A method and apparatus for local decoupling of a three-dimensional face

Abstract

The application provides a local decoupling method and device for a three-dimensional face, comprising the following steps: randomly generating a style vector, specifying a camera parameter, and obtaining a corresponding mapped vector; generating local geometric features of all categories by using a local region geometry generator module; generating respective semantic confidence by using a geometric feature and a semantic prediction network module of each category; obtaining global geometric features by using the semantic confidence and the local geometric features and weighted summation; inputting the global geometric features into a density prediction network module to obtain density information; sending the local geometric features into a local texture generator module to generate local texture features of all categories; obtaining global texture features by using the semantic confidence and the local texture features and weighted summation; inputting the global texture features into a color prediction network module to obtain color information; and finally obtaining a complete semantic mask image and a face image through a ray marching operation, so that the local decoupling method and device for a three-dimensional face guided by a semantic mask are completed through a strategy of local decoupling and global fusion.

Description

Technical Field

[0001] This invention relates to the field of local decoupling technology for three-dimensional faces, and more particularly to a method and apparatus for local decoupling of three-dimensional faces. Background Technology

[0002] Realistic face image generation and editing is a useful topic in image synthesis, widely used in portrait generation and artistic creation. Researchers have devoted considerable effort to improving the quality and resolution of generated facial images. Meanwhile, users desire more interaction and control over the generated images. To enhance the controllability of the generation process, numerous methods for editing face images through different interfaces have been proposed, such as sketching, text, and semantic masks.

[0003] Benefiting from implicit 3D representations of neural radiation fields (NeRF), image synthesis models have made significant progress in transmitting 2D images, shifting the image generation task from 2D to 3D and addressing 3D consistency in perspective transformation. EG3D, StyleNRF, and StyleSDF use implicit 3D representations to improve the quality of 3D face generation. Recently, some NeRF-based face editing methods have shown excellent results in decoupling facial geometry and texture. FENeRF, IDE-3D, and NeRFFaceEditing decouple geometry and texture by using separate geometry and texture networks. These methods use global latent codes to generate global 3D representations, so editing the latent codes will affect the entire face. This inevitably affects non-edited areas when editing local facial regions, and can even lead to inconsistencies in facial identity.

[0004] Therefore, there is an urgent need to propose a new method for local decoupling of three-dimensional faces. Summary of the Invention

[0005] To solve or partially solve the above-mentioned technical problems, the present invention provides a method and apparatus for local decoupling of three-dimensional faces. By designing a local region generator module and a spatial fusion module, the ability of convolutional neural networks to locally decouple three-dimensional faces is effectively improved.

[0006] In a first aspect, the present invention provides a method for local decoupling of a three-dimensional human face, the method comprising:

[0007] Randomly generate style vectors, specify camera parameters, and obtain the corresponding mapped vectors;

[0008] The local region geometry generator module generates local geometric features for all categories.

[0009] Semantic confidence scores are generated using geometric features and semantic prediction network modules for each category;

[0010] Global geometric features are obtained by weighted summation using semantic confidence and local geometric features;

[0011] Global geometric features are input into the density prediction network module to obtain density information;

[0012] The local geometric features are fed into the local texture generator module to generate local texture features for all categories.

[0013] Global texture features are obtained by weighted summation using semantic confidence and local texture features;

[0014] The global texture features are input into the color prediction network module to obtain color information;

[0015] Finally, the complete semantic mask image and face image are obtained through ray marching operation.

[0016] The local geometry generator module consists of 6 fully connected layers, each with a dimension of 64. Each fully connected layer injects a local geometry style vector to control the generation.

[0017] The local texture generator module consists of four fully connected layers, each with a dimension of 64. Each fully connected layer is injected with a local geometric style vector to control the generation.

[0018] According to the local decoupling method for three-dimensional faces provided by the present invention, the local geometry generator module and the local texture generator module together constitute a local generator module.

[0019] The data processing procedure of the local generator module is as follows:

[0020] The first fully connected layer extracts features from the coordinates x of a point in three-dimensional space to obtain the first layer intermediate feature P1;

[0021] The second fully connected layer extracts features from P1 to obtain the intermediate features P2 of the second layer;

[0022] The third fully connected layer extracts features from P2 to obtain the intermediate feature P3 of the third layer.

[0023] The fourth fully connected layer extracts features from P3 to obtain the intermediate feature P4 of the fourth layer.

[0024] The fifth fully connected layer extracts features from P4 to obtain the intermediate feature P5 of the fifth layer.

[0025] The sixth fully connected layer extracts features from P5 to obtain the intermediate feature P6 of the sixth layer.

[0026] The seventh fully connected layer extracts features from P6 to obtain the intermediate feature P7 of the seventh layer.

[0027] The eighth fully connected layer extracts features from P7 to obtain the intermediate feature P8 of the eighth layer.

[0028] The ninth fully connected layer extracts features from P8 to obtain the intermediate feature P9 of the ninth layer.

[0029] The tenth fully connected layer extracts features from P9 to obtain the tenth receptive field feature P. 10 ;

[0030] The semantic prediction network module performs input prediction on P6 to obtain semantic confidence. The semantic confidence is then weighted and summed with P6 to obtain global geometric features. Finally, the semantic confidence is summed with P... 10 Weighted summation yields the global texture features;

[0031] The density prediction network module takes global geometric features as input and predicts the density to obtain the density prediction value.

[0032] The color prediction network module performs input prediction on global texture features to obtain color prediction values;

[0033] According to the local decoupling method for three-dimensional faces provided by the present invention, the local geometry generator and local texture generator modules have the characteristics of being decomposable and fused.

[0034] According to the local decoupling method for three-dimensional faces provided by the present invention, the local geometry generator and local texture generator modules adopt a fully connected layer structure.

[0035] Secondly, the present invention provides a local decoupling device for a three-dimensional human face, the device comprising:

[0036] The mapping module is used to obtain local geometry and texture vectors for all categories;

[0037] The Local Geometry Generator module is used to obtain local geometric features.

[0038] The local texture generator module is used to obtain local texture features.

[0039] The spatial fusion module is used to obtain semantic mask, spatial density and color information by utilizing the semantic prediction network module, density prediction network module and color prediction network module;

[0040] The local geometry generator module can modify the geometry of a category by individually modifying the local geometry vector of that category.

[0041] The local texture generator module can modify the texture of a certain category by modifying the local texture vector of that category individually.

[0042] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the local decoupling method for three-dimensional faces as described in the first aspect.

[0043] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the local decoupling method for three-dimensional faces as described in the first aspect.

[0044] This invention provides a method and apparatus for local decoupling of 3D faces. In the preparation stage, random noise is mapped into local geometric and texture vectors of all categories through a mapping network module. In the application stage, a local region geometry generator module generates local geometric features of all categories. Semantic confidence scores are generated using the geometric features of each category and a semantic prediction network module. Global geometric features are obtained by weighted summation of semantic confidence scores and local geometric features. These global geometric features are then input into a density prediction network module to obtain density information. Local geometric features are then fed into a local texture generator module to generate local texture features of all categories. Global texture features are obtained by weighted summation of semantic confidence scores and local texture features. These global texture features are then input into a color prediction network module to obtain color information. Finally, a complete semantic mask image and face image are obtained through ray marching operations, thereby achieving local decoupling and editing of 3D faces. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the local decoupling method for three-dimensional faces provided by the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the local geometry and texture generator provided by the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of the local decoupling device for a three-dimensional human face provided by the present invention;

[0049] Figure 4This is a schematic diagram of the structure of an electronic device that implements a local decoupling method for three-dimensional human faces, as provided by the present invention.

[0050] Figure label:

[0051] 510: Processor; 520: Communication interface; 530: Memory; 540: Communication bus. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0053] In the local decoupling of 3D faces, the decomposition and fusion of 3D faces are crucial. The rational decomposition and fusion of 3D faces has become a significant challenge.

[0054] To improve the controllability of NeRF-based facial editing, we propose a locally controllable facial generation and editing method for fine-grained local facial region control and decoupling of geometry and texture.

[0055] Two core issues need to be addressed: the decomposition of the global 3D representation and the local 3D region representation, and the fusion of local 3D regions. Decomposing the complete 3D representation into multiple local 3D representations and stably completing the training process is challenging. To address this, we design a generator network with multiple local generators to generate content for each local region. Furthermore, for more flexible control over geometry and texture, we further subdivide the local generators into a geometry network and a texture network, controlled by geometry code and texture code, respectively. Through these designs, our method can modify the geometry and texture of a local region by editing multiple local latent codes without affecting other regions.

[0056] Another key challenge is how to fuse the local 3D representations of all local regions to generate the final facial image. We propose a spatially aware fusion module to accomplish the fusion of multiple local regions. Specifically, each local geometry generator predicts the semantic confidence of spatial points, and the fusion module combines the features of different local generators in a smooth and gentle manner using all confidence scores.

[0057] The following is combined Figures 1-4 This invention describes a method and apparatus for local decoupling of a three-dimensional human face.

[0058] In a first aspect, the present invention provides a method for local decoupling of a three-dimensional human face, the method comprising:

[0059] S11. Randomly generate style vectors, specify camera parameters, and obtain the corresponding mapped vectors;

[0060] Here, style vectors are first randomly generated, camera parameters are specified, and then 8 fully connected layers are input to obtain local geometry vectors and local texture vectors;

[0061] S12. Generate local geometric features of all categories using the local region geometry generator module;

[0062] The assigned geometry generator is used to determine the shape of a 3D human face. We assign a lightweight geometry generator to each local region i of the face. If a 3D point belongs to a local region, the corresponding geometry generator will provide the most information for that point. The generator plays a major role in determining the semantic category and geometric information of the point. As shown in the figure, each local geometry generator contains six fully connected layers with SIREN activation and is powered by geometric latent code. control;

[0063] The first fully connected layer extracts features from the coordinates x of a point in three-dimensional space to obtain the first layer intermediate feature P1;

[0064] The second fully connected layer extracts features from P1 to obtain the intermediate features P2 of the second layer;

[0065] The third fully connected layer extracts features from P2 to obtain the intermediate feature P3 of the third layer.

[0066] The fourth fully connected layer extracts features from P3 to obtain the intermediate feature P4 of the fourth layer.

[0067] The fifth fully connected layer extracts features from P4 to obtain the intermediate feature P5 of the fifth layer.

[0068] The sixth fully connected layer extracts features from P5 to obtain the intermediate feature P6 of the sixth layer.

[0069] The local geometry generator module consists of 6 fully connected layers, each with a dimension of 64. Each fully connected layer injects a local geometry style vector to control the generation.

[0070] Figure 2 The left figure shows the network results of the local geometry generator;

[0071] S13. Generate semantic confidence scores for each category using geometric features and semantic prediction network modules;

[0072] Given a 3D sampling point x, the i-th local geometry generator module Decode it, in geometric vectors Predicting semantic confidence s under control i (x) and geometric features f gi (x);

[0073] The local texture generator module consists of four fully connected layers, each with a dimension of 64. Each fully connected layer is injected with a local geometric style vector to control the generation.

[0074] The seventh fully connected layer extracts features from P6 to obtain the intermediate feature P7 of the seventh layer.

[0075] The eighth fully connected layer extracts features from P7 to obtain the intermediate feature P8 of the eighth layer.

[0076] The ninth fully connected layer extracts features from P8 to obtain the intermediate feature P9 of the ninth layer.

[0077] The tenth fully connected layer extracts features from P9 to obtain the tenth receptive field feature P. 10 ;

[0078] The local region geometry generator module can be described as follows:

[0079]

[0080] Figure 2 The right figure shows the network results of the local texture generator;

[0081] S14. Input the global geometric features into the density prediction network module to obtain density information;

[0082] The spatially aware fusion module is designed for interaction and aggregation among multiple local generators. The fusion module proposed in this invention fuses features from different generators with a soft, adjustable mechanism to generate the entire image. This invention stitches together the semantic confidence scores s of all geometric generators. i (x), and obtain the semantic mask m(x) through softmax activation;

[0083] S15. Feed the local geometric features into the local texture generator module to generate local texture features of all categories;

[0084] The local texture generator can be interpreted as a shader used to fill the color of the geometry generated by the geometry generator. In other words, the texture generator does not participate in or affect the generation of the geometry; the geometry generator is only used to determine the shape of the face, thereby achieving local region decoupling and geometry / texture decoupling. Each texture generator contains four fully connected layers with SIREN activation and is powered by the texture latent code. control.

[0085] Given the viewpoint and viewing direction v, the i-th local texture generator module From geometric latent vectors Decoding texture features

[0086] The local region texture generator module can be described as follows:

[0087]

[0088] The texture features predicted by all local texture generators will be fused in the subsequent fusion module to predict the final color value of the sampled 3D point x.

[0089] S16. Use semantic confidence and local texture features to obtain global texture features by weighted summation;

[0090] Local texture features It is also fused with the semantic mask m(x) to obtain the final global texture feature f. t (x);

[0091] S17. Input the global texture features into the color prediction network module to obtain color information;

[0092] This invention uses a linear layer to process the final global texture features f t (x) is predicted as the color value c(x);

[0093] S18. Finally, the complete semantic mask image and face image are obtained through the ray marching operation;

[0094] This invention provides a method for local decoupling of 3D faces. In the preparation stage, random noise is mapped into local geometric and texture vectors of all categories through a mapping network module. In the application stage, a local region geometry generator module generates local geometric features of all categories. Semantic confidence scores are generated using the geometric features of each category and a semantic prediction network module. Global geometric features are obtained by weighted summation of semantic confidence scores and local geometric features. These global geometric features are then input into a density prediction network module to obtain density information. Local geometric features are then fed into a local texture generator module to generate local texture features of all categories. Global texture features are obtained by weighted summation of semantic confidence scores and local texture features. These global texture features are then input into a color prediction network module to obtain color information. Finally, a complete semantic mask image and face image are obtained through ray marching, thereby achieving local decoupling and editing of 3D faces. Secondly, the local decoupling device for 3D faces provided by this invention is described below. The local decoupling device for 3D faces described below corresponds to the local decoupling method for 3D faces described above. Figure 3 An example is a schematic diagram of a local decoupling device for a 3D human face, as shown below. Figure 3 As shown, the device includes:

[0095] Mapping module 21 is used to obtain local geometry and texture vectors for all categories;

[0096] Local geometry generator module 22 is used to obtain local geometric features using the local geometry generator module;

[0097] Local texture generator module 23 is used to obtain local texture features using the local texture generator module;

[0098] Spatial fusion module 24 is used to obtain semantic mask, spatial density and color information by utilizing semantic prediction network module, density prediction network module and color prediction network module;

[0099] The local geometry generator module can modify the geometry of a category by individually modifying the local geometry vector of that category.

[0100] The local texture generator module can modify the texture of a certain category by modifying the local texture vector of a specific category individually.

[0101] The local geometry generator module consists of 6 fully connected layers, each with a dimension of 64. Each fully connected layer injects a local geometry style vector to control the generation.

[0102] The local texture generator module consists of four fully connected layers, each with a dimension of 64. Each fully connected layer is injected with a local geometric style vector to control the generation.

[0103] According to the local decoupled face editing method provided by the present invention, the local geometry generator module and the local texture generator module together constitute a local generator module.

[0104] The data processing procedure of the local generator module is as follows:

[0105] The first fully connected layer extracts features from the coordinates x of a point in three-dimensional space to obtain the first layer intermediate feature P1;

[0106] The second fully connected layer extracts features from P1 to obtain the intermediate features P2 of the second layer;

[0107] The third fully connected layer extracts features from P2 to obtain the intermediate feature P3 of the third layer.

[0108] The fourth fully connected layer extracts features from P3 to obtain the intermediate feature P4 of the fourth layer.

[0109] The fifth fully connected layer extracts features from P4 to obtain the intermediate feature P5 of the fifth layer.

[0110] The sixth fully connected layer extracts features from P5 to obtain the intermediate feature P6 of the sixth layer.

[0111] The seventh fully connected layer extracts features from P6 to obtain the intermediate feature P7 of the seventh layer.

[0112] The eighth fully connected layer extracts features from P7 to obtain the intermediate feature P8 of the eighth layer.

[0113] The ninth fully connected layer extracts features from P8 to obtain the intermediate feature P9 of the ninth layer.

[0114] The tenth fully connected layer extracts features from P9 to obtain the tenth receptive field feature P. 10 ;

[0115] The semantic prediction network module performs input prediction on P6 to obtain semantic confidence. The semantic confidence is then weighted and summed with P6 to obtain global geometric features. Finally, the semantic confidence is summed with P... 10 Weighted summation yields the global texture features;

[0116] The density prediction network module takes global geometric features as input and predicts the density to obtain the density prediction value.

[0117] The color prediction network module performs input prediction on global texture features to obtain color prediction values;

[0118] According to the locally decoupled face editing method provided by the present invention, the local geometry generator and local texture generator modules have the characteristics of being decomposable and fused.

[0119] According to the locally decoupled face editing method provided by the present invention, the local geometry generator and local texture generator modules adopt a fully connected layer structure.

[0120] Thirdly, Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a local decoupling method for 3D faces. This method includes: randomly generating style vectors, specifying camera parameters, and obtaining the corresponding mapped vectors; generating local geometric features for all categories using a local region geometry generator module; generating semantic confidence scores for each category using geometric features and a semantic prediction network module; obtaining global geometric features by weighted summation of semantic confidence scores and local geometric features; inputting the global geometric features into a density prediction network module to obtain density information; feeding local geometric features into a local texture generator module to generate local texture features for all categories; obtaining global texture features by weighted summation of semantic confidence scores and local texture features; inputting the global texture features into a color prediction network module to obtain color information; and finally obtaining a complete semantic mask image and face image through ray marching operations.

[0121] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] Fourthly, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the local decoupling method for three-dimensional faces provided by the above methods. The method includes: randomly generating style vectors, specifying camera parameters, and obtaining the corresponding mapped vectors; generating local geometric features of all categories using a local region geometry generator module; generating semantic confidence scores for each category using geometric features and a semantic prediction network module; obtaining global geometric features by weighted summation using semantic confidence scores and local geometric features; inputting the global geometric features into a density prediction network module to obtain density information; sending local geometric features into a local texture generator module to generate local texture features of all categories; obtaining global texture features by weighted summation using semantic confidence scores and local texture features; inputting the global texture features into a color prediction network module to obtain color information; and finally obtaining a complete semantic mask image and face image through ray marching operations.

[0123] Fifthly, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for local decoupling of a 3D face provided by the methods described above. This method includes: randomly generating a style vector, specifying camera parameters, and obtaining the corresponding mapped vector; generating local geometric features of all categories using a local region geometry generator module; generating semantic confidence scores for each category using the geometric features and a semantic prediction network module; obtaining global geometric features by weighted summation of the semantic confidence scores and local geometric features; inputting the global geometric features into a density prediction network module to obtain density information; sending the local geometric features into a local texture generator module to generate local texture features of all categories; obtaining global texture features by weighted summation of the semantic confidence scores and local texture features; inputting the global texture features into a color prediction network module to obtain color information; and finally obtaining a complete semantic mask image and face image through raymarching operations.

[0124] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0125] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for local decoupling of a three-dimensional human face, characterized in that, The method includes: Randomly generate style vectors, specify camera parameters, and obtain the corresponding mapped vectors; The local geometry generator module is used to generate local geometric features of all categories. Semantic confidence scores are generated using geometric features and semantic prediction network modules for each category; Global geometric features are obtained by weighted summation using semantic confidence and local geometric features; Global geometric features are input into the density prediction network module to obtain density information; The local geometric features are fed into the local texture generator module to generate local texture features for all categories. Global texture features are obtained by weighted summation using semantic confidence and local texture features; The global texture features are input into the color prediction network module to obtain color information; Finally, the complete semantic mask image and face image are obtained through ray marching operation; The local geometry generator module consists of 6 fully connected layers, each with a dimension of 64. Each fully connected layer injects a local geometry style vector to control the generation. The local texture generator module consists of four fully connected layers, each with a dimension of 64. Each fully connected layer is injected with a local geometric style vector to control the generation. The local geometry generator module can modify the geometry of a category by individually modifying the local geometry vector of that category. The local texture generator module can modify the texture of a certain category by modifying the local texture vector of a specific category individually. The data processing procedures for the local geometry generator module and the local texture generator module are as follows: The first fully connected layer extracts features from the coordinates x of a point in three-dimensional space to obtain the intermediate features P1 of the first layer. The second fully connected layer extracts features from P1 to obtain the intermediate feature P2 of the second layer. The third fully connected layer extracts features from P2 to obtain the intermediate feature P3 of the third layer. The fourth fully connected layer extracts features from P3 to obtain the fourth layer intermediate feature P4; The fifth fully connected layer extracts features from P4 to obtain the fifth layer intermediate feature P5; The sixth fully connected layer extracts features from P5 to obtain the sixth layer intermediate feature P6; The seventh fully connected layer extracts features from P6 to obtain the intermediate feature P7 of the seventh layer; The eighth fully connected layer extracts features from P7 to obtain the intermediate feature P8 of the eighth layer; The ninth fully connected layer extracts features from P8 to obtain the intermediate feature P9 of the ninth layer; The tenth fully connected layer extracts features from P9 to obtain the tenth receptive field feature P10; The semantic prediction network module performs input prediction on P6 to obtain semantic confidence. The semantic confidence is then weighted and summed with P6 to obtain global geometric features, and the semantic confidence is then weighted and summed with P10 to obtain global texture features. The density prediction network module takes global geometric features as input and predicts the density to obtain the density prediction value. The color prediction network module performs input prediction on global texture features to obtain color prediction values.

2. In the local decoupling method for three-dimensional faces according to claim 1, the local geometry generator module and the local texture generator module together constitute the local generator module.

3. The local decoupling method for a three-dimensional face according to any one of claims 1 to 2, characterized in that, The local geometry generator module, local texture generator module, and spatial fusion module have the characteristics of being decomposable and fusionable.

4. A local decoupling device for a three-dimensional human face, characterized in that, The apparatus, applied to the local decoupling method for a three-dimensional face as described in any one of claims 1 to 3, comprises: The mapping module is used to obtain local geometry and texture vectors for all categories; The Local Geometry Generator module is used to obtain local geometric features. The local texture generator module is used to obtain local texture features. The spatial fusion module is used to obtain semantic mask, spatial density, and color information by utilizing the semantic prediction network module, density prediction network module, and color prediction network module.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the local decoupling method for a three-dimensional face as described in any one of claims 1 to 3.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the local decoupling method for a three-dimensional face as described in any one of claims 1 to 3.

Images