Generative adversarial network-based intelligent sliding plate bowl pool form design generation method
Through the GAN-based intelligent method, users can quickly generate skateboard bowl pool models that meet their needs, solving the problems of traditional design complexity and subjectivity and achieving efficient and stable design output.
Patent Information
- Application Number
- CN202510430347.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-10-21
AI Technical Summary
Traditional skateboard bowl design relies on complex professional modeling software, has low design efficiency, and its quality is affected by subjective factors. It has a high threshold and limits the designer's creativity and popularity.
An intelligent method based on generative adversarial networks (GANs) is adopted to generate a skateboard bowl model based on user input constraints. This includes preprocessing image features and generating a depth map by a generator, and quickly outputting a 3D model.
It simplifies the modeling process, lowers the barrier to entry, improves design efficiency and quality stability, and meets user needs.
Smart Images
Figure CN120822261A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of intelligent design of sports facilities, and specifically to a method for generating the shape design of an intelligent skateboard bowl based on GAN (generative adversarial network). Background Art
[0002] Traditional skateboard bowl design relies on specialized modeling software and requires a solid foundation, making the entire design process cumbersome and time-consuming. Even simple adjustments to a complete skateboard bowl model require extensive manual adjustments by the designer. Furthermore, due to the current lack of unified technical specifications and standards, the shape of skateboard bowls relies heavily on the designer's experience, and the quality of the design is heavily influenced by subjective factors, severely restricting design efficiency and stability. For many skateboarding enthusiasts and practitioners, a lack of relevant modeling skills prevents them from engaging in the high technical barriers and complexity of design work, limiting their creativity during the design process.
[0003] Therefore, developing an intelligent skateboard bowl pool design and generation method that can simplify the modeling process and lower the usage threshold is of great value and significance for my country's urban construction and the popularization of skateboarding. Summary of the Invention
[0004] One objective of this application is to propose a method for intelligent bowl shape design based on GAN (Generative Adversarial Network). This method eliminates the need for users to master complex modeling software and techniques. Instead, users simply input constraints based on their needs and site boundary conditions to quickly obtain a skateboard bowl model that meets their requirements, thereby improving design efficiency and ensuring design quality.
[0005] To achieve the above objectives, the present invention proposes an intelligent bowl pool morphology design generation method based on a generative adversarial network, comprising:
[0006] Obtain constraints based on user requirements and site boundary conditions;
[0007] Input the constraint conditions into the preprocessing program to generate input image features to be processed;
[0008] The image features to be processed are fed into the generator of the pre-trained GAN model to generate a depth map containing the elevation information of the design. The 3D model of the skateboard bowl is then generated based on the depth map generated in the previous step.
[0009] The technical solutions provided in the embodiments of the present application may have the following beneficial effects:
[0010] Constraints are obtained based on the user's needs and the site's boundary conditions. These constraints are then fed into a preprocessor to generate input image features. These features are then fed into the generator of a pretrained GAN model to generate a depth map containing the design proposal. Based on the correspondence between pixel brightness and relative elevation in the depth map, a 3D model of the skateboard bowl's upper surface design is output, which in turn rapidly generates a skateboard bowl design model, enabling intelligent skateboard bowl design. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] To more intuitively illustrate the technical solutions in the embodiments of this application, a brief introduction to the drawings provided in the related technical description is now provided. It is obvious that the following drawings only illustrate some of the embodiments of this application, and those skilled in the art can deduce other similar illustrated contents based on these drawings without inventive effort.
[0012] Figure 1 This is a flow chart of a method for intelligent bowl pool morphology design generation based on a generative adversarial network provided in an embodiment of the present application;
[0013] Figure 2 This is a detailed framework of an intelligent bowl pool morphology design generation method based on a generative adversarial network provided in an embodiment of the present application;
[0014] Figure 3 is a typical set of mapping image data sets provided by the embodiments of the present application, wherein a is the image feature 1 to be processed, which represents the depth map containing the design information, and b is the image feature 2 to be processed, which represents the design boundary condition;
[0015] Figure 4 It is the normalized mapping data provided in the embodiment of the present application;
[0016] Figure 5 This is a graph showing the change trend of the loss function during the GAN training process using the pix2pixHD framework provided in an embodiment of the present application;
[0017] Figure 6 This is a flow chart of generating a design solution based on boundary conditions provided in an embodiment of the present application;
[0018] Figure 7 This is a diagram of a method for evaluating model generation capability using Gaussian curvature provided in an embodiment of the present application; DETAILED DESCRIPTION
[0019] To more clearly illustrate the purpose and technical solutions of this application, an embodiment of an intelligent bowl and pool morphology design generation method based on a generative adversarial network is now described in conjunction with the accompanying drawings. It should be noted that the embodiments described are only some examples of this application, not all. Based on the implementation methods of this application, any other implementation solutions obtained by other technicians in this field without engaging in creative work are within the scope of protection of this application.
[0020] Figure 1 A flow chart of an intelligent bowl pool morphology design generation method based on a generative adversarial network provided in an embodiment of the present application.
[0021] like Figure 1 As shown, the method mainly includes the following three steps:
[0022] Step S101: Obtain constraint conditions based on user requirements and site boundary conditions.
[0023] Step S102: Process the constraint conditions to generate input image features to be processed, wherein the image features include the contour and height information of the site boundary.
[0024] In the example of this application, the image features to be processed are generated by the input constraints, including two parts: user needs and site boundaries; the image features to be processed use 0 in the grayscale image to represent the background and 255 in the grayscale image to represent the contour.
[0025] In step S103, the features of the input image to be processed are input into a pre-trained generator to generate a depth map containing design information, and a three-dimensional model of the skateboard bowl is generated according to the depth map.
[0026] In an embodiment of the present application, before the input image features to be processed are input into a pre-trained GAN model generator to generate a depth map containing design scheme information, the following steps are also included: obtaining a design drawing or model, and separating and reconstructing its upper surface; projecting the reconstructed three-dimensional upper surface to the XOY plane of the world coordinate system, generating a two-dimensional depth map containing height information as the image feature 1 to be processed, and extracting the boundary information of the depth map as the image feature 2 to be processed; combining image feature 1 and image feature 2 according to the mapping relationship to generate the image feature to be trained; then dividing the image features to be trained into a training set and a test set, and using the training set to train the pix2pixHD framework to finally obtain a pre-trained generative adversarial network model.
[0027] Furthermore, image feature 1 and image feature 2 are combined into the image feature to be trained based on the mapping relationship. The model learns the mapping relationship between image feature 2 and image feature 1 by learning the image feature to be trained. The training set is used to train the generator and discriminator, while the test set is used to evaluate the generation results of the model.
[0028] Specifically, the evaluation of the model generation results includes the following steps: first, Gaussian curvature analysis is performed on the image generated according to the test set to obtain its first surface quality coefficient; second, the grayscale height is imported and the surface reconstruction is completed using the depth map in the test set to obtain the second surface quality coefficient; finally, the quality coefficient of the generated surface is obtained by calculating the absolute value of the difference between the first surface quality coefficient and the second surface quality coefficient.
[0029] In order to facilitate researchers in this field to understand the above embodiments more clearly, the following will be described in detail with reference to the accompanying drawings.
[0030] Figure 2 The detailed framework of the intelligent bowl pool morphology design generation method based on the generative adversarial network in the embodiment of the present application is shown. As shown in the figure, this method is mainly divided into two parts. Among them, the first part is data processing and generative adversarial network training: collecting original design data such as design drawings or design models and calibrating and cleaning the key elements in the original design data, normalizing the data to obtain the image features to be processed, and training the GAN network based on the pix2pixHD framework. The second part is the design result generation and quality evaluation: inputting the design boundary conditions into the generative adversarial network trained in the first part to obtain the design result of the skateboard bowl pool, performing Gaussian curvature analysis on the output results to evaluate the model performance and verifying the rationality of the design generated by the model based on this indicator.
[0031] in, Figure 2 The first part of data processing and generative adversarial network training involves collecting original design data such as design drawings or design models and calibrating and cleaning the key elements in the original design data. All collected data are processed, and the collected drawing data are reconstructed and the upper surface information is separated. The upper surface information of the collected 3D model is separated; the collected upper surface information is structurally normalized and converted into an 8-bit grayscale image containing surface height information as the image feature to be processed 1, such as Figure 3 (a) shows; the boundary contours of all the image features to be processed 1 are calibrated to obtain the design boundary information of the image features to be processed 2, as shown in Figure 3(b) A total of 282 grayscale image datasets for GAN network training and 241 pairs of mapping images for training the GAN network of the pix2pixHD framework have been obtained. In 8-bit depth grayscale images, 0 represents the lowest point, 255 represents the highest point, and integer values between 0 and 255 represent the height between the lowest and lowest points.
[0032] in, Figure 2 The first part of data processing and generative adversarial network training is to normalize the data to obtain the image features to be processed, such as Figure 4 The data format and size are organized to facilitate model processing. Image pairs are organized based on the correspondence between boundary conditions and grayscale images. Automatic cropping and scaling are performed on the images using the object detection algorithm in the OpenCV computer vision library. This ensures that the non-blank portion (non-zero pixel value) is centered and that all images are of the same size, facilitating unified scaling and cropping during model pre-processing.
[0033] in, Figure 2 The first part of data processing and generative adversarial network training is to train the GAN network based on the pix2pixHD framework, such as Figure 5 As shown in the figure, the loss of the generator and discriminator changes with the increase of training generations during the training of the adversarial generative network. The key parameter of the pix2pixHD framework used in the embodiment of the present application is λFM. When λFM=10, the pix2pixHD algorithm works well. When the loss value stabilizes during training, the training can be stopped.
[0034] in, Figure 2 The second part is the design boundary conditions in the design result generation and quality assessment. The generated adversarial network trained in the first part is input to obtain the design result of the skateboard bowl, such as Figure 6 As shown, the boundary conditions are input into the generator of the adversarial generative network to generate a grayscale image containing the design information, and the modeling of the skateboard bowl model is completed through the automatic modeling program.
[0035] in, Figure 2 In the second part of design result generation and quality assessment, the Gaussian curvature analysis of the output results is used to evaluate the model performance and verify the rationality of the design generated by the model based on this indicator. Furthermore, the evaluation of the model generation capability is as follows Figure 7 Specifically, the evaluation of the surface in this application is based on the surface quality coefficient C. The surface quality coefficient is calculated as follows:
[0036]
[0037] Where n1 and n2 are the number of segments in the U and V directions respectively; g 正 is the number of grids whose Gaussian curvature absolute value is greater than that of the sawtooth part and whose value is positive; g 负 is the number of grids whose absolute value of Gaussian curvature is greater than that of the sawtooth part and whose value is negative; g 非 The surface quality coefficient C ranges from 0 to 1. When C is close to 1, it means that the surface has fewer Gaussian curvature anomalies and the surface is smoother. When C is close to 0, it means that the surface has more Gaussian curvature anomalies and the surface is less smooth.
[0038] Specifically, when using the test set to evaluate the performance of the model, first, the first surface quality coefficient C1 of the surface generated by the model when the input test design boundary is used and the second surface quality coefficient C2 of the control surface in the test set are calculated according to formula (1); secondly, the difference between the first surface quality coefficient C1 of the surface design generated by the model and the second surface quality coefficient C2 of the control surface design in the test set is calculated according to formula (1) to obtain the final generated quality coefficient S. The calculation method of the generated quality coefficient S is as shown in formula (2):
[0039] S=C1-C2(2)
[0040] A larger generation quality coefficient S indicates a better generation result compared to the reference result. Specifically, when the generation quality coefficient S is greater than 0, the generation result is superior to the reference result in terms of surface generation quality; when S is equal to 0, the generation result is consistent with the reference result in terms of surface generation quality; and when S is less than 0, the generation result is inferior to the reference result in terms of surface generation quality.
[0041] Throughout this specification, terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics included in certain embodiments. These terms do not necessarily refer to the same embodiment or example. Related features, structures, materials, or characteristics may be appropriately combined in one or more embodiments. Furthermore, those skilled in the art may combine features from different embodiments or examples in this specification without conflicting applications.
[0042] Although the embodiments of the present application are shown and described herein, these embodiments are merely illustrative and not limiting of the present application. Persons skilled in the art may modify, change or replace the above embodiments within the scope of the present application.
Claims
1. A GAN-based intelligent skateboard park morphology design generation method, comprising the following steps: Step 101: Obtain constraint conditions based on user requirements and site boundary conditions; Step 102: Process the constraint conditions to generate input image features to be processed, wherein the image features include the contour and height information of the site boundary; In step 103 , the features of the input image to be processed are input into a pre-trained generator to generate a depth map containing design information, and a three-dimensional model of the skateboard bowl is generated based on the depth map.
2. The method according to claim 1, wherein the features of step S102 include: Normalize the input user requirements and site boundary conditions to generate a grayscale image containing boundary information as the input image feature to be processed; The grayscale value ranges from 0 to 255 to represent the height of the skate park surface from the lowest to the highest. The brightness value of the pixel in the grayscale image represents the relative elevation of the point projected on the XOY plane of the world coordinates in the modeling space. The brightness value of the point at the highest relative elevation is 255, and the brightness value of the point at the lowest relative elevation is 0.
3. The method according to claim 1, wherein The training process of the GAN model in step S103 includes: Collect skate park design drawings, design models, or related data sets; Normalize the data and convert it into depth map and boundary condition image pairs; The training dataset is used to train the generator and discriminator in the pix2pixHD framework to learn the mapping between the skatepark design boundaries and the corresponding depth maps.
4. The method according to claim 1, wherein The process of generating the three-dimensional model of the skate park in step S104 includes: Utilize the depth map to perform surface reconstruction and generate an upper surface model; Surface features are identified through object detection algorithms and matched to a parametric component library to generate a complete model including bowls and street components.
5. The method according to claim 1, wherein the three-dimensional model generated in step S103 adopts a surface quality coefficient C during testing and actual performance evaluation, and its calculation method formula (1) is: Where n1 and n2 are the number of segments in the U and V directions respectively; g+ is the number of grids whose Gaussian curvature is greater than that of the sawtooth part and whose value is positive; g- is the number of grids whose Gaussian curvature is greater than that of the sawtooth part and whose value is negative; g- is the number of grids at non-target surfaces.
6. When evaluating model performance using a test set, the generated surface quality evaluation method based on the surface quality coefficient C according to claim 5 is to calculate the generated surface quality coefficient S according to formula (2): S=C1-C2(2) Where C1 is the first surface quality coefficient of the surface generated by the model when the test design boundary is input, and C2 is the second surface quality coefficient corresponding to the control surface in the test set. A larger generation quality coefficient S indicates that the generated result is better than the reference result. In particular, when the generation quality coefficient S is greater than 0, it means that the generated result is better than the reference result in terms of surface generation quality. When S is equal to 0, it means that the generated result is consistent with the reference result in terms of the surface generation quality; When S is less than 0, it means that the generated result is inferior to the reference result in terms of surface generation quality.