A Method and System for Attention Recognition in Desensitized Faces Based on Attention Affine Transform
By using an attention-guided feature affine transformation method, the accuracy of facial key point alignment in complex environments is solved, achieving an effective combination of 3D face reconstruction and personal privacy protection, and improving the model's performance and privacy protection in complex environments.
Patent Information
- Application Number
- CN202311852919.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-12-29
AI Technical Summary
Existing technologies struggle to accurately capture robust facial features for facial landmark alignment in complex environments, and traditional methods have limitations in terms of personal privacy protection and resource consumption.
A feature affine transformation method based on attention mechanism is adopted. By combining self-attention mechanism and coordinate attention mechanism, feature information is dynamically fused to perform spatial rotation, translation and scaling transformation to construct a 3D face model. A constrained encoder loss function is designed to erase personal privacy.
It improves the accuracy and robustness of facial landmark alignment in complex environments, while protecting personal privacy without affecting the displayability of research content.
Smart Images

Figure CN117671772B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of 3D face reconstruction using neural network models, specifically involving a method and system for desensitized face attention discrimination based on attention affine transformation. Background Technology
[0002] Facial attention is a crucial topic in social sciences, biomedicine, psychology, and computer vision, exploring how people allocate and focus their attention when faced with another person's face. Methods for detecting attention have therefore spanned multiple disciplines and are constantly evolving with technological advancements. Most researchers use traditional observational methods and questionnaires in controlled experiments to assess visitor attention. In the biomedical field, methods using heart rate and skin conductance measurements are employed to detect changes in facial attention. Furthermore, in psychology, eye-tracking technology is used to precisely capture details of facial attention. However, all of these methods have limitations in terms of ease of use and resource consumption, such as human effort.
[0003] In the field of computer vision, traditional active shape-based face landmark alignment algorithms generate a statistical shape model for face alignment by statistically modeling the shape information of the face. However, this model only models shape features and ignores the texture features of the face, making it difficult to capture and align local details. Subsequently, active appearance algorithms improve the model's ability to capture details by jointly modeling both shape and texture information.
[0004] However, traditional methods mostly rely on manual feature extraction, resulting in insufficient model expressive power. With the development of deep learning, learning-based models can automatically capture and extract features from massive amounts of data, significantly improving the model's expressive power. However, simple cascaded networks and attention mechanisms cannot accurately capture robust facial features for facial landmark alignment in complex environments. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a method and system for desensitized face attention discrimination based on attention affine transformation, which solves the problem that the prior art cannot accurately capture robust face features for face key point alignment in complex environments.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0007] A face attention discrimination method based on attention mechanism-guided feature affine transformation includes the following steps:
[0008] S1: Construct a face attention discrimination model based on attention mechanism-guided feature affine transformation to capture the spatial correlation within face data features, establish long-distance dependencies, guide features to perform affine transformation in latent space, output face deformation parameters, and then construct a 3D model of the face.
[0009] S2: Construct a facial information desensitization model. Based on the facial attention discrimination model constructed in step S1, perform facial rendering to obtain a two-dimensional face. Then, perform personal privacy erasure through facial mask desensitization and erasure.
[0010] S3: Construct a face deviation pose angle estimation model to calculate the deviation angle between the face plane position obtained in step S2 and the fixed camera standard plane under the observation state, estimate the face pose direction, and achieve accurate face attention discrimination.
[0011] The face attention discrimination model includes a self-attention mechanism, a coordinate attention mechanism, a multi-attention fusion method, and a feature space affine transformation mechanism, among which...
[0012] Self-attention mechanism, which is sensitive to extracting important internal features from input information, captures global discriminative features;
[0013] The coordinate attention mechanism mines horizontal and vertical spatial information of the input features respectively, calculates highly relevant information through global context information, and combines it with information that is highly relevant to itself to help the deep network accurately locate and identify targets.
[0014] A multi-attention mechanism fusion method dynamically fuses the output information of self-attention and coordinate attention models, enhances the ability to establish feature representations, extracts more comprehensive information to improve model robustness, and is used to process complex and variable face data;
[0015] The feature space affine transformation mechanism uses an attention fusion method to fuse robust features to guide the extraction of features in the latent space to perform affine transformations. It performs spatial rotation, translation and scaling transformations on the extracted features to correct and reorganize the spatial information and semantic relationships of the features, thereby achieving 3D face reconstruction and dense alignment of facial key points.
[0016] It also includes a constraint encoder loss function for designing constraint model learning, which is used to capture facial semantic and spatial hierarchical information in samples and generate a multi-level hierarchical knowledge matrix.
[0017] The constraint encoder loss function includes
[0018] Point-level loss function is used to capture discrete shape information of different semantic parts at the point level;
[0019] Line-level loss functions are used to construct attraction and repulsion relationships between point pairs of different semantic parts to explore line-level hierarchical information;
[0020] A planar loss function is used to constrain the overall shape structure and area ratio of different semantic parts of the face, learn planar knowledge of different semantic parts, and suppress extreme expression problems.
[0021] Spatial loss function is used to suppress the loss caused by the density of feature points and to constrain surface details.
[0022] The specific process for erasing personal privacy is as follows:
[0023] The 3D face constructed by the face attention module is projected onto a 2D plane using the projection normalized coordinate code method to achieve facial mask rendering, which is then attached to the face of the corresponding image. This achieves the erasure of personal privacy without affecting the displayability and verifiability of the research content.
[0024] The method for constructing the face deviation pose angle estimation model is as follows:
[0025] A fixed camera acquires a face image, extracts several key points from it to determine the planar position of the face mask under the observation state, and calculates its deviation angle from the standard plane of the fixed camera to determine the concentration range of face attention in the input image; specifically:
[0026] S61: Extract facial key points. By extracting three key points located at the left and right corners of the eyes and the tip of the nose of the 3D face model, the face observation plane is determined. Two vectors on the plane are obtained, and then the cross product formula is used to obtain the normal vector of the plane.
[0027] S62: Calculate the deviation angle between the face mask observation plane and the fixed camera standard plane to estimate the face pose direction, and finally determine the concentration range of face attention.
[0028] By fixing the normal vector N of the camera's standard plane... T The following formula can be used to calculate the angle of deviation between the two observation planes:
[0029]
[0030] Where [o,p,q] and [x,y,z] represent the normal vector N of the face observation plane, respectively. O and the normal vector N of the fixed camera standard plane T 3D coordinates, cos θ The cosine value represents the angle between the two normal vectors.
[0031] The face attention discrimination system includes a face image acquisition module, a face attention discrimination module, a face information desensitization module, and a face deviation pose angle estimation module; among which,
[0032] The face image acquisition module includes historical face images and real-time acquired face images. Historical face images are used to train the model, while real-time acquired face images are input into the trained model to output face attention features.
[0033] The face attention discrimination module is used to train a face attention discrimination model based on historical face images and construct a 3D model of the face.
[0034] The facial information desensitization module is used to build a facial information desensitization model and erase personal facial information;
[0035] The face deviation pose angle estimation module is used to build a face deviation pose angle estimation model and estimate the face pose direction from real-time acquired face images, thereby achieving accurate face attention discrimination.
[0036] The face attention discrimination module accurately constructs a 3D model of the face based on 3DMM.
[0037] The real-time acquired face images are processed to extract facial key points, and the planar position inference points are determined based on the relevant positions of the facial key points.
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] 1. This invention utilizes an attention-based discrimination model guided by an attention mechanism and based on feature affine transformation to construct a facial key point alignment model. This model accurately fits the inherent distribution and relationships of facial data, effectively improving the model's representation ability in complex environments such as complex lighting, large poses, and extreme expressions. It provides important technical support for analyzing facial behavioral features and understanding individual preferences, and has broad practical application value.
[0040] 2. This invention can accurately identify facial attention while protecting personal privacy.
[0041] 3. The 3D face constructed by the face attention module is projected onto a 2D plane using the projection normalized coordinate code method to achieve facial mask rendering, which is then attached to the face of the corresponding image. This allows for the erasure of personal privacy without affecting the displayability and verifiability of the research content. Attached Figure Description
[0042] Figure 1 This is an overall network block diagram of the method described in the embodiments of the present invention.
[0043] Figure 2This is the face attention discrimination module of the method described in the embodiments of the present invention.
[0044] Figure 3 These are the qualitative and quantitative results of the algorithm for attitude estimation and deviation angle calculation in this invention. Detailed Implementation
[0045] The structure and working process of the present invention will be further described below with reference to the accompanying drawings.
[0046] A face attention discrimination method based on attention mechanism-guided affine transformation of features includes the following steps:
[0047] S1: Acquire face data and construct a large-scale, multi-pose, multi-view face dataset for training and testing the network module of this invention.
[0048] S2: Using the face attention discrimination module, the intrinsic correlation of features is mined based on affine transformation, which guides the features to undergo affine transformation in the latent space to accurately construct a 3D model of the face.
[0049] Specifically as follows:
[0050] (1a) The coordinate attention mechanism is used to mine the horizontal and vertical spatial information of the input features respectively. The information with high relevance is calculated through global context information and combined with the information with high relevance to itself to promote feature reasoning.
[0051] (1b) The input features are recalibrated using the coupling coefficient relation matrix of the self-attention mechanism and the horizontal and vertical weight coefficients of the coordinate attention mechanism, and the output features of the self-attention mechanism and the coordinate attention mechanism are dynamically fused to obtain the final output information.
[0052] (1c) Using a latent spatial affine transformation mechanism guided by self-attention and coordinate attention, the extracted features are transformed by spatial rotation, translation and scaling in the feature space.
[0053] (1d) The output transformation features are accurately regressed into 62 3DMM parameters through a fully connected network for personalized 3D face initialization. The pose parameters consist of 12 parameters in the tuple [f,R,t]. 2d The image indicates that a 3D face is modeled using 62 parameters, including shape, expression, and pose parameters.
[0054] (1e) A 3D face is initialized using the 3D face identity and expression parameters predicted by the model based on the 3DMM face reconstruction method.
[0055] S3: Design a joint loss function for the parameters of the constrained model learning to better fit the face distribution model of the acquired data. This function is composed of a weighted parametric distance loss function and a two-dimensional facial key point loss function.
[0056] (2a) Weighted parameter distance loss function: used to constrain model learning so that it can gradually and accurately predict face identity and facial expression attribute parameters;
[0057] (2b) Two-dimensional facial key point loss function: Use sparse two-dimensional facial key point loss as a joint constraint function to further constrain the basic contour of the three-dimensional face;
[0058] S4: Using a facial information desensitization module, a personalized 3D face is rendered onto a 2D plane using weak perspective projection to apply a facial mask, thereby erasing personal privacy.
[0059] U = f * P * R * M + t 2d
[0060] Where U represents the 2D position of the 3D face landmark after projection, rotation, and translation. f represents the scaling factor. It is an orthogonal projection matrix, R = [pitch, yaw, roll] is a rotation matrix, and t 2d =[t x ,t y ] is a translation vector.
[0061] S5: Extract facial key points and determine the planar position of the facial mask generated in step S4 under the observation state. In order to obtain a relatively accurate and stable planar position of the face with minimal computing resources, this invention analyzes the relevant positions of facial key points to determine the planar position inference points.
[0062] (3a) First, the nose is located in the center of the face, with eyebrow and eye positions above it, mouth position below it, and facial contour positions to the left and right. It has very obvious anchor point features. Therefore, the present invention selects the tip of the nose as the first plane point.
[0063] (3b) Secondly, attention judgment is usually based on the direction of gaze on the face. Therefore, the present invention selects the corners of the eyes as the second and third plane points.
[0064] (3c) The planar position of the face under observation is determined by the position information of the above three points.
[0065] S6: Based on the planar position of the face under the determined observation state, calculate the deviation angle between the face and the fixed camera standard plane, calculate the face deviation pose angle, and finally determine the concentration range of the face attention.
[0066] The loss function in step S3 can also be designed as a constraint encoder loss function for constrained model learning, used to capture facial semantic and spatial hierarchical information in samples and generate a multi-level hierarchical knowledge matrix.
[0067] The constraint encoder loss function includes
[0068] Point-level loss function is used to capture discrete shape information of different semantic parts at the point level;
[0069] Line-level loss functions are used to construct attraction and repulsion relationships between point pairs of different semantic parts to explore line-level hierarchical information;
[0070] A planar loss function is used to constrain the overall shape structure and area ratio of different semantic parts of the face, learn planar knowledge of different semantic parts, and suppress extreme expression problems.
[0071] Spatial loss function is used to suppress the loss caused by the density of feature points and to constrain surface details.
[0072] The multi-attention mechanism fusion method dynamically integrates the output information of self-attention and coordinate attention models, enhances the feature representation, extracts more comprehensive information to improve model robustness, and better handles complex and ever-changing face data.
[0073] The feature space affine transformation mechanism uses an attention fusion method to fuse robust features to guide the extraction of features in the latent space to perform affine transformations. It performs spatial rotation, translation and scaling transformations on the extracted features to correct and reorganize the spatial information and semantic relationships of the features, thereby achieving 3D face reconstruction and dense alignment of facial key points.
[0074] Specific embodiments, such as Figures 1 to 3 As shown,
[0075] This embodiment discloses a method for desensitized face attention discrimination based on attention mechanism-guided feature affine transformation, including the following steps:
[0076] S1: The model of this invention first acquires two-dimensional face images for training and testing the network modules. The sources of the two-dimensional face data are as follows:
[0077] S11: 300W-LP Training Dataset: This dataset is an extension of the 300W dataset, containing 120,000 fully annotated (68 key points) face images;
[0078] S12: AFLW test dataset: The AFLW
[24] dataset is a large-scale, multi-pose, multi-view face dataset, in which each face has 21 facial landmark annotations;
[0079] S13: AFLW 2000-3D test dataset: AFLW2000-3D
[13] consists of the first 2000 images of the AFLW dataset, and the real annotations of 21 facial key points are extended to 68 through 3DMM;
[0080] S2: An attention-based discrimination module that guides feature affine transformation using an attention mechanism uses the two-dimensional face image X used for network training in step S1 to infer 50 face parameters (identity, expression) corresponding to each face image.
[0081] S3: Based on the 50 face parameters predicted in step S2, initialize the 3DMM model to obtain the 3D reconstruction results of each face:
[0082]
[0083] Where N is a 3D face initialized using 50 3DMM parameters. It is the average three-dimensional human face. M id and M exp These are the identity and expression parameters of a neutral 3D human face. id and p exp The parameters are the 3D face identity and expression parameters predicted by the model, used to achieve personalized 3D face initialization. Among them, M... id and M exp These are the basic identity and expression parameter vectors, respectively.
[0084] S4: Based on the designed face desensitization module, 12 pose parameters are used to perform weak perspective projection on the 3D face reconstruction result generated in step S3, mapping the 3D face onto a 2D plane to achieve dense alignment of facial key points.
[0085]
[0086] Among them, [f,R,T 2d The orientation is determined by 12 parameters, where f represents the scaling factor, P represents the orthographic projection matrix, R represents the rotation matrix, and T represents the rotation matrix. 2d This represents the translation matrix.
[0087] S5: Construct a joint loss function for parameters. Based on the face parameter results predicted in step S2 and the two-dimensional face rendered in step S4, better constrain the learning of the face attention discrimination module in step S2, so that it can better learn the face distribution model of the data obtained in step S1.
[0088] S51: For the facial deformation parameters predicted in step S2, the model learning is constrained by a weighted parametric distance loss function, enabling it to progressively and accurately predict 3DMM facial attribute parameters.
[0089] Lossw =(W gt -W) T P(WW gt )
[0090] Among them, W gt W and W represent the true labels and predicted values of the 62 3DMM face attribute parameters, respectively, and P is a learnable model weight parameter.
[0091] S52: A sparse 2D facial keypoint loss is used as a joint constraint function to supplement the WPDC objective function, further constraining the basic contour of the 3D face.
[0092]
[0093] Among them, f, R and T 2d f represents the pose parameters predicted by the model. gt R gt and The actual labels representing the attitude parameters. id and p exp The parameters are the predicted 3D facial identity and expression parameters. Realistic labels representing 3D facial identity and expression parameters.
[0094] S6: Using the face dataset obtained in step S1 and step S5 as a constraint, train the designed attention discrimination module network model by iterating through steps S2-S4 above.
[0095] S7: For a face image whose attention needs to be determined, the corresponding two-dimensional face mask image can be obtained based on the attention discrimination module trained in step S6 and the face desensitization module in step S4.
[0096] S8: For a face image whose attention needs to be determined, the face observation plane is determined by extracting three relatively stable key points located at the left and right corners of the eyes and the tip of the nose of the corresponding 3D face model, and the deviation angle between the fixed camera standard plane and the 2D face mask image is calculated to determine the direction of attention.
[0097] S9: Automatically obtains the duration of a person's attention to the target object by calculating the deviation angle of the face observation plane from the fixed camera's standard plane in real time. The fixed camera's standard plane is determined by the placement of the vision sensor and the location of the monitored object or attraction.
[0098] S10: When the above method has a large deviation, the present invention can further project the three-dimensional face into a two-dimensional space, and perform face pose estimation based on the coordinates of 68 two-dimensional face key points obtained by projection, so as to obtain the qualitative result of face pose estimation, and further assist in the judgment and calibration of face attention direction in different application scenarios.
[0099] This embodiment combines self-attention and coordinate attention mechanisms to guide the affine transformation of features in the latent space, improving the expressive power of the model in complex environments. For example... Figure 2 As shown, this model improves the sensitivity of extracting important internal features from input information through a self-attention mechanism, effectively capturing global discriminative features. Secondly, it utilizes a coordinate attention mechanism to mine horizontal and vertical spatial information from the input features, calculating highly relevant information through global contextual information. Combining this with information highly relevant to itself promotes accurate target localization and recognition by the deep network, enhancing the model's reasoning ability. Finally, this paper dynamically fuses the output information from the self-attention and coordinate attention modules, using this to guide the extracted features in affine transformation in the latent space. This better corrects and reconstructs the spatial information and semantic relationships of the features, achieving 3D face reconstruction and dense alignment of facial key points.
[0100] The loss function in step S3 can also be designed as a constraint encoder loss function for constrained model learning, used to capture facial semantic and spatial hierarchical information in samples and generate a multi-level hierarchical knowledge matrix.
[0101] The constraint encoder loss function includes
[0102] Point-level loss function is used to capture discrete shape information of different semantic parts at the point level;
[0103] Line-level loss functions are used to construct attraction and repulsion relationships between point pairs of different semantic parts to explore line-level hierarchical information;
[0104] A planar loss function is used to constrain the overall shape structure and area ratio of different semantic parts of the face, learn planar knowledge of different semantic parts, and suppress extreme expression problems.
[0105] Spatial loss function is used to suppress the loss caused by the density of feature points and to constrain surface details.
[0106] The face attention discrimination system includes a face image acquisition module, a face attention discrimination module, a face information desensitization module, and a face deviation pose angle estimation module; among which,
[0107] The face image acquisition module includes historical face images and real-time acquired face images. Historical face images are used to train the model, while real-time acquired face images are input into the trained model to output face attention features.
[0108] The face attention discrimination module is used to train a face attention discrimination model based on historical face images and construct a 3D model of the face.
[0109] The facial information desensitization module is used to build a facial information desensitization model and erase personal facial information;
[0110] The face deviation pose angle estimation module is used to build a face deviation pose angle estimation model and estimate the face pose direction from real-time acquired face images, thereby achieving accurate face attention discrimination.
[0111] The face attention discrimination module accurately constructs a 3D model of the face based on 3DMM.
[0112] The real-time acquired face images are processed to extract facial key points, and the planar position inference points are determined based on the relevant positions of the facial key points.
[0113] Those skilled in the art should understand that variations can be implemented by combining existing technology and the above embodiments. Such variations do not affect the substantive content of this solution and will not be elaborated here.
[0114] It should be understood that this solution is not limited to the specific embodiments described above. Devices and structures not described in detail herein should be understood as being implemented in a manner common to the art. Any person skilled in the art can make many possible variations and modifications to this solution, or modify it into equivalent embodiments, without departing from the scope of this solution, using the methods and techniques disclosed above. This does not affect the substantive content of this solution. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this solution, without departing from its scope, still fall within the protection scope of this solution.
Claims
1. A method for desensitized face attention discrimination based on attention mechanism-guided affine transformation of features, characterized in that: Includes the following steps: S1: Construct a face attention discrimination model based on attention mechanism-guided affine transformation of features to capture the spatial correlation within face data features, establish long-distance dependencies, guide features to undergo affine transformation in the latent space, output face deformation parameters, and then construct a 3D model of the face; the face attention discrimination model includes self-attention mechanism, coordinate attention mechanism, multi-attention fusion method, and feature space affine transformation mechanism, wherein, Self-attention mechanism, which is sensitive to extracting important internal features from input information, captures global discriminative features; The coordinate attention mechanism mines horizontal and vertical spatial information of the input features respectively, calculates highly relevant information through global context information, and combines it with information that is highly relevant to itself to help the deep network accurately locate and identify targets. A multi-attention mechanism fusion method dynamically fuses the output information of self-attention and coordinate attention models, enhances the ability to establish feature representations, extracts more comprehensive information to improve model robustness, and is used to process complex and variable face data; The feature space affine transformation mechanism uses an attention fusion method to fuse robust features to guide the extracted features to perform affine transformations in the latent space. It performs spatial rotation, translation and scaling transformations on the extracted features to correct and reorganize the spatial information and semantic relationships of the features, thereby achieving 3D face reconstruction and dense alignment of facial key points. Design a constraint encoder loss function for constraint model learning to capture facial semantic and spatial hierarchical information in samples and generate a multi-level hierarchical knowledge matrix. The constraint encoder loss function includes, Point-level loss function is used to capture discrete shape information of different semantic parts at the point level; Line-level loss functions are used to construct attraction and repulsion relationships between point pairs of different semantic parts to explore line-level hierarchical information; A planar loss function is used to constrain the overall shape structure and area ratio of different semantic parts of the face, learn planar knowledge of different semantic parts, and suppress extreme expression problems. Spatial loss function is used to suppress the loss caused by the density of feature points and to constrain surface details; S2: Construct a facial information desensitization model. Based on the facial attention discrimination model constructed in step S1, perform facial rendering to obtain a two-dimensional face. Then, perform personal privacy erasure through facial mask desensitization and erasure. S3: Construct a face deviation pose angle estimation model to calculate the deviation angle between the face plane position obtained in step S2 and the fixed camera standard plane under the observation state, estimate the face pose direction, and achieve accurate face attention discrimination.
2. The method for desensitized face attention discrimination based on attention mechanism-guided affine transformation according to claim 1, characterized in that: The specific process for erasing personal privacy is as follows: The 3D face constructed by the face attention module is projected onto a 2D plane using the projection normalized coordinate code method to achieve facial mask rendering, which is then attached to the face of the corresponding image. This achieves the erasure of personal privacy without affecting the displayability and verifiability of the research content.
3. The method for desensitized face attention discrimination based on attention mechanism-guided affine transformation according to claim 1, characterized in that: The method for constructing the face deviation pose angle estimation model is as follows: A fixed camera acquires a face image, extracts several key points from it to determine the planar position of the face mask under the observation state, and calculates its deviation angle from the standard plane of the fixed camera to determine the concentration range of facial attention in the input image; specifically: S61: Extract facial key points. By extracting three key points located at the left and right corners of the eyes and the tip of the nose of the 3D face model, the face observation plane is determined. Two vectors on the plane are obtained, and then the cross product formula is used to obtain the normal vector of the plane. S62: Calculate the deviation angle between the face mask observation plane and the fixed camera standard plane to estimate the face pose direction, and finally determine the concentration range of face attention.
4. The method for desensitized face attention discrimination based on attention mechanism-guided affine transformation according to claim 3, characterized in that: By fixing the normal vector of the camera's standard plane The following formula can be used to calculate the angle of deviation between the two observation planes: in, and These represent the normal vectors of the face observation plane, respectively. and the normal vector of the fixed camera standard plane The three-dimensional coordinates Represents the angle between two normal vectors value.
5. A face attention discrimination system based on the method of any one of claims 1 to 4, characterized in that: It includes a face image acquisition module, a face attention discrimination module, a face information desensitization module, and a face deviation pose angle estimation module; among which, The face image acquisition module includes historical face images and real-time acquired face images. Historical face images are used to train the model, while real-time acquired face images are input into the trained model to output face attention features. The face attention discrimination module is used to train a face attention discrimination model based on historical face images and construct a 3D model of the face. The facial information desensitization module is used to build a facial information desensitization model and erase personal facial information; The face deviation pose angle estimation module is used to build a face deviation pose angle estimation model and estimate the face pose direction from real-time acquired face images, thereby achieving accurate face attention discrimination.
6. The face attention discrimination system according to claim 5, characterized in that: The face attention discrimination module accurately constructs a 3D model of the face based on 3DMM.
7. The face attention discrimination system according to claim 5, characterized in that: The real-time acquired face images are processed to extract facial key points, and the planar position inference points are determined based on the relevant positions of the facial key points.
Citation Information
Patent Citations
Single-image face three-dimensional reconstruction method based on self-alignment double regression
CN114972619A
Speech-driven face synthesis method based on sparse view
CN116385603A
