A dynamic geometry super-resolution method for digital garment simulation detail enhancement

Abstract

The application discloses a dynamic geometric super-resolution method for digital garment simulation detail enhancement, which takes a low-resolution garment simulation sequence as input, realizes high-resolution and high-frequency geometric detail enhancement, and generates a high-fidelity digital garment deformation sequence with vivid and delicate wrinkle details. The method combines the physical dynamic characteristics of cloth and the interaction information between clothes and bodies, is based on supervised deep learning, realizes the extraction of super-resolution features on a non-regular grid body geometric surface, uses the super-resolution features to construct a garment dynamic geometric detail up-sampling implicit function through a super-parameter network, and generates high-resolution geometric details. The method corrects the low-resolution grid body geometric deformation while generating high-resolution geometric details, so that the quality of the generated garment details is greatly improved. The application provides a lightweight and efficient garment detail enhancement method, which helps to apply high-fidelity garment real-time simulation to low-budget and universal application scenarios.

Description

Technical Field

[0001] This invention belongs to the fields of computer graphics and artificial intelligence, and in particular, it is a dynamic geometric super-resolution method for enhancing the details of digital clothing simulation. Background Technology

[0002] Clothing is a core element of virtual environments, and clothing simulation is a core research area in virtual reality. Numerical physical simulation technology has been the mainstream technology for clothing simulation for nearly half a century, yielding a number of excellent results. Some relatively mature digital clothing simulation software, such as Marvelous Designer and Style3D, have been put into commercial use. However, because high-fidelity digital clothing relies on high-resolution geometric representation, the computational cost of existing physical simulation technologies is expensive, and they consume a large amount of data storage and transmission resources, making them unsuitable for low-budget, widely applicable scenarios. Therefore, how to efficiently and energy-savingly generate high-fidelity clothing simulation animations remains a highly challenging and urgent problem.

[0003] In contrast, another approach is to use digital garments with low-resolution geometric representations in physical simulations. This approach is computationally efficient but sacrifices the subtlety and realism of the garment's dynamics. Over the past decade, researchers have explored how to use low-resolution geometric representations in garment simulation calculations, data storage, and transmission. Then, based on low-resolution geometric deformation, they can enhance the high-frequency information of the garment's geometric model surface to generate high-resolution, high-fidelity garment simulation deformation sequences with rich wrinkle details. Examples include: Wang Huaming, 2021, GPU-Based Simulation of Cloth Wrinkles at Submillimeter Levels, ACM Trans.Graph, Article 169; Zhang Meng, Wang Tuanfeng, Ceylan Duygu, Mitra Niloy J., 2021, Deep Detail Enhancement for Any Garment, Computer Graph Forum, Vol.40, 399-411. However, these methods require multiple iterative geometric deformation optimizations based on high-resolution meshes, still consuming significant computational resources. Therefore, the industry urgently needs a new method to generate detailed and rich high-frequency details of fabric folds based on low-resolution clothing simulation deformation sequences, and to stably and efficiently output high-resolution clothing geometric deformation sequences. Summary of the Invention

[0004] The purpose of this invention is to address the challenges of high-fidelity digital clothing relying on high-resolution geometric representation and requiring a large amount of computing resources, while low-resolution geometric representation simulation is efficient but sacrifices the dynamic details and realism of digital clothing. This invention provides a method for enhancing details in low-resolution clothing simulation.

[0005] The technical solution to achieve the purpose of this invention is: a dynamic geometric super-resolution method for enhancing details in digital clothing simulation, the method comprising the following steps:

[0006] Step 1: Generate training data;

[0007] Step 2, construct a dynamic detail augmentation neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0008] Step 3: Using the training data from Step 1, optimize the learnable network parameters in the dynamic detail augmentation neural network;

[0009] Step 4: Model a low-resolution geometric mesh template of the target digital garment, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0010] Step 5: Based on the digital clothing templates obtained in Step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: Based on the correspondence of mesh volumes in texture space, obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex, and the upsampling coefficient s. f And further construct an upsampling matrix

[0011] Step 6, use the target human motion sequence {B} t The low-resolution geometric mesh template constructed in step 4 is used to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0012] Step 7, at time frame t, use the graph neural network from step 2. Calculate the deformation C of the low-resolution geometric mesh obtained in step 6. t All mesh patch super-resolution features

[0013] Step 8: Using the super-resolution features of the mesh patch obtained in Step 7 As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε from step 2. t To achieve low-resolution geometric deformation correction:

[0014] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork Using the upsampling coefficients s in step 5 f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0015] Use the upsampling matrix from step 5 Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0016] Step 9: Based on the prediction results of the current time frame t obtained in Step 8, iteratively predict the high-resolution geometric mesh deformation of the clothing in the next time frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0017] Further, the generation of training data in step 1 specifically includes:

[0018] A variety of different styles of digital clothing are modeled using low-resolution geometric meshes, and high-resolution geometric mesh volumes are generated that are aligned with them in texture space.

[0019] Prepare multiple human motion sequences, and use a physics simulator to generate dynamic clothing low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences under different human motion drives.

[0020] Based on the alignment relationship of the mesh in texture space, and the deformation sequence of the high-resolution geometric mesh generated by the physical simulator, the deformation sequence of the downsampled low-resolution geometric mesh is obtained by mesh resolution downsampling calculation.

[0021] Furthermore, the optimization of learnable network parameters in the dynamic detail augmentation network described in step 3 specifically includes:

[0022] Using the low-resolution geometric mesh deformation sequence of clothing generated by the simulator and the human motion posture as input, the dynamic detail enhancement neural network predicts and outputs the corrected low-resolution geometric mesh deformation of clothing. Based on multiple local blocks obtained by random cropping in texture space, a high-resolution clothing geometric mesh local block {P} with enhanced details is generated;

[0023] Construct the target loss function By correcting the results of geometric deformation After detail enhancement, the local block {P} of the high-resolution clothing geometry mesh approximates the corresponding ground truth value in the training data. and {P * To achieve end-to-end joint optimization of graph neural networks Geometric correction decoder ε and hyperparameter neural subnetwork All learnable parameters.

[0024] Furthermore, the target loss function It is expressed as follows:

[0025]

[0026] in, This indicates the deformation of the corrected low-resolution geometric mesh. constraint functions, Let λ denote the constraint function for the prediction result {P} of the detail augmentation network. c and λ p These are the coefficients of the constraint terms.

[0027] Furthermore, a graph neural network is used in step 7. Generating super-resolution features specifically includes:

[0028] Based on the current frame t and historical frames t-1, the low-resolution geometric mesh volume deformation C of t-2 t C t-1 C t-2 Human movement posture B t B t-1 and historical frame low-resolution geometric deformation correction results Calculate the node eigenvectors and edge eigenvectors of the graph topology, where each graph topology node corresponds one-to-one with a vertex of the low-resolution geometric mesh;

[0029] Using the graph neural network in step 2 Generate low-resolution geometric mesh vertex super-resolution features i represents the vertex i of the low-resolution geometric mesh. This represents the super-resolution features of vertex i of the low-resolution geometric mesh in the current frame t, followed by the super-resolution features of the low-resolution geometric mesh volume vertices. Generate super-resolution features of mesh patches

[0030] Furthermore, the graph neural network used in step 2... Generate low-resolution geometric mesh vertex super-resolution features Specifically, it includes:

[0031] encoder Projecting the node feature vectors and edge feature vectors into the learned latent feature space yields the initial node features q in the latent space. i and edge features e ij , j represents the index of the vertex connected to vertex i;

[0032] Using node features and edge features as graph convolutional neural networks The input is used to iteratively update node and edge features layer by layer based on the graph structure and topology:

[0033]

[0034]

[0035] in, and These represent graph convolutional neural networks. In a fully connected neural network, l represents the index of a vertex in a different cloth layer that collides with point i, and m represents the number of layers in the convolutional network; Let i and y represent the edge features between vertices i and y in the (m-1)th and m-th convolutional layers, respectively. This represents the edge features between vertex i and vertex l in the (m-1)th convolutional layer; Let i and j represent the node features of vertices i and j in the (m-1)th convolutional layer, respectively. Represents the node features of vertex i in the m-th convolutional layer;

[0036] Graph Convolutional Neural Network The node features output from the last layer are input into the super-resolution decoder. Output the super-resolution features of the vertices of the low-resolution geometry for each vertex i of the current frame.

[0037] Furthermore, the super-resolution features of the low-resolution geometric mesh vertices... Generate super-resolution features of mesh patches Specifically, this includes: the super-resolution features of a low-resolution mesh patch f are composed of the super-resolution features of all vertices of the patch arranged in the topological order of the vertices, thereby realizing the super-resolution features of the vertices of the low-resolution geometric mesh. Constructing super-resolution features of mesh patches

[0038] Furthermore, step 9 iteratively predicts the high-resolution geometric mesh deformation of the clothing in the next time frame t+1, specifically including:

[0039] Human motion posture B at time frame t and the next time frame t+1 t B t+1 The deformation C of the corresponding low-resolution geometric mesh generated by simulation in step 6. t-1 C t C t+1 and the low-resolution geometric deformation correction obtained in step 8. As input to the dynamic detail enhancement neural network, it predicts the low-resolution geometric mesh volume deformation correction vector D for the next time frame t+1. t+1 and detail residual vector R t+1 This leads to the generation of high-resolution geometric mesh deformation G for clothing. t+1 .

[0040] On the other hand, a dynamic geometric super-resolution system for enhancing details in digital clothing simulation is provided, the system comprising sequentially executing:

[0041] The first module is used to generate training data;

[0042] The second module is used to construct a dynamic detail-enhancing neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0043] The third module is used to optimize the learnable network parameters in the dynamic detail augmentation neural network using the training data.

[0044] The fourth module is used to: model a low-resolution geometric mesh template of the target digital clothing, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0045] The fifth module is used to: solve the mapping from low-resolution mesh topology to high-resolution mesh topology based on the digital clothing templates obtained in the fourth module; and obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex and the upsampling coefficient s based on the correspondence of mesh volumes in texture space. f And further construct an upsampling matrix

[0046] The sixth module is used to implement: using the target human motion sequence {B t The driver uses the low-resolution geometric mesh template constructed by the fourth module to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0047] The seventh module is used to implement the graph neural network from the second module at time frame t. Calculate the deformation C of the low-resolution geometric mesh obtained from the sixth module. t All mesh patch super-resolution features

[0048] The eighth module is used to implement the super-resolution features of the mesh patches obtained in the seventh module. As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε in the second module. t To achieve low-resolution geometric deformation correction:

[0049] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork With the upsampling coefficients s f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0050] Using the upsampling matrix Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0051] The ninth module is used to iteratively predict the high-resolution geometric mesh deformation of the garment in the next time frame t+1 based on the prediction results of the current time frame t obtained by the eighth module, until the digital garment detail enhancement of all sequence frames is completed.

[0052] On the other hand, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the dynamic geometric super-resolution method for enhancing details in digital clothing simulation.

[0053] Compared with the prior art, the significant advantages of this invention are:

[0054] (1) This invention learns how to extract super-resolution features of dynamic clothing, and achieves high-resolution, high-frequency detail enhancement on the surface of low-resolution clothing mesh, generating a high-resolution clothing simulation deformation sequence with vivid and delicate fold details.

[0055] (2) This invention provides a lightweight and efficient method for enhancing clothing details, which helps to apply high-fidelity clothing real-time simulation to low-budget, universal scenarios, such as smartphones.

[0056] (3) Compared with existing clothing detail enhancement methods, this invention takes into account the impact of low-resolution clothing deformation on clothing details. Based on dynamic clothing super-resolution features, it corrects low-resolution clothing geometric deformation while generating high-resolution details, thereby greatly improving the quality of the generated clothing details.

[0057] (4) This invention is not limited to a specific physical simulation method and can be compatible with different numerical solution physical simulation systems.

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0059] Figure 1 This is a flowchart of the dynamic geometric super-resolution method for enhancing details in digital clothing simulation according to the present invention.

[0060] Figure 2 This is a neural network structure diagram of the dynamic geometric super-resolution method for enhancing details in digital clothing simulation according to the present invention.

[0061] Figure 3 This is a schematic diagram of a low-resolution mesh digital clothing simulation input in one embodiment.

[0062] Figure 4 This is a low-resolution digital garment geometry correction result diagram in one embodiment.

[0063] Figure 5This is a high-resolution digital clothing result image with enhanced detail in one embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0066] In one embodiment, combined Figure 1 A dynamic geometric super-resolution method for enhancing details in digital clothing simulation is provided, the method comprising the following steps:

[0067] Step 1: Generate training data;

[0068] Step 2, construct a dynamic detail augmentation neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0069] Step 3: Using the training data from Step 1, optimize the learnable network parameters in the dynamic detail augmentation neural network;

[0070] Step 4: Model a low-resolution geometric mesh template of the target digital garment, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0071] Step 5: Based on the digital clothing templates obtained in Step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: Based on the correspondence of mesh volumes in texture space, obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex, and the upsampling coefficient s. f And further construct an upsampling matrix

[0072] Step 6, use the target human motion sequence {B} t The low-resolution geometric mesh template constructed in step 4 is used to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0073] Step 7, at time frame t, use the graph neural network from step 2. Calculate the deformation C of the low-resolution geometric mesh obtained in step 6. t All mesh patch super-resolution features

[0074] Step 8: Using the super-resolution features of the mesh patch obtained in Step 7 As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε from step 2. t To achieve low-resolution geometric deformation correction: (Specifically: super-resolution features of the geometric mesh patches in step 6) As the input to the geometry correction decoder ε in step 2, the network output is the geometry correction vector d of all vertices of the mesh patch. f Next, the geometric correction vector for each vertex i of the low-resolution mesh is calculated. in Let be the number of mesh faces connected to vertex i; then, correct the geometric position c of each vertex in the low-resolution mesh to obtain the corrected vertex position. Obtain the low-resolution clothing geometry after correction in the current frame );

[0075] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork Using the upsampling coefficients s in step 5 f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0076] Use the upsampling matrix from step 5 Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0077] Step 9: Based on the prediction results of the current time frame t obtained in Step 8, iteratively predict the high-resolution geometric mesh deformation of the clothing in the next time frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0078] Furthermore, in one embodiment, step 1, generating training data, specifically includes:

[0079] A variety of different styles of digital clothing are modeled using low-resolution geometric meshes, and high-resolution geometric mesh volumes are generated that are aligned with them in texture space.

[0080] Prepare multiple human motion sequences, and use a physics simulator to generate dynamic clothing low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences under different human motion drives.

[0081] Based on the alignment of the mesh in texture space, and using the high-resolution geometric mesh deformation sequence generated by the physics simulator, a downsampled low-resolution geometric mesh deformation sequence is obtained by mesh resolution downsampling calculation.

[0082] Furthermore, in one embodiment, the optimization of learnable network parameters in the dynamic detail enhancement network in step 3 specifically involves using the clothing low-resolution geometric mesh deformation sequence {C} and the high-resolution geometric mesh deformation sequence {G} obtained in step 1. * Downsampled low-resolution geometric mesh volume deformation sequence Given a human motion sequence {B}, train and optimize the learnable network parameters in the dynamic detail augmentation network. Specifically, this includes:

[0083] Using the low-resolution geometric mesh deformation sequence of clothing generated by the simulator and the human motion posture as input, the dynamic detail enhancement neural network predicts and outputs the corrected low-resolution geometric mesh deformation of clothing. Based on multiple local patches obtained by random cropping in texture space, a high-resolution clothing geometric mesh local patch {P} with enhanced details is generated;

[0084] Construct the target loss function By correcting the results of geometric deformation After detail enhancement, the local block {P} of the high-resolution clothing geometry mesh approximates the corresponding ground truth value in the training data. and {P *To achieve end-to-end joint optimization of graph neural networks Geometric correction decoder ε and hyperparameter neural subnetwork All learnable parameters.

[0085] Here, {P *} represents the high-resolution clothing geometry G * Multiple high-resolution local block geometries are obtained by clipping the corresponding local block {P}.

[0086] Preferably, in one embodiment, the target loss function It is expressed as follows:

[0087]

[0088] in, This indicates the deformation of the corrected low-resolution geometric mesh. constraint functions, Let λ denote the constraint function for the prediction result {P} of the detail augmentation network. c and λ p These are the coefficients of the constraint terms.

[0089] Preferably, in the process of neural network optimization, the constraint term coefficient λ is first... c and λ p Let it be λ c =1.9, λ p =0.1, then gradually decrease the coefficient λ c And increase the coefficient λ p Until it reaches 1, while maintaining the coefficient λ. c =λ p =1, until the loss function optimization converges.

[0090] Furthermore, in one embodiment, step 7 uses a graph neural network. Generating super-resolution features specifically includes:

[0091] Based on the current frame t and historical frames t-1, the low-resolution geometric mesh volume deformation C of t-2 t C t-1 C t-2 Human movement posture b t b t-1 and historical frame low-resolution geometric deformation correction results Calculate the node eigenvectors and edge eigenvectors of the graph topology, where each graph topology node corresponds one-to-one with a vertex of the low-resolution geometric mesh;

[0092] Using the graph neural network in step 2 Generate low-resolution geometric mesh vertex super-resolution features i represents the vertex i of the low-resolution geometric mesh. This represents the super-resolution features of vertex i of the low-resolution geometric mesh in the current frame t, followed by the super-resolution features of the low-resolution geometric mesh volume vertices. Generate super-resolution features of mesh patches

[0093] Preferably, in some embodiments, the use of the graph neural network in step 2 is described. Generate low-resolution geometric mesh vertex super-resolution features Specifically, it includes:

[0094] encoder Projecting the node feature vectors and edge feature vectors into the learned latent feature space yields the initial node features q in the latent space. i and edge features e ij , j represents the index of the vertex connected to vertex i;

[0095] Using node features and edge features as graph convolutional neural networks The input is used to iteratively update node and edge features layer by layer based on the graph structure and topology:

[0096]

[0097]

[0098] in, and These represent graph convolutional neural networks. In a fully connected neural network, l represents the index of a vertex in a different cloth layer that collides with point i, and m represents the number of layers in the convolutional network; Let i and y represent the edge features between vertices i and y in the (m-1)th and m-th convolutional layers, respectively. This represents the edge features between vertex i and vertex l in the (m-1)th convolutional layer; Let i and j represent the node features of vertices i and j in the (m-1)th convolutional layer, respectively. Represents the node features of vertex i in the m-th convolutional layer;

[0099] Graph Convolutional Neural Network The node features output from the last layer are input into the super-resolution decoder. Output the super-resolution features of the vertices of the low-resolution geometry for each vertex i of the current frame.

[0100] Preferably, in some embodiments, the super-resolution feature derived from the vertices of the low-resolution geometric mesh is... Generate super-resolution features of mesh patches Specifically, this includes: the super-resolution features of a low-resolution mesh patch f are composed of the super-resolution features of all vertices of the patch arranged in the topological order of the vertices, thereby realizing the super-resolution features of the vertices of the low-resolution geometric mesh. Constructing super-resolution features of mesh patches

[0101] Furthermore, in one embodiment, step 9 iteratively predicts the deformation of the garment's high-resolution geometric mesh at the next time frame t+1, specifically including:

[0102] Human motion posture B at time frame t and the next time frame t+1 t B t+1 The deformation C of the corresponding low-resolution geometric mesh generated by simulation in step 6. t-1 C t C t+1 and the low-resolution geometric deformation correction obtained in step 8. As input to the dynamic detail enhancement neural network, it predicts the low-resolution geometric mesh volume deformation correction vector D for the next time frame t+1. t+1 and detail residual vector R t+1 This leads to the generation of a high-resolution geometric mesh with rich fold details, and the resulting deformation G. t+1 .

[0103] In one embodiment, a dynamic geometric super-resolution system for enhancing details in digital clothing simulation is provided, the system comprising sequentially executing:

[0104] The first module is used to generate training data;

[0105] The second module is used to construct a dynamic detail-enhancing neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0106] The third module is used to optimize the learnable network parameters in the dynamic detail augmentation neural network using the training data.

[0107] The fourth module is used to: model a low-resolution geometric mesh template of the target digital clothing, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0108] The fifth module is used to: solve the mapping from low-resolution mesh topology to high-resolution mesh topology based on the digital clothing templates obtained in the fourth module; and obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex and the upsampling coefficient s based on the correspondence of mesh volumes in texture space. f And further construct an upsampling matrix

[0109] The sixth module is used to implement: using the target human motion sequence {B t The driver uses the low-resolution geometric mesh template constructed by the fourth module to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0110] The seventh module is used to implement the graph neural network from the second module at time frame t. Calculate the deformation C of the low-resolution geometric mesh obtained from the sixth module. t All mesh patch super-resolution features

[0111] The eighth module is used to implement the super-resolution features of the mesh patches obtained in the seventh module. As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε in the second module. t To achieve low-resolution geometric deformation correction:

[0112] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork With the upsampling coefficients s f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0113] Using the upsampling matrix Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0114] The ninth module is used to iteratively predict the high-resolution geometric mesh deformation of the garment in the next time frame t+1 based on the prediction results of the current time frame t obtained by the eighth module, until the digital garment detail enhancement of all sequence frames is completed.

[0115] Specific limitations regarding the dynamic geometric super-resolution system for enhancing details in digital clothing simulation can be found in the limitations of the dynamic geometric super-resolution method for enhancing details in digital clothing simulation described above, and will not be repeated here. Each module in the aforementioned dynamic geometric super-resolution system for enhancing details in digital clothing simulation can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0116] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0117] Step 1: Generate training data;

[0118] Step 2, construct a dynamic detail augmentation neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0119] Step 3: Using the training data from Step 1, optimize the learnable network parameters in the dynamic detail augmentation neural network;

[0120] Step 4: Model a low-resolution geometric mesh template of the target digital garment, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0121] Step 5: Based on the digital clothing templates obtained in Step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: Based on the correspondence of mesh volumes in texture space, obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex, and the upsampling coefficient s. f And further construct an upsampling matrix

[0122] Step 6, use the target human motion sequence {B} t The low-resolution geometric mesh template constructed in step 4 is used to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0123] Step 7, at time frame t, use the graph neural network from step 2. Calculate the deformation C of the low-resolution geometric mesh obtained in step 6. t All mesh patch super-resolution features

[0124] Step 8: Using the super-resolution features of the mesh patch obtained in Step 7 As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε from step 2. t To achieve low-resolution geometric deformation correction:

[0125] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork Using the upsampling coefficients s in step 5 f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0126] Use the upsampling matrix from step 5 Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0127] Step 9: Based on the prediction results of the current time frame t obtained in Step 8, iteratively predict the high-resolution geometric mesh deformation of the clothing in the next time frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0128] For specific limitations on each step, please refer to the limitations of the dynamic geometric super-resolution method for enhancing details in digital clothing simulation above, which will not be repeated here.

[0129] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0130] Step 1: Generate training data;

[0131] Step 2, construct a dynamic detail augmentation neural network, which includes: a graph neural network. Geometric correction decoder ε and hyperparameter neural network The graph neural network Super-resolution features used to generate low-resolution geometric mesh volumes, by the encoder Graph Convolutional Neural Network Super-resolution decoder The structure includes: the geometric correction decoder ε, used to solve for the low-resolution geometric mesh correction vector; and the hyperparameter neural network. Used to construct upsampled implicit fields for clothing dynamic details, including subnetworks Hezi Network The parameters are determined by Calculated;

[0132] Step 3: Using the training data from Step 1, optimize the learnable network parameters in the dynamic detail augmentation neural network;

[0133] Step 4: Model a low-resolution geometric mesh template of the target digital garment, and model a high-resolution geometric mesh template that is aligned with it in texture space;

[0134] Step 5: Based on the digital clothing templates obtained in Step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: Based on the correspondence of mesh volumes in texture space, obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex, and the upsampling coefficient s. f And further construct an upsampling matrix

[0135] Step 6, use the target human motion sequence {B} t The low-resolution geometric mesh template constructed in step 4 is used to simulate and generate a low-resolution geometric mesh deformation sequence {C}. t};

[0136] Step 7, at time frame t, use the graph neural network from step 2. Calculate the deformation C of the low-resolution geometric mesh obtained in step 6. t All mesh patch super-resolution features

[0137] Step 8: Using the super-resolution features of the mesh patch obtained in Step 7 As input, the low-resolution geometric mesh volume deformation correction vector D is solved using the geometric correction decoder ε from step 2. t To achieve low-resolution geometric deformation correction:

[0138] Meanwhile, using mesh patch super-resolution features For input, use the method from step 2 for each facet f. Construct the corresponding subnetwork Using the upsampling coefficients s in step 5 f Using this as input, compute the detail residual vectors of the vertices of the high-resolution geometric mesh. This yields the detail residual vector R of all high-resolution geometric mesh vertices. t ={r t};

[0139] Use the upsampling matrix from step 5 Achieve upsampling of clothing geometric deformation mesh resolution and calculate the dynamic detail enhancement of the high-resolution clothing geometric mesh deformation at the current time frame t:

[0140] Step 9: Based on the prediction results of the current time frame t obtained in Step 8, iteratively predict the high-resolution geometric mesh deformation of the clothing in the next time frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0141] For specific limitations on each step, please refer to the limitations of the dynamic geometric super-resolution method for enhancing details in digital clothing simulation above, which will not be repeated here.

[0142] As a specific example, the invention will be further described in detail in one embodiment.

[0143] This invention proposes a dynamic geometric super-resolution method to achieve high-resolution dynamic detail enhancement of low-resolution digital clothing, such as... Figure 1 As shown, it specifically includes:

[0144] (1) Generate training data: Use low-resolution triangular meshes with an average vertex distance of 30mm to model different styles of digital clothing: T-shirts, trousers, pleated skirts, single-layer lace dresses, and three-layer lace dresses, and obtain high-resolution triangular meshes with an average vertex distance of 10mm that are aligned with them in texture space to model clothing; prepare a 714-frame swing dance human motion sequence and a 794-frame house dance motion sequence with a frame rate of 30 frames per second, and use Marvelous Designer physics simulator to generate low-resolution clothing dynamic geometric motion sequences and high-resolution clothing sequences driven by human motion sequences; calculate the downsampled low-resolution clothing geometric sequence based on the high-resolution geometric sequence according to the alignment relationship of the meshes in texture space.

[0145] (2) Constructing a dynamic detail-enhancing neural network: such as Figure 2 As shown, the network consists of three main parts: a graph neural network. Geometric correction decoder ε and hyperparameter neural network

[0146] (2.1) Graph Neural Networks A parameter-learnable neural network consists of an encoder. A graph convolutional neural network A node feature decoder constitute;

[0147] (2.2) The geometric correction decoder ε is a multi-layer fully connected neural network, which is a parameter-learnable neural network;

[0148] (2.3) Hyperparameter Neural Networks Sub-network Hezi Network constitute, The parameters are determined by Obtained through calculation.

[0149] (3) Training the neural network constructed in step (2): using the low-resolution clothing geometry sequence {C} and the high-resolution clothing geometry sequence {G} obtained in step (1). *}, Downsampling low-resolution clothing geometric sequences Given the human motion sequence {B}, train and optimize the learnable network parameters in the dynamic detail augmentation network;

[0150] Using low-resolution clothing simulation and human motion posture as input, a detail-enhancing neural network predicts and outputs corrected low-resolution clothing geometry. Based on multiple patches obtained by random cropping in texture space, a high-resolution clothing geometric local block {P} with enhanced details is generated;

[0151] By minimizing the loss function End-to-end joint optimization of graph neural networks in augmented detail networks Geometric correction decoder ε and hyperparameter neural network All learnable parameters, loss function It is expressed as follows:

[0152]

[0153] in, Indicates reference downsampling clothing geometry Defined for low-resolution clothing geometry after correction The constraint function, in practical operation, Includes geometric constraints and deformation constraints Two parts:

[0154]

[0155] The expression is as follows:

[0156]

[0157] in, and These represent low-resolution geometric meshes. and The normal vector; The representation is as follows:

[0158]

[0159] Here, Δ represents the Laplacian operator, θ represents the angle between adjacent faces of the mesh; and the mesh is calculated according to Baraff David, Andrew Witkin, 2023, Large steps in cloth simulation, in Seminal Graphics Papers: Pushing the Boundaries, Vol. 2, 767-778. deformation matrix And through the equation:

[0160]

[0161] Make the geometry and The deformations are similar. {P *} represents the high-resolution clothing geometry G *Multiple high-resolution local block geometries are obtained by cropping the corresponding local block {P}. Represents the reference local block geometry {P * The detail deformation constraints defined for the prediction result {P} of the detail augmentation network, in practice, Includes geometric constraints and deformation constraints Two parts:

[0162]

[0163] The expression is as follows:

[0164]

[0165] The normal vector of a local block P of the mesh is drawn on an 80×80 image based on texture mapping, VGG l This represents the latent feature encoding of an image in the l-th layer of a pre-trained neural network VGG (Karen Simonyan and Andrew Zisserman, 2015, VeryDeep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations.). Represented as:

[0166]

[0167] In the process of neural network optimization, the constraint term coefficient λ is first... c and λ p Let it be λ c =1.9, λ p =0.1, then gradually decrease the coefficient λ c And increase the coefficient λ p Until it reaches 1, while maintaining the coefficient λ. c =λ p =1, until the loss function is optimized and converged; the optimized detailed augmented neural network parameters occupy about 64M of storage space, which is suitable for low-budget devices with limited storage space.

[0168] (4) Model a low-resolution triangular mesh template with an average vertex distance of 30mm for the target digital clothing, and model a high-resolution triangular mesh template with an average vertex distance of 10mm aligned with it in texture space; based on the correspondence of the mesh in texture space, obtain the low-resolution mesh face f corresponding to each high-resolution mesh vertex, and the upsampling coefficient s. f And further construct an upsampling matrix

[0169] (5) Using the target human motion sequence {B t The low-resolution triangular mesh template constructed in driving step (4) is used to simulate and generate the geometric deformation sequence {C}. t}, Figure 3 The image shown is a low-resolution simulation of the geometric deformation of clothing in one frame.

[0170] (6) Use the graph neural network from step (2) The low-resolution clothing C obtained in step (5) is calculated. t Super-resolution features at time frame t: based on the current frame t and historical frames t-1 and t-2, the low-resolution clothing C t C t-1 C t-2 B. Movement of the human body t B t-1 Geometric correction results of low-resolution clothing in historical frames Node eigenvectors of computational graph topology and edge feature vectors As a neural network in step (1) Input;

[0171] The feature vector of node i in the current frame t Represented as:

[0172]

[0173] in, For mesh vertices The normal vector, For speed, For acceleration, The clothing vertex at time t With the movement of human body B t The nearest neighbor b on t Interaction vectors:

[0174]

[0175] Where, σ b This is a preset threshold, which can be set to 0.3mm in actual operation. For b t The normal vector; Represents the vertex of the clothing at time t-1. With b t Interaction vectors:

[0176]

[0177] in, respectively The vector magnitude is represented; the influence of historical frame predictions on current frame predictions is also considered. express One of the vertices on, This indicates the result of low-resolution geometric correction. The vertex velocity is obtained through calculation.

[0178] The feature vector of the edge constructed by node i and node j in the current frame t It mainly consists of the edge vectors c of the mesh at time t and t-1. j -c i and side length ||c j -c i ||, and the side length of the garment in its initial static state o. In other words, the feature vector Ultimately, it is represented as:

[0179]

[0180] When node i and node l are located in different layers of the clothing fabric, and the distance between the nodes satisfies σ l If a preset threshold is set, edge features are dynamically added to the graph topology. mode of expression and Similar, but will Replace with σ l .

[0181] Obtain the node feature vectors of the computation graph topology. and edge feature vectors After that, the encoder First, the node feature vectors and edge feature vectors are projected onto the learned projection space to obtain the initial node features q in the latent space. i and edge features e ij i and j represent graph node indices, e ij The edge features formed by nodes i and j are represented; then, the node features and edge features are used as a graph convolutional neural network. The input, graph structure node features and edge features, are passed through a graph convolutional neural network and iteratively updated according to the graph structure topology:

[0182]

[0183]

[0184] in, and Representing neural networks In a multi-layer fully connected neural network, since collisions can occur between different layers of the clothing, l represents the graph node index of the collision point between different layers, and m represents the number of layers in the convolutional network.

[0185] To make graph convolutional neural networks more efficient, in practice, a single convolution operation can be performed on a second-order loop topology domain of a low-resolution mesh. A six-layer convolutional neural network can then effectively extract super-resolution feature vectors. Next, the neural network... The node features qi output from the last layer are input to the node feature decoder. Generate a super-resolution feature vector for each vertex of the low-resolution clothing geometry mesh for the current frame.

[0186] Furthermore, using super-resolution features Constructing super-resolution features of geometric mesh patches The super-resolved feature of a geometric mesh patch f is the super-resolved feature z of all vertices of that patch. i The facets are arranged according to their topological order. For example, the vertex index of a triangular facet is (f 0 f 1 f 2 If the triangle is a tangent, then the super-resolution feature of the triangle is:

[0187]

[0188] (7) Correcting low-resolution clothing geometry: Super-resolution features of geometric mesh patches from step (6) As the input to the geometric correction decoder ε in step (2), the network output is the geometric correction vector d of all vertices of the mesh patch. f For example, if the mesh patch is triangular, then the input and output of the geometry correction decoder ε are:

[0189]

[0190] Next, the geometric correction vector for each vertex i of the low-resolution mesh is calculated. Where |f| is the number of mesh patches connected to vertex i; then, the geometric position c of each vertex in the low-resolution mesh is corrected to obtain the corrected vertex position. Obtain the low-resolution clothing geometry after correction in the current frame like Figure 4 The image shows the result of low-resolution clothing geometry correction in one frame.

[0191] Simultaneously, high-resolution garment geometry deformation with enhanced details is generated: the super-resolution features of the geometric mesh patches in step (6) are then used. As the network in step (2) The input is used to construct the corresponding facet f for each facet. Then, using the high-resolution geometric upsampling coefficients s from step (4) f Using this as input, compute the detail residual vectors of the vertices of the high-resolution clothing geometry mesh. This yields the detail residual vector R of all high-resolution mesh vertices. t ={r t Next, combining the upsampling matrix in step (4) Corrected low-resolution clothing geometry and detail residual R t Calculate high-resolution clothing geometry for dynamic detail enhancement in the target frame. like Figure 5 The image shows a high-resolution clothing simulation result with enhanced details in one frame.

[0192] (8) When solving for high-resolution clothing detail enhancement at time t+1, the human motion posture B in step (5) is used. t+1 The low-resolution clothing geometric deformation C generated by simulation t-1 C t C t+1 And the low-resolution geometric deformation correction obtained from step (7) As the input to the dynamic detail enhancement neural network, the correction vector D predicts time t+1. t+1 and detail residual vector R t+1 Generate high-resolution garment geometry G with rich fold details. t+1 .

[0193] In summary, this invention uses low-resolution clothing simulation sequences as input to achieve high-resolution, high-frequency geometric detail enhancement, generating high-fidelity digital clothing deformation sequences with vivid and delicate fold details. This method combines the physical dynamic characteristics of the fabric and the interaction information between the clothing and the body. Based on supervised deep learning, it extracts super-resolution features from the irregular mesh geometric surface. Through a hyperparameter network, these super-resolution features are used to construct an implicit function for upsampling the dynamic geometric details of the clothing, generating high-resolution geometric details. This method corrects the low-resolution mesh geometric deformation while generating high-resolution geometric details, significantly improving the quality of the generated clothing details. This invention provides a lightweight and efficient method for enhancing clothing details, facilitating the application of high-fidelity real-time clothing simulation in low-budget, widely accessible scenarios.

[0194] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A dynamic geometry super-resolution method for digital garment simulation detail enhancement, characterized in that, The method comprises the following steps: Step 1, generating training data; Step 2, constructing a dynamic detail enhancement neural network, comprising: a graph neural network , a geometry correction decoder , and a hyperparameter neural network ; the graph neural network is used for generating super-resolution features of a low-resolution geometry mesh, composed of an encoder , a graph convolutional neural network , and a super-resolution decoder ; the geometry correction decoder is used for solving a low-resolution geometry mesh correction vector; the hyperparameter neural network is used for constructing a garment dynamic detail up-sampling implicit field, comprising a sub-network and a sub-network , parameters of which are obtained by calculation . Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network; Step 4, modeling the target digital garment low-resolution geometric mesh template, and modeling the high-resolution geometric mesh template aligned with it in the texture space; Step 5, according to each digital clothing template obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the correspondence of mesh body in texture space, obtain the corresponding low-resolution mesh patch of each high-resolution mesh vertex , and the up-sampling coefficient , and further construct the up-sampling matrix ; Step 6, using the target human motion sequence Drive the low resolution geometry mesh template constructed in step 4, simulate the generation of low resolution geometry mesh deformation sequence ; Step 7, at time frame , using the graph neural network in step 2 , compute the low-resolution geometry mesh deformation from step 6 , all mesh face super-feature of step 6 ; Step 8, mesh patch super-feature obtained from step 7 For input, use the geometry correction decoder in step 2 Solve low resolution geometry mesh deformation correction vector Obtain low resolution geometry deformation correction: ; Meanwhile, with the grid patch super-division feature For input, for each patch Use the feature in step 2 Corresponding sub-network is constructed ; With the up-sampling coefficient in step 5 As input, the detail residual vector of the high-resolution geometry grid vertex is calculated , thus obtaining the detail residual vector of all high-resolution geometry grid vertices ; Using the up-sampling matrix in step 5 Implementing garment geometry morphing mesh resolution up-sampling, computing the current time frame Dynamic detail enhanced garment high resolution geometry mesh morphing: ; Step 9, iteratively rollingly predict next time frame garment high resolution geometry mesh deformation based on prediction results from current time frame obtained by step 8 until digital garment detail enhancement of all sequence frames is completed.​ 2. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 1, wherein, Step 1 of generating training data specifically includes: Use low-resolution geometric mesh to model a plurality of different styles of digital garments, and generate high-resolution geometric mesh aligned with it in the texture space; Prepare a plurality of human motion sequences, and use a physical simulator to generate dynamic garment low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences under different human motion driving respectively; according to the alignment relationship of the mesh in the texture space, based on the high-resolution geometric mesh deformation sequence generated by the physical simulator, the down-sampling low-resolution geometric mesh deformation sequence is obtained by grid resolution down-sampling calculation.

3. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 1, wherein, The optimization of the learnable network parameters in the dynamic detail enhancement neural network in step 3 specifically includes: The dynamic detail enhancement neural network takes as input a sequence of garment low-resolution geometry mesh deformations and human motion poses generated by a simulator, and outputs low-resolution geometry deformation corrections , and generates detail-enhanced high-resolution garment geometry mesh local blocks based on multiple local blocks obtained through random cropping in texture space ; Constructing the target loss function by low-resolution geometry deformation correction and high-resolution garment geometry mesh local blocks after detail enhancement approximating the corresponding ground truth in the training data and all learnable parameters of the graph neural network , the geometry correction decoder and the hyperparameter neural subnetwork are optimized end-to-end.

4. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 3, wherein, The target loss function is represented as follows: ； wherein, represents a constraint function for low-resolution geometry morphing correction , represents a constraint function for high-resolution garment geometry mesh local patch after detail enhancement, , and are constraint term coefficients.

5. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 1, wherein, Using a graph neural network in step 7 Generating hyper-feature, specifically comprising: According to the current frame and historical frames Low-resolution geometric mesh deformation Human movement posture and historical frame low-resolution geometric deformation correction results Calculate the node feature vectors and edge feature vectors of the graph topology, and the graph topology nodes correspond one-to-one with the vertices of the low-resolution geometric mesh; using the graph neural network in step 2 generating low resolution geometry mesh vertex super-resolution features , representing low resolution geometry mesh vertices , representing a current frame low resolution geometry mesh vertices super-resolution features, followed by low resolution geometry mesh vertex super-resolution features generating mesh face super-resolution features .

6. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 5, wherein, the using step 2 generating low-resolution geometry mesh body vertex super-resolution features , and specifically comprises: Encoder projecting the node feature vector and the edge feature vector into a feature hidden space learned to obtain an initial node feature of the hidden space and edge feature , represent a vertex index connected with the vertex ; Using node features and edge features as inputs of a graph convolutional neural network , iteratively updating the node features and edge features layer by layer according to the graph structure topology: ； ； wherein, and respectively represent a graph convolutional neural network , a multi-full connection neural network in the graph convolutional neural network, represents different cloth layer vertex indexes colliding with the point ; represents a number of layers of the convolutional network; respectively represent edge features between vertex i and vertex j of the m-1th layer and the mth layer of the convolutional layer; respectively represent edge features between vertex i and vertex j of the m-1th layer of the convolutional layer; ; respectively represent node features of vertex i and vertex j of the m-1th layer of the convolutional layer; represents a node feature of vertex i of the mth layer of the convolutional layer; Graph Convolutional Neural Network The node features output from the last layer are input into the super-resolution decoder. Output the super-resolution features of the vertices of the low-resolution geometry for each vertex i of the current frame. .

7. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 6, wherein, The super-resolution feature of the low-resolution geometric mesh body vertex Generating mesh patch super-resolution feature , specifically comprising: a low-resolution mesh patch The super-resolution feature of the low-resolution geometric mesh body vertex Constructing the super-resolution feature of the mesh patch .

8. The dynamic geometry super-resolution method for digital garment simulation detail enhancement of claim 6, wherein, Iteratively roll forward to predict the next time frame in step 9 a garment high-resolution geometric mesh morphing, comprising in particular: With time frame and the next frame Human movement posture The deformation of the corresponding low-resolution geometric mesh generated by simulation in step 6. and the low-resolution geometric deformation correction obtained in step 8. As input to a dynamic detail augmentation neural network, it predicts the next frame. Low-resolution geometric mesh volume deformation correction vector and detail residual vector This leads to the generation of high-resolution geometric mesh deformation for clothing. .

9. A dynamic geometry super-resolution system for digital garment simulation detail enhancement based on the method of any one of claims 1 to 8, characterized in that, The system comprises the following modules executed in sequence: A first module for generating training data; a second module configured to construct a dynamic detail enhancement neural network, the network comprising: a graph neural network , a geometry correction decoder , and a hyperparameter neural network ; the graph neural network is configured to generate super-resolution features of a low-resolution geometry mesh, and the encoder , a graph convolutional neural network , a super-resolution decoder ; the geometry correction decoder is configured to solve a geometry mesh correction vector of the low-resolution geometry mesh; the hyperparameter neural network is configured to construct a garment dynamic detail up-sampling implicit field, comprising a sub-network and a sub-network , parameters of which are obtained by calculation . A third module for implementing: using the training data, optimizing the learnable network parameters in the dynamic detail enhancement neural network; A fourth module for implementing: modeling the target digital garment low-resolution geometric mesh template, and modeling the high-resolution geometric mesh template aligned with it in the texture space; A fifth module is configured to solve the mapping from the low-resolution mesh topology to the high-resolution mesh topology according to each digital garment template obtained by the fourth module, and obtain the low-resolution mesh patch corresponding to each high-resolution mesh vertex according to the correspondence of the mesh body in the texture space , and the up-sampling coefficient , and further construct the up-sampling matrix . A sixth module is configured to implement: using the target human motion sequence The fourth module is configured to drive the low-resolution geometric mesh body template constructed to simulate the generation of the low-resolution geometric mesh body deformation sequence ; A seventh module is configured to implement: at time frame , using the graph neural network in the second module , calculate the low-resolution geometric mesh body deformation obtained by the sixth module , all mesh patches of the low-resolution geometric mesh body deformation ; an eighth module configured to implement: mesh patch super-resolution features obtained by the seventh module for input, using a geometry correction decoder in the second module solving low resolution geometry mesh deformation correction vectors obtaining low resolution geometry deformation correction: ; Meanwhile, taking the mesh patch super-division features For input, for each patch Using the super-division features in step 2 Corresponding sub-networks are constructed Taking the up-sampling coefficients As input, the detail residual vector of the high-resolution geometry mesh vertex is calculated Thus, the detail residual vector of all high-resolution geometry mesh vertices is obtained ; using the up-sampling matrix Implementing garment geometry morphing mesh resolution up-sampling, computing the current time frame Dynamic detail enhanced garment high resolution geometry mesh morphing: ; a ninth module configured to implement: iteratively predicting garment high-resolution geometry mesh deformation of the next time instance frame based on the prediction result of the current time instance frame obtained by the eighth module until completing digital garment detail enhancement of all sequence frames. ​​ 10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the method of any one of claims 1-8.

Images