex in the vertex set, the normal of each triangular facet is first calculated and its area is recorded. Specifically, the coordinates of the three vertices of the triangular facet are taken, two edge vectors are constructed, and their cross product is performed to obtain the facet normal. Then, the facet normal is normalized to obtain the facet unit normal. At the same time, the facet area is obtained by taking half the magnitude of the cross product result. Subsequently, the facet unit normals of all triangular faces adjacent to each vertex are collected, and the facet area is used as the weight to perform a weighted summation to obtain the vertex normal cumulative vector. The basis for using the facet area as the weight is that the facet with a larger area has a greater impact on the local surface. The orientation is more representative and can reduce the influence of small noise patches. Finally, the cumulative normal vector of the vertex is normalized to obtain the unit normal. When the magnitude of the cumulative normal vector of the vertex is less than the preset minimum magnitude threshold, it is determined that there is degenerate geometry in the neighborhood of the vertex and a replacement assignment is performed. The minimum magnitude threshold is preferably one multiplied by ten to the power of negative eight to take into account numerical stability and avoid misjudgment. The unit normal of the vertex is replaced with the normalized result of the arithmetic mean vector of the unit normals of other valid vertices in its ring neighborhood, so as to obtain the unit normal of each vertex and ensure that the unit normal has continuity and stability when used for subsequent specular reflection proxy quantity calculation.

[0021] S2. Calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex. Extract specular transition rings based on the degree of transition in the neighborhood of the ring and generate a ring buffer set. In an embodiment of the present invention, the calculation of specular reflection proxy and texture brightness, extraction of specular transition bands, and generation of band buffer sets include: Calculate the incident cosine based on the dot product of the unit normal at the vertex and the incident direction of the light source; Set the effective refractive index of the sebum layer, and calculate the normal incident reflection base value based on the effective refractive index of the sebum layer; Specifically, the incident cosine is the cosine of the angle between the unit normal of a vertex and the incident direction of the light source at a given vertex in a 3D face mesh. It is obtained by the dot product of two normalized direction vectors and is used to characterize the geometric relationship between the surface orientation and the direction of the incoming light at that vertex. When the incident cosine is close to one, it indicates that the light source is approximately incident along the outward normal of the surface, making it easier to form concentrated highlights and obtain strong direct illumination. When the incident cosine is close to zero, it indicates that the light source is grazing-incident or nearly tangentially incident, making it easier to experience a sharp increase in the reflection component and a more pronounced highlight pattern. The spatial distribution of the extensor-induced specular transition rings indicates that when the incident cosine is negative, it means that the vertex surface is in a self-occluding or indirect light-receiving state when facing away from the light source, and usually does not produce an effective direct specular reflection contribution. The normal incident reflection base value refers to the reference reflection ratio caused by the refractive index mismatch when light is incident on the air-sebum thin layer interface in a direction perpendicular to the surface. It is determined only by the difference between the effective refractive index of the sebum thin layer and the refractive index of air, and is used to characterize the inherent reflectivity of the interface under the simplest incident conditions. The larger the value, the stronger the reflection of incoming light by the interface.

[0022] In detail, when calculating the incident cosine based on the dot product of the unit normal of a vertex and the incident direction of the light source, the position coordinates of the light source and the spatial coordinates of the corresponding vertex are first obtained. The spatial coordinates of the vertex are then subtracted from the position coordinates of the light source to obtain the incident vector pointing from that vertex to the light source. The incident vector is then normalized to obtain the incident direction of the light source. Normalization involves dividing each component of the incident vector by its length to obtain a direction vector of length one. After obtaining the incident direction of the light source, the dot product of the incident direction and the unit normal of the vertex is performed, and the result is taken as the incident cosine. To ensure numerical stability, the incident cosine is limited to the range of -1 to +1. The basis for this limitation is that the theoretical range of the dot product under ideal unit vector conditions is also from -1 to +1, and exceeding this range is usually introduced by floating-point errors. When the incident cosine is negative, it indicates that the vertex is facing away from the light source and is in a self-occluding or indirect lighting state. To avoid introducing unreasonable light enhancement in subsequent reflection surrogate quantity calculations, it is preferable to truncate the negative incident cosine to zero to only represent the contribution of direct incident. The basis for this truncation is that specular reflection is mainly produced by visible direct incident and the back surface does not form an effective specular highlight under this light source configuration. When setting the effective refractive index of the sebum layer, it is used as a material parameter characterizing the ability of the facial sebum layer to produce interfacial reflection of incident light. It is preferably set between 1.44 and 1.48, which is preferred because this range covers the effective refractive index range of common sebum and stratum corneum surface composite media in the visible to near-infrared bands while taking into account individual differences and measurement errors. After obtaining the effective refractive index of the sebum layer, the normal incident reflection base value is calculated based on the effective refractive index. The formula for calculating the normal incident reflection base value is as follows: In the formula, The normal incident reflection base value refers to the reference value for the reflection ratio when light is incident perpendicularly to the surface at the interface between the air and the sebum layer. The effective refractive index of the sebum layer refers to the refractive index of the sebum layer relative to air. In the formula, '1' refers to an approximate value of the air refractive index and is used as the refractive index of the incident medium. The formula is derived from a simplified form of the Fresnel reflection law of electromagnetic waves at the interface of two non-absorbing media under normal incidence conditions. The physical reason is that the mismatch of refractive indices on both sides of the interface leads to discontinuity in wave impedance, resulting in reflection. The reflection amplitude is related to the difference in refractive indices and is expressed as the square of the amplitude ratio in an intensity sense. Therefore, the formula can be used to obtain a normal incidence reflection basis value that is directly related to the material properties of the sebum layer and can be explained, providing a stable benchmark for constructing specular reflection surrogate quantities based on the incident cosine.

[0023] Based on the normal incident reflection basis value and the incident cosine, the specular reflection surrogate quantity of each vertex is calculated using the Schlick approximation formula. Obtain the texture color data of the vertex, and perform a weighted summation of the red, green, and blue components of the texture color data to obtain the texture brightness; Specifically, the specular reflection surrogate quantity refers to a dimensionless quantity at each vertex of a 3D face mesh that characterizes the intensity of specular reflection produced by the sebum layer interface on incident light. This quantity is determined by the normal incident reflection basis value and the incident cosine. The normal incident reflection basis value reflects the reference reflection capability of air and the sebum layer under perpendicular incidence conditions due to the refractive index mismatch. The incident cosine reflects the geometric relationship between the surface orientation at the vertex and the incident direction of the light source. The specular reflection surrogate quantity increases rapidly as the incident angle gradually changes from near normal incidence to grazing incidence. It is used to express the reason why specular highlights are more likely to change abruptly and form specular transition rings in regions where the normal changes rapidly. It is not directly equivalent to the actual reflectance measurement value, but rather approximates the law of specular reflection changing with the incident angle with low computational complexity, thereby providing an interpretable and stably calculable reflection intensity index in the process of 3D visualization and region partitioning.

[0024] In detail, when calculating the specular surrogate quantity for each vertex based on the normal incident reflection basis value and the incident cosine, the calculated incident cosine and the corresponding normal incident reflection basis value are first obtained for each vertex. The incident cosine is limited to the range of zero to one to retain only the incident geometry contribution when directly illuminated and to avoid introducing unreasonable specular enhancement at back points. Then, the Schlick approximation formula is used to calculate the specular surrogate quantity. The formula for the specular surrogate quantity is as follows: In the formula, Let be the specular reflection surrogate quantity of the i-th vertex, which is a dimensionless quantity used to characterize the trend of specular reflection intensity at the sebum layer interface at the vertex as a function of the incident angle. The normal incident reflection base value, Let be the incident cosine of the i-th vertex. The fifth power in the formula is used to approximate the angle law of rapid rise of specular reflection under grazing incidence conditions and to approximate the shape of the Fresnel reflection curve with lower computational complexity. This makes the specular reflection surrogate quantity approach the normal incidence reflection basis value when the incident cosine is close to one and gradually approach one when the incident cosine decreases to close to zero, satisfying the physical trend that the closer the incident angle is to grazing incidence, the stronger the specular reflection.

[0025] When acquiring vertex texture color data and calculating texture brightness, the texture color data corresponding to each vertex is read from the texture mapping result of the 3D face mesh. The texture color data contains at least red, green, and blue components and is represented by a zero-to-one normalization to ensure consistent scale across different acquisition devices. If the original texture color data is an integer grayscale code, it is preferable to divide it by the maximum grayscale value to achieve normalization. Subsequently, the red, green, and blue components are weighted and summed to obtain the texture brightness. The preferred weights are 0.2126, 0.7152, and 0.0722, respectively. The rationale for this preference is that this set of weights can approximately reflect the sensitivity of the human eye to the brightness contribution of different wavelengths and make the brightness more sensitive to the green component, thereby more stably characterizing the salience of the highlight area. This results in a more obvious transition in texture brightness at the highlight transition ring, providing a stable input for subsequent ring extraction.

[0026] Construct a ring neighborhood for each vertex; The first transition quantity is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the mirror reflection surrogate quantity and the vertex in the neighborhood of the ring; The second transition amount is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the texture brightness and the adjacent vertices in the ring neighborhood. Specifically, a ring neighborhood refers to the smallest first-order adjacency range defined for a given vertex in a 3D face mesh. This range consists of all adjacent vertices directly connected to that vertex via mesh edges. Essentially, it reflects the most direct local surface connectivity around that vertex, used to measure local changes near that vertex on a discrete mesh without introducing long-distance effects across multiple topologies. The first transition quantity refers to the difference between the specular surrogate quantity of each adjacent vertex and the specular surrogate quantity of that vertex within its ring neighborhood, taking the absolute value, and then calculating the arithmetic mean of these absolute differences. The first value is used to characterize the intensity of local abrupt changes in the specular surrogate quantity around the vertex. The larger the value, the more drastic the change in the surrogate quantity at the vertex is relative to the neighborhood, and the more likely it is near the specular transition ring. The second transition quantity is the value obtained by taking the arithmetic mean of the difference between the texture brightness of each adjacent vertex and the texture brightness of the vertex in the same way within a ring neighborhood of the vertex. It is used to characterize the intensity of local abrupt changes in texture brightness around the vertex. The larger the value, the more discontinuous the brightness distribution near the vertex is, and the more likely it corresponds to a specular boundary or saturation region.

[0027] In detail, when constructing a ring neighborhood for each vertex, a vertex adjacency table is established based on the set of triangular facets of the 3D face mesh. Specifically, each triangular facet is traversed and its three vertex indices are read. Any two vertices within the same triangular facet are registered as adjacent vertices, thus forming an adjacency list for each vertex consisting of adjacent vertices that directly share edges with it. This adjacency list is defined as the ring neighborhood of that vertex. To ensure the stability and comparability of the ring neighborhood, it is preferable to perform deduplication processing on the adjacency list to avoid the same adjacent vertex being counted repeatedly due to repeated registration by multiple facets. Adjacent vertices whose spatial distance from the vertex is greater than the upper bound of the preset neighborhood distance are removed to suppress abnormal adjacencies introduced by local mesh breaks or incorrect edge connections. The upper bound of the neighborhood distance is preferably three times the median of the mesh edge length. The rationale for this preference is that the median edge length can reflect the typical sampling density of the mesh, while a three-fold threshold can accommodate normal geometric changes while filtering out obvious abnormal edges.

[0028] After constructing the ring neighborhood of each vertex, the specular surrogate quantity of the target vertex is read, and the specular surrogate quantities of each adjacent vertex in its ring neighborhood are read one by one. The absolute value of the difference between the specular surrogate quantity of the adjacent vertex and the specular surrogate quantity of the target vertex is calculated and accumulated. The accumulated result is then divided by the number of adjacent vertices in the ring neighborhood to obtain the first transition quantity. If the number of adjacent vertices in the ring neighborhood of a vertex is less than the preset minimum neighborhood number, the vertex is determined to be in the mesh boundary or a locally missing region and a replacement process is performed. The minimum neighborhood number is preferably three to match the minimum closed connection condition of the triangular mesh and to avoid the distortion of the transition quantity caused by extremely sparse neighborhoods. The replacement process is preferably set to the median of the first transition quantity of the vertex that has been calculated in its ring neighborhood to maintain robustness.

[0029] When calculating the second transition value, the same neighborhood and traversal method as the first transition value is used. The texture brightness of the target vertex is read and the texture brightness of each adjacent vertex in its ring neighborhood is read one by one. The absolute value of the difference between the texture brightness of the adjacent vertex and the texture brightness of the target vertex is calculated and accumulated. Then, the accumulated result is divided by the number of adjacent vertices in the ring neighborhood to obtain the second transition value. Thus, the first transition value represents the intensity of the change of the specular reflection surrogate quantity on the local mesh, and the second transition value represents the intensity of the change of texture brightness on the local mesh. The two together serve as the basic input for subsequent specular transition ring extraction.

[0030] The product of the first transition and the second transition is used as the ring score; The median and absolute median difference of the ring zone scores of all grid vertices are statistically calculated, and the ring zone threshold is determined based on the median and the absolute median difference. Vertices with ring scores greater than or equal to the ring threshold are taken as constituent vertices of the specular transition rings; It should be noted that the specular transition ring refers to a set of vertices that are continuously distributed in a band on the surface of a 3D face mesh. The mesh positions corresponding to this set often extend spatially along the transition areas of curved surfaces where the normal changes rapidly, such as from the forehead to the bridge of the nose or from the bridge of the nose to the tip of the nose. The reason for its formation is that the surface orientation of these areas changes rapidly with position, causing the angle between them and the incident direction of the light source to cross the critical range of significantly enhanced specular reflection within a short distance. This results in a sudden increase in specular reflection at the interface of the thin sebum layer, which manifests as a steep change in the specular boundary in terms of texture brightness. At the same time, it is more likely to be accompanied by anomalies such as local saturation, holes, or steps in 3D reconstruction or texture sampling. Therefore, in computation, this ring is represented by a band-shaped region in which both the specular reflection surrogate quantity and texture brightness show significant jumps within the neighborhood of the ring and is selected by the ring score. It is used as the result of locating high-risk distortion sources in subsequent four-zone boundary avoidance, visualization attenuation backfilling, and zone-level statistical weighting.

[0031] In detail, the calculated first and second transition values ​​are read for each vertex, and the two are directly multiplied to obtain the ring score of the vertex. The basis for using the product is that only when the specular reflection surrogate quantity changes abruptly locally and the texture brightness changes abruptly locally at the same time, the vertex is more in line with the physical behavior of the specular transition ring. If only one of the changes occurs, it is more likely to correspond to geometric noise or texture noise and not form a stable ring.

[0032] The ring zone scores of all vertices are collected to form a score sequence. These scores are then sorted by value, and the median value is taken as the median ring zone score. For each vertex, the absolute value of the difference between its ring zone score and the median is calculated to form a deviation sequence. This deviation sequence is then sorted by value, and the median value is taken as the absolute median difference. The median and absolute median difference are used because they are insensitive to a very small number of abnormally high scores and can stably depict overall background fluctuations even when there are local reconstruction holes or overexposure leading to extreme scores. The ring zone threshold is determined based on the median and the absolute median difference. The formula for calculating the ring zone threshold is as follows: In the formula, The annular threshold is a threshold used to distinguish between annular and non-annular vertices of specular transitions. The median, The absolute median difference The set coefficient is an amplification factor used to control the strictness of the threshold, and is preferably three. The reason is that when the ring score background is approximately concentrated, this coefficient can place the threshold above the background fluctuation and retain significant abrupt change points, while avoiding misjudging large areas of normal skin as rings. After obtaining the ring threshold, the ring score of each vertex is compared with the ring threshold. If the ring score of the vertex is greater than or equal to the ring threshold, the vertex is marked as a constituent vertex of the specular transition ring. Finally, all the marked vertices form the vertex set of the specular transition ring.

[0033] Calculate the geodesic distance from each vertex in the vertex set to the specular transition ring; Define a buffer radius and define the set of vertices whose geodesic distance is less than or equal to the buffer radius as the annular buffer set; Specifically, geodesic distance refers to the shortest path length required to reach another vertex or a set of vertices on the surface of a 3D face mesh by walking segment by segment along the connected edges of the mesh. It is obtained by accumulating the shortest path length using the length of the mesh edge as the local walking cost, thus reflecting the true proximity relationship propagating along the facial skin surface. It can avoid false neighbors caused by crossing nostril recesses, labial grooves, or geometric crossings when using spatial straight-line distance, thus better conforming to the diffusion mode of specular transition effects on the surface. The ring buffer set refers to a set of neighboring vertices obtained by expanding outward on the mesh surface according to geodesic distance with the specular transition ring as the center. Specifically, it is a set of all vertices whose geodesic distance to the specular transition ring is less than or equal to the preset buffer radius. This set is used to characterize the range of influence of the specular transition ring on the surrounding area that may cause reconstruction bias, texture saturation, or visualization cross-regional contamination. This allows the selection of subsequent four-region segmentation thresholds to actively avoid this high-risk range and allows visualization attenuation backfill to smoothly transition within this range to reduce the interference of false anomalies on the presentation and statistics of the four regions.

[0034] In detail, when calculating the geodesic distance from each vertex in the vertex set to the spectrophotometric transition ring, the 3D face mesh is first converted into a weighted graph structure with vertices as nodes. Specifically, an undirected connection is established for each edge in the mesh, and the Euclidean distance between the two vertices of the edge is used as the edge weight to represent the local walking cost along the mesh surface. Then, the vertices constituting the spectrophotometric transition ring are used as a multi-source starting point set, and shortest path propagation is performed on the weighted graph to obtain the minimum cumulative edge weight from any vertex along the mesh surface to the starting point set. Preferably, the multi-source Dijkstra algorithm is used for this purpose. The initial distances of all ring vertices are set to zero and enqueued simultaneously. The initial distances of non-ring vertices are set to sufficiently large values. In each iteration, the vertex with the smallest current distance is selected and its distance is used to relax the distances of its adjacent vertices by adding the weights of its adjacent edges. This process continues until the minimum distance of all vertices is determined, thereby obtaining the geodesic distance from each vertex to the specular transition ring. The advantage of the geodesic distance over the direct spatial straight-line distance is that it propagates along the surface of the facial mesh and can truly reflect the proximity of the ring's influence on the skin surface without being misled by the short-distance straight-line distance caused by nostril indentation or geometric crossing.

[0035] When setting the buffer radius, the buffer radius is used as a scale parameter to define the influence range of the ring band. It is preferably taken as 0.5% to 1.5% of the face scale, where the face scale is preferably taken as an estimate of the length of the grid bounding box diagonal or the width of the cheekbone. The reason for this preference is that this range can cover the diffusion effect of local reconstruction bias and texture saturation caused by the specular transition ring band on adjacent areas, while avoiding the inclusion of large areas of normal areas into the buffer, which would cause excessive conservatism. When defining the set of vertices whose geodesic distance is less than or equal to the buffer radius as the ring band buffer set, all vertices are traversed and their geodesic distances are compared with the buffer radius. The vertices that meet the condition of being less than or equal to the buffer radius are added to the buffer vertex set and the set is output, thereby obtaining the ring band buffer set.

[0036] S3. Based on the geodesic distance from the vertex to the specular transition ring, construct the suppression weight, and perform weighted statistical analysis on the grid vertices through the suppression weight to establish the left and right axes and the top and bottom axes of the three-dimensional face grid. In an embodiment of the present invention, establishing the left-right axis and the top-bottom axis of the three-dimensional face mesh includes: The geodesic distance from the vertex to the specular transition ring is mapped to a suppression weight using an exponential decay function. Using the suppression weight as a coefficient, the coordinates of the vertex set are weighted and averaged to obtain the weighted centroid; It should be noted that the suppression weight is a dimensionless coefficient set for each vertex in the 3D face mesh. This coefficient is obtained by mapping the geodesic distance from the vertex to the specular transition ring through an exponential decay relationship. It is used to express that the closer a vertex is to the specular transition ring, the more likely it is to be affected by specular reflection transitions and reconstruction biases, and its contribution should be reduced in the global statistical analysis. Conversely, the farther a vertex is from the specular transition ring, the more likely it is to be in a stable diffuse reflection and geometrically reliable region, and its contribution should be retained. Therefore, the suppression weight essentially transforms the spatial influence range of the ring into a statistical calculation. The participation level adjustment makes the subsequent establishment of coordinate axes and segmentation thresholds less affected by high-risk distortion sources; the weighted centroid refers to the center point obtained by weighting the vertex coordinates by using the suppression weight as the contribution coefficient of each vertex when calculating the overall center position of the 3D face mesh, instead of using the equal weighted average of the coordinates of each vertex. This center point is more spatially biased away from the stable region of the specular transition ring, thus weakening the offset effect of abnormal geometry near the ring on the overall center, and thus serving as a more robust base point for the subsequent establishment of the left and right axes and the upper and lower axes, as well as the construction of the left and right segmentation boundaries.

[0037] In detail, the geodesic distance from each vertex in the 3D face mesh to the specular transition ring is obtained as the distance input, and a decay scale parameter is set for the exponential decay function. The decay scale parameter is preferably taken as the buffer radius to ensure that the decay range of the weights is consistent with the spatial scale of the ring buffer set. The buffer radius is used to characterize the influence range of the specular transition ring on the surrounding area, and using the same scale allows the suppression weights to decrease significantly near the ring and enter a smooth transition near the buffer boundary, thus avoiding abrupt weight changes. For each vertex, the suppression weight is calculated based on its geodesic distance. The formula for calculating the suppression weight is: In the formula, Let be the suppression weight of the i-th vertex. Let be the geodesic distance of the i-th vertex. The decay scale parameter controls how quickly the weights decay with distance. The squared term in the formula makes the weights of vertices closer to the ring decrease faster and makes the weights of vertices farther from the ring closer to one. The exponential form provides continuous and monotonic decay and avoids discontinuities caused by segmented thresholds, so that the suppression weights can stably reflect the physical intuition that closer vertices should be suppressed. After obtaining the suppression weights of each vertex, the three-dimensional coordinates of each vertex are read and its three coordinate components are multiplied by the corresponding suppression weight to obtain the weighted coordinate components. The weighted coordinate components of all vertices are then summed and divided by the sum of the suppression weights to obtain the weighted centroid. This makes the weighted centroid closer to the stable region far from the specular transition ring, reducing the pull of geometric bias or texture distortion near the ring on the centroid position, and providing a more robust global base point for the subsequent establishment of the left and right axes and the top and bottom axes.

[0038] Using the suppression weights as coefficients, calculate the weighted covariance matrix of the vertex set relative to the weighted centroid; Perform eigenvalue decomposition on the weighted covariance matrix, and select the unit eigenvectors corresponding to the largest and second largest eigenvalues ​​as the left and right axes and the top and bottom axes, respectively. Specifically, the left-right axis and top-bottom axis of the 3D face mesh refer to two mutually orthogonal main directional axes established for the mesh in the spatial coordinates of the 3D face mesh. They serve as a unified geometric reference for the division of the four facial regions. The left-right axis represents the main extension direction of the face in the horizontal direction, and its direction is most consistent with the lateral extension between the cheekbones. It can stably express the positional differences of facial surface points in the left-right direction as the same scalar projection. The top-bottom axis represents the secondary extension direction of the face in the vertical direction, and its direction is most consistent with the longitudinal extension from the forehead to the chin. It can stably express the positional differences of facial surface points in the top-bottom direction as the same scalar projection. The left-right axis and top-bottom axis are obtained by weighted statistics on the discrete distribution of vertex coordinates relative to the weighted centroid, so that it reflects more the stable region morphology away from the specular transition ring, thereby reducing the pulling effect of geometric offset near the ring on the axial direction.

[0039] In detail, the three-dimensional coordinates of each vertex are obtained, along with the calculated weighted centroid. The weighted centroid is then subtracted from the three-dimensional coordinates of each vertex to obtain a relative coordinate vector. The three components of the relative coordinate vector represent the degree of deviation of the vertex from the overall center in three spatial directions. Subsequently, the suppression weight of each vertex is read and used as a coefficient contributing to the statistical dispersion. The outer product of the relative coordinate vector and its transpose is calculated to obtain a 3x3 matrix, which is then multiplied by the suppression weight. The weighted outer product matrices of all vertices are then summed and divided by the sum of the suppression weights to obtain the weighted covariance matrix. This weighted covariance matrix better reflects the overall shape extension direction of the stable region far from the specular transition ring and reduces the disturbance of the local reconstruction bias near the ring to the estimation of the principal direction. In the specific implementation, the weighted covariance matrix can be calculated using the following formula: In the formula, The weighted covariance matrix is ​​used to describe the spatial discrete distribution of the 3D face mesh in a weighted sense. The suppression weight for the i-th vertex is used to adjust the contribution level of each vertex. Let be the coordinates of the i-th vertex. As the weighted centroid, the transpose operation is used to construct the outer product to quantify the common changes in different directions. When performing eigenvalue decomposition on the weighted covariance matrix, numerical linear algebra is used to solve for the eigenvalues ​​and eigenvectors of the weighted covariance matrix. The obtained eigenvalues ​​are sorted from largest to smallest. The unit eigenvector corresponding to the largest eigenvalue is used as the left and right axes, and the unit eigenvector corresponding to the second largest eigenvalue is used as the top and bottom axes. The selection is based on the fact that the direction corresponding to the largest eigenvalue indicates that the weighted variance of the face mesh is the largest in that direction, that is, the shape extension is the most significant. It is usually consistent with the left and right expansion direction. The direction corresponding to the second largest eigenvalue indicates the second most significant extension direction, which is usually consistent with the top and bottom expansion direction. To ensure that the axis vectors are unit vectors, it is preferable to normalize the selected eigenvectors and adjust their length to one, thereby obtaining the left and right axes and top and bottom axes of the 3D face mesh, which are used for subsequent vertex projection, determination of left and right segmentation boundaries, and search for the optimal top and bottom segmentation thresholds.

[0040] S4. Project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds, and generate a four-region mask based on the left and right segmentation boundaries and the top and bottom segmentation thresholds. In an embodiment of the present invention, determining the left and right dividing lines and the upper and lower dividing thresholds, and generating a four-region mask based on the left and right dividing lines and the upper and lower dividing thresholds includes: Using the weighted centroid as the base point, construct left and right dividing planes with the normal to the left and right axes, and use the left and right dividing planes as the left and right dividing lines; Search for the optimal upper and lower segmentation thresholds along the projection directions of the upper and lower axes, and construct an objective function that includes a weighted balance term and a belt-crossing penalty term; The value that minimizes the objective function is selected as the upper and lower segmentation threshold; Specifically, the left and right segmentation boundary refers to a spatial segmentation boundary on the 3D face mesh used to distinguish the left and right regions. It is implemented as a left-right segmentation plane passing through the weighted centroid and with its normal vector aligned with the left and right axes. Any vertex in the mesh can be determined to be on either side of this segmentation plane by projecting its coordinates relative to the weighted centroid's displacement vector along the left and right axes. This stably divides the entire face mesh into left and right parts, serving as the common boundary between the upper left and lower left regions, and the upper right and lower right regions in the subsequent four-region mask. The upper and lower segmentation threshold refers to the threshold in the upper and lower axis projection coordinates. A scalar threshold is selected to distinguish between the upper and lower regions. Specifically, the displacement vector of each vertex coordinate relative to the weighted centroid is projected onto the upper and lower axes to obtain the upper and lower projection values. If the upper and lower projection values ​​are greater than or equal to the threshold, the vertex is determined to be located in the upper region. If the upper and lower projection values ​​are less than the threshold, the vertex is determined to be located in the lower region. The threshold is determined by simultaneously considering the coverage balance of the upper and lower sides under the meaning of suppression weight and the avoidance requirement of the dividing line away from the ring buffer set. This ensures that the upper and lower segmentation is neither overly biased to one side nor too biased to the other side. It also reduces the risk of cross-regional contamination near the high-light transition ring.

[0041] In detail, the established left and right axis direction vectors are obtained, and the weighted centroid is taken as the point through which the plane passes. The left and right axis direction vectors are taken as the normal of the plane, and the weighted centroid is taken as a point on the plane, thus obtaining the left and right dividing plane. The left and right dividing plane divides the projection of the 3D face mesh in the left and right axis directions into positive and negative sides and serves as the left and right dividing line. In the specific implementation, for any vertex, the coordinates are subtracted from the weighted centroid and the dot product is taken with the left and right axes to obtain the left and right projection values. The left and right projection values ​​are determined to be less than zero and are determined to be the left region, while the left and right projection values ​​are greater than or equal to zero and are determined to be the right region.

[0042] When searching for the optimal upper and lower segmentation thresholds along the projection directions of the upper and lower axes, the upper and lower projection values ​​of each vertex are first calculated. This is done by subtracting the weighted centroid from the vertex coordinates and taking the dot product with the upper and lower axes. The suppression weights and annular buffer sets corresponding to each vertex are then obtained. Based on this, the search interval for the upper and lower segmentation thresholds is set to the minimum to maximum value of the upper and lower projection values. Preferably, this interval is divided into 100 to 300 candidate threshold points at equal intervals to balance search accuracy and computational complexity. The preferred number of candidate threshold points is 200, based on the fact that the number of vertices in a face mesh is usually large and the upper and lower projection values ​​are continuously distributed; 200 points can obtain a stable threshold without introducing a complex optimizer. For each candidate threshold point, an objective function is calculated, and the smallest one is selected as the optimal upper and lower segmentation threshold. The objective function consists of a weighted balance term and a cross-band penalty term, and its calculation formula is as follows: In the formula, The candidate upper and lower segmentation thresholds, This is the objective function value corresponding to the threshold. This represents the set of vertices whose vertical projection values ​​are greater than or equal to a threshold. This represents the set of vertices whose vertical projection values ​​are less than a threshold. For vertex suppression weights, For ring-band buffer sets, These are the vertical projection values ​​of the vertex. The weighting coefficient for the penalty of wearing belts. The parameter is the band-penetration sensitive width. The weighted balance term uses the absolute value of the difference between the sum of the suppression weights on the upper and lower sides to make the upper and lower sides as balanced as possible under the meaning of the suppression weights after the upper and lower segments, so as to avoid the segmentation line being pulled by a few abnormal points or local defects, resulting in the upper or lower side covering too small an area. The band-penetration penalty term exponentially scores the proximity of the upper and lower projection values ​​of the vertices in the ring buffer set to the threshold. It is used to quantify the band-penetration degree when the segmentation line is close to the ring buffer set and to make the threshold move away from the ring buffer set to reduce the risk of cross-regional contamination caused by the high-light transition ring. The parameter is preferably set to take the band-penetration sensitive width parameter as one-third of the buffer radius in the upper and lower axis projection scale so that the penalty mainly acts on the buffer points near the segmentation line. The band-penetration penalty weight coefficient is preferably taken as 0.5 to 3 to achieve a stable compromise between balancing the segmentation area and avoiding the ring. Finally, all candidate threshold points are traversed and their objective function values ​​are compared. The threshold that minimizes the objective function is selected as the optimal upper and lower segmentation threshold.

[0043] A four-region mask is generated based on the left and right segmentation boundaries and the upper and lower segmentation thresholds; In detail, when generating a four-region mask based on the left and right segmentation boundaries and the upper and lower segmentation thresholds, the left and right segmentation planes corresponding to the left and right segmentation boundaries are first obtained, and the upper and lower segmentation thresholds are also obtained. Simultaneously, the weighted centroid, left and right axes, and upper and lower axes of the 3D face mesh are obtained for projection discrimination of vertices. For each vertex in the vertex set, the displacement vector of the vertex relative to the weighted centroid is calculated, and the displacement vector is projected onto the left and right axes and the upper and lower axes respectively to obtain left and right projection values ​​and upper and lower projection values. The sign of the left and right projection values ​​determines which side of the left and right segmentation plane the vertex is located on. A left and right projection value less than zero is determined to be a left vertex, while a left and right projection value greater than or equal to... A vertex is determined to be on the right side based on a zero value. The vertex is then positioned in the upper or lower region based on the relationship between its vertical projection value and the vertical segmentation threshold. A vertex with a vertical projection value greater than or equal to the threshold is considered an upper vertex, while a vertex with a vertical projection value less than the threshold is considered a lower vertex. After determining the left and right sides and the upper and lower sides for each vertex, vertices that simultaneously satisfy both the left and upper side criteria are added to the first mask set to form the upper left region mask. Vertices that simultaneously satisfy both the right and upper side criteria are added to the second mask set to form the upper right region mask. Vertices that simultaneously satisfy both the left and lower side criteria are added to the third mask set to form the upper left region mask. The lower region mask adds vertices that simultaneously satisfy both the right-side and lower-side criteria to the fourth mask set to form the lower-right region mask, resulting in a four-region mask composed of four vertex sets. To enable the four-region mask to be directly used for region assignment of triangular faces, it is preferable to generate face masks synchronously at the face level. Specifically, for each triangular face, the region assignment of its three vertices is read. If all three vertices belong to the same region, the triangular face is assigned to the face mask of that region. If the triangular face crosses the left-right dividing line or the up-down dividing threshold, resulting in the three vertices not all belonging to the same region, the majority vote result of the region assignment of the three vertices of the triangular face is preferred. The region assignment of this patch is to avoid unassigned patches and ensure the continuity of region boundaries during visualization rendering. Furthermore, to avoid mask jitter caused by numerical errors near the segmentation boundary, a tolerance band for projection determination is preferably set, and vertices whose absolute values ​​of left and right projection values ​​are less than the tolerance band and vertices whose absolute values ​​of the difference between the upper and lower projection values ​​and the upper and lower segmentation thresholds are less than the tolerance band are marked as boundary vertices. The tolerance band is preferably one-tenth of the median of the grid side length in the corresponding axial direction projection scale, so as to ensure that the boundary vertices only appear in a very narrow range and do not change the overall four-zone division structure, thereby obtaining a four-zone mask that can be stably used for subsequent ring attenuation backfilling and zone-level statistical calculations.

[0044] S5. Based on geodesic distance, perform ring attenuation backfilling on the vertex scalar to be visualized to obtain the visualized scalar, and calculate the statistical index of each region according to the suppression weight and the four-region mask. In an embodiment of the present invention, annular attenuation backfilling is performed on the vertex scalar to be visualized based on geodesic distance, and statistical indices for each region are calculated according to suppression weights and a four-region mask, including: Attenuation coefficients are constructed based on the geodesic distance from the vertex to the high-light transition ring. Within the ring neighborhood of each vertex, the difference between the value of the value and the attenuation coefficient is used as a weight to perform a weighted average of the scalar values ​​of the vertices to be visualized within the neighborhood, thus obtaining the non-ring reference value. Based on the attenuation coefficient, linear interpolation is performed between the vertex scalar to be visualized and the non-circular reference value to obtain the visualized scalar; Specifically, the attenuation coefficient refers to a dimensionless weighting factor set for each vertex in the 3D face mesh. This factor is obtained by mapping the geodesic distance from the vertex to the specular transition ring through an exponential attenuation relationship. It is used to characterize the probability and intensity of the influence of the specular transition ring on the vertex. The smaller the geodesic distance, the closer the vertex is to the specular transition ring, and the larger the attenuation coefficient. Therefore, in subsequent backfilling processing, it tends to weaken the original scalar value of the vertex to be visualized and introduce a reference value. The larger the geodesic distance, the farther the vertex is from the ring, and the smaller the attenuation coefficient. Therefore, it tends to retain the original scalar value of the vertex to be visualized and avoid unnecessary smoothing of stable regions. The non-ring reference value refers to the reference scalar value calculated for a vertex in its ring neighborhood. This reference scalar value is obtained by... The scalar values ​​of the vertex to be visualized within the neighborhood are obtained by weighted averaging. The weights used are one minus the attenuation coefficient of the adjacent vertices to suppress the contribution of neighborhood points near the ring zone and highlight stable neighborhood points far from the ring zone. Thus, the non-ring zone reference value represents the neighborhood baseline level of the vertex when it is not disturbed by the ring zone anomaly and is used to backfill out anomalies near the ring zone. The visualization scalar refers to the vertex scalar output obtained by linearly interpolating the original vertex to be visualized scalar value and the non-ring zone reference value according to the attenuation coefficient. This output is more biased towards the reference value near the ring zone to reduce false anomalies introduced by specular saturation, holes or steps, and closer to the original vertex to be visualized scalar value far from the ring zone to maintain the true distribution and detail continuity. Finally, it is used as the input data for heatmap rendering and calculation of four-zone statistical indicators.

[0045] In detail, when constructing the attenuation coefficient based on the geodesic distance from the vertex to the specular transition ring, the geodesic distance from each vertex in the 3D face mesh to the specular transition ring is first obtained as the distance input. An attenuation scale parameter is then set to control the attenuation range. The preferred attenuation scale parameter is the buffer radius, ensuring that the attenuation range matches the spatial scale of the ring buffer set. This is because the buffer radius characterizes the influence range of the specular transition ring on the surrounding area, and using the same scale allows the attenuation to significantly increase near the ring and rapidly decrease away from it, thus avoiding unnecessary smoothing of the stable region. Subsequently, the attenuation coefficient is calculated for each vertex. The formula for calculating the attenuation coefficient is: Wherein, the attenuation coefficient is a dimensionless quantity used at the apex to control the backfill intensity. This represents the attenuation coefficient of the i-th vertex. Let represent the geodesic distance from the i-th vertex to the specular transition ring. The attenuation scale parameter is represented by the formula which uses exponential attenuation because a smaller geodesic distance indicates that the vertex is closer to the hyperspectral transition ring and is more likely to be affected by hyperspectral saturation or reconstruction bias. Therefore, a stronger backfill weight is required. As the geodesic distance increases, the vertex is more likely to be in a stable region, so the backfill weight should be reduced rapidly to maintain the true distribution of the original scalar. The squared term is used to make the attenuation change closer to the ring more sensitive and the attenuation change farther from the ring smoother, thus forming a continuous and monotonic spatial transition.

[0046] When calculating the non-annular reference value within the annular neighborhood of each vertex, an annular neighborhood is constructed for the target vertex, and the attenuation coefficients of each adjacent vertex in the neighborhood and the scalar of the vertex to be visualized are read. The attenuation coefficient of the adjacent vertex is subtracted from the value of the value of the adjacent vertex, and the scalar of the vertex to be visualized of each adjacent vertex in the neighborhood is weighted and averaged to obtain the non-annular reference value. The reason for using the attenuation coefficient as the weight is that the larger the attenuation coefficient, the closer it is to the annular zone and the less it should be used as a reference source. Therefore, subtracting the attenuation coefficient can naturally suppress the contribution of the neighborhood points near the annular zone to the reference value and make the reference value mainly determined by the neighborhood points far away from the annular zone. After obtaining the non-annular reference value, the attenuation coefficient is used as the interpolation weight for each vertex. The visualization scalar with a larger attenuation coefficient is more biased towards the non-annular reference value to achieve the backfilling of anomalies near the annular zone. The visualization scalar with a smaller attenuation coefficient is more biased towards the original scalar of the vertex to be visualized to maintain the details of the stable region. Thus, a visualization scalar with a continuous transition near the annular zone and less prone to cross-regional contamination is obtained.

[0047] The visualization scalar is divided into four regions based on the four-region mask; Within each region, the visual scalar is weighted according to the suppression weight to obtain the region mean as a statistical indicator. Specifically, the statistical index refers to the regional summary quantity calculated after the four-region mask division is completed and the visual scalar of each vertex is obtained, in order to summarize the scalar distribution level of each region at the regional scale. This summary quantity is obtained by weighted averaging of all the visual scalars of vertices belonging to the same region. The suppression weight is used as the contribution coefficient of each vertex to reduce the influence of pseudo-anomalies caused by saturation, holes or steps near the specular transition ring, so that the statistical index can better represent the typical numerical level on the stable surface of the region and facilitate the comparability analysis between the four regions. The statistical index is preferably output in the form of region mean, but it can also be extended to region variance, discriminant digits or over-threshold percentage under the same weight system to describe the degree of dispersion and extremes in the region. However, their common purpose is to compress the vertex-level visual scalar into an interpretable regional-level numerical expression and output it synchronously with the four-region facial visualization map to support evaluation and comparison.

[0048] In detail, the process involves obtaining the vertex sets corresponding to the top-left, top-right, bottom-left, and bottom-right regions in the four-region mask, and acquiring the visualization scalar and suppression weight for each vertex. The vertex index is used as the association key to assign the visualization scalar to the vertex set of its respective region. Specifically, all vertices are traversed, and the region affiliation of each vertex in the four-region mask is queried. The visualization scalar of that vertex is written into the scalar list of the corresponding region, thus obtaining the visualization scalar sets for each of the four regions. Within each region, a weighted average of the visualization scalars based on the suppression weights is calculated to obtain the region mean. For any target region, all vertices within that region are traversed, and the suppression weight and visualization scalar of each vertex are read. The suppression weight and visualization scalar are multiplied, and the products are summed to obtain a weighted sum. Simultaneously, the suppression weights within the region are summed to obtain a weighted sum. Finally, the weighted sum is divided by the weighted sum to obtain the region's mean. The mean values ​​of the four regions are used as regional statistical indicators. The reason for using suppression weights is that suppression weights can reduce the contribution of high-risk vertices near the hyperglycemic transition ring to the statistical results, so that the mean values ​​of the four regions reflect the level of stable regions far away from the ring and reduce the dominance of false anomalies in regional statistics. Thus, the mean values ​​of the four regions are obtained. Based on the 3D face mesh, the visualization scalar of each vertex is mapped to color intensity and rendered on the mesh surface with continuous color levels. At the same time, the segmentation boundaries corresponding to the four region masks are superimposed to form a four-region facial visualization map. The segmentation boundaries are preferably drawn with the boundary vertices at the intersection of the four regions or the set of cross-region face edges to ensure that the boundary position is consistent with the four-region mask. Then, the mean values ​​of the four regions are output in the form of structured data and associated with the four-region facial visualization map for storage or display, so as to obtain a four-region facial visualization output that can simultaneously express vertex-level distribution and regional summary results.

[0049] like Figure 2The diagram shown is a functional block diagram of a facial four-zone division visualization device based on a three-dimensional model provided in an embodiment of the present invention.

[0050] In this embodiment, the functions of each module / unit are as follows: The mesh normal module is used to obtain a 3D face mesh containing a set of vertices and calculate the unit normal of each vertex in the vertex set. The ring extraction module is used to calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex, extract specular transition rings based on the degree of transition of specular reflection proxy and texture brightness in a ring neighborhood, and generate a ring buffer set. The coordinate construction module is used to construct suppression weights based on the geodesic distance from the vertex to the specular transition ring. The suppression weights are used to perform weighted statistical analysis on the mesh vertices to establish the left and right axes and the top and bottom axes of the 3D face mesh. The four-region segmentation module is used to project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, and combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds. Based on the left and right segmentation boundaries and the top and bottom segmentation thresholds, a four-region mask is generated. The visualization statistics module is used to perform ring attenuation backfilling on the vertex scalars to be visualized based on geodesic distance to obtain the visualization scalars, and calculate the statistical indicators of each region according to the suppression weight and the four-region mask.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visualization method for facial four-region division based on a 3D model, characterized in that, include: S1. Obtain a 3D face mesh containing a set of vertices, and calculate the unit normal of each vertex in the vertex set; S2. Calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex. Extract specular transition rings based on the degree of transition in the neighborhood of the ring and generate a ring buffer set. S3. Based on the geodesic distance from the vertex to the specular transition ring, construct the suppression weight, and perform weighted statistical analysis on the grid vertices through the suppression weight to establish the left and right axes and the top and bottom axes of the three-dimensional face grid. S4. Project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds, and generate a four-region mask based on the left and right segmentation boundaries and the top and bottom segmentation thresholds. S5. Based on geodesic distance, perform ring attenuation backfilling on the vertex scalar to be visualized to obtain the visualized scalar, and calculate the statistical index of each region according to the suppression weight and the four-region mask.

2. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Calculate specular reflection proxy and texture brightness, including: The incident cosine is calculated by the dot product of the unit normal of the vertex and the incident direction of the light source, where the incident direction of the light source is the direction vector obtained by subtracting the position vector of the corresponding vertex from the position vector of the light source and normalizing it. Set the effective refractive index of the sebum layer, and calculate the normal incident reflection base value based on the effective refractive index of the sebum layer; Based on the normal incident reflection basis value and the incident cosine, the specular reflection surrogate quantity of each vertex is calculated using the Schlick approximation formula. Obtain the texture color data of the vertex, and perform a weighted summation of the red, green, and blue components of the texture color data to obtain the texture brightness.

3. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Specular transition bands are extracted based on the degree of transition between specular reflection proxy and texture brightness in a ring neighborhood, including: Construct a ring neighborhood for each vertex; The first transition quantity is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the target vertex and the adjacent vertices in the ring neighborhood of the mirror reflection surrogate quantity. The second transition amount is obtained by calculating the arithmetic mean of the absolute values ​​of the differences between the texture brightness and the target vertex in the neighborhood of the ring. The product of the first transition and the second transition is used as the ring score; The median and absolute median difference of the ring zone scores of all grid vertices are statistically calculated, and the ring zone threshold is determined based on the median and the absolute median difference. Vertices with ring scores greater than or equal to the ring threshold are used as constituent vertices of specular transition rings.

4. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Generate a ring-banded buffer set, including: Calculate the geodesic distance from each vertex in the vertex set to the specular transition ring; Define a buffer radius, and define the set of vertices whose geodesic distance is less than or equal to the buffer radius as the annular buffer set.

5. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Establish the left and right axes and the top and bottom axes of the 3D face mesh, including: The geodesic distance from the vertex to the specular transition ring is mapped to a suppression weight using an exponential decay function. Using the suppression weight as a coefficient, the coordinates of the vertex set are weighted and averaged to obtain the weighted centroid; Using the suppression weights as coefficients, calculate the weighted covariance matrix of the vertex set relative to the weighted centroid; Perform eigenvalue decomposition on the weighted covariance matrix, and select the unit eigenvectors corresponding to the largest and second largest eigenvalues ​​as the left and right axes and the top and bottom axes, respectively.

6. The facial four-region division visualization method based on a three-dimensional model according to claim 5, characterized in that, Projecting the vertex set onto the left and right axes and top and bottom axes of the 3D face mesh, and combining this with the annular buffer set, determines the left and right segmentation boundaries and the top and bottom segmentation thresholds, including: Using the weighted centroid as the base point, construct left and right dividing planes with the normal to the left and right axes, and use the left and right dividing planes as the left and right dividing lines; The optimal upper and lower segmentation thresholds are searched along the projection directions of the upper and lower axes, and an objective function containing a weighted balance term and a cross-penetration penalty term is constructed. The weighted balance term is used to measure the difference in the sum of the suppression weights of the regions on both sides of the upper and lower segmentation thresholds. The cross-penetration penalty term is used to measure the degree to which the segmentation line corresponding to the upper and lower segmentation thresholds crosses the annular buffer set. The value that minimizes the objective function is selected as the upper and lower segmentation threshold.

7. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Based on geodesic distance, annular attenuation backfilling is performed on the vertex scalar to be visualized, including: Attenuation coefficients are constructed based on the geodesic distance from the vertex to the high-light transition ring. Within the ring neighborhood of each vertex, the difference between the attenuation coefficients of the vertex and its neighboring vertex is used as a weight to perform a weighted average of the scalar values ​​of the vertices to be visualized in the neighborhood, thus obtaining the non-ring reference value. Based on the attenuation coefficient, linear interpolation is performed between the vertex scalar to be visualized and the non-annular reference value to obtain the visualized scalar.

8. The facial four-region division visualization method based on a three-dimensional model according to claim 1, characterized in that, Statistical indices for each region are calculated based on suppression weights and the four-region mask, including: The visualization scalar is divided into four regions based on the four-region mask; Within each region, the visualized scalar is weighted according to the suppression weight to obtain the region mean as a statistical indicator.

9. A facial four-region division visualization device based on a three-dimensional model, characterized in that, The apparatus, used in the facial four-region segmentation visualization method based on a three-dimensional model according to any one of claims 1-8, comprises: The mesh normal module is used to obtain a 3D face mesh containing a set of vertices and calculate the unit normal of each vertex in the vertex set. The ring extraction module is used to calculate specular reflection proxy and texture brightness based on the incident direction of the light source and the unit normal of the vertex, extract specular transition rings based on the degree of transition of specular reflection proxy and texture brightness in a ring neighborhood, and generate a ring buffer set. The coordinate construction module is used to construct suppression weights based on the geodesic distance from the vertex to the specular transition ring. The suppression weights are used to perform weighted statistical analysis on the mesh vertices to establish the left and right axes and the top and bottom axes of the 3D face mesh. The four-region segmentation module is used to project the vertex set onto the left and right axes and the top and bottom axes of the 3D face mesh, and combine the ring buffer set to determine the left and right segmentation boundaries and the top and bottom segmentation thresholds. Based on the left and right segmentation boundaries and the top and bottom segmentation thresholds, a four-region mask is generated. The visualization statistics module is used to perform ring attenuation backfilling on the vertex scalars to be visualized based on geodesic distance to obtain the visualization scalars, and calculate the statistical indicators of each region according to the suppression weight and the four-region mask.

Images