Search

CN-121305092-B - Multi-mode collaborative design method for clothing driven by mass ornaments

CN121305092BCN 121305092 BCN121305092 BCN 121305092BCN-121305092-B

Abstract

The invention discloses a multi-mode collaborative design method for a public ornament driven clothing, which comprises the steps of receiving an ornament plane image uploaded by a user, extracting an ornament main body outline through an image segmentation network, executing three-level attribute classification based on outline features in parallel to generate a structural feature matrix M, inputting the feature matrix M into a copyright filtering module, outputting a clothing design constraint set, encoding the constraint set C into an embedded vector with the same dimension as the feature matrix M, splicing the embedded vector after LayerNorm is standardized, inputting the embedded vector into an antagonism network, outputting a matching degree score S, performing process decision based on the matching degree score S, including combining the complexity of the shape and the unit price of the fabric to generate a cost optimization model, performing an associated anti-wrinkle fabric on the metal ornament, enabling a laser cutting process for the special-shaped outline, and outputting a cutting scheme and a bill of materials which can be directly put into production.

Inventors

  • LI JIANMIN
  • ZHONG YING
  • ZHU SHUNZHI
  • HUANG ZHICAI

Assignees

  • 厦门理工学院

Dates

Publication Date
20260512
Application Date
20251210

Claims (9)

  1. 1. A method for multi-modal collaborative design of a mass-driven apparel, comprising: S1, receiving an ornament plane image uploaded by a user, extracting an ornament main body outline through an image segmentation network, and generating a structural feature matrix M; s2, inputting the feature matrix M into a copyright filtering module and outputting a clothing design constraint set, wherein the copyright filtering module comprises the following specific implementation steps: s2.1 calculating element originality index , wherein, Representing the unique scores of the same class elements in the copyright library, Representing visual similarity to the public domain element; s2.2, sampling from a pre-trained cultural theme vector database, generating symbols through style migration network fusion, and generating element originality indexes When the element replacement engine is triggered and executed, a basic outline is reserved for a separable structure in the feature matrix M and the symbol is injected, and a visual twin is generated for an inseparable structure through CycleGAN, wherein the separable structure is detected through outline curvature change rate variance, and the separable structure is judged to be separable when the variance is smaller than a threshold value eta; S2.3 when element originality index When the method is used, a knowledge graph is called, and a clothing design constraint set C is output by matching with a history element collocation rule according to the association relation of elements in the knowledge graph; S3, coding the constraint set C into an embedded vector with the same dimension as the feature matrix M, splicing after LayerNorm is standardized, and inputting to generate an countermeasure network, wherein the method specifically comprises the following steps: s3.1, mapping the ornament characteristics to a clothing potential space through a training projection matrix, taking the constraint set C as a query vector, taking the projected characteristics as a key vector, and calculating an attention weight matrix; S3.2, generating a candidate clothing scheme by using a diffusion model, outputting a matching degree score S, and rendering a virtual try-on sequence, wherein the denoising process of the diffusion model takes a constraint set C as a condition, and shielding similar feature vectors in a copyright library in a latent space sampling stage; S4, carrying out process decision based on matching degree score S, wherein the process decision comprises the steps of generating a cost optimization model by combining the complexity of the layout and the unit price of the fabric, carrying out correlation crease-resistant fabric on metal ornaments, enabling a laser cutting process by special-shaped contours, and outputting a cutting scheme and a bill of materials which can be put into production directly, wherein the process decision comprises the following steps: in the fabric matching rule, the anti-wrinkle fabric is blended when the hardness of the ornament material is more than 5H, and the design of the buffer lining is started by the brittle material; The cost optimization model specifically comprises the following steps: , wherein, Represents the consumption of the i-th fabric, The unit price is represented by the formula, Representing a model complexity weight factor, complexity representing model processing complexity, and n representing the total sample size.
  2. 2. The method according to claim 1, wherein the step S1 further comprises performing three-level attribute classification based on the contour features in parallel to generate a structured feature matrix M, the three-level attribute classification specifically comprising: Material classification, namely identifying metal, fabric or plastic materials by adopting a convolutional neural network, wherein the fabric materials are subdivided into knitting, tatting and lace subclasses; Structure classification, namely detecting geometric features by applying Hough transformation and judging regular geometric bodies or special-shaped structures by combining a classifier; And theme classification, namely fusing visual features of the image and text labels of the user, and outputting sub-cultural theme codes through an attention mechanism.
  3. 3. The method of multi-modal collaborative design for a garment according to claim 2, wherein the three-level attribute classification further comprises: The texture classification sub-model adopts MobileNetV network architecture, and the training dataset comprises 200 texture surface macro textures; The structure classification extracts straight lines and circular arc feature vectors through Hough transformation, and inputs the straight lines and circular arc feature vectors into a support vector machine classifier; the topic classification adopts bimodal fusion, specifically, the image ResNet features and the text BERT embedded vector are spliced, and topic codes are extracted through a multi-head self-attention layer.
  4. 4. The method for collaborative design of multiple modes for a garment according to claim 1, wherein in the element originality index calculation, the homogeneous element uniqueness score is calculated Wherein N represents the number of samples with the characteristic distance from the target element smaller than a threshold tau in the copyright library; Visual similarity Through cosine similarity calculation: ; Wherein, the Representing the characteristic vector of the ornament element ResNet-50, Represents a reference feature vector of elements in the public domain, The text label of the indicating ornament is embedded, The text description of the public domain is represented embedded, The meaning of the representation is the visual feature similarity weight, The meaning of the representation is text feature similarity weight.
  5. 5. The method for collaborative design of a multi-modal garment according to claim 1, wherein the element replacement engine is embodied as: The separability detection comprises the steps of calculating the variance of the curvature change rate of the profile, and judging that the profile is a separable structure when the variance is smaller than a threshold value eta, wherein the threshold value eta is determined by a K-means clustering profile curvature data set, and the value range is [0.2,0.5]; performing style migration by adopting CycleGAN networks to ensure that Euclidean distance between output and original elements is larger than a copyright safety threshold f, adding a structure constraint loss function during CycleGAN network training to ensure that the twin retains original space topological characteristics, wherein the structure constraint loss function is as follows: , wherein, Representing an initial mask of the image, Representing a binary mask.
  6. 6. The method of claim 1, wherein the mapping the decoration feature to the clothing potential space is specifically to establish a projection relation M' = M, wherein, Representing a projection matrix trained by contrast learning Full connection mapping is performed on all sub-features of the feature matrix M.
  7. 7. A crowd-sourced, multi-modal collaborative design system for apparel, characterized in that the method for multi-modal collaborative design for apparel according to any of claims 1-6, in particular comprises: the multi-mode input module is configured to receive the ornament plane image uploaded by the user, extract the ornament main body outline through the image segmentation network, and execute three-level attribute classification in parallel based on outline features to generate a structural feature matrix M; The copyright filter engine module is configured to input the feature matrix M into the copyright filter module and output a clothing design constraint set, and the copyright filter module comprises the specific implementation steps of calculating element originality indexes , wherein, Representing the unique scores of the same class elements in the copyright library, Representing visual similarity to elements in the public domain, when element originality index Triggering an element replacement engine when >0.8, reserving a basic outline for a separable structure, injecting a symbol, generating a visual twin body for an inseparable structure through CycleGAN, sampling the symbol from a pre-trained cultural theme vector database, and fusing through a style migration network, wherein the separable structure is detected through a profile curvature change rate variance, the variance is judged to be separable when < threshold eta, the inseparable structure generates the visual twin body through CycleGAN, the Euclidean distance between the visual twin body and an original element is larger than the threshold f, and when the element originality index is obtained When the method is used, a knowledge graph matching history collocation rule is called to output a garment design constraint set C; The cross-modal generator module is configured to encode the constraint set C into an embedded vector with the same dimension as the feature matrix M, splice the embedded vector after LayerNorm is standardized, and input the embedded vector to generate an countermeasure network, and specifically comprises the steps of mapping ornament features to a garment potential space through a training projection matrix, taking the constraint set C as a query vector, taking the projected features as key vectors, calculating an attention weight matrix, generating a candidate garment scheme by utilizing a diffusion model, outputting a matching degree score S, and rendering a virtual try-on sequence, wherein the denoising process of the diffusion model takes the constraint set C as a condition, and shielding similar feature vectors in a copyright library in a latent space sampling stage; The production conversion interface module is configured to perform process decision based on the matching degree score S, and comprises the steps of combining the complexity of the layout and the unit price of the fabric to generate a cost optimization model, performing the correlation crease-resistant fabric on the metal ornaments, enabling a laser cutting process on the special-shaped contours, and outputting a cutting scheme and a bill of materials which can be put into production directly, wherein the process decision comprises the following steps: in the fabric matching rule, the anti-wrinkle fabric is blended when the hardness of the ornament material is more than 5H, and the design of the buffer lining is started by the brittle material; The cost optimization model specifically comprises the following steps: , wherein, Represents the consumption of the i-th fabric, The unit price is represented by the formula, Representing a model complexity weight factor, complexity representing model processing complexity, and n representing the total sample size.
  8. 8. A computer program product, characterized in that it has stored thereon a computer program which, when executed by a processor, implements the method according to any of claims 1-6.
  9. 9. A computing system comprising a processor and a memory, the processor configured to perform the method of any of claims 1-6.

Description

Multi-mode collaborative design method for clothing driven by mass ornaments Technical Field The invention relates to the technical field of computer-aided garment design, in particular to a multi-mode collaborative design method for a garment driven by an ornament of a public. Background In the current clothing collocation field, the traditional process generally takes clothing as a dominant one, namely, a clothing main body is selected or designed firstly, and then ornaments, such as jewelry, accessories and the like, are adapted based on clothing styles and functions so as to meet the aesthetic and practical requirements of the public. However, with the advent of mass culture communities (e.g., cosplay role plays, blind box collection derivative products, etc.), accessories have exceeded apparel as a core element, which emphasize the uniqueness and symbolism of certain accessories (e.g., character-marking props or limited version dolls), thereby deriving a "jewelry-guided" demand pattern. In this mode, the existing standardized clothing cannot be matched with the highly customized or themed ornaments, so that problems of disharmony of collocation, disharmony of vision and the like are caused, and urgent requirements for custom-designing and producing clothing based on ornament characteristics (such as shapes, materials and themes) are further stimulated, so that the uniformity of overall modeling and market response efficiency are improved. Therefore, the invention designs the multi-mode collaborative design method for the clothing driven by the crowd ornaments, and the clothing matched with the ornaments can be designed under the condition of not violating the protection of related versions so as to meet the requirement scene of the crowd on special clothing. Disclosure of Invention In order to overcome the defects, according to the first aspect of the invention, a multi-mode collaborative design method for a garment driven by an ornament of a public is provided, the problem of adaptation and imbalance between the ornament and the garment in cultural scenes of the public such as role playing, tide playing and collection is solved, and the specific steps are as follows: S1, receiving an ornament plane image uploaded by a user, extracting an ornament main body outline through an image segmentation network, and generating a structural feature matrix M; S2, inputting the feature matrix M into a copyright filtering module and outputting a clothing design constraint set C, wherein the copyright filtering module comprises the following specific implementation steps: s2.1 calculating element originality index , wherein,Representing the unique scores of the same class elements in the copyright library,Representing visual similarity to the public domain element; S2.2, sampling from a pre-trained cultural theme vector database, generating symbols through style migration network fusion, and generating element originality indexes When the element replacement engine is triggered and executed, a basic outline is reserved for a separable structure in the feature matrix M and the symbol is injected, and a visual twin is generated for an inseparable structure through CycleGAN, wherein the separable structure is detected through outline curvature change rate variance, and the separable structure is judged to be separable when the variance is smaller than a threshold value eta; S2.3 when element originality index When the method is used, a knowledge graph is called, and matching history elements are matched according to the association relation of the elements in the knowledge graph to obtain matching rules; S3, coding the constraint set C into an embedded vector with the same dimension as the feature matrix M, splicing after LayerNorm standardization, and inputting to generate an countermeasure network, wherein the method specifically comprises an adaptation strategy consisting of a feature projection layer, an attention distribution layer and a conditional diffusion model: s3.1, mapping the ornament characteristics to a clothing potential space through a training projection matrix, taking the constraint set C as a query vector, taking the projected characteristics as a key vector, and calculating an attention weight matrix; S3.2, generating a candidate clothing scheme by using a diffusion model, outputting a matching degree score S, and rendering a virtual try-on sequence, wherein the denoising process of the diffusion model takes a constraint set C as a condition, and shielding similar feature vectors in a copyright library in a latent space sampling stage; S4, carrying out process decision based on the matching degree score S, further, combining the complexity of the layout and the unit price of the fabric to generate a cost optimization model, carrying out the correlation crease-resistant fabric on the metal ornaments, enabling the laser cutting process on the special-shaped profile, and outputting a cutting scheme an