A single-view human body reconstruction method based on geometric perception and consistency constraint

Abstract

The present application relates to a single-view human body reconstruction method based on geometric perception and consistency constraint, and belongs to the field of computer vision, computer aided design and computer graphics. Firstly, based on the initial human body normal map and human body mesh generated by the single-view human body image in the dressed human body data set, geometric consistency modeling is carried out through normal confidence estimation and adaptive normal fusion to obtain vertex-level fusion normal; secondly, edge perception gradient optimization is carried out on the vertex-level fusion normal and human body mesh based on edge perception smoothing and bilateral gradient optimization of local features to obtain optimization momentum; then, topology level decoupling optimization is carried out on the optimization momentum and initial human body mesh through multi-level structure optimization and local detail enhancement to obtain geometric displacement and detail displacement; finally, the dressed human body reconstruction result with global anatomical robustness and high-quality wrinkle details is obtained by superimposing the geometric displacement and the detail displacement to the human body mesh and performing re-meshing.

Description

Technical Field

[0001] This invention relates to a single-view human body reconstruction method based on geometric perception and consistency constraints, belonging to the fields of computer vision, computer-aided design and graphics. Background Technology

[0002] Single-view clothing human reconstruction requires only one RGB image, resulting in low computational cost, low complexity, and flexible applications, with broad potential in fields such as virtual reality, digital humans, film and television special effects, medical analysis, and sports science. Known methods mainly fall into four categories: explicit mesh, implicit function, neural radiation field, and 3D Gaussian methods for 3D human reconstruction. Implicit function-based methods model human geometry using continuous fields, capable of expressing complex topological structures and detailed surfaces at arbitrary resolutions, and possess strong expressive power for local human deformations. However, they can produce local artifacts, structural blurring, and pose deviations for depth ambiguity and occluded areas. Neural radiation field-based methods learn volume density and color distribution in space and simulate the interaction between light and the human body, generating new perspective images with complex lighting, reflection, and transparency effects. However, they suffer from slow convergence speed, poor real-time performance, and geometric stability issues, making it difficult to meet the demands of rapid human reconstruction and real-time rendering. The 3D Gaussian-based method represents the human body surface as a large number of learnable 3D Gaussian elements and uses Gaussian splashing for differentiable rendering, which can significantly improve training efficiency and rendering speed, but is prone to geometric floating, edge blurring, and local structural discontinuities. The explicit mesh-based method directly constructs the human body surface geometry using vertices, edges, and faces, modeling the human body shape through explicit topological structures. This results in clear geometric structures, stable topological relationships, and good structural interpretability and editability. For example, Wu (<Neural Information Processing Systems>, 2024) uses a multi-view diffusion model to generate high-resolution multi-view images and normal maps, and proposes an efficient and consistent mesh reconstruction algorithm that quickly generates explicit 3D meshes with rich surface details and realistic textures while ensuring the stability of the human body's topological structure. Li (…<Conference on Computer Vision and Pattern Recognition > Sun (2025) utilized a cross-scale, multi-view diffusion model to collaboratively generate multi-view human body priors. Combined with explicit remeshing and human body sculpting techniques based on human body template initialization, fine-grained deformation correction on explicit geometry effectively resolved geometric ambiguities, achieving explicit human body mesh reconstruction with realistic details. Sun (<AAAI Conference on Artificial Intelligence>, 2026) used a geometry-aware latent diffusion model to generate high-quality multi-view normals. Combined with a progressive mesh optimization strategy including shape-aware deformation alignment and global-to-local detail refinement, it guided the parametric human body model to gradually deform, accurately restoring fine-grained clothing geometric folds, achieving high-quality explicit dressed human body mesh reconstruction.

[0003] The normal field prior generated from single-view human images exhibits significant viewpoint dependence and spatial heterogeneity. Existing methods, employing predefined static weights, fail to effectively model the local reliability of the normal prior and lack explicit cross-view geometric consistency constraints. Limited by isotropic smoothing, existing mesh optimization methods are insufficient in modeling local geometric anisotropy and high-curvature regions, making it difficult to effectively preserve sharp geometric features of the surface and easily leading to degradation of high-frequency edge details and excessive smoothing of the structure. Furthermore, in the deformation and reconstruction of dressed human bodies, the geometry of the human surface exhibits multi-scale hierarchical characteristics. Existing methods are mostly limited by the single constraint of global topology, failing to combine human anatomical structure for hierarchical decoupling optimization, and making it difficult to simultaneously consider the geometric consistency of the overall structure and the refinement of the local topology. To this end, this invention first generates an initial human body normal map and human body mesh based on single-view images in a dressed human body dataset. Geometric consistency modeling is then performed through normal confidence estimation and adaptive normal fusion to obtain vertex-level fused normals. Secondly, edge-aware gradient optimization is applied to the vertex-level fused normals and human body mesh using edge-aware smoothing based on local features and bilateral gradient optimization to obtain optimized momentum. Then, topological decoupling optimization is performed on the optimized momentum and initial human body mesh through multi-level structure optimization and local detail enhancement to obtain geometric displacement and detail displacement. Finally, by superimposing the geometric displacement and detail displacement onto the human body mesh and remeshing it, a dressed human body reconstruction result with both global anatomical robustness and high-quality wrinkle details is obtained. Summary of the Invention

[0004] This invention provides a single-view human body reconstruction method based on geometric perception and consistency constraints, which solves the problems of inconsistent structural constraints, easy degradation of edge details and lack of local topological details caused by unreliable geometric priors, and can generate more accurate, realistic and detailed clothing human body models.

[0005] The technical solution of this invention is: a single-view human body reconstruction method based on geometric perception and consistency constraints, comprising the following steps:

[0006] Step 1: Generating the initial human normal map based on single-view images from the dressed human body dataset. and human body grid Geometric consistency modeling is performed through normal confidence estimation and adaptive normal fusion to obtain vertex-level fused normals. ;

[0007] Step 2: Edge-aware smoothing and bilateral gradient optimization based on local features for human body mesh Merge normals with vertex level Edge-aware gradient optimization is performed to obtain the optimized momentum. ;

[0008] Step 3: Optimize momentum through multi-level structural optimization and local detail enhancement. and the initial human body mesh Perform topological decoupling optimization to obtain geometric displacement. With detail displacement ;

[0009] Step 4: Geometric displacement and detailed displacement Overlay onto human body mesh The remesh was then performed to obtain high-quality clothing human reconstruction results. .

[0010] The specific process of Step 1 is as follows:

[0011] First, an initial human normal map is generated based on the diffusion of single-view human images from the dressed human dataset. A depth-separable convolutional residual encoder is used to... Encoding as high-dimensional encoded features Enhanced features are obtained through cross-view feature fusion. And mapped to pixel-level confidence maps. ;

[0012] Then, based on the human body mesh Through perspective any valid pixel Establish adjacent view pairs using depth information and camera extrinsic parameters Pixel correspondence set Define the adjacent view normal consistency loss ,in Indicates perspective medium pixel The unit normal vector of the corresponding surface point in the world coordinate system. Represents the spatial reprojection mapping operator;

[0013] Finally, the normal map of the pixel domain is obtained through projection relationships. and confidence plot Spatial sampling is performed to explicitly associate 2D features with 3D mesh vertices. This yields a set of normals from multiple perspectives. and the corresponding vertex-level normal confidence. Define perspective lower vertex Fusion weights ,in For visibility weight, As the basic weight of perspective, Weighted fusion of normals from multiple viewpoints is performed based on the direction of the line of sight. ,in Representing the normalized weights, we obtain the vertex-level fused normals. .

[0014] The specific process of Step 2 is as follows:

[0015] Step 2.1 Edge-aware smoothing;

[0016] First, based on vertex-level fusion normals Rendering yields the target rendering normal map. and target mask Construct the basic reconstruction objective function ;in For normal loss, for Mask loss;

[0017] Then, extract the vertices. Discrete Gaussian curvature Absolute value of mean curvature and dihedrals The local geometric features are normalized, and then... and Constructing a binary edge mask The marginal probability is obtained by combining normalized features. , build Adaptive regularization strength ;

[0018] Finally, combining Cotangent weight Laplace as regularization term Add to the basic objective function In the middle, by taking the derivative with respect to the vertex coordinates ,in As vertex The discrete Laplace differential coordinates at the point are used to obtain the spatially adaptive geometric gradient. .

[0019] Step 2.2 Bilateral gradient optimization;

[0020] First, the geometric gradient is calculated using an exponential moving average. Converted into momentum The vertex normals of the SMPL-X template are extracted and K-means clustering is performed to obtain stable sampling directions. ;

[0021] Then, the sampling direction With normal Perform dot product calculations to construct the geometric guiding factor. ,by As the initial signal, a weighted implicit thermal diffusion is performed on the mesh surface using a sparse matrix system, and the translational sliding term is obtained respectively. With normalization term ;

[0022] Finally, by aggregating sampling features and And then normalize it, specifically as follows: ,in For aggregation weight operators, A constant term is used to ensure numerical stability. The optimized momentum is obtained. .

[0023] The specific process of Step 3 is as follows:

[0024] Step 3.1 Multi-level structure optimization;

[0025] First, set the skinning weight vector of the human mesh vertex to Determine the joint index based on the maximum weight of vertex skinning. The set of mesh vertices is divided into non-overlapping hierarchical subsets according to a rule from the core to the fine tip. ;

[0026] Then, structural regularization constraints are applied to the human body mesh based on skeletal topology, anatomical proportions, and component connections to construct a structure that includes skeletal length constraints. Consistency of proportions Connection constraints of components Global structural prior loss ;

[0027] Finally, through and optimize momentum and combined with hierarchical activators Second-order moment estimation Learning rate and local reference side length The spatial adaptive geometric displacement of the vertex is obtained. in This is a constant term.

[0028] Step 3.2 Enhance local details;

[0029] First, regarding high-frequency changes Hierarchical structure, constructing vertex feature vectors that include local feature embeddings, vertex positions, and normal vectors. ;

[0030] Then, the vertex feature vectors Mapped to a higher-order nonlinear feature space, the normalized original displacement residuals are predicted by multilayer perceptron regression. and the corresponding spatial confidence scalar and by scaling factor Limited to the maximum displacement range;

[0031] Finally, Decomposed into normal components and tangential components Combining spatial confidence scalar Obtain detailed displacement , For normal weights, For tangential weights.

[0032] The present invention also provides a single-view human body reconstruction system based on geometric perception and consistency constraints, the system comprising: a module for executing the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0033] The present invention also 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 executes the program to implement the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0034] The present invention also 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 single-view human body reconstruction method based on geometric perception and consistency constraints.

[0035] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0036] The beneficial effects of this invention are:

[0037] 1. Existing methods directly treat the normal field as a deterministic geometric constraint, introducing global weight constraints only at the viewpoint level. This ignores the reliability differences and geometric consistency between normal fields, leading to inconsistent cross-viewpoint geometric constraints and incomplete human reconstruction results. This invention improves the reliability and consistency of geometric priors in single-viewpoint clothing human reconstruction by designing a normal confidence estimation and cross-viewpoint geometric consistency constraint method, and generating consistent fused normals through adaptive normal fusion.

[0038] 2. Existing methods for mesh optimization driven by differentiable rasterization rely solely on isotropic approaches. This often leads to blurring and noise in local geometry and high-curvature regions, resulting in loss of detail, degraded edge details, and overly smoothed structures. This invention addresses this by employing edge-aware smoothing based on local features and normal-guided bilateral gradient optimization, which suppresses noise while preserving geometric wrinkle details.

[0039] 3. Existing methods for human body deformation and reconstruction primarily rely on global topological constraints, which limit the model's ability to express local structures and cause loss of detail, making it difficult to achieve a balance between overall structural consistency and local refinement. This invention employs multi-level structural optimization and local detail enhancement to perform topological decoupling optimization of the human body. While ensuring global structural geometric consistency, it achieves precise restoration of local topological details, improving the quality of single-view clothing-reconstruction results. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention;

[0041] Figure 2 This is a flowchart of the normal confidence estimation process of the present invention;

[0042] Figure 3 This is a flowchart of the edge-aware smoothing process of the present invention;

[0043] Figure 4 This is a flowchart of the bilateral gradient optimization process of the present invention;

[0044] Figure 5 This is a flowchart illustrating the enhancement of local details in this invention;

[0045] Figure 6 This is a flowchart of the regrinding process of the present invention;

[0046] Figure 7 This is an example diagram of the three-dimensional human body reconstruction result of the present invention. Detailed Implementation

[0047] Example 1: As Figures 1-7 As shown, a single-view human reconstruction method based on geometric perception and consistency constraints includes:

[0048] Step 1: Generating the initial human normal map based on single-view images from the dressed human body dataset. and human body grid Geometric consistency modeling is performed through normal confidence estimation and adaptive normal fusion to obtain vertex-level fused normals. ;

[0049] Step 2: Edge-aware smoothing and bilateral gradient optimization based on local features for human body mesh Merge normals with vertex level Edge-aware gradient optimization is performed to obtain the optimized momentum. ;

[0050] Step 3: Optimize momentum through multi-level structural optimization and local detail enhancement. and the initial human body mesh Perform topological decoupling optimization to obtain geometric displacement. With detail displacement ;

[0051] Step 4: Geometric displacement and detailed displacement Overlay onto human body mesh The remesh was then performed to obtain high-quality clothing human reconstruction results. .

[0052] The specific process of Step 1 is as follows:

[0053] First, an initial human normal map is generated based on the diffusion of single-view human images from the dressed human dataset. A depth-separable convolutional residual encoder is used to... Encoding as high-dimensional encoded features Enhanced features are obtained through cross-view feature fusion. And mapped to pixel-level confidence maps. ;

[0054] Then, based on the human body mesh Through perspective any valid pixel Establish adjacent view pairs using depth information and camera extrinsic parameters Pixel correspondence set Define the adjacent view normal consistency loss ,in Indicates perspective medium pixel The unit normal vector of the corresponding surface point in the world coordinate system. Represents the spatial reprojection mapping operator;

[0055] Finally, the normal map of the pixel domain is obtained through projection relationships. and confidence plot Spatial sampling is performed to explicitly associate 2D features with 3D mesh vertices. This yields a set of normals from multiple perspectives. and the corresponding vertex-level normal confidence. Define perspective lower vertex Fusion weights ,in For visibility weight, As the basic weight of perspective, Weighted fusion of normals from multiple viewpoints is performed based on the direction of the line of sight. ,in Representing the normalized weights, we obtain the vertex-level fused normals. .

[0056] The specific process of Step 2 is as follows:

[0057] Step 2.1 Edge-aware smoothing;

[0058] First, based on vertex-level fusion normals Rendering yields the target rendering normal map. and target mask Construct the basic reconstruction objective function ;in For normal loss, for Mask loss;

[0059] Then, extract the vertices. Discrete Gaussian curvature Absolute value of mean curvature and dihedrals The local geometric features are normalized, and then... and Constructing a binary edge mask The marginal probability is obtained by combining normalized features. , build Adaptive regularization strength ;

[0060] Finally, combining Cotangent weight Laplace as regularization term Add to the basic objective function In the middle, by taking the derivative with respect to the vertex coordinates ,in As vertex The discrete Laplace differential coordinates at the point are used to obtain the spatially adaptive geometric gradient. .

[0061] Step 2.2 Bilateral gradient optimization;

[0062] First, the geometric gradient is calculated using an exponential moving average. Converted into momentum The vertex normals of the SMPL-X template are extracted and K-means clustering is performed to obtain stable sampling directions. ;

[0063] Then, the sampling direction With normal Perform dot product calculations to construct the geometric guiding factor. ,by As the initial signal, a weighted implicit thermal diffusion is performed on the mesh surface using a sparse matrix system, and the translational sliding term is obtained respectively. With normalization term ;

[0064] Finally, by aggregating sampling features and And then normalize it, specifically as follows: ,in For aggregation weight operators, A constant term is used to ensure numerical stability. The optimized momentum is obtained. .

[0065] The specific process of Step 3 is as follows:

[0066] Step 3.1 Multi-level structure optimization;

[0067] First, set the skinning weight vector of the human mesh vertex to Determine the joint index based on the maximum weight of vertex skinning. The set of mesh vertices is divided into non-overlapping hierarchical subsets according to a rule from the core to the fine tip. ;

[0068] Then, structural regularization constraints are applied to the human body mesh based on skeletal topology, anatomical proportions, and component connections to construct a structure that includes skeletal length constraints. Consistency of proportions Connection constraints of components Global structural prior loss ;

[0069] Finally, through and optimize momentum and combined with hierarchical activators Second-order moment estimation Learning rate and local reference side length The spatial adaptive geometric displacement of the vertex is obtained. in This is a constant term.

[0070] Step 3.2 Enhance local details;

[0071] First, regarding high-frequency changes Hierarchical structure, constructing vertex feature vectors that include local feature embeddings, vertex positions, and normal vectors. ;

[0072] Then, the vertex feature vectors Mapped to a higher-order nonlinear feature space, the normalized original displacement residuals are predicted by multilayer perceptron regression. and the corresponding spatial confidence scalar and by scaling factor Limited to the maximum displacement range;

[0073] Finally, Decomposed into normal components and tangential components Combining spatial confidence scalar Obtain detailed displacement , For normal weights, For tangential weights.

[0074] The present invention also provides a single-view human body reconstruction system based on geometric perception and consistency constraints, the system comprising: a module for executing the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0075] The present invention also 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 executes the program to implement the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0076] The present invention also 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 single-view human body reconstruction method based on geometric perception and consistency constraints.

[0077] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the single-view human body reconstruction method based on geometric perception and consistency constraints.

[0078] Example 2: Figure 1 As shown, the specific steps of this method are as follows:

[0079] Step 1: Generating the initial human normal map based on single-view images from the dressed human body dataset. and human body grid Geometric consistency modeling is performed through normal confidence estimation and adaptive normal fusion to obtain vertex-level fused normals. ;

[0080] Step 1 specifically includes: First, generating a normal map based on the single-view image diffusion from the clothing human dataset. ,like Figure 2 As shown, a depthwise separable convolutional residual encoder is used for feature extraction. By cascading 3×3 convolution, batch normalization, and ReLU activation, the normal map is mapped into a high-dimensional encoded feature with 64 channels. .right Perform linear transformations to obtain the query matrices respectively. Key matrix Sum matrix Calculate cross attention Obtain cross-perspective context features The number of attention heads is 4, and the dimension of each head is 16. Upsampled to the original resolution and corresponding viewpoint Residual fusion yields cross-view consistency enhancement features. Through confidence regression, proceed sequentially. Convolutional layers and the Sigmoid activation function are mapped to pixel-level confidence maps. ;

[0081] Then, based on the human body mesh , perspective any valid pixel The depth information and camera extrinsic parameters are used to establish adjacent viewpoint pairs. Pixel correspondence set Define the adjacent view normal consistency loss ,in Indicates perspective medium pixel The unit normal vector of the corresponding surface point in the world coordinate system. Represents the spatial reprojection mapping operator;

[0082] Finally, the normal map of the pixel domain is obtained through projection relationships. and confidence plot Spatial sampling is performed to explicitly associate 2D features with 3D mesh vertices. Thus, a set of multi-view normals is obtained. and the corresponding vertex-level normal confidence. Define perspective lower vertex Fusion weights ,in For visibility weight, As the basic weight of perspective, Weighted fusion of normals from multiple viewpoints is performed based on the direction of the line of sight. ,in This represents the normalized weights, ultimately yielding the vertex-level fused normals. .

[0083] Step 2: Edge-aware smoothing and bilateral gradient optimization based on local features for human body mesh Merge normals with vertex level Edge-aware gradient optimization is performed to obtain the optimized momentum. Step 2 includes:

[0084] Step 2.1: First, based on vertex-level fusion normals... Rendering yields the target rendering normal map. and target mask Construct the basic reconstruction objective function , and These represent the initial mesh rendering normals and contour masks, respectively. As for perspective weight, For the number of effective pixels, Represents a pixel;

[0085] Then, as Figure 3 As shown, extract the vertices. Discrete Gaussian curvature Absolute value of mean curvature and dihedrals The local geometric features are obtained, and feature processing is performed to obtain normalized features, which are then used in marginal probability estimation according to... The weighted fusion is performed based on the weighted allocation to obtain a preliminary score. ,use and Determine potential edge points to obtain a binary edge mask. ,pass and Obtain marginal probabilities , build Adaptive regularization strength ;

[0086] Finally, combining Combined with the cotangent weight Laplace as a regularization term Add to the basic objective function In the middle, by taking the derivative with respect to the vertex coordinates ,in As vertex The discrete Laplace differential coordinates at the point are used to ultimately generate the spatially adaptive geometric gradient. .

[0087] Step 2.2, as follows Figure 4 As shown, the exponential moving average (EMA) is used to... Converted into momentum The vertex normals of the SMPL-X template are extracted and K-means clustering is performed to obtain stable sampling directions. ;

[0088] Then, the sampling direction With normal Constructing geometric guiding factors by performing dot product calculations ,by As the initial signal, a weighted implicit thermal diffusion is performed on the mesh surface using a sparse matrix system, and the translational sliding term is solved. With normalization term ;

[0089] Finally, by aggregating sampling features and And then normalize it, specifically as follows: ,in For aggregation weight operators, As a constant term, the optimized momentum is obtained. .

[0090] Step 3: Optimize momentum through multi-level structural optimization and local detail enhancement. and the initial human body mesh Perform topological decoupling optimization to obtain geometric displacement. With detail displacement Step 3 includes:

[0091] Step 3.1: First, set the skinning weight vector of all mesh vertices to... Determine the joint index based on the maximum weight of vertex skinning. The set of mesh vertices is divided into non-overlapping hierarchical subsets according to a rule from the core to the fine tip. ;

[0092] Then, structural regularization constraints are applied to the human body mesh based on skeletal topology, anatomical proportions, and component connections to construct a structure that includes skeletal length constraints. Consistency of proportions Connection constraints of components Global structural prior loss Among them, skeleton consistency weight proportional constraint weight Component connection weight ;

[0093] Finally, through and optimize momentum and combined with hierarchical activators Second-order moment estimation Learning rate and local reference side length The spatial adaptive geometric displacement of the vertex is obtained. , This is a constant term.

[0094] Step 3.2, Targeting high-frequency changes Hierarchical structure, such as Figure 5 As shown, the vertex feature vector is obtained by encoding local features, vertex positions, and normal vectors through an encoder. ;

[0095] Then, the vertex feature vectors After feature processing and mapping to a higher-order nonlinear feature space, we obtain Through multilayer perceptron Displacement regression and confidence prediction yield the original displacement residuals and the corresponding spatial confidence scalar And by scaling factor Limiting to the maximum displacement range Inside;

[0096] Finally, Decomposed into normal components and tangential components , combined Obtain detailed displacement .

[0097] Step 4: Geometric displacement and detailed displacement Overlay onto human body mesh The remesh was then performed to obtain high-quality clothing human reconstruction results. Specifically:

[0098] like Figure 6 As shown, through Geometric displacement and high-frequency detail displacement Overlay onto human body mesh The vertices are remeshed by splitting the long sides, folding the short sides, and flipping the sides to obtain a fine human body mesh. The result is as follows Figure 7 As shown.

[0099] In summary, compared with related known methods, the method of the present invention has higher reconstruction accuracy and more stable topological structure in terms of human body geometric consistency, edge structure preservation, and high-frequency clothing detail reconstruction.

[0100] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A single-view human body reconstruction method based on geometric perception and consistency constraints, characterized in that: Includes the following steps: Step 1: Based on the initial human normal map and human mesh generated from the single-view images in the clothing human dataset, geometric consistency modeling is performed through normal confidence estimation and adaptive normal fusion to obtain vertex-level fused normals; Step 2: Edge-aware smoothing and bilateral gradient optimization based on local features are performed on the human mesh and vertex-level fusion normals to obtain optimized momentum. Step 3: Perform topological decoupling optimization on the optimized momentum and the initial human body mesh through multi-level structural optimization and local detail enhancement to obtain geometric displacement and detail displacement; Step 4: Overlay the geometric displacement and detail displacement onto the human body mesh and remesh it to obtain a high-quality clothing human body reconstruction result.

2. The single-view human body reconstruction method based on geometric perception and consistency constraints according to claim 1, characterized in that: The specific process of Step 1 is as follows: First, an initial human normal map is generated based on the diffusion of single-view human images from the dressed human dataset. A depth-separable convolutional residual encoder is used to... Encoding as high-dimensional encoded features Enhanced features are obtained through cross-view feature fusion. And mapped to pixel-level confidence maps. ; Then, based on the human body mesh Through perspective any valid pixel Establish adjacent view pairs using depth information and camera extrinsic parameters Pixel correspondence set Define the adjacent view normal consistency loss ,in Indicates perspective medium pixel The unit normal vector of the corresponding surface point in the world coordinate system. Represents the spatial reprojection mapping operator; Finally, the normal map of the pixel domain is obtained through projection relationships. and confidence plot Spatial sampling is performed to explicitly associate 2D features with 3D mesh vertices. This yields a set of normals from multiple perspectives. and the corresponding vertex-level normal confidence. Define perspective lower vertex Fusion weights ,in For visibility weight, As the basic weight of perspective, Weighted fusion of normals from multiple viewpoints is performed based on the direction of the line of sight. ,in Representing the normalized weights, we obtain the vertex-level fused normals. .

3. The single-view human body reconstruction method based on geometric perception and consistency constraints according to claim 1, characterized in that: Step 2, the edge-aware gradient optimization includes: Step 2.1 Edge-aware smoothing; First, based on vertex-level fusion normals Rendering yields the target rendering normal map. and target mask Construct the basic reconstruction objective function ;in For normal loss, for Mask loss; Then, extract the vertices. Discrete Gaussian curvature Absolute value of mean curvature and dihedrals The local geometric features are normalized, and then... and Constructing a binary edge mask The marginal probability is obtained by combining normalized features. , build Adaptive regularization strength ; Finally, combining Cotangent weight Laplace as regularization term Add to the basic objective function In the middle, by taking the derivative with respect to the vertex coordinates ,in As vertex The discrete Laplace differential coordinates at the point are used to obtain the spatially adaptive geometric gradient. ; Step 2.2 Bilateral gradient optimization; First, the geometric gradient is calculated using an exponential moving average. Converted into momentum The vertex normals of the SMPL-X template are extracted and K-means clustering is performed to obtain stable sampling directions. ; Then, the sampling direction With normal Perform dot product calculations to construct the geometric guiding factor. ,by As the initial signal, a weighted implicit thermal diffusion is performed on the mesh surface using a sparse matrix system to obtain the translational sliding term. With normalization term ; Finally, by aggregating sampling features and And then normalize it, specifically as follows: ,in For aggregation weight operators, A constant term is used to ensure numerical stability. The optimized momentum is obtained. .

4. The single-view human body reconstruction method based on geometric perception and consistency constraints according to claim 1, characterized in that: The topology hierarchical decoupling optimization in Step 3 includes: Step 3.1 Multi-level structure optimization; First, set the skinning weight vector of the human mesh vertex to Determine the joint index based on the maximum weight of vertex skinning. The set of mesh vertices is divided into non-overlapping hierarchical subsets according to a rule from the core to the fine tip. ; Then, structural regularization constraints are applied to the human body mesh based on skeletal topology, anatomical proportions, and component connections to construct a structure that includes skeletal length constraints. Consistency of proportions Connection constraints of components Global structural prior loss ; Finally, through and optimize momentum and combined with hierarchical activators Second-order moment estimation Learning rate and local reference side length The spatial adaptive geometric displacement of the vertex is obtained. ,in For constant terms; Step 3.2 Enhance local details; First, regarding high-frequency changes Hierarchical structure, constructing vertex feature vectors that include local feature embeddings, vertex positions, and normal vectors. ; Then, the vertex feature vectors Mapped to a higher-order nonlinear feature space, the normalized original displacement residuals are predicted by multilayer perceptron regression. and the corresponding spatial confidence scalar and by scaling factor Limited to the maximum displacement range; Finally, Decomposed into normal components and tangential components Combining spatial confidence scalar Obtain detailed displacement , For normal weights, For tangential weights.

5. A single-view human body reconstruction system based on geometric perception and consistency constraints, characterized in that, The system includes a module for performing the single-view human reconstruction method based on geometric perception and consistency constraints as described in any one of claims 1 to 4.

6. 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 single-view human body reconstruction method based on geometric perception and consistency constraints as described in any one of claims 1 to 4.

7. 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 single-view human body reconstruction method based on geometric perception and consistency constraints as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the single-view human body reconstruction method based on geometric perception and consistency constraints as described in any one of claims 1 to 4.

Images