No-reference quality assessment method and device based on three-dimensional face geometry

Abstract

The present invention discloses a reference-free quality assessment method and device based on three-dimensional face geometry, which relates to the field of computer vision. The method includes: reconstructing a three-dimensional face mesh model; calculating the Euclidean distance between the mapping key points and the extraction key points to obtain a geometric consistency score; dividing the face region using a three-dimensional face segmentation algorithm, calculating the Gaussian curvature to obtain a regional curvature score; calculating the smoothness of the model surface, detecting whether there are unnatural protrusions or defects on the model surface, and obtaining a smoothness score; fusing the geometric consistency score, the regional curvature score and the smoothness score according to a weighted ratio, and outputting a comprehensive quality score of the three-dimensional face mesh model. The present invention does not need to rely on a standard face model in a database, and can comprehensively evaluate the quality of a single reconstructed three-dimensional face mesh model based on face geometry features and regional analysis, and is suitable for personalized reconstruction scenarios such as intelligent beauty and precision medicine.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer vision and three-dimensional face reconstruction, and in particular to a reference-free quality assessment method and device based on three-dimensional face geometric structure. Background Art

[0002] 3D face quality assessment is a key link in evaluating the quality of reconstruction results and is extremely important in practical applications. Most existing 3D face quality assessment methods rely on comparison with standard face models in the database. When faced with actual application scenarios such as photographing and reconstructing the face of a new user, there is no corresponding 3D face model reference. In addition, the current assessment algorithm has problems such as poor generalization and strong dependence on reference data, and cannot achieve comprehensive evaluation of a single reconstructed face, which limits its applicability in actual scenarios (such as personalized customization, skin health testing, etc.). Summary of the invention

[0003] The purpose of the present invention is to provide a reference-free quality assessment method and device based on the three-dimensional face geometric structure. By analyzing the geometric characteristics of the three-dimensional face mesh model, the quality of the reconstructed three-dimensional face mesh model can be objectively evaluated without relying on the standard face model in the database. The method is suitable for the evaluation and optimization of high-precision three-dimensional face models in personalized skin care, intelligent beauty diagnosis, virtual makeup trial and related fields.

[0004] The present invention adopts the following technical solution:

[0005] On the one hand, a no-reference quality assessment method based on 3D face geometry structure, comprising:

[0006] The 3D face mesh model reconstruction step is to obtain the corresponding 3D point cloud data based on the acquired face depth map through the internal and external parameter matrices of the acquisition device, and reconstruct the 3D face mesh model based on the 3D point cloud data;

[0007] The geometric consistency score acquisition step is to extract 68 key points of the face based on the collected two-dimensional image of the face, map these key points to the three-dimensional point cloud space through the internal and external parameter matrices of the acquisition device, and obtain the mapping key points; at the same time, extract 68 key points from the three-dimensional point cloud data to obtain the extracted key points; calculate the Euclidean distance between the mapped key points and the extracted key points and perform linear normalization to obtain the distance set between all key points, and obtain the geometric consistency score based on the average value of the normalized distance set;

[0008] The step of obtaining the regional curvature score is to use a 3D face segmentation algorithm to divide the 3D face mesh model into 7 regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region; the average Gaussian curvature of each region is calculated respectively, and linearly normalized to generate the quality score of each region; the quality score of each region is weighted according to a preset weight to synthesize the total regional curvature score;

[0009] The smoothness score acquisition step is to construct the vertex adjacency relationship of the 3D face mesh model, evaluate the geometric difference between each vertex and its neighboring vertices, and use the mean of the geometric differences as the local smoothness; accumulate the local smoothness of all vertices and take the average value to obtain the smoothness score;

[0010] In the comprehensive quality score output step, the geometric consistency score, regional curvature score and smoothness score are integrated according to a weighted ratio to output the comprehensive quality score of the 3D face mesh model.

[0011] Preferably, in the step of reconstructing the three-dimensional face mesh model, the acquisition device is a structured light camera; the internal and external parameter matrices of the acquisition device are calculated by a camera calibration method; and the three-dimensional face mesh model is reconstructed using the Songbai reconstruction algorithm.

[0012] Preferably, the step of obtaining the geometric consistency score specifically includes:

[0013] Extract 68 2D key points of a 2D image , these key points represent the locations of feature points of the face on a two-dimensional plane, as follows:

[0014] ;

[0015] in, Indicates key points;

[0016] By collecting the internal and external parameter matrix of the equipment, the two-dimensional key points Map into 3D point cloud space to get mapping key points ,as follows:

[0017] ;

[0018] in, Indicates Mapping key points;

[0019] The 68 key points of the 3D point cloud data are extracted by the 3D key point extraction algorithm to obtain the extracted key points. ,as follows:

[0020] ;

[0021] in, Indicates Extract key points;

[0022] Calculate the Euclidean distance between the mapped keypoints and the corresponding points of the extracted keypoints ,as follows:

[0023] ;

[0024] Get the distance set between all key points ,as follows:

[0025] ;

[0026] Distance Set Perform linear normalization to obtain ,as follows:

[0027] ;

[0028] in, is the minimum distance in the distance set, is the maximum distance in the distance set; get the normalized distance set ,as follows:

[0029] ;

[0030] Finally, the normalized distance set is used to calculate the geometric consistency score ,as follows:

[0031] .

[0032] Preferably, the step of obtaining the regional curvature score specifically includes:

[0033] The 3D face segmentation algorithm is used to divide the 3D face mesh model into the forehead area , left eye area , right eye area , left cheek area , right cheek area , nose area and mouth area There are 7 partitioned regions in total; after partitioning, each region contains a set of vertices ,as follows:

[0034] ;

[0035] in, Indicates area The vertices;

[0036] For each segmented area , calculate the average Gaussian curvature of all vertices in the corresponding area ,as follows:

[0037] ;

[0038] in, Represents a vertex The Gaussian curvature of Indicates area The number of vertices in ;

[0039] The average Gaussian curvature of each region is linearly normalized to generate a normalized score ,as follows:

[0040] ;

[0041] in, is the minimum area curvature, is the maximum regional curvature;

[0042] Set weights for each area based on its impact on the overall quality , weighted calculation to obtain the final regional curvature score ,as follows:

[0043] .

[0044] Preferably, the smoothness score acquisition step specifically includes:

[0045] Construct an adjacency graph for each vertex of the 3D face mesh model, and record each vertex and its directly adjacent vertices; let the vertex set be , the mesh face set is , each face consists of three vertices; the adjacency relationship can be represented as a graph ,in is the edge formed by each pair of adjacent vertices;

[0046] Set Vertex and its neighbor vertices , then the Euclidean distance between vertices for:

[0047] ;

[0048] The Euclidean distance Represents a vertex and its neighbor vertices Differences in position in three-dimensional space;

[0049] For each vertex , calculate the mean geometric difference between it and its neighboring vertices as the local smoothness value of the vertex; specifically, assuming that the vertex Total neighbor nodes, and the set of neighbor vertices is , then the vertex The local smoothness It is defined as the average of the Euclidean distances between all neighbor vertices and it, as follows:

[0050] ;

[0051] in, Represents a vertex and neighbor vertices The Euclidean distance between

[0052] By calculating the local smoothness of all vertices The average value of the overall mesh smoothness score ,as follows:

[0053] ;

[0054] in, Represents the total number of vertices of the 3D face mesh model.

[0055] Preferably, the comprehensive quality score output step specifically includes:

[0056] The geometric consistency score , regional curvature fraction and smoothness score Weighted according to the set weights to obtain the final comprehensive quality score ,as follows:

[0057] ;

[0058] in, , and are the weights of the regional curvature score, geometric consistency score, and smoothness score, respectively, and , and The sum of is equal to 1.

[0059] On the other hand, a reference-free quality assessment device based on three-dimensional face geometry structure comprises:

[0060] A 3D face mesh model reconstruction module is used to obtain the corresponding 3D point cloud data based on the acquired face depth map through the internal and external parameter matrices of the acquisition device, and reconstruct the 3D face mesh model based on the 3D point cloud data;

[0061] The geometric consistency score acquisition module is used to extract 68 key points of the face based on the collected two-dimensional image of the face, map these key points to the three-dimensional point cloud space through the internal and external parameter matrices of the acquisition device, and obtain the mapping key points; at the same time, extract 68 key points from the three-dimensional point cloud data to obtain the extracted key points; calculate the Euclidean distance between the mapped key points and the extracted key points and perform linear normalization processing to obtain the distance set between all key points, and obtain the geometric consistency score based on the average value of the normalized distance set;

[0062] The regional curvature score acquisition module is used to divide the 3D face mesh model into 7 regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region, by using a 3D face segmentation algorithm; calculate the average Gaussian curvature of each region respectively, and perform linear normalization on it to generate the quality score of each region; weight the quality score of each region according to a preset weight to synthesize the total regional curvature score;

[0063] The smoothness score acquisition module is used to construct the vertex adjacency relationship of the 3D face mesh model, evaluate the geometric difference between each vertex and its neighboring vertices, and use the mean of the geometric differences as the local smoothness; the local smoothness of all vertices is accumulated and averaged to obtain the smoothness score;

[0064] The comprehensive quality score output module is used to fuse the geometric consistency score, regional curvature score and smoothness score in a weighted ratio to output the comprehensive quality score of the three-dimensional face mesh model.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The present invention is based on the two-dimensional image and depth map of the face captured by a structured light camera, and generates a three-dimensional face mesh model through a face reconstruction algorithm; on this basis, a reference-free quality assessment method and device based on face geometric features and regional analysis are designed to accurately reflect the quality of the three-dimensional face model. Specifically, firstly, by extracting the key points of the face in the two-dimensional image and mapping them to the reconstructed three-dimensional face mesh model, the accurate alignment of the key point positions is ensured, and then the three-dimensional face key points are extracted using the existing algorithm, and then the distance between the key points obtained by the two methods is calculated, and the distance is linearly normalized, and the geometric consistency score is obtained by calculating the average value of the normalized distance and taking 1-; then the 3D face is divided into 7 regions including the forehead area, the left and right eye areas, the left and right cheek areas, the nose area and the mouth area by the three-dimensional face segmentation algorithm, and the average Gaussian curvature of each area is calculated, and it is normalized to obtain the regional curvature score; then by analyzing the smoothness of the surface of the three-dimensional face mesh model, whether there are unnatural protrusions or defects on the model surface is detected, and the smoothness quality score is calculated; finally, the geometric consistency score, the regional curvature score and the smoothness score are weighted and fused, and the overall quality evaluation result of the model is output. The present invention can comprehensively evaluate the quality of a single reconstructed three-dimensional face mesh model without relying on the standard face model in the database, and is suitable for various personalized reconstruction scenarios, especially in the fields of intelligent beauty and precision medicine. It has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A flowchart of a reference-free quality assessment method based on three-dimensional face geometry provided by an embodiment of the present invention;

[0068] Figure 2 A detailed flowchart of a reference-free quality assessment method based on three-dimensional face geometry provided by an embodiment of the present invention;

[0069] Figure 3 A schematic diagram of a flow chart of a step of obtaining a geometric consistency score provided in an embodiment of the present invention;

[0070] Figure 4 A structural block diagram of a reference-free quality assessment device based on three-dimensional face geometry provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0072] See also Figure 1As shown, this embodiment provides a reference-free quality assessment method based on three-dimensional face geometric structure, including the following steps.

[0073] In step S101, a three-dimensional face mesh model reconstruction step is performed based on the acquired face depth map, and the corresponding three-dimensional point cloud data is obtained through the internal and external parameter matrices of the acquisition device, and a three-dimensional face mesh model is reconstructed based on the three-dimensional point cloud data.

[0074] Specifically, for three-dimensional acquisition devices (such as structured light cameras, etc.), they have built-in RGB cameras and depth cameras, so they can collect the original facial data color map (two-dimensional image) and depth map. For the depth map, the internal and external parameter matrix of the three-dimensional acquisition device is calculated in combination with the camera calibration method. The three-dimensional point cloud data of the corresponding face can be obtained through calculation, and then the reconstruction algorithm such as Poisson reconstruction is used to obtain the final high-precision three-dimensional face mesh model.

[0075] In the step S102 for obtaining the geometric consistency score, 68 key points of the face are extracted based on the collected two-dimensional image of the face, and these key points are mapped to the three-dimensional point cloud space through the internal and external parameter matrices of the acquisition device to obtain the mapped key points; at the same time, 68 key points are extracted from the three-dimensional point cloud data to obtain the extracted key points; the Euclidean distance between the mapped key points and the extracted key points is calculated and linearly normalized to obtain the distance set between all the key points, and the geometric consistency score is obtained based on the average value of the normalized distance set.

[0076] First, for the collected 2D face image data, 68 key points of the face are extracted using algorithms such as dlib. Then, these key points are mapped to the 3D point cloud through the camera's internal and external parameter matrix. At the same time, 68 key points are directly extracted from the 3D face point cloud through algorithms such as 3DDFA_v3. Next, the Euclidean distance between the corresponding key points of the mapping method and the extraction method is calculated and linearly normalized to obtain the distance set between all key points. The final geometric consistency score is obtained by calculating the average value of the normalized distance and subtracting it from the natural number 1.

[0077] The calculation of the geometric consistency score aims to evaluate the geometric accuracy of the 3D face mesh model by matching the 2D face key points with the 3D point cloud key points. Specifically, the 68 2D key points of the 2D image are extracted using the dlib algorithm. , these key points represent the locations of feature points of the face on a two-dimensional plane, as follows:

[0078] ;

[0079] in, Indicates key points;

[0080] By collecting the internal and external parameter matrix of the device such as the camera, the two-dimensional key points Map into 3D point cloud space to get mapping key points ,as follows:

[0081] ;

[0082] in, Indicates Mapping key points;

[0083] 68 key points of 3D point cloud data are extracted through 3D key point extraction algorithms such as 3DDFA_v3 to obtain extracted key points. ,as follows:

[0084] ;

[0085] in, Indicates Extract key points;

[0086] Calculate the Euclidean distance between the mapped keypoints and the corresponding points of the extracted keypoints ,as follows:

[0087] ;

[0088] Get the distance set between all key points ,as follows:

[0089] ;

[0090] Distance Set Perform linear normalization to obtain ,as follows:

[0091] ;

[0092] in, is the minimum distance in the distance set, is the maximum distance in the distance set; get the normalized distance set ,as follows:

[0093] ;

[0094] Finally, the normalized distance set is used to calculate the geometric consistency score ,as follows:

[0095]

[0096] in, It indicates the matching degree of each key point, and the average value of all key points reflects the overall consistency.

[0097] In the regional curvature score acquisition step S103, a three-dimensional face mesh model is divided into seven regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region, using a three-dimensional face segmentation algorithm; the average Gaussian curvature of each region is calculated respectively, and linearly normalized to generate a quality score for each region; the quality score of each region is weighted according to a preset weight to synthesize a total regional curvature score.

[0098] Taking into account the significant differences in the average Gaussian curvature of different face regions (for example, the average Gaussian curvature of the cheek is smaller, while the average Gaussian curvature of the nose is larger), the three-dimensional face segmentation algorithm is first used to divide the face model into seven regions, namely the forehead region, left and right eye regions, left and right cheek regions, nose region and mouth region.

[0099] Forehead area , left eye area , right eye area , left cheek area , right cheek area , nose area and mouth area There are 7 partitioned regions in total; after partitioning, each region contains a set of vertices ,as follows:

[0100] ;

[0101] in, Indicates area The vertices;

[0102] For each segmented area , calculate the average Gaussian curvature of all vertices in the corresponding area ,as follows:

[0103] ;

[0104] in, Represents a vertex The Gaussian curvature of Indicates area The number of vertices in ;

[0105] The average Gaussian curvature of each region is linearly normalized to generate a normalized score ,as follows:

[0106] ;

[0107] in, is the minimum area curvature, is the maximum regional curvature;

[0108] Set weights for each area based on its impact on the overall quality , weighted calculation to obtain the final regional curvature score ,as follows:

[0109] .

[0110] in, =1.

[0111] In the step S104 of obtaining the smoothness score, the vertex adjacency relationship of the three-dimensional face mesh model is constructed, the geometric difference between each vertex and its neighboring vertices is evaluated, and the mean of the geometric differences is used as the local smoothness; the local smoothness of all vertices is accumulated and the average value is taken to obtain the smoothness score.

[0112] In a three-dimensional face mesh model, smoothness is an important indicator to measure the continuity and surface quality of the face mesh surface. A smooth mesh surface has a more uniform vertex distribution and smaller geometric differences, while an unsmooth surface may have protrusions, broken lines, or unnatural curvature changes. In order to evaluate the smoothness of the mesh, the present invention constructs the adjacency relationship between mesh vertices, analyzes the geometric differences between each vertex and its neighboring vertices, and then calculates the local smoothness. Finally, by summarizing the local smoothness of all vertices, the overall mesh smoothness score is obtained.

[0113] The mesh model consists of vertices and faces, and each vertex is connected to its adjacent vertices through edges. In order to calculate the smoothness of the mesh, we first need to build an adjacency graph for each vertex, recording each vertex and its directly adjacent vertices. Let the vertex set be , the mesh face set is , each face consists of three vertices. The adjacency relationship can be represented as a graph ,in is an edge formed by each pair of adjacent vertices.

[0114] Next, for each vertex in the mesh , calculate the geometric difference between it and its neighbor vertices, usually measured by Euclidean distance.

[0115] Specifically, we first construct an adjacency relationship graph for each vertex of the 3D face mesh model, and record each vertex and its directly adjacent vertices; let the vertex set be , the mesh face set is , each face consists of three vertices; the adjacency relationship can be represented as a graph ,in is the edge formed by each pair of adjacent vertices;

[0116] Set Vertex and its neighbor vertices , then the Euclidean distance between vertices for:

[0117] ;

[0118] The Euclidean distance Represents a vertex and its neighbor vertices Differences in position in three-dimensional space;

[0119] For each vertex , calculate the mean geometric difference between it and its neighboring vertices as the local smoothness value of the vertex; specifically, assuming that the vertex Total neighbor nodes, and the set of neighbor vertices is , then the vertex The local smoothness It is defined as the average of the Euclidean distances between all neighbor vertices and it, as follows:

[0120] ;

[0121] in, Represents a vertex and neighbor vertices The Euclidean distance between

[0122] By calculating the local smoothness of all vertices The average value of the overall mesh smoothness score ,as follows:

[0123] ;

[0124] in, Represents the total number of vertices of the 3D face mesh model.

[0125] In the comprehensive quality score output step S105, the geometric consistency score, the regional curvature score and the smoothness score are integrated according to a weighted ratio to output the comprehensive quality score of the three-dimensional face mesh model.

[0126] Specifically, the geometric consistency score , regional curvature fraction and smoothness score Weighted according to the set weights to obtain the final comprehensive quality score ,as follows:

[0127] ;

[0128] in, , and are the weights of the regional curvature score, geometric consistency score, and smoothness score, respectively, and , and The sum of is equal to 1.

[0129] In this embodiment, a reference-free quality assessment method based on the geometric structure of a three-dimensional face is built based on the Python 3.8 development environment, and the trimesh 3.9.0 library is integrated to implement three-dimensional mesh processing (such as calling through the library function trimesh.curvature.discrete_gaussian_curvature_measure), and Open3D is combined for three-dimensional visualization. The core calculation uses a 20mm local sampling radius for curvature analysis, and the key points such as the nose tip (No. 50), the left eye (No. 30), and the right eye (No. 40) are located through hard-coded vertex indexes. When performing quality assessment, the geometric consistency is obtained by the algorithm np.linalg.norm. The feature distances such as the eye distance and the nose-eye distance are then calculated, and the normalized value score is obtained to obtain the geometric consistency score; the regional curvature analysis is based on the standard deviation of the Gaussian curvature of the seven partitions to obtain the total regional curvature score; the surface smoothness is calculated by the mean of the vertex neighborhood position difference to obtain the smoothness score. When calculating the comprehensive quality score, the weight of the geometric consistency score can be 45%, the weight of the regional curvature score can be 35%, and the weight of the smoothness score can be 20%. The final output comprehensive quality score average is approximately 0.85-0.92, which is consistent with the scoring range obtained by human eye perception.

[0130] See also Figure 4 As shown, the present invention also discloses a reference-free quality assessment device based on three-dimensional face geometric structure, comprising:

[0131] The 3D face mesh model reconstruction module 401 is used to obtain the corresponding 3D point cloud data based on the acquired face depth map through the internal and external parameter matrices of the acquisition device, and reconstruct the 3D face mesh model based on the 3D point cloud data;

[0132] The geometric consistency score acquisition module 402 is used to extract 68 key points of the face based on the collected two-dimensional image of the face, map these key points to the three-dimensional point cloud space through the internal and external parameter matrix of the acquisition device, and obtain the mapping key points; at the same time, extract 68 key points from the three-dimensional point cloud data to obtain the extracted key points; calculate the Euclidean distance between the mapped key points and the extracted key points and perform linear normalization processing to obtain the distance set between all key points, and obtain the geometric consistency score based on the average value of the normalized distance set;

[0133] The regional curvature score acquisition module 403 is used to divide the 3D face mesh model into 7 regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region, by using a 3D face segmentation algorithm; calculate the average Gaussian curvature of each region respectively, and perform linear normalization on it to generate a quality score of each region; weight the quality score of each region according to a preset weight to synthesize a total regional curvature score;

[0134] The smoothness score acquisition module 404 is used to construct the vertex adjacency relationship of the three-dimensional face mesh model, evaluate the geometric difference between each vertex and its neighboring vertices, and use the mean of the geometric differences as the local smoothness; accumulate the local smoothness of all vertices and take the average value to obtain the smoothness score;

[0135] The comprehensive quality score output module 405 is used to fuse the geometric consistency score, the regional curvature score and the smoothness score according to a weighted ratio, and output the comprehensive quality score of the three-dimensional face mesh model.

[0136] The reference-free quality assessment method based on three-dimensional face geometry proposed in this invention shows significant technical advantages and practical value in intelligent beauty diagnosis scenarios. By integrating the multi-dimensional evaluation mechanism of geometric consistency, regional curvature and surface smoothness, this method can achieve accurate quality diagnosis of a single sample without relying on a standard database, effectively solving the problem of existing methods' dependence on databases. Specifically: geometric consistency detection can automatically identify symmetry deviations of key facial points (such as zygomatic distance differences exceeding the threshold), thereby indicating potential model distortion or reconstruction errors; regional curvature analysis can locate facial pathological features (such as abnormal curvature of nasolabial grooves may indicate tissue atrophy), providing a quantitative basis for diagnosis; surface smoothness assessment can be used to quantify postoperative repair effects, such as reflecting the degree of healing by detecting changes in the sharpness of suture edges. This method shows strong adaptability in personalized feature scenarios such as asymmetry and unilateral facial paralysis after cosmetic surgery, with a detection accuracy of 0.15 mm, providing reliable technical support for medical beauty applications such as injection simulation and repair effect tracking. Typical application scenarios include: guiding the recalibration of 3D scanning equipment based on key point offset detection, identifying hyaluronic acid filling boundary defects through curvature continuity scoring, and quantifying the smoothness change after zygomatic surgery using curvature heat map. Its regional segmentation analysis mechanism can effectively distinguish between real pathological features and reconstruction errors, and provides reliable guarantee for the quality controllability of 3D face models in scenarios such as intelligent beauty diagnosis and virtual makeup adaptation.

[0137] The specific implementation of each module of a reference-free quality assessment device based on three-dimensional face geometry is the same as the reference-free quality assessment method based on three-dimensional face geometry, and this embodiment will not repeat the description.

[0138] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that the specific embodiments described in this application are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A reference-free quality assessment method based on three-dimensional face geometry, characterized in that: include: The 3D face mesh model reconstruction step is to obtain the corresponding 3D point cloud data based on the acquired face depth map through the internal and external parameter matrices of the acquisition device, and reconstruct the 3D face mesh model based on the 3D point cloud data; The geometric consistency score acquisition step is to extract 68 key points of the face based on the collected two-dimensional image of the face, map these key points to the three-dimensional point cloud space through the internal and external parameter matrix of the acquisition device, and obtain the mapped key points; at the same time, extract 68 key points from the three-dimensional point cloud data to obtain the extracted key points; The Euclidean distance between the mapped key points and the extracted key points is calculated and linearly normalized to obtain the distance set between all key points. The geometric consistency score is obtained based on the average value of the normalized distance set. The step of obtaining the regional curvature score is to use a 3D face segmentation algorithm to divide the 3D face mesh model into 7 regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region; the average Gaussian curvature of each region is calculated respectively, and linearly normalized to generate the quality score of each region; the quality score of each region is weighted according to a preset weight to synthesize the total regional curvature score; The smoothness score acquisition step is to construct the vertex adjacency relationship of the 3D face mesh model, evaluate the geometric difference between each vertex and its neighboring vertices, and use the mean of the geometric differences as the local smoothness; accumulate the local smoothness of all vertices and take the average value to obtain the smoothness score; In the comprehensive quality score output step, the geometric consistency score, regional curvature score and smoothness score are integrated according to a weighted ratio to output the comprehensive quality score of the 3D face mesh model.

2. The reference-free quality assessment method based on three-dimensional face geometry according to claim 1, characterized in that: In the step of reconstructing the three-dimensional face mesh model, the acquisition device is a structured light camera; the internal and external parameter matrices of the acquisition device are calculated by a camera calibration method; and the three-dimensional face mesh model is reconstructed using the Songbai reconstruction algorithm.

3. The reference-free quality assessment method based on three-dimensional face geometry according to claim 1, characterized in that: The step of obtaining the geometric consistency score specifically includes: Extract 68 2D key points of a 2D image , these key points represent the locations of feature points of the face on a two-dimensional plane, as follows: ； in, Indicates key points; By collecting the internal and external parameter matrix of the equipment, the two-dimensional key points Map into 3D point cloud space to get mapping key points ,as follows: ； in, Indicates Mapping key points; The 68 key points of the 3D point cloud data are extracted by the 3D key point extraction algorithm to obtain the extracted key points. ,as follows: ； in, Indicates Extract key points; Calculate the Euclidean distance between the mapped keypoints and the corresponding points of the extracted keypoints ,as follows: ； Get the distance set between all key points ,as follows: ； Distance Set Perform linear normalization to obtain ,as follows: ； in, is the minimum distance in the distance set, is the maximum distance in the distance set; get the normalized distance set ,as follows: ； Finally, the normalized distance set is used to calculate the geometric consistency score ,as follows: 。 4. The reference-free quality assessment method based on 3D face geometry according to claim 1, characterized in that: The step of obtaining the regional curvature score specifically includes: The 3D face segmentation algorithm is used to divide the 3D face mesh model into the forehead area , left eye area , right eye area , left cheek area , right cheek area , nose area and mouth area There are 7 partitioned regions in total; after partitioning, each region contains a set of vertices ,as follows: ； in, Indicates area The vertices; For each segmented area , calculate the average Gaussian curvature of all vertices in the corresponding area ,as follows: ； in, Represents a vertex The Gaussian curvature of Indicates area The number of vertices in ; The average Gaussian curvature of each region is linearly normalized to generate a normalized score ,as follows: ； in, is the minimum area curvature, is the maximum regional curvature; Set weights for each area based on its impact on the overall quality , weighted calculation to obtain the final regional curvature score ,as follows: 。 5. The reference-free quality assessment method based on 3D face geometry according to claim 1, characterized in that: The step of obtaining the smoothness score specifically includes: Construct an adjacency graph for each vertex of the 3D face mesh model, and record each vertex and its directly adjacent vertices; let the vertex set be , the mesh face set is , each face consists of three vertices; the adjacency relationship can be represented as a graph ,in is the edge formed by each pair of adjacent vertices; Set Vertex and its neighbor vertices , then the Euclidean distance between vertices for: ； The Euclidean distance Represents a vertex and its neighbor vertices Differences in position in three-dimensional space; For each vertex , calculate the mean of the geometric differences between it and its neighboring vertices as the local smoothness value of the vertex; specifically, assuming that the vertex Total neighbor nodes, and the set of neighbor vertices is , then the vertex The local smoothness It is defined as the average of the Euclidean distances between all neighbor vertices and it, as follows: ； in, Represents a vertex and neighbor vertices The Euclidean distance between By calculating the local smoothness of all vertices The average value of the overall mesh smoothness score ,as follows: ； in, Represents the total number of vertices of the 3D face mesh model.

6. The reference-free quality assessment method based on 3D face geometry according to claim 1, characterized in that: The comprehensive quality score output step specifically includes: The geometric consistency score , regional curvature fraction and smoothness score Weighted according to the set weights to obtain the final comprehensive quality score ,as follows: ； in, , and are the weights of the regional curvature score, geometric consistency score, and smoothness score, respectively, and , and The sum of is equal to 1.

7. A reference-free quality assessment device based on three-dimensional face geometry, characterized in that: include: A 3D face mesh model reconstruction module is used to obtain the corresponding 3D point cloud data based on the acquired face depth map through the internal and external parameter matrices of the acquisition device, and reconstruct the 3D face mesh model based on the 3D point cloud data; The geometric consistency score acquisition module is used to extract 68 key points of the face based on the collected two-dimensional image of the face, map these key points to the three-dimensional point cloud space through the internal and external parameter matrix of the acquisition device, and obtain the mapped key points; at the same time, extract 68 key points from the three-dimensional point cloud data to obtain the extracted key points; The Euclidean distance between the mapped key points and the extracted key points is calculated and linearly normalized to obtain the distance set between all key points. The geometric consistency score is obtained based on the average value of the normalized distance set. The regional curvature score acquisition module is used to divide the 3D face mesh model into 7 regions, namely, forehead region, left eye region, right eye region, left cheek region, right cheek region, nose region and mouth region, by using a 3D face segmentation algorithm; calculate the average Gaussian curvature of each region respectively, and perform linear normalization on it to generate the quality score of each region; weight the quality score of each region according to a preset weight to synthesize the total regional curvature score; The smoothness score acquisition module is used to construct the vertex adjacency relationship of the 3D face mesh model, evaluate the geometric difference between each vertex and its neighboring vertices, and use the mean of the geometric differences as the local smoothness; the local smoothness of all vertices is accumulated and averaged to obtain the smoothness score; The comprehensive quality score output module is used to fuse the geometric consistency score, regional curvature score and smoothness score in a weighted ratio to output the comprehensive quality score of the three-dimensional face mesh model.

Images