Clothing pattern intelligent design method and system based on parameterized generative adversarial network

Abstract

This invention discloses an intelligent design method and system for clothing patterns based on parametric generative adversarial networks, relating to the field of clothing design technology. The method includes: acquiring the basic graphic type and initial growth rule parameters input by the user, constructing an initial pattern growth state, and inputting it into a pre-trained generator; during the pattern growth process, receiving text instructions input by the user in real time, converting them into feature vectors, and generating growth control parameters including rule adjustments, morphological intervention, and texture density based on the current pattern state; the generator iteratively growing and updating the pattern according to these parameters until the user terminates or a preset number of steps is reached, ultimately outputting pattern data adapted to the clothing pattern. This method, by integrating text semantics and pattern growth state, enables users to dynamically control the pattern's shape, texture, and growth direction, supports growth process backtracking and pattern adaptation, and enhances the intelligence, personalization, and flexibility of clothing pattern design.

Description

Technical Field

[0001] This invention relates to the field of clothing design technology, and more specifically, to a method and system for intelligent design of clothing patterns based on parametric generative adversarial networks. Background Technology

[0002] Apparel pattern design is undergoing a technological leap from static creation to dynamic evolution. Traditional models that rely on manual drawing or template reuse are unable to meet the complex demands for pattern growth, interactivity, and parameter controllability in personalized customization. Naturally growing patterns, such as mushroom textures and fern branching, follow underlying evolutionary rules. However, existing generation models mostly focus on single static output results and lack the ability to dynamically model from initial elements to complete patterns. While reaction-diffusion systems can simulate biological texture generation, they cannot respond to real-time intervention from text commands. The Lin system can depict branching growth patterns, but it is difficult to adaptively match the irregular boundaries of apparel patterns.

[0003] The disconnect in cross-modal interaction further highlights the technical bottlenecks: semantic instructions such as having petals gather towards the collar are difficult to translate into quantifiable generation parameters, resulting in a semantic gap between user creativity and algorithm output. At the same time, clothing patterns need to be adapted to irregular carriers such as sleeves and bodies. Existing models handle growth boundaries mechanically and rigidly, often resulting in pattern distortion or unbalanced white space. The breakthrough of parametric generative adversarial networks lies in embedding growth rules (reaction diffusion parameters, Lin rewriting rules) into the generator parameter module. Through multimodal feature alignment technology, it achieves a precise mapping of text semantics, growth parameters, and visual form, giving the pattern a biological-like dynamic evolution ability. At the same time, through the boundary awareness mechanism, it adapts to the characteristics of clothing carriers, constructing a new design paradigm that combines intelligent growth and controllable interaction.

[0004] Therefore, existing technologies suffer from problems such as difficulty in converting textual semantics into growth parameters, lack of dynamic pattern evolution, and poor boundary adaptability. Summary of the Invention

[0005] To overcome the problems of existing technologies, such as difficulty in converting text semantics into growth parameters, lack of dynamic pattern evolution, and poor boundary adaptability, this invention discloses an intelligent design method and system for clothing patterns based on parametric generative adversarial networks, which can effectively solve the above-mentioned technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A smart design method for clothing patterns based on parametric generative adversarial networks includes:

[0008] Obtain the initial seed parameters input by the user, which include the basic graphic type and initial growth rule parameters;

[0009] The initial state of pattern growth is constructed based on the initial seed parameters, and the initial state is input into the generator of the pre-trained parameterized generative adversarial network.

[0010] During the pattern growth process, text commands input by the user are received in real time. These text commands are used to indicate the pattern growth direction, texture features, or morphological features.

[0011] The text instructions are converted into feature vectors, and growth control parameters are generated by combining them with the current pattern growth state. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters.

[0012] The growth control parameters are input into the generator, and the generator iteratively grows and updates the current pattern to obtain a new pattern growth state.

[0013] Repeat the steps of receiving text instructions, generating growth control parameters, and iteratively updating growth until the user terminates growth or the preset number of growth steps is reached.

[0014] Obtain the final pattern growth state and output the corresponding clothing pattern data.

[0015] Preferably, the step of constructing the initial state of pattern growth based on the initial seed parameters includes:

[0016] An initial graphic matrix is ​​generated in a two-dimensional grid based on the basic graphic type, which includes points, lines, or polygons.

[0017] The initial growth rule parameters are used to initialize the concentration field parameters of the reaction diffusion system or the string parameters of the Ringel system, forming the initial state matrix of pattern growth.

[0018] Preferably, converting the text instruction into a feature vector includes:

[0019] The text instructions are encoded using a pre-trained multimodal model to obtain text feature vectors;

[0020] The text feature vector is aligned with the visual feature vector of the current pattern growth state to obtain a fused feature vector, which is used as the feature vector corresponding to the text instruction.

[0021] Preferably, the step of generating growth control parameters based on the current pattern growth state includes:

[0022] Extract the morphological features of the current pattern growth state, including contour complexity, texture distribution density, and symmetry;

[0023] Based on the matching degree between the feature vector and the morphological feature, the rule adjustment weight, morphological intervention area and texture density threshold are determined.

[0024] The growth control parameters are generated by adjusting the weights according to the rules to produce reaction diffusion rate parameters or Lin's rewrite rule parameters, by generating local coordinate offset parameters according to the morphological intervention region, and by generating pixel fill density parameters according to the texture density threshold.

[0025] Preferably, the training process of the parameterized generative adversarial network includes:

[0026] Construct a training dataset, which includes natural growth pattern sequences, corresponding text descriptions, and growth parameter records;

[0027] Construct a generator and a discriminator. The generator includes a parameterized rule module and an iterative growth module. The discriminator is used to distinguish between generated patterns and real patterns.

[0028] The generator and discriminator are trained alternately using the similarity between the generated pattern and the real pattern, the matching degree of the text instruction, and the growth continuity as loss functions, until the loss function converges.

[0029] Preferably, the parameterization rule module of the generator is used to receive growth control parameters and output corresponding reaction-diffusion equation parameters, Ringelmann system rewriting rules, or local morphology adjustment coefficients;

[0030] The iterative growth module is used to iteratively update the current pattern state based on the output of the parameterized rule module.

[0031] Preferably, the iterative growth and update of the current pattern through the generator includes:

[0032] The iterative growth module of the generator performs growth operations on local regions of the current pattern based on growth control parameters. The growth operations include diffusion growth, branching growth, texture filling, or shape deformation.

[0033] During the growth operation, the pattern boundary conditions are detected in real time. When the growth area touches the preset boundary, the growth direction is automatically adjusted or the growth rate is reduced.

[0034] Preferably, it further includes:

[0035] During the iterative growth and update process, the growth control parameters and corresponding pattern states at each step are recorded to form growth sequence data;

[0036] Upon receiving the user's backtracking instruction, the pattern growth state is restored to the specified step based on the growth sequence data, and growth is restarted from that state.

[0037] Preferably, the output of the corresponding clothing pattern data includes:

[0038] Convert the final pattern growth state into a vector graphics format or a bitmap format;

[0039] Based on preset clothing pattern parameters, the pattern data is adapted and adjusted to generate pattern data adapted to a specified clothing part, including sleeves, body, or collar.

[0040] Preferably, the intelligent design system for clothing patterns based on parametric generative adversarial networks includes:

[0041] An initial parameter receiving module is used to obtain the initial seed parameters input by the user, the initial seed parameters including the basic graphic type and initial growth rule parameters;

[0042] An initial state construction module is used to construct the initial state of pattern growth based on the initial seed parameters;

[0043] A parameterized generative adversarial network includes a generator and a discriminator. The generator is used to receive the pattern growth state and growth control parameters, and output the iteratively updated pattern growth state.

[0044] The instruction processing module is used to receive text instructions input by the user in real time and convert the text instructions into feature vectors;

[0045] The control parameter generation module is used to generate growth control parameters by combining the current pattern growth state with the feature vector. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters.

[0046] An iterative growth module is used to input the growth control parameters into the generator to drive the pattern to perform iterative growth and update.

[0047] The results output module is used to output the corresponding clothing pattern data after growth has ended.

[0048] Compared with existing technologies, the beneficial effects of this invention are as follows: This method addresses the challenge of converting text semantics into growth parameters. It employs a pre-trained multimodal model to encode text instructions into feature vectors, which are then aligned and fused with the current pattern visual feature vectors. Combined with morphological feature matching, it generates rules for adjusting and intervening in morphological parameters, achieving a precise mapping from text semantics to growth parameters. This bridges the semantic gap between natural language descriptions and pattern generation parameters, allowing users to efficiently control the pattern's evolution direction through intuitive text. Secondly, to overcome the limitation of patterns lacking dynamic evolution, an iterative growth mechanism is constructed based on a parameterized generative adversarial network. Starting with initial seed parameters, the generator's parameterized rule module and iterative growth module, combined with real-time generated control parameters, execute growth operations such as diffusion and branching, and record the results. The growth sequence supports backtracking adjustments, forming a dynamic evolution process from the initial state to the final form. This enables the pattern to have the continuity and controllability of natural growth, meeting users' needs for dynamic pattern creation. Finally, to address the issue of poor boundary adaptability, boundary conditions are detected in real time during iterative growth. When a preset boundary is reached, the direction or decay rate is automatically adjusted. Furthermore, the pattern is adapted and adjusted according to the garment pattern parameters during output, generating data that adapts to sleeves, body, and other parts. This ensures both coordination with the boundary during pattern growth and precise matching of the final pattern with garment parts, improving the adaptability of the pattern in actual garment applications. In summary, this method comprehensively solves the pain points of existing technologies through semantic conversion mechanisms, dynamic growth frameworks, and boundary adaptation strategies, enhancing the intelligence, interactivity, and practicality of garment pattern design. Attached Figure Description

[0049] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other embodiments can be derived from the provided drawings without creative effort.

[0050] Figure 1 This is a diagram illustrating the steps of the method of the present invention;

[0051] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0052] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0053] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;

[0054] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0055] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] Example 1

[0057] Please see Figure 1 A smart design method for clothing patterns based on parametric generative adversarial networks includes:

[0058] Obtain the initial seed parameters input by the user, which include the basic graphic type and initial growth rule parameters;

[0059] The initial state of pattern growth is constructed based on the initial seed parameters, and the initial state is input into the generator of the pre-trained parameterized generative adversarial network.

[0060] During the pattern growth process, text commands input by the user are received in real time. These text commands are used to indicate the pattern growth direction, texture features, or morphological features.

[0061] The text instructions are converted into feature vectors, and growth control parameters are generated by combining them with the current pattern growth state. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters.

[0062] The growth control parameters are input into the generator, and the generator iteratively grows and updates the current pattern to obtain a new pattern growth state.

[0063] Repeat the steps of receiving text instructions, generating growth control parameters, and iteratively updating growth until the user terminates growth or the preset number of growth steps is reached.

[0064] Obtain the final pattern growth state and output the corresponding clothing pattern data.

[0065] The construction of the initial state for pattern growth based on the initial seed parameters includes:

[0066] An initial graphic matrix is ​​generated in a two-dimensional grid based on the basic graphic type, which includes points, lines, or polygons.

[0067] The initial growth rule parameters are used to initialize the concentration field parameters of the reaction diffusion system or the string parameters of the Ringel system, forming the initial state matrix of pattern growth.

[0068] The step of converting the text instruction into a feature vector includes:

[0069] The text instructions are encoded using a pre-trained multimodal model to obtain text feature vectors;

[0070] The text feature vector is aligned with the visual feature vector of the current pattern growth state to obtain a fused feature vector, which is used as the feature vector corresponding to the text instruction.

[0071] The generation of growth control parameters based on the current pattern growth state includes:

[0072] Extract the morphological features of the current pattern growth state, including contour complexity, texture distribution density, and symmetry;

[0073] Based on the matching degree between the feature vector and the morphological feature, the rule adjustment weight, morphological intervention area and texture density threshold are determined.

[0074] The growth control parameters are generated by adjusting the weights according to the rules to produce reaction diffusion rate parameters or Lin's rewrite rule parameters, by generating local coordinate offset parameters according to the morphological intervention region, and by generating pixel fill density parameters according to the texture density threshold.

[0075] The training process of the parameterized generative adversarial network includes:

[0076] Construct a training dataset, which includes natural growth pattern sequences, corresponding text descriptions, and growth parameter records;

[0077] Construct a generator and a discriminator. The generator includes a parameterized rule module and an iterative growth module. The discriminator is used to distinguish between generated patterns and real patterns.

[0078] The generator and discriminator are trained alternately using the similarity between the generated pattern and the real pattern, the matching degree of the text instruction, and the growth continuity as loss functions, until the loss function converges.

[0079] The parameterization rule module of the generator is used to receive growth control parameters and output the corresponding reaction-diffusion equation parameters, Ringelmann system rewriting rules, or local morphology adjustment coefficients.

[0080] The iterative growth module is used to iteratively update the current pattern state based on the output of the parameterized rule module.

[0081] The iterative growth and update of the current pattern through the generator includes:

[0082] The iterative growth module of the generator performs growth operations on local regions of the current pattern based on growth control parameters. The growth operations include diffusion growth, branching growth, texture filling, or shape deformation.

[0083] During the growth operation, the pattern boundary conditions are detected in real time. When the growth area touches the preset boundary, the growth direction is automatically adjusted or the growth rate is reduced.

[0084] It also includes: during the iterative growth update process, recording the growth control parameters and corresponding pattern states at each step to form growth sequence data;

[0085] Upon receiving the user's backtracking instruction, the pattern growth state is restored to the specified step based on the growth sequence data, and growth is restarted from that state.

[0086] The output corresponding clothing pattern data includes:

[0087] Convert the final pattern growth state into a vector graphics format or a bitmap format;

[0088] Based on preset clothing pattern parameters, the pattern data is adapted and adjusted to generate pattern data adapted to a specified clothing part, including sleeves, body, or collar.

[0089] Before using the intelligent design method for clothing patterns based on parametric generative adversarial networks, it is necessary to ensure that the development environment is set up. First, install the necessary development tools, such as the Python programming environment. It is recommended to use Python 3.7 or above to ensure good support for various deep learning frameworks and libraries. At the same time, install the deep learning framework PyTorch or TensorFlow. Here, we take PyTorch as an example. You can follow the instructions in the official documentation and combine them with your computer's hardware configuration, such as whether you have an NVIDIA GPU and the corresponding CUDA version, to make full use of hardware acceleration and improve computing efficiency. In addition, some auxiliary libraries need to be installed, such as NumPy for data processing, OpenCV for image manipulation, and torchvision for building neural networks.

[0090] The specific steps in implementation are as follows:

[0091] Suppose a designer uses software specifically developed for clothing pattern design. In the software's interface, there is a parameter input area. The designer first selects from the drop-down menu of the basic graphic type, such as selecting polygon as the basic graphic type. For the initial growth rule parameters, the software provides corresponding input boxes. The designer can input parameters such as the number of sides of the polygon and the initial growth direction angle. These parameters will serve as the starting conditions for the entire pattern generation, laying the foundation for pattern growth.

[0092] Once the basic graphic type is determined to be a polygon, according to the mathematical definition of a polygon, an initial graphic matrix is ​​generated in a two-dimensional mesh by calculating the coordinates of each vertex of the polygon. For example, for a regular hexagon, the coordinates of the six vertices are determined in a two-dimensional coordinate system, and the grid positions corresponding to these coordinates are marked as regions of the polygon, thus forming the initial graphic matrix. For a reaction-diffusion system, its concentration field parameters need to be initialized. Assuming the Gray-Scott model is used, this model has two main concentration field variables, u and v, and corresponding reaction-diffusion coefficients. Based on the initial growth rule parameters, the initial concentration distributions of u and v are determined. For example, u... The initial concentration of is set to a higher value within the polygonal region, while the initial concentration of v is set to a lower value. Simultaneously, the value of the reaction diffusion coefficient is determined. These parameters collectively form the concentration field parameters of the reaction diffusion system. Combined with the initial graphic matrix, this forms the initial state matrix for pattern growth. In the case of a Linn system, string parameters are determined based on the initial growth rule parameters. For example, if the initial string for a Linn system is A, the rewriting rule is A→AB. Complex pattern structures are generated by continuously applying the rewriting rule. The initial string and rewriting rule parameters constitute the string parameters of the Linn system, which, together with the initial graphic matrix, form the initial state matrix for pattern growth.

[0093] A pre-trained CLIP (Contrastive Language - Image Pretraining) multimodal model is used to encode text instructions. The CLIP model includes an image encoder (such as ViT or CNN) and a text encoder (Transformer). When a text instruction is input, such as adding a branch to a pattern, the text encoder maps the text instruction to a text feature vector. At the same time, it extracts a visual feature vector from the current pattern growth state, for example, by using a convolutional neural network to extract features from the pattern. Then, a cross-modal attention mechanism is used to align the text feature vector and the visual feature vector. Specifically, the text feature vector is used as the query, and the visual feature vector is used as the key and value. By calculating attention weights, the text features and visual features are fused to obtain a fused feature vector. This fused feature vector is used as the feature vector corresponding to the text instruction to generate growth control parameters.

[0094] When extracting morphological features of the current pattern growth state, contour complexity is measured by calculating the ratio of the pattern contour's perimeter to its area; a larger ratio indicates a more complex contour. Texture distribution density is determined by counting the number of texture elements per unit area. Symmetry is calculated by analyzing the pattern's symmetry in the frequency domain using methods such as Fourier transform of the image. Relevant parameters are determined based on the matching degree between feature vectors and morphological features. For example, the matching degree is measured by calculating the cosine similarity between the feature vector and morphological feature vectors such as contour complexity, texture distribution density, and symmetry. A high matching degree indicates a strong correlation between the text command and the current pattern state. Weights are adjusted according to the matching degree rules. If the text command emphasizes increasing... If the current pattern has low contour complexity and the matching degree calculation shows high correlation, then increase the coefficient related to branch growth in the reaction diffusion rate parameter, or adjust the Lin rewrite rule to make the generated pattern more prone to branching. For morphological intervention areas, determine the areas that need to be adjusted based on the matching degree. For example, when the matching degree shows high correlation with the lower right corner of the pattern, take the lower right corner as the morphological intervention area, generate local coordinate offset parameters, and adjust the pattern in that area. Determine the texture density threshold based on the matching degree. If the matching degree indicates that the texture density needs to be increased, and the current texture distribution density is low, lower the texture density threshold, thereby increasing the pixel fill density parameter during the generation process to make the pattern texture denser.

[0095] The generator's iterative growth module performs growth operations on the current pattern based on growth control parameters. For example, when the rule adjustment parameter in the growth control parameters indicates diffusion growth, for a reaction-diffusion system, according to the adjusted reaction-diffusion rate parameter, elements in the pattern, such as color and shape, diffuse outwards within a local area of ​​the pattern according to the reaction-diffusion equation, thus achieving diffusion growth. When performing branch growth, new branch structures are generated on the basis of the current pattern at positions that meet certain conditions, based on the local coordinate offset parameter in the growth control parameters. When performing texture filling, texture is filled in a specified area according to a certain pattern and density based on the texture density parameter. When performing morphological deformation, stretching, twisting, and other deformation operations are performed on a local area of ​​the pattern according to the morphological intervention parameter. During the growth operation, the pattern boundary conditions are detected in real time. By setting a boundary detection function, when the growth area of ​​the pattern touches a preset boundary, such as when the pattern growth exceeds the set two-dimensional grid range, the growth direction is automatically adjusted, such as reversing the growth direction or rotating it by a certain angle, or attenuating the growth rate to gradually reduce the growth rate, so as to avoid the pattern growing excessively beyond the boundary.

[0096] Users can terminate the growth process by clicking the "Stop Growth" button on the design software interface, or the system will automatically terminate growth when a preset number of growth steps is reached. After growth terminates, the final pattern growth state is obtained and converted into a vector graphics format, such as SVG. When converting to a bitmap format, such as PNG, relevant graphics conversion algorithms are used to convert each element in the pattern, such as points, lines, and polygons, into vector graphics descriptions supported by SVG. When converting to a bitmap format, such as PNG, a corresponding PNG image file is generated based on the pattern's pixel information. The pattern data is then adapted and adjusted according to preset garment pattern parameters. For example, if the sleeve is a specific rectangular area with a certain degree of curvature, the generated pattern data is scaled, rotated, and distorted based on the sleeve's shape and size to perfectly fit the sleeve's shape, generating pattern data adapted to the sleeve. Similarly, for different garment parts such as the body and collar, the pattern is adapted and adjusted accordingly based on their respective pattern parameters, ultimately outputting pattern data adapted to the specified garment part for application in actual garment production.

[0097] Example 2

[0098] Please see Figure 2 A smart design system for clothing patterns based on parametric generative adversarial networks, including:

[0099] An initial parameter receiving module is used to obtain the initial seed parameters input by the user, the initial seed parameters including the basic graphic type and initial growth rule parameters;

[0100] An initial state construction module is used to construct the initial state of pattern growth based on the initial seed parameters;

[0101] A parameterized generative adversarial network includes a generator and a discriminator. The generator is used to receive the pattern growth state and growth control parameters, and output the iteratively updated pattern growth state.

[0102] The instruction processing module is used to receive text instructions input by the user in real time and convert the text instructions into feature vectors;

[0103] The control parameter generation module is used to generate growth control parameters by combining the current pattern growth state with the feature vector. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters.

[0104] An iterative growth module is used to input the growth control parameters into the generator to drive the pattern to perform iterative growth and update.

[0105] The results output module is used to output the corresponding clothing pattern data after growth has ended.

[0106] The intelligent clothing pattern design system based on parametric generative adversarial networks adopts a layered architecture, mainly consisting of a user interaction layer, a business logic layer, and a data storage layer. The user interaction layer is responsible for interacting with designers, receiving user input, and displaying the generated results. The business logic layer contains various functional modules, which are responsible for processing business logic and realizing the core functions of pattern design. The data storage layer is used to store training data, generated pattern data, and system-related configuration information. The various modules transmit and interact with each other through interfaces to ensure the efficient operation of the system.

[0107] The initial parameter receiving module provides an intuitive user interface, such as a graphical parameter input panel. The input panel includes drop-down menus for users to select the basic shape type, including points, lines, and polygons. For initial growth rule parameters, a text input box is provided, allowing users to enter specific values. For example, for polygonal basic shapes, users can enter parameters such as the number of sides and the initial rotation angle. After the user completes the input, the module performs a preliminary validation, checking the format of the parameters, such as whether the number of sides is a positive integer and whether the angle value is within a reasonable range. If the parameter format is incorrect, a prompt box will pop up to inform the user to make corrections. After successful validation, the parameters are passed to the initial state construction module.

[0108] After receiving the initial seed parameters, the initial state construction module, if the basic graphic type is a point, randomly selects one or more grid points in the 2D grid as initial points to form the initial graphic matrix; if it is a line, it draws line segments in the 2D grid based on the input starting point coordinates and line length, direction, and other parameters to determine the initial graphic matrix; if it is a polygon, it draws a polygon in the 2D grid based on the polygon's vertex coordinate information to obtain the initial graphic matrix. For a reaction-diffusion system, concentration field parameters are set according to the initial growth rule parameters, such as setting the initial concentration distribution of reactants and products, as well as reaction and diffusion coefficients. Assuming the use of... The Gray-Scott model determines the initial concentrations of two substances, u and v, as well as their reaction coefficients k and diffusion coefficients D1 and D2, based on the input parameters, forming concentration field parameters. These parameters are combined with the initial pattern matrix to construct the initial state for pattern growth. If it is a Lynch system, the initial string and rewriting rules are determined based on the initial growth rule parameters. For example, if the initial string is F, the rewriting rule is F→F+F−F. The initial state for pattern growth is constructed using these parameters. After the module completes the initial state construction, it passes the results to the generator and instruction processing module of the parameterized generative adversarial network to prepare for pattern generation and instruction processing.

[0109] Parametric Generative Adversarial Network (GAN): The generator employs a multi-layer neural network structure. The parametric rule module consists of multiple fully connected layers, receiving growth control parameters. When it receives the reaction-diffusion rate parameter from the rule adjustment parameters, the fully connected layers calculate and output the corresponding reaction-diffusion equation parameters, such as adjusting the reaction coefficient and diffusion coefficient. When it receives the Lynch rewriting rule parameters, it outputs the rewriting rules for the Lynch system. The iterative growth module contains convolutional and deconvolutional layers, iteratively updating the current pattern state based on the output of the parametric rule module. For example, when the parametric rule module outputs parameters for diffusion growth, the iterative growth module uses convolution operations to perform diffusion growth in a local region of the pattern according to the reaction-diffusion equation, thus increasing the density of elements in the pattern, such as color and shape. The pattern spreads outwards. The discriminator uses a convolutional neural network structure to distinguish between generated and real patterns. It extracts features of the pattern, such as texture and shape, through convolutional layers, and then performs classification judgment through fully connected layers, outputting a probability value representing the likelihood that the input pattern is a real pattern. During training, the similarity between the generated and real patterns, the matching degree of the text instruction, and the growth continuity are used as loss functions. The generator and discriminator are trained alternately. For example, the similarity between the generated and real patterns is measured by indicators such as mean squared error. The parameters of the generator and discriminator are continuously adjusted through the backpropagation algorithm, so that the pattern generated by the generator is closer and closer to the real pattern, while the discriminator can more accurately distinguish between real and generated patterns, until the loss function converges.

[0110] The instruction processing module listens in real time to the text instructions entered by the user in the text input box of the interactive interface. When the user enters an instruction, such as increasing the complexity of the pattern, a pre-trained multimodal model, such as the CLIP model, is used to encode the text instruction. The text encoder in the CLIP model converts the text instruction into a text feature vector. At the same time, it extracts a visual feature vector from the current pattern growth state. For example, it uses a convolutional neural network to extract features from the pattern to obtain a visual feature vector. Then, a cross-modal attention mechanism is used to align the text feature vector and the visual feature vector, and the text features and visual features are fused to obtain a fused feature vector, which is used as the feature vector corresponding to the text instruction and passed to the control parameter generation module.

[0111] The control parameter generation module extracts morphological features from the current pattern growth state, calculates the pattern contour using contour detection algorithms such as the Canny algorithm, and measures contour complexity by calculating the ratio of contour perimeter to area. It determines texture distribution density by counting the number of texture elements per unit area. It analyzes the pattern's symmetry in the frequency domain using methods such as Fourier transform of the image to calculate symmetry. Relevant parameters are determined based on the matching degree between feature vectors and morphological features. The matching degree is measured by calculating the cosine similarity between the feature vectors and morphological feature vectors such as contour complexity, texture distribution density, and symmetry. Weights are adjusted according to the matching degree rules; when the matching degree is high and the text instruction emphasizes increasing complexity, weights are adjusted accordingly. Increase the coefficients related to complex structure generation in the reaction diffusion rate parameter, or adjust the Lin rewrite rule to make the generated pattern more complex; determine the morphological intervention area based on the matching degree. For example, when the matching degree shows a high correlation with the upper left corner of the pattern, the upper left corner is taken as the morphological intervention area, and a local coordinate offset parameter is generated to adjust the pattern in that area; determine the texture density threshold based on the matching degree. If the matching degree indicates that the texture density needs to be increased, and the current texture distribution density is low, the texture density threshold is reduced, thereby increasing the pixel fill density parameter during the generation process to make the pattern texture denser. The generated growth control parameters are passed to the iterative growth module to drive the iterative growth update of the pattern.

[0112] The iterative growth module inputs growth control parameters to the generator's iterative growth module. When the growth control parameters indicate branch growth, the iterative growth module generates new branch structures at specified positions based on the local coordinate offset parameters in the parameters, according to the current pattern. For example, at the edge of the pattern, new branch lines are generated according to certain angle and length rules. During the growth operation, the pattern boundary conditions are detected in real time. By setting a boundary detection function, when the growth area of ​​the pattern touches the preset boundary, if the growth exceeds the set two-dimensional grid range, the growth direction is automatically adjusted by rotating the growth direction by a certain angle or attenuating the growth rate to gradually reduce the growth rate, so as to avoid the pattern growing excessively outside the boundary. After each iterative growth update, the new pattern growth status is fed back to the instruction processing module and the control parameter generation module for the next round of instruction processing and parameter generation.

[0113] The output module, upon termination of growth, converts the final pattern growth state into a vector graphics format, such as SVG, or a bitmap format, such as PNG. When converting to SVG, a graphics conversion algorithm is used to convert each element in the pattern, such as points, lines, and polygons, into SVG. The format supports vector graphics descriptions, which are saved as SVG files. When converted to PNG format, a corresponding PNG image file is generated based on the pixel information of the pattern. According to preset clothing pattern parameters, the pattern data is adapted and adjusted. For example, for the sleeve, the pattern is scaled, rotated, and distorted according to the shape of the sleeve, such as cylindrical or conical, and its size (length, circumference, etc.), so that the pattern can perfectly fit the shape of the sleeve, generating pattern data adapted to the sleeve. Similarly, for different clothing parts such as the body and collar, the pattern is adapted and adjusted accordingly according to their respective pattern parameters. The final generated clothing pattern data can be stored in the local file system, categorized into folders according to different design projects, or stored in a database for easy subsequent query and management. A display window is provided on the user interface to show the generated clothing pattern data to the user in the form of an image. The user can zoom in, zoom out, rotate, etc., to view the pattern effect more clearly.

[0114] The same or similar labels correspond to the same or similar parts;

[0115] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0116] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for intelligent design of clothing patterns based on parametric generative adversarial networks, characterized in that, include: Obtain the initial seed parameters input by the user, which include the basic graphic type and initial growth rule parameters; The initial state of pattern growth is constructed based on the initial seed parameters, and the initial state is input into the generator of the pre-trained parameterized generative adversarial network. During the pattern growth process, text commands input by the user are received in real time. These text commands are used to indicate the pattern growth direction, texture features, or morphological features. The text instructions are converted into feature vectors, and growth control parameters are generated by combining them with the current pattern growth state. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters. The growth control parameters are input into the generator, and the generator iteratively grows and updates the current pattern to obtain a new pattern growth state. Repeat the steps of receiving text instructions, generating growth control parameters, and iteratively updating growth until the user terminates growth or the preset number of growth steps is reached. Obtain the final pattern growth state and output the corresponding clothing pattern data.

2. The method according to claim 1, characterized in that, The construction of the initial state for pattern growth based on the initial seed parameters includes: An initial graphic matrix is ​​generated in a two-dimensional grid based on the basic graphic type, which includes points, lines, or polygons. The initial growth rule parameters are used to initialize the concentration field parameters of the reaction diffusion system or the string parameters of the Ringel system, forming the initial state matrix of pattern growth.

3. The method according to claim 1, characterized in that, The step of converting the text instruction into a feature vector includes: The text instructions are encoded using a pre-trained multimodal model to obtain text feature vectors; The text feature vector is aligned with the visual feature vector of the current pattern growth state to obtain a fused feature vector, which is used as the feature vector corresponding to the text instruction.

4. The method according to claim 1, characterized in that, The generation of growth control parameters based on the current pattern growth state includes: Extract the morphological features of the current pattern growth state, including contour complexity, texture distribution density, and symmetry; Based on the matching degree between the feature vector and the morphological feature, the rule adjustment weight, morphological intervention area and texture density threshold are determined. The growth control parameters are generated by adjusting the weights according to the rules to produce reaction diffusion rate parameters or Lin's rewrite rule parameters, by generating local coordinate offset parameters according to the morphological intervention region, and by generating pixel fill density parameters according to the texture density threshold.

5. The method according to claim 1, characterized in that, The training process of the parameterized generative adversarial network includes: Construct a training dataset, which includes natural growth pattern sequences, corresponding text descriptions, and growth parameter records; Construct a generator and a discriminator. The generator includes a parameterized rule module and an iterative growth module. The discriminator is used to distinguish between generated patterns and real patterns. The generator and discriminator are trained alternately using the similarity between the generated pattern and the real pattern, the matching degree of the text instruction, and the growth continuity as loss functions, until the loss function converges.

6. The method according to claim 5, characterized in that, The parameterization rule module of the generator is used to receive growth control parameters and output the corresponding reaction-diffusion equation parameters, Ringelmann system rewriting rules, or local morphology adjustment coefficients. The iterative growth module is used to iteratively update the current pattern state based on the output of the parameterized rule module.

7. The method according to claim 1, characterized in that, The iterative growth and update of the current pattern through the generator includes: The iterative growth module of the generator performs growth operations on local regions of the current pattern based on growth control parameters. The growth operations include diffusion growth, branching growth, texture filling, or shape deformation. During the growth operation, the pattern boundary conditions are detected in real time. When the growth area touches the preset boundary, the growth direction is automatically adjusted or the growth rate is reduced.

8. The method according to claim 1, characterized in that, Also includes: During the iterative growth and update process, the growth control parameters and corresponding pattern states at each step are recorded to form growth sequence data; Upon receiving the user's backtracking instruction, the pattern growth state is restored to the specified step based on the growth sequence data, and growth is restarted from that state.

9. The method according to claim 1, characterized in that, The output corresponding clothing pattern data includes: Convert the final pattern growth state into a vector graphics format or a bitmap format; Based on preset clothing pattern parameters, the pattern data is adapted and adjusted to generate pattern data adapted to a specified clothing part, including sleeves, body, or collar.

10. A smart design system for clothing patterns based on parametric generative adversarial networks, used to implement the method according to any one of claims 1-9, characterized in that, include: An initial parameter receiving module is used to obtain the initial seed parameters input by the user, the initial seed parameters including the basic graphic type and initial growth rule parameters; An initial state construction module is used to construct the initial state of pattern growth based on the initial seed parameters; A parameterized generative adversarial network includes a generator and a discriminator. The generator is used to receive the pattern growth state and growth control parameters, and output the iteratively updated pattern growth state. The instruction processing module is used to receive text instructions input by the user in real time and convert the text instructions into feature vectors; The control parameter generation module is used to generate growth control parameters by combining the current pattern growth state with the feature vector. The growth control parameters include rule adjustment parameters, morphological intervention parameters, and texture density parameters. An iterative growth module is used to input the growth control parameters into the generator to drive the pattern to perform iterative growth and update. The results output module is used to output the corresponding clothing pattern data after growth has ended.

Images