User generated content quality control system and method based on generative adversarial network

Abstract

The invention relates to the technical field of content quality control, in particular to a user generated content quality control system and method based on a generative adversarial network, and provides a 3D model generation and quality control system. Through the synergistic effect of content monitoring, rendering generation, quality monitoring, generator and discriminator feedback, content generation, user creation, quality evaluation and other modules, the deep learning and adversarial generative network technology is utilized to effectively detect and generate a legal and high-quality 3D model; the system not only ensures the content security, but also improves the creation experience through topological refinement and user interface optimization; the innovative method provides a brand new way for 3D content creation, the generation efficiency and quality are remarkably improved, and rapid development and technical innovation of related industries are expected to be promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content quality control, in particular to a user-generated content quality control system based on generative adversarial networks and a method thereof. BACKGROUND

[0002] With the rapid development of 3D technology, user-generated content (UGC) plays an increasingly important role in game development, virtual reality, and augmented reality. However, ensuring the quality of user-generated 3D content has always been a major challenge in the industry. Traditional 3D content creation methods usually require professional modeling software and high technical barriers, which greatly limits the possibility of ordinary users participating in 3D content creation.

[0003] In recent years, 3D content generation methods based on deep learning have made significant progress. In particular, generative adversarial networks (GAN) have shown great potential in the field of 3D model generation. However, existing GAN-based 3D content generation methods still have some key problems. First, the generated 3D models often lack details and realism, making it difficult to meet the needs of high-quality applications. Second, existing methods are difficult to effectively control the quality of generated content, and are prone to produce unreasonable or physically inconsistent 3D structures. Third, the user's creative intent is difficult to accurately integrate into the automatic generation process, resulting in a large gap between the generated results and the user's expectations.

[0004] In addition, existing 3D content quality control methods mostly focus on post-processing stages, lacking real-time guidance and feedback during the creation process. This not only reduces the creation efficiency, but also may lead to users spending a lot of time creating content that does not meet the requirements. At the same time, the detection and filtering of illegal or inappropriate content also face great challenges, especially in the 3D field, existing content review technologies are difficult to effectively deal with complex three-dimensional structures and textures.

[0005] In view of the above problems, there is an urgent need for a 3D content generation system that can balance content quality control and user creativity. The system should be able to automatically generate high-quality 3D models while allowing users to make personalized creations, and provide real-time quality assessment and optimization suggestions during the creation process. In addition, the system also needs to have effective illegal content detection capabilities to ensure the legality and appropriateness of the generated content. SUMMARY

[0006] The user-generated content quality control system based on generative adversarial networks and the method thereof proposed by the present application are designed to solve the above technical problems. Through innovative module design and algorithm optimization, the system effectively improves the generation quality and creation efficiency of 3D content, while ensuring the legality and appropriateness of the content.

[0007] The application provides a user-generated content quality control system based on a generative adversarial network, comprising:

[0008] a content monitoring module for:

[0009] detecting illegal content in a 3D model submitted by a user using a deep convolutional neural network;

[0010] generating an illegal content discrimination result;

[0011] a rendering generation module in communication with the content monitoring module, for:

[0012] receiving the illegal content discrimination result sent by the content monitoring module;

[0013] generating a preliminary 3D model that meets semantic and physical rationality based on the illegal content discrimination result using a generative adversarial network;

[0014] a quality monitoring module in communication with the rendering generation module, for:

[0015] receiving the preliminary 3D model generated by the rendering generation module;

[0016] mapping the preliminary 3D model into a spatial topological relationship graph;

[0017] discriminating the topological rationality and texture details of the preliminary 3D model based on the spatial topological relationship graph;

[0018] generating a rationality score;

[0019] a generator feedback module in communication with the quality monitoring module, for:

[0020] receiving the rationality score generated by the quality monitoring module;

[0021] training the generative adversarial network using gradient descent method based on the rationality score;

[0022] a discriminator feedback module in communication with the quality monitoring module and the generator feedback module, for:

[0023] receiving the rationality score generated by the quality monitoring module;

[0024] optimizing the discriminator using gradient descent method based on the rationality score;

[0025] a content generation module in communication with the generator feedback module and the discriminator feedback module, for:

[0026] receiving the optimization results of the generator feedback module and the discriminator feedback module;

[0027] Based on the optimization result, a high-quality 3D model meeting the user's demand is generated.

[0028] A user creation module is in communication connection with the content generation module, configured to:

[0029] Provide a geometry drawing interface, a texture drawing and editing interface, and an edge drawing interface of the 3D model;

[0030] Receive the user's creation input;

[0031] A quality evaluation module is in communication connection with the content generation module and the user creation module, configured to:

[0032] Receive the 3D model generated by the content generation module and the user creation input of the user creation module;

[0033] Map the 3D model to a spatial topological relationship graph through topological refinement;

[0034] Evaluate the legitimacy and quality rationality score of the 3D model;

[0035] A suggestion module is in communication connection with the quality evaluation module and the user creation module, configured to:

[0036] Receive the evaluation result of the quality evaluation module;

[0037] Based on the evaluation result, provide creation suggestions to the user;

[0038] Optimize the creation interface of the user creation module.

[0039] As a preferred, the content monitoring module comprises:

[0040] A feature extraction unit is configured to extract deep features from the 3D model submitted by the user;

[0041] An illegal content identification unit is in communication connection with the feature extraction unit, configured to identify illegal content in the deep features based on a pre-trained deep learning model;

[0042] A result generation unit is in communication connection with the illegal content identification unit, configured to generate an illegal content discrimination result.

[0043] As a preferred, the rendering generation module comprises:

[0044] A conditional encoder is configured to encode the user-defined content features into a conditional vector;

[0045] A style encoder is configured to encode the user-defined style features into a style vector;

[0046] a generation network, communicatively connected with the conditional encoder and the style encoder, configured to generate a 3D model based on the conditional vector and the style vector;

[0047] a rendering network, communicatively connected with the generation network, configured to render the 3D model generated by the generation network to generate a preliminary 3D model conforming to semantics and physical rationality.

[0048] Preferably, the quality monitoring module comprises:

[0049] a topology mapping unit, configured to convert the preliminary 3D model into an undirected graph model, and use an adjacency matrix to represent the topological relationship between graphs;

[0050] a feature extraction unit, communicatively connected with the topology mapping unit, configured to extract topological features and texture features from the undirected graph model;

[0051] a discriminant network, communicatively connected with the feature extraction unit, configured to determine the rationality of the 3D model based on the topological features and the texture features, and generate a rationality score.

[0052] Preferably, the content generation module comprises:

[0053] a geometry module, configured to generate the geometry of the 3D model based on the geometry information provided by the user;

[0054] a texture module, communicatively connected with the geometry module, configured to generate the texture of the 3D model based on the texture information provided by the user;

[0055] an edge drawing module, communicatively connected with the geometry module and the texture module, configured to map the line texture drawn by the edge drawing to the surface of the 3D model based on the edge drawn by the user and the texture prompt;

[0056] a style transfer module, communicatively connected with the geometry module, the texture module and the edge drawing module, configured to perform style transfer on the generated 3D model to maintain the consistency of the model style.

[0057] Preferably, the geometry module comprises:

[0058] a graph attention network, configured to learn context information from the initial grid graph, and generate a geometry feature map in combination with a user-defined feature map;

[0059] a differential rendering unit, communicatively connected with the graph attention network, configured to convert the geometry feature map into a topological transformation, and optimize the geometry of the initial generated grid through back propagation.

[0060] Preferably, the user creation module comprises:

[0061] a geometry drawing interface for receiving user-drawn 3D model geometry;

[0062] a texture drawing and editing interface for receiving user-drawn and edited 3D model texture, wherein the texture drawing and editing interface supports generation of texture map, normal map and mask map;

[0063] an edge drawing interface for receiving user-drawn 3D model edge;

[0064] a real-time feedback unit in communication with the geometry drawing interface, the texture drawing and editing interface and the edge drawing interface, for displaying quality assessment results and creation suggestions in real time during user creation process.

[0065] Preferably, the quality assessment module comprises:

[0066] a topology refinement unit for topology refinement of the generated 3D model to generate a high-precision spatial topology graph;

[0067] a multi-dimension assessment unit in communication with the topology refinement unit, for assessing the quality of the 3D model from multiple dimensions including topology rationality, physical consistency, semantic structure and aesthetic features;

[0068] a comprehensive scoring unit in communication with the multi-dimension assessment unit, for generating a comprehensive quality score based on the multi-dimension assessment results.

[0069] Preferably, the system further comprises a data enhancement module for:

[0070] data augmentation of high-quality real 3D models to generate diversified training samples;

[0071] embedding quality labels on real and generated 3D models to construct a labeled dataset for training discriminators.

[0072] a user-generated content quality control method based on generative adversarial networks, using the system, comprising the following steps:

[0073] S1. using a deep convolutional neural network to detect illegal content in a 3D model submitted by a user, generating an illegal content discrimination result;

[0074] S2. based on the illegal content discrimination result, using a generative adversarial network to generate a preliminary 3D model that conforms to semantic and physical rationality;

[0075] S3. mapping the preliminary 3D model to a spatial topology graph, and based on the spatial topology graph, discriminating the topology rationality and map details of the preliminary 3D model, and generating a rationality score;

[0076] S4. Based on the rationality score, training the generative adversarial network and optimizing the discriminator using gradient descent method;

[0077] S5. Based on the trained generative adversarial network and optimized discriminator, generating high-quality 3D models that meet user needs;

[0078] S6. Providing a geometric structure drawing interface, texture drawing and editing interface, and edge drawing interface of the 3D model, receiving user's creative input;

[0079] S7. Topological refinement mapping of the generated 3D model to a spatial topological relationship graph, evaluating the legality and quality rationality score of the 3D model;

[0080] S8. Based on the evaluation results, providing creative suggestions to users and optimizing the creation interface;

[0081] S9. Repeat steps S5 to S8 until a 3D model that meets the quality requirements is generated.

[0082] The system of the present application realizes significant technical breakthroughs and improvements in many aspects. First, by introducing a content monitoring module of deep convolutional neural network, the system can effectively detect and filter illegal or inappropriate 3D content, greatly improving the safety and compliance of generated content. This not only protects users, but also provides protection for the healthy development of the entire 3D content ecosystem.

[0083] Secondly, the rendering generation module of the present application adopts advanced generative adversarial network technology, which can generate high-quality 3D models that meet semantic and physical rationality. The innovation of this module lies in that it not only considers the visual effect of the model, but also pays attention to the physical rationality, so as to generate more realistic and credible 3D content. This method greatly improves the quality and application value of the generated content, providing better content support for game development, virtual reality and other fields.

[0084] Furthermore, the quality monitoring module of the present application realizes accurate discrimination of 3D model topological rationality and texture details through innovative spatial topological relationship graph mapping technology. This method not only improves the accuracy of quality evaluation, but also provides clear direction guidance for subsequent optimization process. In this way, the system can continuously improve the quality of generated content to meet the high standard requirements of different application scenarios.

[0085] Moreover, the generator feedback module and the discriminator feedback module of the present application form a closed-loop optimization system, which continuously optimizes the generative adversarial network and the discriminator through gradient descent method. This dynamic optimization mechanism enables the system to adaptively improve the generation capability and continuously output high-quality 3D content. At the same time, this mechanism also enables the system to better adapt to different users' creative styles and needs, providing personalized content generation services.

[0086] The collaborative design of the content generation module and the user creation module of the present application skillfully balances the relationship between automatic generation and user creation. This design not only improves the creation efficiency, but also preserves the user's creative freedom, making the generated content both high-quality and personalized. In particular, the introduction of the real-time feedback unit provides immediate quality evaluation and optimization suggestions for users, greatly improving user experience and creation effect.

[0087] Finally, the quality evaluation module of the present application adopts a multi-dimensional evaluation method, comprehensively considering the topological rationality, physical consistency, semantic structure and aesthetic features of 3D models. This comprehensive evaluation method ensures that the generated content meets high standards in all dimensions, laying a solid foundation for subsequent application and display.

[0088] In summary, the system proposed by the present application realizes the comprehensive improvement and effective control of 3D content generation quality through the organic combination and collaborative work of various modules. It not only solves many problems existing in the prior art, but also makes significant progress in content security, generation quality, user experience and other aspects. This innovative method opens up new possibilities for the field of 3D content creation, and is expected to promote the rapid development and technological progress of related industries. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is the main logic block diagram of the system of the present application;

[0090] Figure 2 is the sub-flow chart of the content monitoring module of the present application;

[0091] Figure 3 is the sub-flow chart of the rendering generation module of the present application;

[0092] Figure 4 is the sub-flow chart of the quality monitoring module of the present application;

[0093] Figure 5 is the sub-flow chart of the content generation module of the present application;

[0094] Figure 6 is the sub-flow chart of the user creation module of the present application;

[0095] Figure 7A subflow chart of the quality assessment module of the present application; DETAILED DESCRIPTION

[0096] Please refer to the attached Figures 1-7 The present application provides a user-generated content quality control system based on a generative adversarial network and a method thereof. The system includes multiple functional modules that work together to achieve high-quality, personalized 3D content generation and effective quality control.

[0097] Firstly, the system of the present application includes a content monitoring module 1. It detects illegal content in the 3D model submitted by the user and generates an illegal content discrimination result. Specifically, the content monitoring module 1 can detect 3D models involving illegal content. The present application uses a multi-view projection conversion technology to convert 3D models into a standardized 2D view set, and then uses a pre-trained VGG16 network for feature extraction and classification. The system first renders the 3D model from multiple preset angles (typically configured as 8 horizontal angles and 3 vertical angles, a total of 24 projection views) to generate a standardized 2D image sequence. These images contain geometric structure, texture and surface detail information of the model. Subsequently, each 2D projection image is fed into a fine-tuned VGG16 network for feature extraction, and the extracted feature vectors are fused through a spatiotemporal attention layer, and finally classified and discriminated through a three-layer fully connected network (dimensions are 1024, 512 and 64 respectively). In practice, we found that when the above-mentioned view conversion technology is combined, and the number of convolution layers of the model reaches 13 layers, and the number of fully connected layers is 3, the illegal content detection accuracy can reach more than 95%. This multi-view projection method effectively solves the conversion problem from 3D structure to 2D feature, enabling mature 2D image classification networks such as VGG16 to be efficiently applied to 3D content review.

[0098] This module uses a deep convolutional neural network to detect illegal content in the 3D model submitted by the user and generates an illegal content discrimination result. Specifically, the content monitoring module 1 can detect 3D models involving illegal content. For example, the module can use a pre-trained VGG16 network as a feature extractor, and then connect a fully connected layer for classification. In practice, we found that when the number of convolution layers of the model reaches 13 layers, and the number of fully connected layers is 3, the detection accuracy can reach more than 95%.

[0099] In communication connection with the content monitoring module 1 is the rendering generation module 2. This module receives the illegal content discrimination result sent by the content monitoring module 1, and based on the result, uses a generative adversarial network to generate a preliminary 3D model that meets semantic and physical rationality. In a preferred embodiment of the present application, the rendering generation module 2 adopts a conditional generative adversarial network (cGAN) structure. The loss function of the generator G and the discriminator D can be represented as:

[0100] ,

[0101] wherein, is the input condition, is the real sample, is the random noise. By minimizing this loss function, the system can generate high-quality 3D model rendering effects. Practice shows that when the generator adopts the U-Net structure and the discriminator adopts the PatchGAN structure, the best rendering effect can be achieved.

[0102] Next, the quality monitoring module 3 is in communication connection with the rendering generation module 2. This module receives the preliminary 3D model generated by the rendering generation module 2, maps it into a spatial topological relationship graph, and based on this graph, judges the topological rationality and the mapping details of the preliminary 3D model, finally generates a rationality score. In an embodiment of the present application, the spatial topological relationship graph is represented by an adjacency matrix, and the matrix element A_ij represents the connection relationship between node i and node j. The calculation of the rationality score can use the following formula:

[0103] ,

[0104] wherein, is the topological rationality score, is the mapping detail score, is the weight coefficient. Experience shows that when takes a value between 0.6 and 0.7, a relatively balanced evaluation result can be obtained.

[0105] The generator feedback module 4 is in communication connection with the quality monitoring module 3, which receives the rationality score generated by the quality monitoring module 3, and based on this score, uses the gradient descent method to train the generative adversarial network. In the training process, the improved Wasserstein GAN (WGAN) algorithm is adopted, and its loss function is as follows:

[0106] ,

[0107] wherein, is the gradient penalty coefficient, usually taking a value of 10. This improvement can effectively improve the training stability and generation quality of the model.

[0108] In communication connection with the quality monitoring module 3 and the generator feedback module 4 is the discriminator feedback module 5. This module also receives the rationality score generated by the quality monitoring module 3, and based on this score, uses the gradient descent method to optimize the discriminator. In the optimization process, we adopt a dynamic learning rate adjustment strategy, with an initial learning rate of 0.0002, and a decay of half of the original value every 50 epochs. This strategy can quickly converge in the early stage of training, and can more finely adjust the model parameters in the later stage.

[0109] The content generation module 6 is communicatively connected with the generator feedback module 4 and the discriminator feedback module 5. This module receives the optimization results of the generator feedback module 4 and the discriminator feedback module 5, and generates high-quality 3D models that meet user needs based on these results. In a preferred embodiment of the present application, the content generation module 6 adopts a progressive growing of GANs structure. This structure starts from low resolution, gradually increases the number of network layers and output resolution, and can generate 3D models with higher quality and more detailed features.

[0110] The user creation module 7 is communicatively connected with the content generation module 6, which provides a geometric structure drawing interface, a texture drawing and editing interface, and an edge drawing interface for 3D models, and receives user creation inputs. This design realizes a human-computer collaborative creation mode, which not only takes advantage of the automatic generation capability of AI, but also retains the user's creation initiative. In practice, we found that when the response time of the interface is controlled within 100 ms, the user experience is best.

[0111] The quality assessment module 8 is communicatively connected with the content generation module 6 and the user creation module 7. It receives the 3D models generated by the content generation module 6 and the user creation inputs of the user creation module 7, maps the 3D models to a spatial topology relationship graph through topological refinement, and evaluates the legality and quality of the 3D models. The evaluation process adopts a multi-dimensional evaluation method, including topological rationality, physical consistency, semantic structure, and aesthetic features. The score of each dimension is obtained by weighted average to get the final comprehensive quality score.

[0112] Finally, the suggestion module 9 is communicatively connected with the quality assessment module 8 and the user creation module 7. This module receives the evaluation results of the quality assessment module 8, provides creation suggestions to the user based on the results, and optimizes the creation interface of the user creation module 7. For example, when the topological structure score of a certain area is low, the system will suggest the user to adjust the geometric structure of that area; when the texture details are insufficient, the system will prompt the user to increase the texture details.

[0113] Through the cooperative work of the above modules, the system of the present application can effectively control the quality of user-generated content while providing personalized creation experience. This method can not only be applied to the field of 3D modeling, but also can be extended to game development, virtual reality and other fields, and has wide application prospects.

[0114] In the system of the present invention, the quality monitoring module 3 includes a topology mapping unit 31, a feature extraction unit 32, and a discriminant network 33. The main function of the topology mapping unit 31 is to convert the preliminary 3D model into an undirected graph model and use an adjacency matrix to represent the topological relationship between the graphs. In a preferred embodiment of the present invention, the topology mapping unit 31 uses an improved graph convolutional network (GCN) algorithm to achieve this function. Specifically, for a given 3D model, the system first extracts its vertex and edge information and then constructs the adjacency matrix A. The calculation formula of the adjacency matrix is ​​as follows:

[0115] ,

[0116] In order to improve the expressive power of the model, the present invention also introduces a self-connection mechanism, that is, adding an identity matrix I to the diagonal of the adjacency matrix A. The adjacency matrix processed in this way can better capture the information of the node itself.

[0117] The feature extraction unit 32 is connected to the topology mapping unit 31. The main task of this unit is to extract topological features and mapping features from the undirected graph model. In the embodiment of the present invention, the feature extraction unit 32 adopts a multi-scale feature extraction method. For topological features, the system uses a multi-layer graph convolutional network. The output of each layer can be expressed as:

[0118] ,

[0119] in, is the adjacency matrix after adding self-connection, yes The degree matrix of It is The feature matrix of the layer, Array, is the activation function. By stacking multiple layers of graph convolution, the system can capture topological features of different scales.

[0120] For texture features, this paper uses an improved ResNet architecture. In practice, we found that a ResNet depth of 50 layers achieves excellent feature extraction results. Furthermore, to accommodate texture inputs of varying sizes, the system also introduces a spatial pyramid pooling layer, enabling the network to process input textures of any size.

[0121] The feature extraction unit 32 is communicatively connected with the discriminative network 33. The discriminative network 33 discriminates the rationality of the 3D model based on the extracted topological features and texture features, and generates a rationality score. In a preferred embodiment of the present application, the discriminative network 33 adopts a fully connected network structure enhanced by an attention mechanism. The attention mechanism can help the network focus on more important features, improving the accuracy of discrimination. The calculation formula of the rationality score is as follows:

[0122]

[0123] wherein, and are topological features and texture features, and are corresponding attention weights, , , and are learnable parameters, is a sigmoid activation function.

[0124] In the system of the present application, the content generation module 6 includes a geometric structure module 61, a texture module 62, an edge drawing module 63 and a style transfer module 64. The main function of the geometric structure module 61 is to generate the geometric structure of the 3D model based on the geometric structure information provided by the user. In an embodiment of the present application, the geometric structure module 61 adopts a 3D shape representation method based on implicit function. Specifically, the system uses a deep neural network f to approximate the signed distance function (SDF) of a 3D object:

[0125]

[0126] wherein, is a coordinate point in 3D space, is a network parameter. By optimizing the network parameter , the system can generate a 3D geometric structure that meets the user's needs.

[0127] The texture module 62 is communicatively connected with the geometric structure module 61. This module generates the texture of the 3D model based on the texture information provided by the user. In a preferred embodiment of the present application, the texture module 62 adopts a method based on neural texture synthesis. The system first uses a generation network G to generate a texture feature map:

[0128]

[0129] wherein, is random noise, is the texture condition provided by the user. Then, the system uses a rendering network​​​ Apply the texture feature map to the 3D model:

[0130] ,

[0131] where, is the geometric information of the 3D model. This method can generate high-quality and controllable 3D model textures.

[0132] The edge drawing module 63 is communicatively connected with the geometry module 61 and the texture module 62. The main task of this module is to map the edge-drawn line texture to the surface of the 3D model based on the user-drawn edges and texture hints. In the embodiment of the present application, the edge drawing module 63 adopts an optimization method based on differential rendering. The system defines an objective function:

[0133] ,

[0134] where, is the edge matching loss, is the smooth regularization term, is the balance coefficient. By minimizing this objective function, the system can generate 3D model textures that conform to the user-drawn edges.

[0135] Finally, the style transfer module 64 is communicatively connected with the geometry module 61, the texture module 62, and the edge drawing module 63. The main function of this module is to perform style transfer on the generated 3D model to maintain the consistency of the model style. In the preferred embodiment of the present application, the style transfer module 64 adopts a StyleGAN structure based on 3D adaptation. The system first uses a mapping network to map the input style code to the intermediate latent space W:

[0136] ,

[0137] Then, the system uses a series of adaptive instance normalization (AdaIN) layers to control the style during the generation process:

[0138] ,

[0139] where, is the feature map, and are the scaling and bias parameters calculated from In this way, the system can achieve precise style control of the 3D model.

[0140] Through the cooperative work of the above modules, the system of the present application can generate high-quality, style-consistent 3D models while ensuring the creative freedom of users. This method not only improves the efficiency of 3D content creation, but also greatly improves the quality and diversity of generated content.

[0141] In a preferred embodiment of the present application, the geometry module 61 includes a graph attention network 611 and a differential rendering unit 612. The main function of the graph attention network 611 is to learn the context information from the initial mesh graph and generate a geometric feature map in combination with the user-defined feature map. This method can better capture the global and local geometric features of the 3D model, thereby generating more reasonable and fine 3D structures.

[0142] Specifically, the graph attention network 611 adopts an improved graph attention mechanism. For a given node Its updated features can be represented as:

[0143] ,

[0144] Where, represents the neighbor set of node , is the attention coefficient, is a learnable weight matrix, is a nonlinear activation function. The calculation formula of the attention coefficient is as follows:

[0145] ,

[0146] Where, is a learnable attention vector, denotes the concatenation operation. In this way, the system can adaptively assign different importance to different neighbor nodes, thereby better capturing complex geometric structures.

[0147] In communication with the graph attention network 611 is the differential rendering unit 612. The main task of this unit is to convert the geometric feature map into a topological transformation and optimize the geometric shape of the initial generated mesh through backpropagation. In an embodiment of the present application, the differential rendering unit 612 adopts a differentiable rendering method based on ray tracing. The system defines a rendering loss function:

[0148] ,

[0149] Where, refers to the pixel value of the target image, and is the pixel value rendered by the system. By minimizing this loss function, the system can backpropagate the gradient to the geometric parameters of the 3D model, thus achieving precise control over the shape of the model.

[0150] The system of the present application also includes a user creation module 7, which includes a geometry drawing interface 71, a texture drawing and editing interface 72, an edge drawing interface 73, and a real-time feedback unit 74. The geometry drawing interface 71 provides an intuitive 3D modeling tool, allowing users to create and edit the basic shape of the 3D model through simple operations. In an embodiment of the present application, the interface adopts an implicit function-based modeling method, and users can create complex 3D structures by adding, deleting, or deforming basic geometric bodies.

[0151] The texture drawing and editing interface 72 supports users to draw and edit textures directly on the surface of the 3D model. This interface not only supports traditional 2D texture mapping, but also supports the generation of normal maps and mask maps. Preferably, the system adopts a projection-based texture drawing method, allowing users to draw directly in the 3D view, greatly improving the intuitiveness and efficiency of texture creation.

[0152] The edge drawing interface 73 allows users to draw edge lines directly on the 3D model. In a preferred embodiment of the present application, the interface adopts a view-based line drawing technique. The system first projects the 2D lines drawn by the user onto the surface of the 3D model, and then uses an optimization algorithm to smoothly integrate these lines into the geometry of the model. This method can generate edge lines that both conform to the user's intention and maintain 3D consistency.

[0153] The real-time feedback unit 74 is in communication connection with the geometry drawing interface 71, the texture drawing and editing interface 72, and the edge drawing interface 73. The main function of this unit is to display quality assessment results and creation suggestions in real time during the user creation process. In order to achieve real-time feedback, the system adopts a lightweight evaluation network and a caching mechanism. The evaluation network uses a knowledge distillation technique to extract key knowledge from a complete quality evaluation model, building a small network with higher computational efficiency. At the same time, the system caches the results of common creation operations, further improving the feedback speed. Practice shows that when the feedback delay is controlled within 100ms, the user experience is best.

[0154] The system of the present application also includes a quality assessment module 8, which includes a topology refinement unit 81, a multi-dimensional assessment unit 82, and a comprehensive scoring unit 83. The main task of the topology refinement unit 81 is to perform topology refinement on the generated 3D model, generating a high-precision spatial topology relationship graph. In response to the demand for real-time interaction, the present application adopts an adaptive multi-resolution refinement strategy rather than global tetrahedral subdivision. The system first converts the 3D model into a base tetrahedral mesh, and then prioritizes according to the user interaction area and visual importance, performing local refinement only on the current focus area of creation. At the same time, the system uses a GPU-accelerated parallel computing architecture to distribute the subdivision calculation to hundreds of processing cores for simultaneous execution.

[0155] The sub-process uses an optimized iterative formula:

[0156] ,

[0157] wherein, is the vertex to be subdivided, is the neighbor vertex set of vertex , denotes the number of neighbor vertices.

[0158] In communication with the topology refinement unit 81 is the multi-dimensional assessment unit 82. This unit assesses the quality of the 3D model from multiple dimensions such as topology rationality, physical consistency, semantic structure, and aesthetic features. In the preferred embodiment of the present application, the multi-dimensional assessment unit 82 uses an evaluation method based on multi-task learning. The system defines a multi-task loss function:

[0159] ,

[0160] wherein, is the evaluation loss of each dimension, is the corresponding weight coefficient. By jointly optimizing this multi-task loss function, the system can simultaneously learn the evaluation criteria of multiple dimensions, thereby achieving more comprehensive and accurate quality assessment.

[0161] The multi-dimensional assessment unit 82 is in communication with the comprehensive scoring unit 83. The main function of the comprehensive scoring unit 83 is to generate a comprehensive quality score based on the multi-dimensional assessment results. In the embodiment of the present application, the comprehensive scoring unit 83 uses a scoring method based on fuzzy logic. The system first converts the assessment results of each dimension into a fuzzy set, and then calculates the final comprehensive score through fuzzy inference rules. This method can better handle the uncertainty and fuzziness in the assessment process, resulting in a more reasonable comprehensive score.

[0162] In addition, the system of the present application also includes a data augmentation module 9. The main function of this module is to perform data augmentation on high-quality real 3D models, generate diverse training samples, and embed quality labels on real and generated 3D models to build a labeled dataset for training the discriminator. In the preferred embodiment of the present application, the data augmentation module 9 adopts a data augmentation method based on a generative adversarial network. The system uses a conditional generative adversarial network to generate diverse 3D model variants:

[0163] ,

[0164] wherein, is random noise, is a control condition (such as category, attribute, etc.), is the generated 3D model. By adjusting the control condition , the system can generate 3D models of various styles and features, greatly increasing the diversity of training data.

[0165] At the same time, the data augmentation module 9 also implements an automatic quality labeling function. The system uses a pre-trained quality evaluation model to score the generated 3D models and embed the score results as quality labels into the models. This method not only improves the quality of the training data, but also provides more accurate supervision signals for the training of the discriminator.

[0166] Finally, the present application also provides a user-generated content quality control method based on a generative adversarial network. The method includes the following steps:

[0167] S1. Using a deep convolutional neural network to detect illegal content in the 3D model submitted by the user, and generating an illegal content discrimination result;

[0168] S2. Based on the illegal content discrimination result, using a generative adversarial network to generate a preliminary 3D model that meets semantic and physical rationality;

[0169] S3. Mapping the preliminary 3D model to a spatial topology relationship graph, and based on the spatial topology relationship graph, discriminating the topology rationality and texture details of the preliminary 3D model to generate a rationality score;

[0170] S4. Based on the rationality score, using gradient descent method to train the generative adversarial network and optimize the discriminator;

[0171] S5. Based on the trained generative adversarial network and the optimized discriminator, generating a high-quality 3D model that meets the user's demand;

[0172] S6. Providing a geometric structure drawing interface, a texture drawing and editing interface, and an edge drawing interface of the 3D model, and receiving the user's creative input;

[0173] S7. Topologically refine the generated 3D model and map it to a spatial topology graph, and evaluate the legitimacy and quality of the 3D model to obtain a rationality score;

[0174] S8. Based on the evaluation result, provide creation suggestions to the user and optimize the creation interface;

[0175] S9. Repeat steps S5 to S8 until a 3D model meeting the quality requirements is generated.

[0176] Through this method, the system of the present application can effectively control the quality of user-generated content while providing personalized creation experience. This method can not only be applied to the field of 3D modeling, but also be extended to game development, virtual reality and other fields, and has wide application prospects.

[0177] It should be noted that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A user-generated content quality control system based on a generative adversarial network, characterized in that: include: Content monitoring module, used to: Use deep convolutional neural networks to detect illegal content in user-submitted 3D models; Generate illegal content identification results; A rendering generation module, in communication with the content monitoring module, is configured to: receiving the illegal content identification result sent by the content monitoring module; Based on the illegal content identification result, a preliminary 3D model that is consistent with semantics and physical rationality is generated using a generative adversarial network; The quality monitoring module is in communication with the rendering generation module and is used to: Receiving the preliminary 3D model generated by the rendering generation module; Mapping the preliminary 3D model into a spatial topological relationship graph; Based on the spatial topological relationship diagram, determining the topological rationality and mapping details of the preliminary 3D model; Generate a plausibility score; A generator feedback module is in communication with the quality monitoring module and is used to: Receiving a rationality score generated by the quality monitoring module; Based on the rationality score, training the adversarial generative network using gradient descent; A discriminator feedback module is communicatively connected to the quality monitoring module and the generator feedback module, and is configured to: Receiving a rationality score generated by the quality monitoring module; Based on the rationality score, optimizing the discriminator using gradient descent; A content generation module, in communication with the generator feedback module and the discriminator feedback module, configured to: Receiving optimization results of the generator feedback module and the discriminator feedback module; Based on the optimization results, a high-quality 3D model that meets user needs is generated; A user creation module is in communication with the content generation module and is used to: Provides 3D model geometry drawing interface, texture drawing and editing interface, and edge drawing interface; Receive creative input from users; A quality assessment module, in communication with the content generation module and the user creation module, configured to: receiving a 3D model generated by the content generation module and a user creation input from the user creation module; Topologically refine the 3D model and map it into a spatial topological relationship graph; A reasonableness score to assess the legality and quality of the 3D model; A suggestion module, in communication with the quality assessment module and the user creation module, configured to: receiving an evaluation result of the quality evaluation module; Providing creative suggestions to the user based on the evaluation results; Optimize the creation interface of the user creation module.

2. The system according to claim 1, wherein: The content monitoring module includes: A feature extraction unit, used to extract depth features from the 3D model submitted by the user; an illegal content identification unit, communicatively connected to the feature extraction unit, and configured to identify illegal content in the deep features based on a pre-trained deep learning model; The result generating unit is connected to the illegal content identifying unit for communication and is used to generate an illegal content identification result.

3. The system according to claim 1, wherein: The rendering generation module includes: A conditional encoder for encoding user-defined content features into a conditional vector; Style encoder, used to encode user-defined style features into a style vector; a generation network, communicatively connected to the conditional encoder and the style encoder, for generating a 3D model based on the conditional vector and the style vector; The rendering network is connected to the generation network for rendering the 3D model generated by the generation network to generate a preliminary 3D model that conforms to semantic and physical rationality.

4. The system according to claim 1, wherein: The quality monitoring module includes: A topological mapping unit, used to convert the preliminary 3D model into an undirected graph model and use an adjacency matrix to represent the topological relationship between graphs; A feature extraction unit, in communication with the topology mapping unit, configured to extract topological features and mapping features from the undirected graph model; A discriminant network is connected to the feature extraction unit for discriminating the rationality of the 3D model based on the topological features and the mapping features, and generating a rationality score.

5. The system according to claim 1, wherein: The content generation module includes: A geometry module, used to generate the geometry of a 3D model based on the geometry information provided by the user; a texture module, in communication with the geometry module, for generating a texture of the 3D model based on texture information provided by a user; an edge drawing module, in communication with the geometry module and the texture module, for mapping the edge drawn line texture to the 3D model surface based on the edge and texture hints drawn by the user; A style transfer module is in communication with the geometry module, the texture module and the edge rendering module, and is used to transfer the style of the generated 3D model to maintain the consistency of the model style.

6. The system according to claim 5, characterized in that The geometric structure module includes: Graph Attention Network, which is used to learn contextual information from the initial grid graph and generate a geometric feature map in combination with the user-defined feature map; A differential rendering unit is connected to the graph attention network, and is used to convert the geometric feature graph into a topological transformation and optimize the geometry of the initial generated mesh through back propagation.

7. The system according to claim 1, wherein: The user creation module includes: A geometry drawing interface for receiving the 3D model geometry drawn by the user; A texture drawing and editing interface, used to receive 3D model textures drawn and edited by a user, wherein the texture drawing and editing interface supports the generation of texture maps, normal maps, and mask maps; Edge drawing interface, used to receive the 3D model edge drawn by the user; A real-time feedback unit is connected to the geometric structure drawing interface, the texture drawing and editing interface, and the edge drawing interface for displaying quality assessment results and creation suggestions in real time during the user's creation process.

8. The system according to claim 1, wherein: The quality assessment module includes: Topology refinement unit, used to refine the generated 3D model topology and generate a high-precision spatial topology relationship diagram; a multi-dimensional evaluation unit, in communication with the topology refinement unit, for evaluating the quality of the 3D model from multiple dimensions, including topological rationality, physical consistency, semantic structure, and aesthetic features; The comprehensive scoring unit is in communication with the multi-dimensional evaluation unit and is configured to generate a comprehensive quality score based on the multi-dimensional evaluation result.

9. The system according to claim 1, wherein: Also includes data augmentation modules for: Perform data augmentation on high-quality real 3D models to generate diverse training samples; Embed quality labels on real and generated 3D models to construct an annotated dataset for training the discriminator.

10. A method for controlling user-generated content quality based on a generative adversarial network, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Use a deep convolutional neural network to detect illegal content in user-submitted 3D models and generate illegal content identification results; S2. Based on the illegal content identification results, generate a preliminary 3D model that is semantically and physically reasonable using a generative adversarial network; S3. Mapping the preliminary 3D model into a spatial topological relationship diagram, determining the topological rationality and mapping details of the preliminary 3D model based on the spatial topological relationship diagram, and generating a rationality score; S4. Based on the rationality score, use gradient descent to train the adversarial generative network and optimize the discriminator; S5. Generate high-quality 3D models that meet user needs based on the trained GAN and optimized discriminator. S6. Provide a 3D model geometry drawing interface, a texture drawing and editing interface, and an edge drawing interface to receive user creative input; S7. Topologically refine the generated 3D model and map it into a spatial topological relationship diagram, and evaluate the legitimacy and quality rationality score of the 3D model; S8. Based on the evaluation results, provide users with creative suggestions and optimize the creation interface; S9. Repeat steps S5 to S8 until a 3D model that meets quality requirements is generated.

Images