Dynamic geometric super-resolution method for digital clothing simulation detail enhancement

Abstract

The invention discloses a dynamic geometric super-resolution method for digital clothing simulation detail enhancement, which takes a low-resolution clothing simulation sequence as input to realize high-resolution and high-frequency geometric detail enhancement and generate a high-fidelity digital clothing deformation sequence with vivid and fine wrinkle details. According to the method, the physical and dynamic characteristics of cloth and interaction information of clothes and a body are combined, based on supervised deep learning, super-division characteristics are extracted on the geometric surface of an irregular grid body, the super-division characteristics are used for constructing an up-sampling implicit function of the dynamic geometric details of the clothes through a hyper-parameter network, and the high-resolution geometric details are generated. According to the method, the geometric deformation of the low-resolution grid body is corrected while the high-resolution geometric details are generated, so that the quality of the generated clothing details is greatly improved. The invention provides a lightweight and efficient clothing detail enhancement method, which is helpful for applying high-fidelity clothing real-time simulation to low-budget and general application scenes.

Description

Technical Field

[0001] The invention belongs to the technical field of computer graphics and artificial intelligence, and in particular to a dynamic geometric super-resolution method for enhancing the simulation details of digital clothing. Background Art

[0002] Clothing is the core element of the virtual environment, and clothing simulation is the core research content of virtual reality. Physical simulation technology based on numerical solution has been the mainstream technology for clothing simulation in the past half century. A number of excellent results have emerged in the research of this technology, and some relatively mature digital clothing simulation software has been put into commercial application, such as Marvelous Designer, Style3D, etc. However, since high-fidelity digital clothing relies on high-resolution geometric expression, the existing physical simulation technology is computationally expensive and occupies a large amount of data storage and transmission resources, making it difficult to apply to low-budget, inclusive application scenarios. How to generate high-fidelity clothing simulation animation efficiently and energy-savingly is a very challenging and urgent problem.

[0003] In contrast, another solution is to use digital clothing expressed in low-resolution geometry to participate in physical simulation. This solution is computationally efficient, but sacrifices the delicacy and realism of the dynamic performance of clothing. In the past decade, researchers have been exploring how to use low-resolution geometry to participate in clothing simulation calculations, data storage and transmission, and then based on low-resolution geometric deformation, enhance the high-frequency information on the surface of the clothing geometric model to generate high-resolution and high-fidelity clothing simulation deformation sequences with rich wrinkle details, such as: 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, which still consume a lot of computing resources. Therefore, the industry urgently needs a new method to generate delicate and rich high-frequency details of cloth wrinkles 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 the present invention is to provide a method for enhancing details of low-resolution clothing simulation to address the difficulties that high-fidelity digital clothing relies on high-resolution geometric expression and requires a large amount of computing resources, while the simulation calculation of low-resolution geometric expression is efficient but sacrifices the dynamic details and realism of the digital clothing.

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

[0006] Step 1, generate training data;

[0007] Step 2: Build a dynamic detail enhancement neural network, which includes: Graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0008] Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

[0009] Step 4, modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in the texture space;

[0010] Step 5: According to the digital clothing templates obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0011] Step 6: Use the target human motion sequence {B t} drives the low-resolution geometric mesh template constructed in step 4 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 in step 2 Calculate the low-resolution geometric mesh deformation C obtained in step 6 t All mesh patch super-resolution features

[0013] Step 8: Super-resolution features of mesh patches obtained in step 7 As input, use the geometric correction decoder ε in step 2 to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction:

[0014] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s in step 5 f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0015] Use the upsampling matrix from step 5 Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0016] Step 9, based on the prediction result of the current moment frame t obtained in step 8, iteratively and rollingly predict the deformation of the clothing high-resolution geometric mesh body of the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0017] Furthermore, the step 1 of generating training data specifically includes:

[0018] Modeling multiple styles of digital clothing using low-resolution geometric meshes and generating high-resolution geometric meshes that align with them in texture space;

[0019] Prepare a variety of human motion sequences, and use a physical simulator to generate low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences of dynamic clothing driven by different human motions;

[0020] 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, a downsampled low-resolution geometric mesh deformation sequence is obtained by mesh resolution downsampling calculation.

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

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

[0023] Constructing the target loss function By correcting the geometric deformation The local patch {P} of the high-resolution clothing geometry mesh after detail enhancement approximates the corresponding true value in the training data and {P *}, achieving end-to-end joint optimization of graph neural networks Geometric Correction Decoder ε and Hyperparameter Neural Subnetwork All learnable parameters of .

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

[0025]

[0026] in, Represents the deformation of the low-resolution geometric mesh after correction The constraint function is represents the constraint function for the detail enhancement network prediction result {P}, λ c and λ p is the constraint coefficient.

[0027] Furthermore, in step 7, we use graph neural network Generate super-resolution features, including:

[0028] According to the low-resolution geometric mesh deformation C of the current frame t and the historical frames t-1 and t-2 t , C t-1 , C t-2 、Human body movement posture B t , B t-1 , and the historical frame low-resolution geometric deformation correction results Calculate the node feature vector and edge feature vector of the graph topology, where the graph topology nodes correspond one-to-one to the low-resolution geometric mesh vertices;

[0029] Use the graph neural network from step 2 Generate low-resolution geometric mesh vertex super-resolution features i represents the low-resolution geometric mesh vertex i, Represents the super-resolution feature of vertex i of the low-resolution geometric mesh of the current frame t, and then the super-resolution feature of the vertex of the low-resolution geometric mesh Generate mesh patch super-resolution features

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

[0031] Encoder Project the node feature vector and edge feature vector into the learned feature latent space to obtain the initial node feature q in the latent space i and edge feature e ij , j represents the vertex index connected to vertex i;

[0032] Using node features and edge features as graph convolutional neural networks The node features and edge features are updated layer by layer iteratively according to the graph structure topology:

[0033]

[0034]

[0035] in, and Represents graph convolutional neural network A multi-fully connected neural network in , l represents the vertex index of different cloth layers that collide with point i, and m represents the number of layers of the convolutional network; Respectively represent the edge features between vertex i and vertex, in the m-1th and mth convolutional layers, Represents the edge feature between vertex i and vertex l in the m-1th convolutional layer; They represent the node features of vertex i and vertex j in the m-1th convolutional layer respectively. Represents the node features of vertex i in the mth convolutional layer;

[0036] Graph Convolutional Neural Network The node features output by the last layer are input to the super-resolution decoder Output the super-resolution features of the low-resolution geometric mesh vertex of the current frame for each vertex i of the low-resolution geometric mesh

[0037] Furthermore, the super-resolution feature of the low-resolution geometric mesh vertex Generate mesh patch super-resolution features Specifically, the super-resolution features of a low-resolution mesh face f are composed of the super-resolution features of all vertices of the face in the topological order of the face vertices, thereby realizing the super-resolution features of the low-resolution geometric mesh body vertices. Constructing super-resolution features of mesh patches

[0038] Furthermore, in step 9, the deformation of the clothing high-resolution geometric mesh at the next frame t+1 is predicted iteratively and rollingly, specifically including:

[0039] The human body motion posture B at time frame t and next time frame t+1 t , B t+1 , the corresponding low-resolution geometric mesh deformation C 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 the input of the dynamic detail enhancement neural network, the low-resolution geometric mesh deformation correction vector D of the next frame t+1 is predicted t+1 and the detail residual vector R t+1 , and then generate clothing high-resolution geometric mesh deformation G t+1 .

[0040] In another aspect, a dynamic geometric super-resolution system for digital clothing simulation detail enhancement is provided, the system comprising:

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

[0042] The second module is used to build a dynamic detail enhancement neural network, which includes: graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0043] The third module is used to implement: using the training data, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

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

[0045] The fifth module is used to achieve: according to each digital clothing template obtained in the fourth module, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0046] The sixth module is used to implement: using the target human motion sequence {B t} drives 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: at time frame t, use the graph neural network in the second module Calculate the low-resolution geometric mesh deformation C obtained in the sixth module t All mesh patch super-resolution features

[0048] The eighth module is used to implement: the mesh patch super-resolution features obtained in the seventh module As input, the geometric correction decoder ε in the second module is used to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction:

[0049] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0050] Using the upsampling matrix Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0051] The ninth module is used to implement: based on the prediction result of the current moment frame t obtained by the eighth module, iteratively and rollingly predict the deformation of the high-resolution geometric mesh of the clothing at the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0052] On the other hand, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the dynamic geometric super-resolution method for detail enhancement of digital clothing simulation is implemented.

[0053] Compared with the prior art, the present invention has the following significant advantages:

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

[0055] (2) The present invention provides a lightweight and efficient clothing detail enhancement method, which helps to apply high-fidelity clothing real-time simulation to low-budget and affordable scenarios, such as smartphones.

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

[0057] (4) The present 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 is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The present invention is a flow chart of a dynamic geometric super-resolution method for detail enhancement of digital clothing simulation.

[0060] Figure 2 This is a neural network structure diagram of the dynamic geometric super-resolution method for digital clothing simulation detail enhancement of the present invention.

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

[0062] Figure 4 A low-resolution digital clothing geometry correction result diagram in one embodiment.

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

[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0066] In one embodiment, in combination Figure 1 , a dynamic geometric super-resolution method for digital clothing simulation detail enhancement is provided, the method comprising the following steps:

[0067] Step 1, generate training data;

[0068] Step 2: Build a dynamic detail enhancement neural network, which includes: Graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0069] Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

[0070] Step 4, modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in the texture space;

[0071] Step 5: According to the digital clothing templates obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0072] Step 6: Use the target human motion sequence {B t} drives the low-resolution geometric mesh template constructed in step 4 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 in step 2 Calculate the low-resolution geometric mesh deformation C obtained in step 6 t All mesh patch super-resolution features

[0074] Step 8: Super-resolution features of mesh patches obtained in step 7 As input, use the geometric correction decoder ε in step 2 to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction: (Specifically: super-resolution features of geometric mesh patches in step 6 As the input of 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, calculate the geometric correction vector for each vertex i of the low-resolution mesh in is the number of mesh patches connected to vertex i; then, correct the geometric position c of each vertex of the low-resolution mesh to obtain the corrected vertex position Get the corrected low-resolution clothing geometry of the current frame );

[0075] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s in step 5 f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0076] Use the upsampling matrix from step 5 Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0077] Step 9, based on the prediction result of the current moment frame t obtained in step 8, iteratively and rollingly predict the deformation of the clothing high-resolution geometric mesh body of the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0078] Further, in one embodiment, the step 1 of generating training data specifically includes:

[0079] Modeling multiple styles of digital clothing using low-resolution geometric meshes and generating high-resolution geometric meshes that align with them in texture space;

[0080] Prepare a variety of human motion sequences, and use a physical simulator to generate low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences of dynamic clothing driven by different human motions;

[0081] 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, a downsampled low-resolution geometric mesh deformation sequence is obtained by mesh resolution downsampling calculation.

[0082] Further, in one embodiment, the optimization of the learnable network parameters in the dynamic detail enhancement network in step 3 is specifically: using the clothing low-resolution geometric mesh deformation sequence {C} and the high-resolution geometric mesh deformation sequence {G *} Downsampled low-resolution geometry mesh deformation sequence and human motion sequences {B}, training and optimizing the learnable network parameters in the dynamic detail enhancement network. Specifically, it includes:

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

[0084] Constructing the target loss function By correcting the geometric deformation The local patch {P} of the high-resolution clothing geometry mesh after detail enhancement approximates the corresponding true value in the training data and {P *}, achieving end-to-end joint optimization of graph neural networks Geometric Correction Decoder ε and Hyperparameter Neural Subnetwork All learnable parameters of .

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

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

[0087]

[0088] in, Represents the deformation of the low-resolution geometric mesh after correction The constraint function is represents the constraint function for the detail enhancement network prediction result {P}, λ c and λ p is the constraint coefficient.

[0089] Here, preferably, in the process of neural network optimization, the constraint coefficient λ is first c and λ p Set to λ c =1.9,λ p =0.1, and then gradually reduce the coefficient λ c And increase the coefficient λ p Until it is 1, and keep the coefficient λ c =λ p =1 until the loss function optimization converges.

[0090] Further, in one embodiment, in step 7, a graph neural network is used Generate super-resolution features, including:

[0091] According to the low-resolution geometric mesh deformation C of the current frame t and the historical frames t-1 and t-2 t , C t-1 , C t-2 、Human body movement posture t , b t-1 , and the historical frame low-resolution geometric deformation correction results Calculate the node feature vector and edge feature vector of the graph topology, where the graph topology nodes correspond one-to-one to the low-resolution geometric mesh vertices;

[0092] Use the graph neural network from step 2 Generate low-resolution geometric mesh vertex super-resolution features i represents the low-resolution geometric mesh vertex i, Represents the super-resolution feature of vertex i of the low-resolution geometric mesh of the current frame t, and then the super-resolution feature of the vertex of the low-resolution geometric mesh Generate mesh patch super-resolution features

[0093] Here preferably, in some embodiments, the graph neural network in step 2 is used Generate low-resolution geometric mesh vertex super-resolution features Specifically include:

[0094] Encoder Project the node feature vector and edge feature vector into the learned feature latent space to obtain the initial node feature q in the latent space i and edge feature e ij , j represents the vertex index connected to vertex i;

[0095] Using node features and edge features as graph convolutional neural networks The node features and edge features are updated layer by layer iteratively according to the graph structure topology:

[0096]

[0097]

[0098] in, and Represents graph convolutional neural network A fully connected neural network in , l represents the vertex index of different cloth layers that collide with point i, and m represents the number of layers of the convolutional network; Respectively represent the edge features between vertex i and vertex, in the m-1th and mth convolutional layers, Represents the edge feature between vertex i and vertex l in the m-1th convolutional layer; They represent the node features of vertex i and vertex j in the m-1th convolutional layer respectively. Represents the node features of vertex i in the mth convolutional layer;

[0099] Graph Convolutional Neural Network The node features output by the last layer are input to the super-resolution decoder Output the super-resolution features of the low-resolution geometric mesh vertex of the current frame for each vertex i of the low-resolution geometric mesh

[0100] Preferably, in some embodiments, the low-resolution geometric mesh vertex super-resolution features Generate mesh patch super-resolution features Specifically, the super-resolution features of a low-resolution mesh face f are composed of the super-resolution features of all vertices of the face in the topological order of the face vertices, thereby realizing the super-resolution features of the low-resolution geometric mesh body vertices. Constructing super-resolution features of mesh patches

[0101] Furthermore, in one embodiment, the iterative rolling prediction of the clothing high-resolution geometric mesh deformation at the next time frame t+1 in step 9 specifically includes:

[0102] The human body motion posture B at time frame t and next time frame t+1 t , B t+1 , the corresponding low-resolution geometric mesh deformation C 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 the input of the dynamic detail enhancement neural network, the low-resolution geometric mesh deformation correction vector D of the next frame t+1 is predicted t+1 and the detail residual vector R t+1 , and then generate clothing with rich wrinkle details of high-resolution geometric mesh deformation G t+1 .

[0103] In one embodiment, a dynamic geometric super-resolution system for digital clothing simulation detail enhancement is provided, the system comprising:

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

[0105] The second module is used to build a dynamic detail enhancement neural network, which includes: graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0106] The third module is used to implement: using the training data, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

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

[0108] The fifth module is used to achieve: according to each digital clothing template obtained in the fourth module, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0109] The sixth module is used to implement: using the target human motion sequence {B t} drives 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: at time frame t, use the graph neural network in the second module Calculate the low-resolution geometric mesh deformation C obtained in the sixth module t All mesh patch super-resolution features

[0111] The eighth module is used to implement: the mesh patch super-resolution features obtained in the seventh module As input, the geometric correction decoder ε in the second module is used to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction:

[0112] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0113] Using the upsampling matrix Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0114] The ninth module is used to implement: based on the prediction result of the current moment frame t obtained by the eighth module, iteratively and rollingly predict the deformation of the high-resolution geometric mesh of the clothing at the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

[0115] For the specific definition of the dynamic geometric super-resolution system for digital clothing simulation detail enhancement, please refer to the definition of the dynamic geometric super-resolution method for digital clothing simulation detail enhancement in the above text, which will not be repeated here. Each module in the above-mentioned dynamic geometric super-resolution system for digital clothing simulation detail enhancement can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[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 when the processor executes the computer program, the following is achieved:

[0117] Step 1, generate training data;

[0118] Step 2: Build a dynamic detail enhancement neural network, which includes: Graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0119] Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

[0120] Step 4, modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in the texture space;

[0121] Step 5: According to the digital clothing templates obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0122] Step 6: Use the target human motion sequence {B t} drives the low-resolution geometric mesh template constructed in step 4 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 in step 2 Calculate the low-resolution geometric mesh deformation C obtained in step 6 t All mesh patch super-resolution features

[0124] Step 8: Super-resolution features of mesh patches obtained in step 7 As input, use the geometric correction decoder ε in step 2 to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction:

[0125] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s in step 5 f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0126] Use the upsampling matrix from step 5 Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0127] Step 9, based on the prediction result of the current moment frame t obtained in step 8, iteratively and rollingly predict the deformation of the clothing high-resolution geometric mesh body of the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

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

[0129] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the computer program implements:

[0130] Step 1, generate training data;

[0131] Step 2: Build a dynamic detail enhancement neural network, which includes: Graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain;

[0132] Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network;

[0133] Step 4, modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in the texture space;

[0134] Step 5: According to the digital clothing templates obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix

[0135] Step 6: Use the target human motion sequence {B t} drives the low-resolution geometric mesh template constructed in step 4 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 in step 2 Calculate the low-resolution geometric mesh deformation C obtained in step 6 t All mesh patch super-resolution features

[0137] Step 8: Super-resolution features of mesh patches obtained in step 7 As input, use the geometric correction decoder ε in step 2 to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction:

[0138] At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s in step 5 f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t};

[0139] Use the upsampling matrix from step 5 Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t:

[0140] Step 9, based on the prediction result of the current moment frame t obtained in step 8, iteratively and rollingly predict the deformation of the clothing high-resolution geometric mesh body of the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

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

[0142] As a specific example, the present invention is further described in detail in one of the embodiments.

[0143] This paper 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, specifically including:

[0144] (1) Generate training data: Use low-resolution triangular meshes with an average vertex distance of 30 mm to model digital clothing of different styles: T-shirts, trousers, pleated skirts, single-layer lace dresses, and three-layer lace long skirts, and obtain high-resolution triangular mesh clothing modeling with an average vertex distance of 10 mm aligned with them in texture space; prepare 714 frames of swing dance human motion sequence and 794 frames of 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 geometry motion sequence and high-resolution clothing sequence driven by the human motion sequence; according to the alignment relationship of the meshes in texture space, calculate the downsampled low-resolution clothing geometry sequence based on the high-resolution geometry sequence.

[0145] (2) Constructing a dynamic detail enhancement neural network: Figure 2 As shown in Figure 2, the network consists of three main parts: Graph Neural Network Geometric Correction Decoder ε and Hyperparameter Neural Networks

[0146] (2.1) Graph Neural Networks is a parameterized learnable neural network consisting of an encoder A graph convolutional neural network and 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 Network By sub-network and subnetwork constitute, The parameters are Obtained by 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 *}, Downsampled low-resolution clothing geometry sequence and human motion sequences {B}, training to optimize the learnable network parameters in the dynamic detail enhancement network;

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

[0151] By minimizing the loss function End-to-end joint optimization of detail enhancement networks in graph neural networks Geometric Correction Decoder ε and Hyperparameterized Neural Networks All learnable parameters, loss function It is expressed as follows:

[0152]

[0153] in, Represents the reference downsampled clothing geometry The defined corrected low-resolution clothing geometry The constraint function, in actual operation, Include geometric constraints and deformation constraints Two parts:

[0154]

[0155] The expression is as follows:

[0156]

[0157] in, and Represents low-resolution geometric meshes and The normal vector of The representation is as follows:

[0158]

[0159] Here, Δ represents the Laplacian operator, and θ represents the angle operation 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. The deformation matrix And through the equation:

[0160]

[0161] Make the geometry and The deformation is similar. *} represents the high-resolution clothing geometry G *The geometries of multiple high-resolution local blocks obtained by cropping the corresponding local blocks {P}, Represents the reference local block geometry {P *} defined by the detail deformation constraint for the detail enhancement network prediction result {P}. In actual operation, Include geometric constraints and deformation constraints Two parts:

[0162]

[0163] The expression is as follows:

[0164]

[0165] The normal vector of the local block P of the mesh is drawn on an image of size 80×80 according to the texture mapping, VGG l Represents the hidden feature encoding of the image in the lth layer of the pre-trained neural network VGG (Karen Simonyan and Andrew Zisserman, 2015, Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations.). It is expressed as:

[0166]

[0167] In the process of neural network optimization, the constraint coefficient λ is first c and λ p Set to λ c =1.9,λ p =0.1, and then gradually reduce the coefficient λ c And increase the coefficient λ p Until it is 1, and keep the coefficient λ c =λ p =1 until the loss function is optimized and converged; the optimized detail enhancement 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 30 mm for the target digital clothing, and model a high-resolution triangular mesh template with an average vertex distance of 10 mm aligned with it in the texture space; according to the corresponding relationship of the mesh in the 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 the upsampling matrix

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

[0170] (6) Use the graph neural network in step (2) The low-resolution clothing C obtained in step (5) is calculated t Super-resolution features at time frame t: Based on the low-resolution clothing C of the current frame t and historical frames t-1 and t-2 t , C t-1 , C t-2 、Sports B t , B t-1 , Geometric correction results of low-resolution clothing in historical frames Compute node feature vectors of graph topology and edge eigenvector As step (1) in the graph neural network Input;

[0171] Feature vector of node i in current frame t It is expressed as:

[0172]

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

[0174]

[0175] Among them, σ b is a preset threshold, which can be set to 0.3mm in actual operation. for b t The normal vector of Represents the clothing vertex at time t-1 With b t The interaction vector is:

[0176]

[0177] in, respectively The vector modulus length represented by ; taking into account the impact of historical frame prediction on the current frame prediction, express A vertex on Represents the result of low-resolution geometric correction The vertex velocity is obtained by calculation.

[0178] The feature vector of the edge constructed by node i and node j in the current frame t Mainly composed of the edge vector c of the grid at time t and time t-1 j -c i and side length || c j -c i ||, and the length of the side of the garment in its initial static state o It means that the feature vector Finally it is expressed as:

[0179]

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

[0181] Get the node feature vector of the computational graph topology and edge eigenvector Afterwards, the encoder First, the node feature vector and edge feature vector are projected into the learned projection space to obtain the initial node feature q in the latent space. i and edge feature e ij , i and j represent graph node indices, e ij represents the edge features formed by nodes i and j; then, the node features and edge features are used as graph convolutional neural networks The input, graph structure node features and edge features are iteratively updated according to the graph structure topology through the graph convolutional neural network:

[0182]

[0183]

[0184] in, and Represents neural network A multi-layer fully connected neural network in . Since collisions may occur between different layers of clothing, l represents the graph node index of collision points in different layers, and m represents the number of layers of the convolutional network.

[0185] In order to make the operation of graph convolutional neural network more efficient, in actual operation, a convolution operation can be performed on the 2-level ring topology domain of the low-resolution grid body, and the super-resolution feature vector can be effectively extracted through a 6-layer convolutional neural network; then, the neural network The node feature qi output by the last layer is input to the node feature decoder Generates the current frame super-resolution feature vector for each vertex output of the low-resolution clothing geometry mesh

[0186] Furthermore, using super-resolution features Constructing super-resolution features of geometric mesh patches The super-resolution feature of a geometric mesh patch f is the super-resolution feature z of all vertices of the patch i The vertex index of a triangle patch is (f 0 , f 1 , f 2 ), then the super-resolution feature of the triangular patch is:

[0187]

[0188] (7) Correcting low-resolution clothing geometry: Super-resolution features of the geometric mesh patches in step (6) As the input of 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 , for example, if the mesh patch is a triangle, the output of the input of the geometric correction decoder ε is:

[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 of the low-resolution mesh is corrected to obtain the corrected vertex position Get the corrected low-resolution clothing geometry of the current frame like Figure 4 Shown is one frame of low-resolution clothing geometry correction results.

[0191] At the same time, generate high-resolution clothing geometric deformation with enhanced details: super-resolution features of geometric mesh patches in step (6) As the network in step (2) Input, construct the corresponding Then, the high-resolution geometry is upsampled by the factor s in step (4). f As input, calculate the detail residual vector of the high-resolution clothing geometric mesh vertex Thus, the detail residual vector R of all high-resolution mesh vertices is obtained. t = {r t}; Then, combine the upsampling matrix in step (4) Corrected low-resolution clothing geometry and the detail residual R t , computing high-resolution clothing geometry for dynamic detail enhancement in the target frame like Figure 5 Shown is a high-resolution garment simulation result after detail enhancement in one frame.

[0192] (8) When solving the high-resolution clothing detail enhancement at time t+1, the human body motion posture B in step (5) t+1 , low-resolution clothing geometric deformation C generated by simulation t-1 , C t , C t+1 , and the low-resolution geometric deformation correction obtained by step (7) As the input of the dynamic detail enhancement neural network, the correction vector D at time t+1 is predicted t+1 and the detail residual vector R t+1 , generating high-resolution clothing geometry deformation G with rich wrinkle details t+1 .

[0193] In summary, the present invention uses a low-resolution clothing simulation sequence as input to achieve high-resolution, high-frequency geometric detail enhancement, and generate a high-fidelity digital clothing deformation sequence with vivid and delicate wrinkle details. This method combines the physical dynamic characteristics of the cloth and the interaction information between the clothing and the body, and based on supervised deep learning, it achieves the extraction of super-resolution features on the geometric surface of irregular grid bodies, and uses the super-resolution features through a hyperparameter network to construct an implicit function for sampling the dynamic geometric details of clothing to generate high-resolution geometric details. This method corrects the low-resolution grid geometric deformation while generating high-resolution geometric details, so that the quality of the generated clothing details is greatly improved. The present invention provides a lightweight and efficient clothing detail enhancement method, which helps to apply high-fidelity clothing real-time simulation to low-budget, inclusive application scenarios.

[0194] The above shows and describes 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 by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A dynamic geometric super-resolution method for detail enhancement of digital clothing simulation, characterized in that: The method comprises the following steps: Step 1, generate training data; Step 2: Build a dynamic detail enhancement neural network, which includes: Graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain; Step 3, using the training data in step 1, optimizing the learnable network parameters in the dynamic detail enhancement neural network; Step 4, modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in the texture space; Step 5: According to the digital clothing templates obtained in step 4, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix Step 6: Use the target human motion sequence {B t } drives the low-resolution geometric mesh template constructed in step 4 to simulate and generate a low-resolution geometric mesh deformation sequence {C t }; Step 7: At time frame t, use the graph neural network in step 2 Calculate the low-resolution geometric mesh deformation C obtained in step 6 t All mesh patch super-resolution features Step 8: Super-resolution features of mesh patches obtained in step 7 As input, use the geometric correction decoder ε in step 2 to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction: At the same time, the mesh patch super-resolution feature As input, for each patch f, use the Construct the corresponding sub-network With the upsampling factor s in step 5 f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t }; Use the upsampling matrix from step 5 Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t: Step 9, based on the prediction result of the current moment frame t obtained in step 8, iteratively and rollingly predict the deformation of the clothing high-resolution geometric mesh body of the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

2. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 1, characterized in that: Step 1 of generating training data specifically includes: Modeling multiple styles of digital clothing using low-resolution geometric meshes and generating high-resolution geometric meshes that align with them in texture space; Prepare a variety of human motion sequences, and use a physical simulator to generate low-resolution geometric mesh deformation sequences and high-resolution geometric mesh deformation sequences of dynamic clothing driven by different human motions; 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, a downsampled low-resolution geometric mesh deformation sequence is obtained by mesh resolution downsampling calculation.

3. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 1, characterized in that: The optimization of the learnable network parameters in the dynamic detail enhancement network described in step 3 specifically includes: Taking the low-resolution clothing geometric mesh deformation sequence generated by the simulator and the human body motion posture as input, the dynamic detail enhancement neural network predicts and outputs the corrected clothing low-resolution geometric mesh deformation Based on the multiple local blocks obtained by random cropping in the texture space, a high-resolution clothing geometric mesh local block {P} with enhanced details is generated; Constructing the target loss function By correcting the geometric deformation The local patch {P} of the high-resolution clothing geometry mesh after detail enhancement approximates the corresponding true value in the training data and {P * }, achieving end-to-end joint optimization of graph neural networks Geometric Correction Decoder ε and Hyperparameter Neural Subnetwork All learnable parameters of .

4. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 3, characterized in that: The objective loss function It is expressed as follows: in, Represents the deformation of the low-resolution geometric mesh after correction The constraint function is represents the constraint function for the detail enhancement network prediction result {P}, λ c and λ p is the constraint coefficient.

5. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 1, characterized in that: Using Graph Neural Networks in Step 7 Generate super-resolution features, including: According to the low-resolution geometric mesh deformation C of the current frame t and the historical frames t-1 and t-2 t ,C t-1 ,C t-2 、Human body movement posture B t ,B t-1 , and the historical frame low-resolution geometric deformation correction results Calculate the node feature vector and edge feature vector of the graph topology, where the graph topology nodes correspond one-to-one to the low-resolution geometric mesh vertices; Use the graph neural network from step 2 Generate low-resolution geometric mesh vertex super-resolution features i represents the low-resolution geometric mesh vertex i, Represents the super-resolution feature of vertex i of the low-resolution geometric mesh of the current frame t, and then the super-resolution feature of the vertex of the low-resolution geometric mesh Generate mesh patch super-resolution features 6. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 5, characterized in that: The graph neural network used in step 2 Generate low-resolution geometric mesh vertex super-resolution features Specifically include: Encoder Project the node feature vector and edge feature vector into the learned feature latent space to obtain the initial node feature q in the latent space i and edge feature e ij , j represents the vertex index connected to vertex i; Using node features and edge features as graph convolutional neural networks The node features and edge features are updated layer by layer iteratively according to the graph structure topology: in, and Represents graph convolutional neural network A multi-fully connected neural network in , l represents the vertex index of different cloth layers that collide with point i, and m represents the number of layers of the convolutional network; Respectively represent the edge features between vertex i and vertex j in the m-1th and mth convolutional layers, Represents the edge feature between vertex i and vertex l in the m-1th convolutional layer; They represent the node features of vertex i and vertex j in the m-1th convolutional layer respectively. Represents the node features of vertex i in the mth convolutional layer; Graph Convolutional Neural Network The node features output by the last layer are input to the super-resolution decoder Output the super-resolution features of the low-resolution geometric mesh vertex of the current frame for each vertex i of the low-resolution geometric mesh 7. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 6, characterized in that: The low-resolution geometric mesh vertex super-resolution features Generate mesh patch super-resolution features Specifically, the super-resolution features of a low-resolution mesh face f are composed of the super-resolution features of all vertices of the face in the topological order of the face vertices, thereby realizing the super-resolution features of the low-resolution geometric mesh body vertices. Constructing super-resolution features of mesh patches 8. The dynamic geometric super-resolution method for digital clothing simulation detail enhancement according to claim 6, characterized in that: In step 9, the deformation of the high-resolution geometric mesh of the clothing at the next frame t+1 is predicted iteratively and rollingly, including: The human body motion posture B at time frame t and next time frame t+1 t ,B t+1 , the corresponding low-resolution geometric mesh deformation C 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 the input of the dynamic detail enhancement neural network, the low-resolution geometric mesh deformation correction vector D of the next frame t+1 is predicted t+1 and the detail residual vector R t+1 , and then generate clothing high-resolution geometric mesh deformation G t+1 .

9. A dynamic geometric super-resolution system for digital clothing simulation detail enhancement based on the method according to any one of claims 1 to 8, characterized in that: The system includes sequentially executing: The first module is used to generate training data; The second module is used to build a dynamic detail enhancement neural network, which includes: graph neural network Geometric Correction Decoder ε and Hyperparameterized Neural Networks The graph neural network Used to generate super-resolution features of low-resolution geometric meshes by the encoder Graph Convolutional Neural Networks Super Resolution Decoder The geometric correction decoder ε is used to solve the low-resolution geometric grid correction vector; the hyperparameter neural network Used to construct implicit fields for upsampling clothing dynamic details, including sub-networks and subnetwork The parameters are Calculate and obtain; The third module is used to implement: using the training data, optimizing the learnable network parameters in the dynamic detail enhancement neural network; The fourth module is used to achieve: modeling a low-resolution geometric mesh template of the target digital clothing, and modeling a high-resolution geometric mesh template aligned with it in texture space; The fifth module is used to achieve: according to each digital clothing template obtained in the fourth module, solve the mapping from low-resolution mesh topology to high-resolution mesh topology: according to the corresponding relationship of the mesh body in the 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 the upsampling matrix The sixth module is used to implement: using the target human motion sequence {B t } drives the low-resolution geometric mesh template constructed by the fourth module to simulate and generate a low-resolution geometric mesh deformation sequence {C t }; The seventh module is used to implement: at time frame t, use the graph neural network in the second module Calculate the low-resolution geometric mesh deformation C obtained in the sixth module t All mesh patch super-resolution features The eighth module is used to implement: the mesh patch super-resolution features obtained in the seventh module As input, the geometric correction decoder ε in the second module is used to solve the low-resolution geometric mesh deformation correction vector D t , obtain low-resolution geometric deformation correction: At the same time, the mesh patch super-resolution feature As input, for each face f, use the Construct the corresponding sub-network With the upsampling factor s f As input, calculate the detail residual vector of the high-resolution geometric mesh vertex Thus, the detail residual vector R of all high-resolution geometric mesh vertices is obtained. t = {r t }; Using the upsampling matrix Implement clothing geometry deformation mesh resolution upsampling, and calculate clothing high-resolution geometry mesh deformation with dynamic detail enhancement at the current frame t: The ninth module is used to implement: based on the prediction result of the current moment frame t obtained by the eighth module, iteratively and rollingly predict the deformation of the high-resolution geometric mesh of the clothing at the next moment frame t+1 until the digital clothing detail enhancement of all sequence frames is completed.

10. A computer 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 computer program, the method according to any one of claims 1 to 8 is implemented.

Images