A niche accessory-driven clothing multi-modal collaborative design method

Abstract

The application discloses a kind of minority accessories driven clothing multimodal collaborative design method, specifically includes: receiving accessory plan image uploaded by user, accessory main contour is extracted by image segmentation network, and three-level attribute classification is executed in parallel based on the contour feature, and structured feature matrix M is generated;The feature matrix M is input into copyright filtering module, and the output clothing design constraint set;The constraint set C is coded into the embedding vector of the same dimension as the feature matrix M, and is spliced after LayerNorm standardization, input into generative adversarial network, and output matching degree score S;Process decision is made based on matching degree score S, including: combining pattern complexity and fabric unit price to generate cost optimization model;Metal accessories are associated with wrinkle-resistant fabrics, and laser cutting process is enabled for irregular contours;Output cutting scheme and bill of materials that can be directly put into production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided fashion design, and particularly relates to a minority accessory-driven fashion multi-modal collaborative design method. BACKGROUND

[0002] In the current field of clothing matching, the traditional process is usually dominated by clothing, that is, the main body of the clothing is selected or designed first, and then accessories such as jewelry and accessories are matched based on the style and function of the clothing to meet the aesthetic and practical needs of the public. However, with the rise of minority cultural groups (such as Cosplay role-playing, blind box collection derivative products, etc.), accessories have become the core element beyond clothing, and these groups emphasize the uniqueness and symbolic meaning of specific accessories (such as role signature props or limited edition dolls), thereby deriving a "accessory-first" demand mode. Under this mode, existing standardized clothing often cannot match these highly customized or themed accessories, leading to problems such as mismatching and visual discord, which in turn stimulates the urgent need for customized design and production of clothing based on accessory characteristics (such as shape, material, theme) to improve the unity of the overall look and market response efficiency. Therefore, the present application designs a minority accessory-driven fashion multi-modal collaborative design method, which can design clothing that matches accessories without violating relevant version protection, to meet the needs of special clothing scenarios for minority groups. SUMMARY

[0003] To address the above problems, according to a first aspect of the present application, a minority accessory-driven fashion multi-modal collaborative design method is provided to solve the problem of mismatching between accessories and clothing in role-playing, toy collecting and other minority cultural scenarios. The specific steps are as follows:

[0004] S1, receiving a user-uploaded accessory planar image, extracting an accessory main contour through an image segmentation network, and generating a structured feature matrix M;

[0005] S2, inputting the feature matrix M into a copyright filtering module to output a clothing design constraint set C, and the copyright filtering module specifically implements the following steps:

[0006] S2.1, calculating an element originality index wherein, represents the uniqueness score of the same element in the copyright library, represents the visual similarity with public domain elements;

[0007] S2.2, sampling from a pre-trained cultural theme vector database and generating symbols through a style transfer network, when the element originality index When the value is greater than 0.8, the element replacement engine is triggered. When the element replacement engine is executed, the basic contours of separable structures in the feature matrix M are preserved and the symbols are injected. For non-separable structures, visual twins are generated using CycleGAN. The separable structures are determined to be separable when the variance of the contour curvature change rate is less than the threshold η. The non-separable structures are used to generate visual twins using CycleGAN, and the Euclidean distance between the original elements and the twins is greater than the threshold f.

[0008] S2.3, When the element's originality index When needed, the knowledge graph is invoked, and matching rules are obtained by matching historical elements based on the relationships between elements in the knowledge graph.

[0009] S3. Encode the constraint set C into an embedding vector with the same dimension as the feature matrix M, normalize it using LayerNorm, concatenate them, and input them into a generative adversarial network. Specifically, this includes an adaptation strategy composed of a feature projection layer, an attention allocation layer, and a conditional diffusion model.

[0010] S3.1. By training the projection matrix, the features of the accessories are mapped to the latent space of the clothing. The constraint set C is used as the query vector, the projected features are used as the key vector, and the attention weight matrix is ​​calculated.

[0011] S3.2. Generate candidate clothing schemes using a diffusion model and output a matching score S, and render a virtual try-on sequence; the denoising process of the diffusion model is based on a constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage;

[0012] S4. Based on the matching score S, make process decisions. Further, combine pattern complexity and fabric unit price to generate cost optimization model; associate metal ornaments with anti-wrinkle fabric, and use laser cutting technology for irregular contours; output cutting schemes and bill of materials that can be directly put into production.

[0013] Furthermore, step S1 also includes generating a structured feature matrix M by performing three-level attribute classification in parallel based on contour features. The three-level attribute classification specifically includes:

[0014] Material classification: Convolutional neural networks are used to identify metal, fabric, or plastic materials, with fabric materials further subdivided into knitted, woven, and lace subcategories;

[0015] Structural classification: Hough transform is applied to detect geometric features, and a classifier is used to determine whether the geometric object is regular or irregular.

[0016] Topic classification: It integrates image visual features with user text tags and outputs subculture topic codes through an attention mechanism.

[0017] Furthermore, the three-level attribute classification further includes:

[0018] The material classification sub-model adopts the MobileNetV3 network architecture, and the training dataset contains 200 kinds of material surface micro textures;

[0019] Structural classification extracts feature vectors of straight lines and circular arcs through Hough transform, and inputs them into a support vector machine classifier;

[0020] Topic classification employs a dual-modal fusion approach, specifically by concatenating ResNet features of images with BERT embedding vectors of text, followed by topic encoding extraction through a multi-head self-attention layer.

[0021] Furthermore, in the calculation of the element originality index: the uniqueness score of elements within the same category... , where N represents the number of samples in the copyright library whose feature distance to the target element is less than the threshold τ;

[0022] Visual similarity Calculated using cosine similarity:

[0023]

[0024] in, This represents the ResNet-50 feature vector of the jewelry element. Represents the baseline feature vector of public domain elements. This indicates that the text tag for the accessory is embedded. This indicates a public domain text description embedding. The meaning represented is the visual feature similarity weight. The meaning represented is the text feature similarity weight.

[0025] Furthermore, the specific implementation process of the element replacement engine is as follows:

[0026] Separability detection: Calculate the variance of the rate of change of contour curvature. If the variance is less than the threshold η, it is determined to be a separable structure. The threshold η is determined by K-means clustering of the contour curvature dataset, and the value range is [0.2, 0.5].

[0027] Visual twin generation of inseparable structures: A CycleGAN network is used to perform style transfer, ensuring that the Euclidean distance between the output and the original elements is greater than the copyright security threshold f; a structural constraint loss function is added during CycleGAN network training to ensure that the twin retains the original spatial topological features; the structural constraint loss function is: ,in, Indicates the initial mask. This represents a twin mask.

[0028] Furthermore, the mapping of accessory features to the clothing latent space specifically involves: establishing a projection relationship M'= M, where This represents the projection matrix trained through contrastive learning. Perform a fully connected mapping on all sub-features of the feature matrix M.

[0029] Furthermore, the process decisions include:

[0030] In the fabric matching rules, when the hardness of the jewelry material in the matching score S is >5H, it is associated with a blended wrinkle-resistant fabric, and a cushioning lining design is enabled for brittle materials; when the hardness of the jewelry material in the matching score S is higher than the standard hardness threshold, it is associated with a wrinkle-resistant fabric.

[0031] The cost optimization model is specifically as follows: ,in, This represents the consumption of the i-th type of fabric. Indicates the unit price. This represents the pattern complexity weighting factor, where Complexity represents the pattern processing complexity, and n represents the total number of samples.

[0032] According to a second aspect of the present invention, a niche accessories-driven multimodal collaborative design system for clothing is proposed, comprising the following modules:

[0033] The multimodal input module is configured to receive flat images of jewelry uploaded by users, extract the main outline of the jewelry through an image segmentation network, perform three-level attribute classification in parallel based on the outline features, and generate a structured feature matrix M.

[0034] The copyright filtering engine module is configured to input the feature matrix M into the copyright filtering module and output a set of clothing design constraints. The specific implementation steps of the copyright filtering module are as follows: calculating the originality index of elements. ,in This represents the uniqueness score of similar elements in the copyright database. Indicates visual similarity to elements in the public domain; when the element's originality index... When the element originality index is >0.8, the element replacement engine is triggered. For separable structures, the basic outline is preserved and symbols are injected. For inseparable structures, visual twins are generated using CycleGAN. These symbols are sampled from a pre-trained cultural theme vector database and fused through a style transfer network. Separable structures are determined to be separable by detecting the variance of the outline curvature change rate; a variance < threshold η indicates separability. Inseparable structures generate visual twins using CycleGAN, where the Euclidean distance from the original element is greater than a threshold f. When the element originality index... At that time, the knowledge graph is invoked to match historical elements to obtain matching rules;

[0035] The cross-modal generator module is configured to encode the constraint set C into an embedding vector with the same dimension as the feature matrix M, which is then concatenated after LayerNorm normalization and input into a generative adversarial network. Specifically, it includes: mapping the accessory features to the clothing latent space by training a projection matrix, using the constraint set C as the query vector and the projected features as the key vector, and calculating the attention weight matrix; generating candidate clothing schemes using a diffusion model and outputting a matching score S, and rendering a virtual try-on sequence; the denoising process of the diffusion model is conditional on the constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage.

[0036] The production conversion interface module is configured to make process decisions based on the matching score S, including: generating a cost optimization model by combining pattern complexity and fabric unit price; associating metal ornaments with wrinkle-resistant fabrics; and enabling laser cutting technology for irregular contours; and outputting cutting schemes and bills of materials that can be directly put into production.

[0037] According to a third aspect of the present invention, a computer program product is provided, on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0038] The above-described one or more technical solutions in the embodiments of this application have at least one of the following technical effects:

[0039] This invention achieves high-precision preprocessing of accessories and clothing by integrating an image segmentation network and a three-level classifier. The image segmentation network accurately extracts the outline and detailed features of the clothing, while the three-level classifier further refines the attribute labels, laying a data foundation for subsequent adaptation analysis. The originality index calculation dynamically adjusts the element replacement strategy by quantifying the scarcity of design elements and their correlation with cultural symbols, ensuring that the modified accessories both conform to niche cultural aesthetics and retain originality. The feature projection layer maps multi-dimensional attributes to a low-dimensional space, the attention allocation layer focuses on key areas (such as the interaction between clothing texture and accessory structure), and the conditional diffusion model generates highly realistic adaptation schemes based on user preferences.

[0040] The system connects a small-batch fabric database with a process parameter library to match physical properties (such as elasticity and gloss) with production process constraints in real time, and outputs feasible design decisions. Attached Figure Description

[0041] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0042] Figure 1 A flowchart illustrating a niche accessories-driven multimodal collaborative design method for clothing is shown according to an embodiment of the present invention.

[0043] Figure 2 A schematic diagram of a niche accessories-driven multimodal collaborative design system for clothing modules is shown according to an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the structure of a computer system suitable for implementing the electronic devices of the present application embodiments. Detailed Implementation

[0045] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] Figure 1 A flowchart illustrating a niche accessories-driven multimodal collaborative design method for clothing, according to an embodiment of the present invention, is shown. Figure 1 As shown:

[0048] S1. Receive a flat image of the jewelry uploaded by the user, extract the main outline of the jewelry through an image segmentation network, and perform three-level attribute classification in parallel based on the outline features, including:

[0049] Material classification: Convolutional neural networks are used to identify metal, fabric, or plastic materials, with fabric materials further subdivided into knitted, woven, and lace subcategories;

[0050] Structural classification: Hough transform is applied to detect geometric features, and a classifier is used to determine whether the geometric object is regular or irregular.

[0051] Topic classification: It integrates image visual features with user text tags and outputs subculture topic codes through an attention mechanism.

[0052] The third-level attribute classification further includes:

[0053] The material classification sub-model adopts the MobileNetV3 network architecture, and the training dataset contains 200 kinds of material surface micro textures;

[0054] Structural classification extracts feature vectors of straight lines and circular arcs through Hough transform, and inputs them into a support vector machine classifier;

[0055] Topic classification employs a dual-modal fusion approach, specifically by concatenating ResNet features of images with BERT embedding vectors of text, followed by topic encoding extraction through a multi-head self-attention layer.

[0056] Finally, a structured feature matrix M = [contour vector, material encoding, theme encoding] is generated;

[0057] The image segmentation network accurately segments images of accessories and clothing into key parts and extracts highly recognizable features. A three-level classifier further refines the classification process for each part, enabling the system to accurately classify and label different types of clothing and accessories, as well as their combinations. This process significantly improves the accuracy and efficiency of image processing, providing clear and accurate foundational data for subsequent adaptation.

[0058] S2. Input the feature matrix M into the copyright filtering module and output the clothing design constraint set C. The specific implementation steps of the copyright filtering module are as follows:

[0059] S2.1 Calculate the originality index of elements ,in, This represents the uniqueness score of similar elements in the copyright database. Indicates visual similarity to public domain elements; weighting coefficients of 0.7 and 0.3 were determined based on regression analysis of 2000 historical design infringement cases.

[0060] Uniqueness score of elements of the same type , where N represents the number of samples in the copyright library whose feature distance to the target element is less than the threshold τ;

[0061] Visual similarity Calculated using cosine similarity:

[0062]

[0063] in, This represents the ResNet-50 feature vector of the jewelry element. Represents the baseline feature vector of public domain elements. This indicates that the text tag for the accessory is embedded. This indicates a public domain text description embedding. The meaning represented is the visual feature similarity weight. The meaning represented is the text feature similarity weight;

[0064] S2.2, When the element's originality index When the value is greater than 0.8, the element replacement engine is triggered. For separable structures, the basic outline is preserved and symbols are injected. For non-separable structures, visual twins are generated using CycleGAN. The symbols are sampled from a pre-trained cultural theme vector database and fused through a style transfer network. Separable structures are determined to be separable by detecting the variance of the rate of change of contour curvature. If the variance is less than a threshold η, the structure is considered separable. The threshold η is obtained experimentally. For non-separable structures, visual twins are generated using CycleGAN. The Euclidean distance between the generated elements and the original elements is greater than a threshold f, which is obtained experimentally.

[0065] The elements sampled from the vector database in this invention are mostly common, highly repetitive, or high-risk symbols. If the value of a certain element is calculated... High originality, meaning it is neither identical to existing copyrighted elements nor like public domain elements, but a fresh, unique, and worthwhile creative symbol. Therefore, this invention is deemed to be... When the value is >0.8, highly original elements are replaced in the main outline of the garment as the main design elements, while the original mediocre or repetitive elements are replaced.

[0066] During the replacement process, a style transfer network is used to maintain the overall structural consistency of the original outline, while symbolic features of highly original elements are injected into the retained outline to form an enhanced design subject. If the original outline is inseparable, or if the two structures differ too much, a visual twin with features of both is generated using CycleGAN.

[0067] By replacing original elements with highly original symbols, the level of innovation in generated clothing can be significantly improved, while reducing the risk of repetitive designs caused by the use of low-original elements, thus achieving higher visual differentiation and content generation quality.

[0068] The specific implementation process of the element replacement engine is as follows:

[0069] Separability detection: Calculate the variance of the rate of change of contour curvature. If the variance is less than the threshold η, it is determined to be a separable structure. The threshold η is determined by K-means clustering of the contour curvature dataset, and the value range is [0.2, 0.5].

[0070] The specific steps of contour curvature change rate variance detection include: obtaining the contour using edge detection methods and calculating the curvature of each point; calculating the curvature change rate, i.e., the change in curvature between adjacent points of the contour; and calculating the variance V of the curvature change rate, implemented using the formula: V = ,in, This represents the rate of change of curvature at the i-th point. This represents the mean of the rate of change of curvature, and number is the number of contour points.

[0071] If the variance is less than the threshold η, the structure is judged to be a separable structure, that is, the outline of the graphic or object is relatively smooth and the boundary changes little; if the variance is greater than the threshold η, the structure is judged to be an inseparable structure; Visual twin generation of inseparable structures: CycleGAN network is used to perform style transfer to ensure that the Euclidean distance between the output and the original element is greater than the copyright security threshold f.

[0072] Style transfer aims to apply a style to a target image. The steps include: source image (the original image containing the content to be transferred); target style (an image containing the style to be transferred to the source image); and using a CycleGAN network, the generator transforms the source image into an image with the target style while preserving the content features of the source image. A key aspect of this process is ensuring that the style-transferred image is visually significantly different from the original image to avoid copyright infringement. To achieve this, the Euclidean distance between the style-transferred image and the original image needs to be calculated, and this distance must be greater than a set threshold f.

[0073] Euclidean distance refers to the "distance" between two images in pixel space, which can be calculated using the following formula: D = , where x and y are the pixel values ​​of the two images, and num is the total number of pixels; to ensure that the style-transferred image is not too similar to the original image, a copyright security threshold f needs to be set, requiring the Euclidean distance to be greater than f.

[0074] During the training of the CycleGAN network, a structural constraint loss function is incorporated to ensure that the twins retain the original spatial topological features; the structural constraint loss function is: ,in, Indicates the initial mask. Indicates a twin mask;

[0075] S2.3, When the element's originality index At that time, the knowledge graph is called to match historical elements to obtain matching rules, and the overall clothing design constraint set of step S2 is output, where the clothing design constraint set C={pattern code, color scheme, pattern parameters};

[0076] The originality index calculation and element replacement execution work together to enhance the intelligence and personalization of the adaptation process. The originality index calculation generates an innovation index by analyzing the matching degree and style differences between accessories and clothing, measuring the uniqueness and originality of the accessory-clothing combination. The element replacement engine then automatically adjusts based on this index, replacing unsuitable elements to ensure that the final combination not only meets the user's needs but also possesses high visual appeal and personalization. This process not only improves the compatibility of accessories and clothing but also avoids excessive imitation of existing works, ensuring copyright protection and innovation.

[0077] S3. Encode the constraint set C into embedding vectors of the same dimension as the feature matrix M, normalize them using LayerNorm, concatenate them, and input them into the generative adversarial network. Specifically, this includes:

[0078] S3.1. By training the projection matrix, the features of the accessories are mapped to the latent space of the clothing. Using the constraint set C as the query vector and the projected features as the key vector, the attention weights of the query vector and the key vector are calculated using a dot product attention mechanism to obtain the attention weight matrix. Specifically, this includes establishing the projection relationship M' = M, where This represents the projection matrix trained through contrastive learning. Perform a fully connected mapping on all sub-features of the feature matrix M;

[0079] In some embodiments, multi-objective optimization can also be performed, and the Pareto optimal solution set can be calculated using the attention weight matrix: min[E(W),T(W)],W ;

[0080] E(W) =

[0081] T(W) = max[TS( )],1 i m

[0082] Where W represents the attention weight matrix, E(W) represents the total system energy consumption, and T(W) represents the total task completion time. Represents the Pareto optimal solution set This represents the element in the i-th row and j-th column of the weight matrix. This represents the computational energy consumption of task i on resource j. Let T(W) represent the energy consumption for communication between task i and resource j, n represent the total number of tasks, m represent the total number of resources, and T(W) represent the total time. TS( ) represents resources The completion time of the last subtask. This represents the i-th processor or production resource;

[0083] After calculating the optimal solution set as described above, the projection matrix is ​​further iteratively updated:

[0084] =M'+

[0085] M' represents the projection matrix. Indicates the learning rate. This represents the gradient of the objective function J with respect to the projection matrix M', reflecting the optimization direction. Represents a multi-objective loss function. This represents the projection matrix after iteration;

[0086] S3.2. Generate candidate clothing schemes using a diffusion model and output a matching score S, and render a virtual try-on sequence; the denoising process of the diffusion model is based on a constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage;

[0087] The feature projection layer, attention allocation layer, and conditional diffusion model together constitute a more refined adaptation strategy. The feature projection layer uses deep learning algorithms to project the visual features of clothing and accessories into a multi-dimensional space, helping the system better understand the relationships between their shape, color, material, and other features. The attention allocation layer dynamically adjusts the weight of each feature, intelligently optimizing every detail for different matching needs, improving the naturalness and harmony of the overall effect. The conditional diffusion model incorporates external control factors (such as user preferences and scene requirements) into the generation process through a conditional mechanism, further enhancing the system's flexibility and user customization experience.

[0088] S4. Based on the matching score S, make process decisions. Further, combine pattern complexity and fabric unit price to generate cost optimization model; associate metal ornaments with anti-wrinkle fabric, and use laser cutting technology for irregular contours; output cutting schemes and bill of materials that can be directly put into production.

[0089] The process decisions include:

[0090] In the fabric matching rules, when the hardness of the jewelry material in the matching score S is >5H, it is associated with a blended wrinkle-resistant fabric, and a cushioning lining design is enabled for brittle materials; when the hardness of the jewelry material in the matching score S is higher than the standard hardness threshold, it is associated with a wrinkle-resistant fabric.

[0091] The cost optimization model is specifically as follows: ,in, This represents the consumption of the i-th type of fabric. Indicates the unit price. The complexity factor represents the pattern complexity weighting factor, n represents the total number of samples, Complexity represents the pattern processing complexity, and the pattern processing complexity is quantified by the number of contour curvature change points, the number of seams, and the number of special process steps, and the unit of measurement is the standardized complexity unit (SCU).

[0092] A specific calculation example is as follows: If the ornament is an irregular metal badge (the aforementioned irregular structure):

[0093] Contour curvature change point: 32 SCU 15 SCU;

[0094] Laser cutting process required: 8 SCUs;

[0095] Cushioning liner design: 5 SCU;

[0096] Total pattern processing complexity: 28 SCU;

[0097] Substitute into the cost model: =10 (yuan / SCU) Complexity cost = 280 yuan.

[0098] Finally, the decision output connects a small-batch fabric database and a process parameter library, closely integrating actual fabric and process information with the adaptation plan. This provides users with more precise customization suggestions, offering specific advice based on the compatibility of clothing and accessories, and adjusting design plans in real time based on the latest data in the database. This solves the problem of mismatch between accessories and clothing in niche cultural scenarios such as cosplay and collectible toys, demonstrating significant technical effectiveness.

[0099] Furthermore, such as Figure 2 The diagram illustrates a modular design system for niche accessories-driven multimodal collaborative design of clothing according to an embodiment of the present invention, specifically including:

[0100] The multimodal input module is configured to receive flat images of jewelry uploaded by users, extract the main outline of the jewelry through an image segmentation network, perform three-level attribute classification in parallel based on the outline features, and generate a structured feature matrix M.

[0101] The copyright filtering engine module is configured to input the feature matrix M into the copyright filtering module and output the clothing design constraint set C. The specific implementation steps of the copyright filtering module are as follows: calculate the originality index of the elements. ,in, This represents the uniqueness score of similar elements in the copyright database. Indicates visual similarity to elements in the public domain; when the element's originality index... When the element originality index is >0.8, the element replacement engine is triggered. For separable structures, the basic outline is preserved and symbols are injected. For inseparable structures, visual twins are generated using CycleGAN. These symbols are sampled from a pre-trained cultural theme vector database and fused through a style transfer network. Separable structures are determined to be separable by detecting the variance of the outline curvature change rate; a variance < threshold η indicates separability. Inseparable structures generate visual twins using CycleGAN, where the Euclidean distance from the original element is greater than a threshold f. When the element originality index... At that time, the knowledge graph is invoked to match historical elements to obtain matching rules;

[0102] The cross-modal generator module is configured to encode the constraint set C into an embedding vector with the same dimension as the feature matrix M, which is then concatenated after LayerNorm normalization and input into a generative adversarial network. Specifically, it includes: mapping the accessory features to the clothing latent space by training a projection matrix, using the constraint set C as the query vector and the projected features as the key vector, and calculating the attention weight matrix; generating candidate clothing schemes using a diffusion model and outputting a matching score S, and rendering a virtual try-on sequence; the denoising process of the diffusion model is conditional on the constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage.

[0103] The production conversion interface module is configured to make process decisions based on the matching score S. Furthermore, it combines pattern complexity and fabric unit price to generate a cost optimization model; it associates metal ornaments with wrinkle-resistant fabrics, and uses laser cutting technology for irregular contours; it outputs cutting schemes and bills of materials that can be directly put into production.

[0104] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system 300 suitable for implementing electronic devices according to embodiments of the present application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0105] like Figure 3 As shown, the computer system 300 includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the system 300. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0106] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a liquid crystal display (LCD) and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 310 as needed so that computer programs read from it can be installed into storage section 308 as needed.

[0107] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application. It should be noted that the computer-readable storage medium of this application can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0108] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0109] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0110] The modules described in the embodiments of this application can be implemented in software or in hardware.

[0111] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: receive a flat image of a jewelry uploaded by a user; extract the main outline of the jewelry through an image segmentation network; perform three-level attribute classification in parallel based on the outline features to generate a structured feature matrix M; input the feature matrix M into a copyright filtering module and output a clothing design constraint set; encode the constraint set C into an embedding vector of the same dimension as the feature matrix M, standardize it using LayerNorm, concatenate them, input it into a generative adversarial network, and output a matching score S; make process decisions based on the matching score S, including: generating a cost optimization model by combining pattern complexity and fabric unit price; associating metal jewelry with wrinkle-resistant fabric; and using laser cutting technology for irregularly shaped outlines; and outputting a cutting scheme and bill of materials that can be directly put into production.

[0112] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A niche accessories-driven multimodal collaborative design method for clothing, characterized in that, include: S1. Receive the flat image of the jewelry uploaded by the user, extract the main outline of the jewelry through an image segmentation network, and generate a structured feature matrix M; S2. Input the feature matrix M into the copyright filtering module and output the clothing design constraint set. The specific implementation steps of the copyright filtering module are as follows: S2.1 Calculate the originality index of elements ,in, This represents the uniqueness score of similar elements in the copyright database. Indicates visual similarity to elements in the public domain; S2.

2. Sample from a pre-trained cultural theme vector database, fuse them through a style transfer network to generate symbols, and when the element originality index... When the value is greater than 0.8, the element replacement engine is triggered. When the element replacement engine is executed, the basic contours of separable structures in the feature matrix M are preserved and the symbols are injected. For non-separable structures, visual twins are generated using CycleGAN. The separable structures are determined to be separable when the variance of the contour curvature change rate is less than the threshold η. The non-separable structures are used to generate visual twins using CycleGAN, and the Euclidean distance between the original elements and the twins is greater than the threshold f. S2.3, When the element's originality index When the knowledge graph is invoked, the clothing design constraint set C is output by matching historical element matching rules based on the relationship between elements in the knowledge graph. S3. Encode the constraint set C into embedding vectors of the same dimension as the feature matrix M, normalize them using LayerNorm, concatenate them, and input them into the generative adversarial network. Specifically, this includes: S3.

1. By training the projection matrix, the features of the accessories are mapped to the latent space of the clothing. The constraint set C is used as the query vector, the projected features are used as the key vector, and the attention weight matrix is ​​calculated. S3.

2. Generate candidate clothing schemes using a diffusion model and output a matching score S, and render a virtual try-on sequence; the denoising process of the diffusion model is based on a constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage; S4. Based on the matching score S, make process decisions, including: optimizing the model by combining pattern complexity and fabric unit price; associating metal ornaments with wrinkle-resistant fabrics; using laser cutting technology for irregular contours; outputting cutting schemes and bills of materials that can be directly put into production; the process decisions include: In the fabric matching rules, when the hardness of the jewelry material is >5H, it is associated with a blended wrinkle-resistant fabric, and brittle materials are designed with a cushioning lining; when the hardness of the jewelry material is higher than the standard hardness threshold, it is associated with a wrinkle-resistant fabric. The cost optimization model is specifically as follows: ,in, This represents the consumption of the i-th type of fabric. Indicates the unit price. This represents the pattern complexity weighting factor, where Complexity represents the pattern processing complexity, and n represents the total number of samples.

2. The multimodal collaborative design method for clothing according to claim 1, characterized in that, Step S1 further includes generating a structured feature matrix M by performing three-level attribute classification in parallel based on contour features. The three-level attribute classification specifically includes: Material classification: Convolutional neural networks are used to identify metal, fabric, or plastic materials, with fabric materials further subdivided into knitted, woven, and lace subcategories; Structural classification: Hough transform is applied to detect geometric features, and a classifier is used to determine whether the geometric object is regular or irregular. Topic classification: It integrates image visual features with user text tags and outputs subculture topic codes through an attention mechanism.

3. The multimodal collaborative design method for clothing according to claim 2, characterized in that, The three-level attribute classification further includes: The material classification sub-model adopts the MobileNetV3 network architecture, and the training dataset contains 200 kinds of material surface micro textures; Structural classification extracts feature vectors of straight lines and circular arcs through Hough transform, and inputs them into a support vector machine classifier; Topic classification employs a dual-modal fusion approach, specifically by concatenating ResNet features of images with BERT embedding vectors of text, followed by topic encoding extraction through a multi-head self-attention layer.

4. The multimodal collaborative design method for clothing according to claim 1, characterized in that, In the calculation of the element originality index: the uniqueness score of elements of the same type , where N represents the number of samples in the copyright library whose feature distance to the target element is less than the threshold τ; Visual similarity Calculated using cosine similarity: ； in, This represents the ResNet-50 feature vector of the jewelry element. Represents the baseline feature vector of public domain elements. This indicates that the text tag for the accessory is embedded. This indicates a public domain text description embedding. The meaning represented is the visual feature similarity weight. The meaning represented is the text feature similarity weight.

5. The multimodal collaborative design method for clothing according to claim 1, characterized in that, The specific implementation process of the element replacement engine is as follows: Separability detection: Calculate the variance of the rate of change of contour curvature. If the variance is less than the threshold η, it is determined to be a separable structure. The threshold η is determined by K-means clustering of the contour curvature dataset, and the value range is [0.2, 0.5]. Visual twin generation of inseparable structures: Style transfer is performed using CycleGAN network to ensure that the Euclidean distance between the output and the original element is greater than the copyright security threshold f; a structural constraint loss function is added during CycleGAN network training to ensure that the twin retains the original spatial topological features; The structural constraint loss function is: ,in, Indicates the initial mask. This represents a twin mask.

6. The multimodal collaborative design method for clothing according to claim 1, characterized in that, The specific step of mapping the features of the accessories to the latent space of the clothing is: establishing a projection relationship M'= M, where This represents the projection matrix trained through contrastive learning. Perform a fully connected mapping on all sub-features of the feature matrix M.

7. A niche accessories-driven multimodal collaborative design system for clothing, characterized in that, The multimodal collaborative design method for clothing according to any one of claims 1-6 specifically includes: The multimodal input module is configured to receive flat images of jewelry uploaded by users, extract the main outline of the jewelry through an image segmentation network, perform three-level attribute classification in parallel based on the outline features, and generate a structured feature matrix M. The copyright filtering engine module is configured to input the feature matrix M into the copyright filtering module and output a set of clothing design constraints. The specific implementation steps of the copyright filtering module are as follows: calculating the originality index of elements. ,in, This represents the uniqueness score of similar elements in the copyright database. Indicates visual similarity to elements in the public domain; when the element's originality index... When the element originality index is >0.8, the element replacement engine is triggered. For separable structures, the basic outline is preserved and symbols are injected. For inseparable structures, visual twins are generated using CycleGAN. These symbols are sampled from a pre-trained cultural theme vector database and fused through a style transfer network. Separable structures are determined to be separable by detecting the variance of the outline curvature change rate; a variance < threshold η indicates separability. Inseparable structures generate visual twins using CycleGAN, where the Euclidean distance from the original element is greater than a threshold f. When the element originality index... At that time, the knowledge graph is invoked to match historical matching rules and output the clothing design constraint set C; The cross-modal generator module is configured to encode the constraint set C into an embedding vector with the same dimension as the feature matrix M, which is then concatenated after LayerNorm normalization and input into a generative adversarial network. Specifically, it includes: mapping the accessory features to the clothing latent space by training a projection matrix, using the constraint set C as the query vector and the projected features as the key vector, and calculating the attention weight matrix; generating candidate clothing schemes using a diffusion model and outputting a matching score S, and rendering a virtual try-on sequence; the denoising process of the diffusion model is conditional on the constraint set C, and similar feature vectors in the copyright library are masked during the latent space sampling stage. The production conversion interface module is configured to make process decisions based on a matching score S, including: combining pattern complexity and fabric unit price to generate a cost optimization model; associating metal ornaments with wrinkle-resistant fabrics; and enabling laser cutting for irregular contours; outputting production-ready cutting schemes and bills of materials; the process decisions include: In the fabric matching rules, when the hardness of the jewelry material is >5H, it is associated with a blended wrinkle-resistant fabric, and brittle materials are designed with a cushioning lining; when the hardness of the jewelry material is higher than the standard hardness threshold, it is associated with a wrinkle-resistant fabric. The cost optimization model is specifically as follows: ,in, This represents the consumption of the i-th type of fabric. Indicates the unit price. This represents the pattern complexity weighting factor, where Complexity represents the pattern processing complexity, and n represents the total number of samples.

8. A computer program product, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.

9. A computing system, characterized in that, It includes a processor and a memory, the processor being configured to perform the method as described in any one of claims 1-6.

Images