A garment style design creative generation system and method

Abstract

This invention relates to the field of clothing design technology and discloses a clothing style design creative generation system and method. The system includes a local attribute editing module and a texture rendering module. The two modules achieve data interaction and collaborative work through a standardized data interface protocol. The local attribute editing module includes a semantic segmentation network, a sketch inversion unit, and an attribute editing unit. The texture rendering module includes an encoding unit and an FcF generator. This system solves the limitations of local attribute editing through the local attribute editing module. In this module, the semantic segmentation network with a Swin Transformer architecture finely separates the target regions of the sleeves and torso. The sketch inversion unit maps the segmented sketches to the latent space to obtain latent vectors. The attribute editing unit uses Sparse PCA to reduce the dimensionality and decouple the latent vectors, independently applying perturbations along the principal component directions to achieve independent and controllable editing of the geometric attributes of specific regions, improving editing accuracy.

Description

Technical Field

[0001] This invention relates to the field of clothing design technology, and in particular to a system and method for generating creative clothing style designs. Background Technology

[0002] Clothing design is an organic fusion of creative expression and functional needs. Traditional design processes rely heavily on the designer's personal experience, resulting in problems such as long design cycles, limited creativity, and low efficiency.

[0003] With the development of artificial intelligence technology, intelligent design methods are gradually being applied to the field of clothing design, but existing technologies still have many shortcomings: 1. Limitations of local attribute editing: Existing intelligent design mostly adopts a global generation mode, which lacks fine control over local parts; during editing, it is easy to cause situations where adjusting one part affects other areas, such as changing the sleeve length but accidentally changing the neckline shape; at the same time, the related feature information is complex and redundant, resulting in insufficient editing precision, making it difficult to achieve independent and controllable adjustments only to specific areas such as sleeves and torso, and failing to meet the core needs of designers to accurately modify local details; 2. Poor texture rendering: Traditional texture mapping methods are difficult to adapt to the complex geometric structure of clothing, and are prone to texture distortion and breakage. In addition, they lack the simulation of real light and shadow effects, resulting in insufficient realism in the generated clothing images. For periodic textures with repeating patterns such as grids and stripes, existing methods are difficult to present their global distribution patterns. When applied on a large scale, they are prone to poor texture connection, blurred details, and lack of layering, failing to achieve the presentation effect expected by designers. Summary of the Invention

[0004] This invention aims to at least solve the technical problems existing in the prior art. To this end, this invention proposes a system and method for generating creative ideas for clothing style design, which solves the problems of limitations in local attribute editing and poor texture rendering effects in existing intelligent clothing design methods.

[0005] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is characterized in that it includes a local attribute editing module and a texture rendering module, the two modules realizing data interaction and collaborative work through a standardized data interface protocol; The local attribute editing module includes a semantic segmentation network, a sketch inversion unit, and an attribute editing unit. The semantic segmentation network consists of an encoding network and a thinning network. After receiving the clothing sketch, the encoding network extracts features, and the thinning network separates the target regions of the sleeves and torso. The sketch inversion unit maps the segmented sketch to the latent space to obtain latent vectors. The attribute editing unit performs dimensionality reduction and decoupling on the latent vectors and applies perturbation along the principal component direction to output the edited clothing sketch. The texture rendering module includes an encoding unit and an FcF generator. The FcF generator is embedded with a fast Fourier convolutional layer. After receiving the edited sketch and the texture patch input by the user, the encoding unit fuses the features of the sketch and the texture patch through a two-way feature interaction mechanism. The FcF generator generates a clothing image by modeling through spatial frequency domain dual-path features.

[0006] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system further includes a preprocessing module, a data management module, an interaction module, and an aesthetic evaluation module. The preprocessing module includes a semantic segmentation unit that uses a DeepLab network to generate a coarse positioning image to complete the initial positioning of the clothing foreground. The data management module stores clothing sketch datasets, texture patch libraries, model parameters, and intermediate design data, supporting data reading, writing, and updating. The aesthetic evaluation module uses a Siamese network to score and evaluate the attributes of the generated clothing images. The interaction module includes a parameter adjustment unit and a result display unit. The parameter adjustment unit adjusts the perturbation amplitude and resolution, and the result display unit displays the aesthetic score and the attribute evaluation results.

[0007] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the semantic segmentation network adopts a Swing Transformer architecture, the encoding network performs downsampling through Patch Merging and outputs multi-scale feature maps, and the encoding network adopts a moving window multi-head self-attention mechanism to divide the input feature map into non-overlapping windows, the windows in adjacent Swing Transformer blocks move along a preset pixel distance, and a mask is applied to the attention to avoid invalid cross-region interactions.

[0008] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the thinning network includes three thinning blocks, each of which fuses multi-resolution feature maps by combining skip connections and adjacent connections, and each thinning block is provided with weighted convolution, pixel recombining upsampling and element-wise addition layers, and outputs the segmentation results of sleeves, torso and background through feature fusion.

[0009] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the weighted convolution is divided into two branches, one branch convolves the feature map and then uses the ReLU activation function, and the other branch uses the sigmoid function to activate the trainable parameters; the pixel recombination upsampling method upsamples the low-resolution feature map into a high-resolution feature map.

[0010] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the sketch inversion unit constructs an autoencoder using a VGG encoder and a pre-trained StyleGAN2 generator to map the segmented sketch and the complete sketch to the latent space to obtain the corresponding latent vector.

[0011] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the attribute editing unit uses Sparse PCA to reduce the dimensionality and decouple the latent vectors, solves the principal component direction through an alternating optimization algorithm, and applies a perturbation within a preset range along the principal component direction to achieve directional editing of the geometric attributes of the sleeves and torso.

[0012] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the encoding unit adopts a bidirectional feature interaction mechanism, the parameter generator includes two levels of 1×1 convolutional layers and activation functions, and a batch normalization layer is connected in series after the activation functions. The feature interaction layer realizes the bidirectional fusion of sketch features and texture patch features through affine transformation.

[0013] According to some embodiments of the first aspect of the present invention, a clothing style design creative generation system is provided, wherein the fast Fourier convolutional layer of the texture rendering module is divided into global Fourier units and local Fourier units, which act on feature channels of different proportions respectively, and realizes the modeling of the global distribution law of periodic textures through spatial frequency domain dual-path feature modeling.

[0014] According to some embodiments of the second aspect of the present invention, a method for generating clothing style design ideas is performed by a clothing style design idea generation system according to some embodiments of the first aspect. The method includes the following steps: Step 1: Receive the clothing sketch or foreground image processed by the preprocessing module from the user, and start the local attribute editing process; Step 2: Extract features through the encoding network of the semantic segmentation network, refine the network to separate the target regions of the sleeve and torso, and output the segmentation results; Step 3: The autoencoder, constructed by the VGG encoder and the pre-trained StyleGAN2 generator, maps the segmented sketch and the complete sketch to the latent space to obtain the corresponding latent vectors. Step 4: Use Sparse PCA to reduce the dimensionality and decouple the latent vectors, solve the principal component direction, apply a preset range perturbation along the direction, edit the geometric properties of the sleeves and torso, and output the edited clothing sketch. Step 5: Receive the texture patch input by the user, start the texture rendering process, and fuse the features of the edited sketch and texture patch through the bidirectional feature interaction mechanism of the encoding unit. Step 6: Use an FcF generator with embedded Fast Fourier Convolutional layers to model the spatial frequency domain dual-path features and generate clothing images; Step 7: Based on the clothing image generated by the interactive module, the user adjusts the parameters and repeats steps 4 to 6 to optimize the design result.

[0015] A clothing style design idea generation system and method according to some embodiments of the present invention has at least the following beneficial effects: 1. This clothing style design creative generation system solves the limitations of local attribute editing through a local attribute editing module. The semantic segmentation network finely separates the target regions of sleeves and torso. The sketch inversion unit maps the segmented sketch to the latent space to obtain latent vectors. The attribute editing unit reduces the dimensionality of the latent vectors and decouples them, and applies perturbations independently along the principal component direction to achieve independent and controllable editing of the geometric attributes of specific regions. This avoids adjusting one part from affecting other regions, reduces feature redundancy, and improves editing accuracy.

[0016] 2. This clothing style design creative generation system relies on the texture rendering module to improve the texture rendering effect. Its encoding unit deeply integrates the features of the edited sketch and texture patch through a two-way feature interaction mechanism, and embeds the FcF generator of the fast Fourier convolution layer to perform spatial frequency domain dual-path feature modeling. It adapts to the complex geometric structure of clothing, effectively avoids texture distortion, breakage and poor connection, simulates real light and shadow effects, and makes periodic textures present a global distribution pattern, thereby improving the realism and detail of clothing images.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a diagram illustrating the overall architecture of the clothing style design creativity generation system of the present invention. Figure 2 This is a flowchart of the clothing style design idea generation method of the present invention; Figure 3 This is a schematic diagram of the local attribute editing module structure of the present invention; Figure 4 This is a schematic diagram of the bidirectional feature interaction mechanism of the encoding unit of the texture rendering module of the present invention; Figure 5 This is a schematic diagram of the spatial frequency domain dual-path structure of the fast Fourier convolutional layer of the present invention; Figure 6 This diagram illustrates the semantic segmentation functions of the preprocessing module and the local attribute editing module in this invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, left, right, front, and back, are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the module or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0021] In the description of this invention, the use of "first" and "second" is for the purpose of distinguishing technical features only, and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0022] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0023] First embodiment: like Figures 1-6 As shown in the figure, this embodiment of the invention provides a clothing style design creative generation system.

[0024] A clothing style design creative generation system includes a local attribute editing module and a texture rendering module. The two modules achieve data interaction and collaborative work through a standardized data interface protocol. The local attribute editing module includes a semantic segmentation network, a sketch inversion unit, and an attribute editing unit. The semantic segmentation network consists of an encoding network and a thinning network. After receiving the clothing sketch, the encoding network extracts features, and the thinning network separates the target regions of the sleeves and torso. The sketch inversion unit maps the segmented sketch to the latent space to obtain latent vectors. The attribute editing unit performs dimensionality reduction and decoupling on the latent vectors and applies perturbation along the principal component direction to output the edited clothing sketch. The texture rendering module includes an encoding unit and an FcF generator. The FcF generator is embedded with a fast Fourier convolutional layer. After receiving the edited sketch and the texture patch input by the user, the encoding unit fuses the features of the sketch and the texture patch through a two-way feature interaction mechanism. The FcF generator generates a clothing image by modeling through spatial frequency domain dual-path features.

[0025] This embodiment aims to solve the core problems of lack of fine control in local attribute editing and easy distortion of texture rendering in the prior art. By building a complete system architecture including core modules and supporting modules, clarifying the collaborative logic and key technical parameters of each module, it provides a precise and efficient basic implementation solution for the generation of clothing style ideas.

[0026] The clothing style design creative generation system includes a core module and supporting modules. The core modules are the local attribute editing module and the texture rendering module. The supporting modules include the preprocessing module, the data management module, the interaction module, and the aesthetic evaluation module. Each module realizes data interaction and collaborative work through the RESTful API standardized data interface protocol. The data management module is responsible for storing clothing sketch datasets, texture patch libraries, trained model parameters, and intermediate data in the design process, and supports data reading, writing, and updating.

[0027] The semantic segmentation unit of the preprocessing module uses the DeepLab network to generate a coarse mask image. First, the input original clothing sketch is converted to grayscale to remove background noise. Then, multi-scale features are extracted through dilated convolution of the DeepLab network to accurately capture the outline of the main body of the clothing and complete the initial foreground localization.

[0028] The semantic segmentation network of the local attribute editing module adopts the Swing Transformer architecture, which is functionally independent of and works collaboratively with the semantic segmentation unit of the preprocessing module.

[0029] Encoding network: Divided into 5 stages, stage 1 has a patch size of 1 and an output size of [missing information]. Stage 2 achieves downsampling through Patch Merging, with an output size of [size missing]. The output size of stage 3 is The output size of stage 4 is Phase 5 does not perform Patch Merging, and the output size remains unchanged. The encoding network employs a moving window multi-head self-attention mechanism to divide the input feature map into... Non-overlapping The attention output of each window (M=8) is calculated using a formula: ; in For querying the matrix, The key matrix, For value matrices, For feature dimension, It is a relative position offset matrix with dimension . ; Window movement within adjacent Swing Transformer blocks Pixels are used to mask attention and prevent invalid interactions across regions.

[0030] The network refinement consists of three refinement blocks. Each refinement block fuses multi-resolution feature maps using a combination of skip connections and adjacent connections. Each refinement block includes weighted convolutions, pixel re-sampling, and element-wise addition layers. The weighted convolutions are divided into two branches, one of which processes the feature maps... After convolution, a ReLU activation function is used; the other branch uses the sigmoid function on the trainable parameters. The activation operation is performed, and the pixel recombination upsampling method upsamples the low-resolution feature map into a high-resolution feature map, and finally outputs the fine segmentation results of the sleeve, torso and background. The training process of the semantic segmentation network uses the cross-entropy loss function, the formula of which is: ; in For the number of categories, For the first Output of each category For the true class probability distribution, To predict the probability distribution of categories, the training process is configured with a learning rate of 0.001, a batch size of 8, a momentum of 0.9, and a weight decay of 0.0002. After completing 100 training rounds, the trained model is output.

[0031] The local attribute editing module includes a semantic segmentation network, a sketch inversion unit, and an attribute editing unit, which enables independent and controllable editing of the geometric attributes of the sleeves and torso.

[0032] Sketch Inversion Unit: An autoencoder is constructed using a VGG encoder and a pre-trained StyleGAN2 generator to map the segmented sketch and the complete sketch to the latent space, obtaining the corresponding latent vectors. The VGG encoder consists of five VGG blocks, each containing... Convolutional layers Convolutional layers and The system uses max pooling layers, each convolutional layer employs the ReLU activation function, and finally, a fully connected layer adjusts the output dimension to match the 512-dimensional latent vector.

[0033] Dimensionality Reduction, Decoupling, and Editing: The attribute editing unit uses Sparse PCA to reduce the dimensionality and decouple the latent vectors. The optimization objective function is: ; The constraints are: ; in for 3D normalized latent vector matrix, Where p is the number of samples, and p is the number of variables. for Principal component direction matrix, for 3D sparse load matrix, The ridge regression penalty coefficient (with a value of 0.01) For the first The sparsity intensity coefficients of each principal component (with a value of 10.0) are used; the principal component directions are solved using an alternating optimization algorithm, and the sparsity intensity coefficients are applied along these directions. Perturbations within the range enable attribute editing. For example, perturbations are applied only to the latent vector corresponding to the sleeve along the principal component direction, while keeping the latent vector corresponding to the torso unchanged, thus achieving sleeve length adjustment without changing the neckline shape.

[0034] The texture rendering module receives the edited sketch and the texture patch input by the user, and generates clothing images with realistic lighting effects through a two-way feature interaction mechanism and an FcF generator with an embedded fast Fourier convolution layer.

[0035] Bidirectional Feature Interaction: The encoding unit adopts a bidirectional feature interaction mechanism, and the parameter generator contains two levels. Convolutional layers and ReLU activation function; first stage The number of output channels of the convolutional layer is the same as the number of input feature map channels, the second stage The number of output channels of the convolutional layer matches the number of channels of the input feature map. A batch normalization layer is then concatenated after the ReLU activation function. The feature interaction layer performs an affine transformation, with the following formulas: ; ; in The feature map of the sketch, For the feature map of the texture patch, This is the mean calculation function. This is the variance calculation function. These are the affine transformation parameters corresponding to the texture patch. These are the affine transformation parameters corresponding to the sketch.

[0036] FcF Generator and Loss Function: Global Channel Allocation Ratio of Fast Fourier Convolutional Layers Global Fourier units operate on 1 / 4 of the feature channels, while local Fourier units operate on 3 / 4 of the feature channels. Through dual-path feature modeling in both the spatial and frequency domains, the global distribution pattern of periodic textures is modeled. The intermediate VGG layer with frequency domain perceptual loss is selected... The total loss function of the texture rendering module is: ; in To combat the losses, For frequency domain sensing loss, For the loss of style, .

[0037] The interactive module includes a parameter adjustment unit and a result display unit. The parameter adjustment unit uses a slider to adjust the perturbation amplitude and resolution, with adjustment steps of 0.05 and 64×64, respectively. The result display unit uses the React framework to implement a 2×3 grid layout, and supports the simultaneous display of aesthetic scores and attribute evaluation results when comparing and viewing.

[0038] The aesthetic evaluation module uses a Siamese network, and the total loss function of the Siamese network is: ; in For Euclid's loss, For sorting loss, Activate loss for aesthetic attributes Take 0.2, Take 0.8.

[0039] This embodiment also provides a method for generating creative ideas for clothing style design, including the following steps: Step 1: Receive the clothing sketch or foreground image processed by the preprocessing module from the user, and start the local attribute editing process; Step 2: Extract features through the encoding network of the semantic segmentation network, refine the network to separate the target regions of the sleeve and torso, and output the segmentation results; Step 3: The autoencoder, constructed by the VGG encoder and the pre-trained StyleGAN2 generator, maps the segmented sketch and the complete sketch to the latent space to obtain the corresponding latent vectors. Step 4: Use Sparse PCA to reduce the dimensionality and decouple the latent vectors, solve for the principal component direction, apply a preset range perturbation along the direction, edit the geometric properties of the sleeves and torso, and output the edited clothing sketch. Step 5: Receive the texture patch input by the user, start the texture rendering process, and fuse the features of the edited sketch and texture patch through the bidirectional feature interaction mechanism of the encoding unit; Step 6: Use an FcF generator with embedded Fast Fourier Convolutional layers to model spatial frequency domain dual-path features and generate clothing images; Step 7: Based on the clothing image generated by the interactive module, the user adjusts the parameters and repeats steps 4 to 6 to optimize the design result.

[0040] Second embodiment: This embodiment addresses the problem of insufficient editing accuracy caused by cross-regional interference and high-dimensional feature redundancy in local editing within the basic system. By optimizing the semantic segmentation network structure, adjusting key parameters of Sparse PCA, and refining editing operations, the accuracy and stability of local attribute editing are further improved.

[0041] Based on the semantic segmentation network of the first embodiment, the thinning blocks are optimized to address the problem of insufficient edge segmentation accuracy. An adaptive weight adjustment mechanism is added to the weight convolution of each thinning block to dynamically adjust the convolution kernel weights according to the texture complexity of the input sketch. The moving window size M of the encoding network is set to 16, and the moving window of adjacent blocks moves... Pixels improve the segmentation accuracy of large-sized clothing sketches.

[0042] The training process uses a hybrid loss function, the formula of which is: ; in For the number of categories, For the first Output of each category For the true class probability distribution, To predict the probability distribution of the categories, the training parameters were adjusted to a learning rate of 0.0005, a batch size of 16, and 150 training epochs.

[0043] To address the issue of feature redundancy in high-dimensional latent spaces, the Sparse PCA parameters were adjusted; sleeve attribute editing was also implemented. Set the value to 0.3 and edit the torso attributes. Set the value to 0.5 to optimize the ridge regression penalty coefficient in the objective function. We set it to 0.005 to reduce overfitting of high-dimensional data.

[0044] During the latent vector mapping process, an attention mechanism layer is added after the VGG encoder to enhance the feature extraction of key clothing parts; the autoencoder is trained using the MSE loss function, the formula of which is: ; in For the original potential vector, For the reconstructed latent vector, The number of samples is 12; the training batch size is set to 12, the learning rate is 0.001, and the training run is 200 rounds.

[0045] Multiple principal component directions are set for sleeve attributes, corresponding to sleeve length, sleeve width, and sleeve type adjustments respectively; an adjustment is applied along the first principal component direction. Disturbance adjustment sleeve length, applied along the direction of the second principal component 0.4 disturbance adjustment sleeve width, applied along the direction of the third principal component. 0.6 disturbance adjusts sleeve shape.

[0046] For the torso attribute, the principal component directions correspond to garment length, waist circumference, and neckline style, respectively, and the perturbation range is set to... 0.4 to 0.7; Through optimization, the interference of editing operations on non-target areas is reduced by more than 30%, and the accuracy of target attribute editing is improved to more than 85%.

[0047] Third embodiment: This embodiment focuses on the problems of texture breakage, unrealistic lighting and shadows, and poor global consistency of periodic textures in existing texture rendering. By optimizing the bidirectional feature interaction mechanism, adjusting the parameters of the fast Fourier convolutional layer, and improving the generator structure, the realism and global coherence of texture rendering are improved.

[0048] First stage of parameter generator The number of output channels of the convolutional layer is set to 1.2 times the number of input feature map channels. (Second stage) The number of output channels of the convolutional layer is set to 0.8 times the number of input feature map channels; the ReLU activation function is replaced with LeakyReLU to avoid neuron death.

[0049] In the affine transformation of the feature interaction layer, an adaptive mean and variance adjustment mechanism is introduced to dynamically adjust the feature similarity between the sketch and the texture patch. and The calculation range is defined; the affine transformation formula remains: ; ; The meanings of the symbols are the same as in Example 1.

[0050] Adjust the global channel allocation ratio of the Fast Fourier Convolutional Layer to The global Fourier unit operates on 1 / 3 of the feature channels, while the local Fourier unit operates on 2 / 3 of the feature channels.

[0051] Increase the intermediate layer of VGG with frequency domain sensing loss In the total loss function Adjusted to 5. Adjust to 4, the formula is: ; in To combat the losses, For frequency domain sensing loss, This is a loss of style.

[0052] To reduce gradient vanishing, residual connection blocks are added to the FcF generator; the output layer of the generator uses the Tanh activation function to map the pixel values ​​of the output image to the range [0,1].

[0053] The training process used the AdamW optimizer with a learning rate of 0.0001, weight decay of 0.0001, batch size of 8, and 300 training rounds. After optimization, the texture breakage rate was reduced by 40%, and the lighting realism score was improved by 25%.

[0054] Fourth embodiment: This embodiment addresses the issues of poor user interaction and insufficient accuracy of aesthetic evaluation in the basic system by enhancing the functionality of the interaction module, optimizing the aesthetic evaluation network structure, and improving the module collaboration mechanism, thereby strengthening the system's practicality and design loop.

[0055] Preset parameter templates have been added, providing various parameter combinations such as casual daily life, formal business, and fashion trends; the grid layout of the result display unit has been optimized to 3×3, and an image magnification function has been added, with a maximum magnification of 200%.

[0056] The step size for parameter adjustment has been optimized to a perturbation amplitude of 0.02 and a resolution of 32×32; a real-time preview function has been added, displaying the adjustment effect within 0.5 seconds after the user adjusts the parameters.

[0057] Optimize the Siamese network structure by adding two convolutional layers; adjust the weight parameters of the total loss function. Take 0.3, Taking 0.7, the formula is: ; in Euclidean loss is used for regression optimization; To mitigate ranking loss and improve the accuracy of ranking evaluation; Activate the loss function for aesthetic attributes to enhance the capture of aesthetic features; For sorting loss weights, Activate loss weights for aesthetic attributes.

[0058] A user feedback mechanism has been added, allowing users to rate the generated results from 1 to 5 points and add labels. The system uses this data as additional training data to update the aesthetic evaluation model. The training adopts an incremental training method, updating the model once every 1,000 pieces of user feedback data are accumulated.

[0059] The interaction module and the aesthetic evaluation module establish real-time data interaction, and the aesthetic evaluation module dynamically adjusts the evaluation parameters based on user operation behavior; the data management module adds the function of classifying and storing design results, classifying them according to dimensions such as aesthetic score and attribute type, so as to facilitate user retrieval and access.

[0060] This clothing design creative generation system addresses the limitations of local attribute editing through a local attribute editing module. In this module, the semantic segmentation network of the Swin Transformer architecture finely separates the target regions of the sleeves and torso. The sketch inversion unit maps the segmented sketch to the latent space to obtain latent vectors. The attribute editing unit uses Sparse PCA to reduce the dimensionality of the latent vectors and decouples them, applying perturbations independently along the principal component direction. This enables independent and controllable editing of the geometric attributes of specific regions, avoiding the impact of adjusting one part on other regions, reducing feature redundancy, and improving editing accuracy.

[0061] Furthermore, this clothing style design creative generation system improves the texture rendering effect by relying on the texture rendering module. Its encoding unit deeply integrates the features of the edited sketch and texture patch through a two-way feature interaction mechanism, and embeds an FcF generator with a fast Fourier convolution layer to perform spatial frequency domain dual-path feature modeling. This adapts to the complex geometric structure of clothing, effectively avoids texture distortion, breakage and poor connection, simulates real light and shadow effects, and allows periodic textures to present a global distribution pattern, improving the realism and detail of clothing images. This solves the problems of limited local attribute editing and poor texture rendering effect in existing intelligent clothing design methods.

[0062] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A garment style design idea generation system, characterized by: It includes a local attribute editing module and a texture rendering module. The two modules achieve data interaction and collaborative work through a standardized data interface protocol. The local attribute editing module includes a semantic segmentation network, a sketch inversion unit, and an attribute editing unit. The semantic segmentation network consists of an encoding network and a thinning network. After receiving the clothing sketch, the encoding network extracts features, and the thinning network separates the target regions of the sleeves and torso. The sketch inversion unit maps the segmented sketch to the latent space to obtain latent vectors. The attribute editing unit performs dimensionality reduction and decoupling on the latent vectors and applies perturbation along the principal component direction to output the edited clothing sketch. The texture rendering module includes an encoding unit and an FcF generator. The FcF generator is embedded with a fast Fourier convolutional layer. After receiving the edited sketch and the texture patch input by the user, the encoding unit fuses the features of the sketch and the texture patch through a two-way feature interaction mechanism. The FcF generator generates a clothing image by modeling through spatial frequency domain dual-path features.

2. The garment style design creative generation system according to claim 1, wherein: It also includes a preprocessing module, a data management module, an interaction module, and an aesthetic evaluation module. The preprocessing module includes a semantic segmentation unit that uses a DeepLab network to generate a coarse positioning image to complete the initial positioning of the clothing foreground. The data management module stores the clothing sketch dataset, texture patch library, model parameters, and intermediate design data, supporting data reading, writing, and updating. The aesthetic evaluation module uses a Siamese network to score and evaluate the attributes of the generated clothing images. The interaction module includes a parameter adjustment unit and a result display unit. The parameter adjustment unit adjusts the perturbation amplitude and resolution, and the result display unit displays the aesthetic score and the attribute evaluation results.

3. The garment style design creative generation system according to claim 1, wherein: The semantic segmentation network adopts the Swing Transformer architecture. The encoding network implements downsampling through Patch Merging and outputs multi-scale feature maps. The encoding network adopts a moving window multi-head self-attention mechanism to divide the input feature map into non-overlapping windows. Windows in adjacent Swing Transformer blocks move along a preset pixel distance, and a mask is applied to the attention to avoid invalid cross-region interactions.

4. The garment style design creative generation system according to claim 1, wherein: The thinning network comprises three thinning blocks. Each thinning block fuses multi-resolution feature maps by combining skip connections and adjacent connections. Each thinning block is equipped with weighted convolution, pixel re-sampling, and element-wise addition layers. The segmentation results of sleeves, torso, and background are output through feature fusion.

5. The garment style design creative generation system according to claim 4, wherein: The weighted convolution is divided into two branches: one branch convolves the feature map and then uses the ReLU activation function, and the other branch uses the sigmoid function to activate the trainable parameters; the pixel recombination upsampling method upsamples the low-resolution feature map into a high-resolution feature map.

6. The garment style design creative generation system according to claim 1, wherein: The sketch inversion unit uses a VGG encoder and a pre-trained StyleGAN2 generator to construct an autoencoder, which maps the segmented sketch and the complete sketch to the latent space to obtain the corresponding latent vector.

7. The garment style design creative generation system according to claim 6, wherein: The attribute editing unit uses Sparse PCA to reduce the dimensionality of the latent vectors and decouples them. It solves the principal component directions through an alternating optimization algorithm and applies a perturbation within a preset range along the principal component directions to achieve directional editing of the geometric attributes of the sleeves and torso.

8. The garment style design creative generation system according to claim 1, wherein: The encoding unit adopts a bidirectional feature interaction mechanism. The parameter generator contains two levels of 1×1 convolutional layers and activation functions. After the activation functions, a batch normalization layer is concatenated. The feature interaction layer achieves bidirectional fusion of sketch features and texture patch features through affine transformation.

9. The garment style design creative generation system according to claim 1, wherein: The fast Fourier convolutional layer of the texture rendering module is divided into global Fourier units and local Fourier units, which act on different proportions of feature channels. Through spatial frequency domain dual-path feature modeling, the global distribution law of periodic textures is modeled.

10. A method for generating creative ideas for clothing style design, characterized in that: This method is performed by the clothing style design idea generation system according to any one of claims 1-9, and the method includes the following steps: Step 1: Receive the clothing sketch or foreground image processed by the preprocessing module from the user, and start the local attribute editing process; Step 2: Extract features through the encoding network of the semantic segmentation network, refine the network to separate the target regions of the sleeve and torso, and output the segmentation results; Step 3: The autoencoder, constructed by the VGG encoder and the pre-trained StyleGAN2 generator, maps the segmented sketch and the complete sketch to the latent space to obtain the corresponding latent vectors. Step 4: Use Sparse PCA to reduce the dimensionality and decouple the latent vectors, solve the principal component direction, apply a preset range perturbation along the direction, edit the geometric properties of the sleeves and torso, and output the edited clothing sketch. Step 5: Receive the texture patch input by the user, start the texture rendering process, and fuse the features of the edited sketch and texture patch through the bidirectional feature interaction mechanism of the encoding unit. Step 6: Use an FcF generator with embedded Fast Fourier Convolutional layers to model the spatial frequency domain dual-path features and generate clothing images; Step 7: Based on the clothing image generated by the interactive module, the user adjusts the parameters and repeats steps 4 to 6 to optimize the design result.

Images