Garment fitting method and system based on human body three-dimensional model

Abstract

The invention discloses a garment fitting method and system based on a human body three-dimensional model, and belongs to the technical field of virtual fitting. The method comprises the following steps: receiving and analyzing abstract wearing intention description input by a user; generating a target multi-clothing combination scheme; constructing a clothing combination template library to form a core template set; generating a processing scheme suitable for the current combination; sequentially processing matching between each piece of clothes and the human body three-dimensional model; establishing a personal preference file; carrying out personalized fine adjustment on a processing scheme generated by template matching; carrying out scenario optimization on the processing scheme after personalized fine adjustment; and generating a multi-garment try-on effect based on the processing scheme after scenario optimization. According to the method, the clothing combination template library is constructed, the processing scheme is generated by adopting a three-level matching mechanism, clothing matching and collision detection are sequentially carried out according to the stacking sequence, and the problem of matching distortion caused by the fact that multiple clothing stacking sequences cannot be processed in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of virtual fitting technology, and in particular to a method and system for fitting clothing based on a three-dimensional human body model. Background Technology

[0002] With the rapid development of e-commerce and consumers' increasing demands for online shopping experiences, virtual try-on technology has become an important development direction in the apparel retail sector. By using 3D modeling and virtual reality technology to enable digital try-on of clothing, the limitations of traditional online shopping—the inability to physically try on clothes—can be effectively overcome, providing consumers with a more intuitive shopping experience.

[0003] Existing virtual fitting systems construct 3D models of the human body and clothing, and use physics engines to match individual garments and demonstrate the fitting effect. These systems typically employ spring-mass models or finite element analysis methods to handle clothing deformation, enabling the matching of individual garments with human body models.

[0004] In existing technologies, virtual try-on systems mainly focus on the try-on effect of single garments, which has the following problems: First, there is a lack of clear definition and processing mechanism for the stacking order of multiple garments, making it impossible to realistically reproduce the matching effect of multiple layers of clothing such as underwear and outerwear in daily wear; second, when trying on multiple layers of clothing, the space occupied by inner garments to outer garments is often ignored, resulting in unreasonable penetration or deformation when matching outer garments; finally, due to insufficient accuracy in layer matching, users find it difficult to evaluate the overall effect of wearing multiple garments together through the system, limiting its practical application value in scenarios such as clothing matching recommendations and virtual wardrobes. Summary of the Invention

[0005] This application provides a clothing try-on method and system based on a three-dimensional human body model, which solves the problems in the prior art that cannot effectively handle the inaccurate matching, penetration phenomenon and low practicality of matching caused by the stacking order of multiple clothing layers. It achieves accurate matching, realistic visual effect and enhanced matching application capability for multi-layer clothing try-on.

[0006] This application provides a clothing try-on method based on a three-dimensional human body model, including: receiving and parsing an abstract description of the user's wearing intention, and mapping the intention description into visual targets of body modification, style emotion and scene fit.

[0007] Using the visual target as multiple constraints, generate target multi-clothing combination schemes that satisfy all visual constraints, and output the optimized target try-on scheme;

[0008] Acquire user-selected multi-person clothing combinations and 3D human body model data;

[0009] Build a clothing combination template library to form a core template set;

[0010] Based on the user's selection of multiple clothing combinations, a template is matched from the clothing combination template library to generate a processing solution suitable for the current combination;

[0011] According to the stacking order rules defined in the template matching processing scheme, the matching of each garment with the human body 3D model is processed sequentially;

[0012] Collect users' fitting history, adjustment preferences, and evaluation feedback to build personal preference profiles;

[0013] By combining personal preference profiles, the processing scheme generated by template matching can be fine-tuned.

[0014] Identify the current fitting scenario and optimize the personalized, fine-tuned processing solution in a contextualized manner;

[0015] Based on the contextualized optimized processing scheme, multiple clothing try-on effects are generated.

[0016] Furthermore, the steps of receiving and parsing the abstract clothing intention description input by the user, and mapping the intention description to visual targets of body modification, style emotion, and scene fit include:

[0017] Receive text describing the user's intended clothing style submitted through a text input interface;

[0018] A pre-trained natural language processing model is used to perform word segmentation and semantic parsing on the descriptive text to extract semantic units from the descriptive text;

[0019] The semantic units are mapped and matched with a predefined visual target concept library;

[0020] For body modification, the mapping generates quantitative modification instructions regarding the direction of contour changes for specific parts of the human body;

[0021] For style sentiment, the mapping generates style vectors that describe the style attributes of clothing;

[0022] For scene fit, the mapping generates scene type identifiers and a list of constraints related to scene fit;

[0023] The quantization modification instructions, style vectors, scene type identifiers, and constraint list are integrated and encapsulated into a visual target data object.

[0024] Furthermore, the steps of generating target multi-clothing combination schemes that satisfy all visual constraints, and outputting the optimized target try-on scheme, using the visual target as multiple constraints, include:

[0025] The visual target data object is parsed into multiple independent constraints, and a pre-built knowledge graph is loaded.

[0026] Using the style vector and scene constraints in the visual target data as initial query conditions, multiple initial clothing combination candidate schemes are formed;

[0027] For each initial candidate solution, calculate its satisfaction score on each visual constraint dimension, and select the set of solutions with a comprehensive score exceeding the preset score threshold as the target multi-clothing combination solution.

[0028] Physical simulations were performed on each target multi-clothing combination scheme to simulate the shape and layering relationship of clothing under gravity, human body collision and interaction between clothing;

[0029] Output an optimized target fitting solution that satisfies both visual and physical constraints.

[0030] Furthermore, the steps to build a clothing combination template library and form a core template set include:

[0031] Extract clothing type combination data, layering order rule data, matching priority data, and deformation parameter data;

[0032] The extracted clothing type combination data, layering order rule data, matching priority data, and deformation parameter data are integrated into a core template set;

[0033] A hierarchical storage structure is adopted, with template subsets arranged in descending order of usage frequency, template subsets categorized by applicable scenarios, and template subsets filtered by user rating score thresholds.

[0034] Generate metadata records for each template;

[0035] Link the core template set with metadata records to enable the template retrieval process;

[0036] Initialize the template library storage space and verify the integrity of the integrated data.

[0037] Furthermore, the steps of matching templates from the clothing combination template library to generate a processing solution suitable for the current combination include:

[0038] Read the user's selected multi-person clothing combination data;

[0039] Query the clothing combination template library and perform precise layer matching;

[0040] When an exact match fails, a similarity matching hierarchy is initiated to obtain the cosine similarity score between the user's clothing type combination identifier and the template's clothing type combination identifier, and the template with the highest similarity score is selected as the candidate template.

[0041] When all similarity matching scores are below the preset similarity threshold, rule derivation hierarchy is performed to extract the stacked order rule sequence and the matching priority value list from the core template set;

[0042] A new stacked order rule sequence is generated by combining logical reasoning algorithms with extracted rule sequences;

[0043] The new stacking order rule sequence, user clothing type combination identifier, and deformation parameter threshold range are merged to form a new processing scheme, which is then output.

[0044] Furthermore, following the stacking order rules defined in the template matching processing scheme, the steps for matching each garment with the 3D human body model sequentially include:

[0045] Extract the stacking order rule sequence from the processing scheme generated according to template matching;

[0046] Starting from the first element of the stacking order sequence, select the first garment data, including the garment mesh model and size parameters;

[0047] Map the selected clothing mesh model onto the surface of the 3D human body model, and adjust the position of the clothing mesh model;

[0048] Calculate collision detection between clothing mesh model and human 3D model, and identify overlapping areas;

[0049] When processing multiple garments, the matching results of the previous garment are accumulated and used as the basis for the matching of the current garment.

[0050] Generate a complete matching model, verify the effectiveness of the matching parameter adjustment, and confirm the non-overlapping interference region through distance calculation.

[0051] Furthermore, the steps to obtain users' fitting history, adjust preferences, and provide feedback to build a personal preference profile include:

[0052] Extract historical fitting data, analyze and adjust preference data, and collect evaluation feedback data;

[0053] The historical data of fitting, the data of adjusting preferences and the data of evaluation feedback are classified and stored. The historical data of fitting is sorted by timestamp, the data of adjusting preferences is grouped by parameter type, and the data of evaluation feedback is stratified by score level.

[0054] Obtain preference statistics, aggregate the average adjustment magnitude from the preference adjustment data, and calculate the weighted average score from the evaluation feedback data;

[0055] Build a personal preference profile structure.

[0056] Furthermore, the steps for personalizing the template matching processing scheme by incorporating individual preference profiles include:

[0057] Read the processing scheme generated by template matching, and parse the stacking order rules, matching priority settings and deformation parameter values;

[0058] Extract user preference parameters from personal preference profiles, including preferred values ​​for layer spacing, clothing fit requirements, and acceptable range for deformation effects;

[0059] The parameters in the processing scheme generated by template matching are compared and analyzed with the user preference parameters to identify the parameter items that need to be adjusted;

[0060] The treatment plan was adjusted item by item based on the comparative analysis results;

[0061] Establish an adjustment constraint mechanism to perform consistency checks on the adjusted processing plan;

[0062] The fine-tuned processing solution is stored as a new solution version, associated with the corresponding user ID and preference profile version number.

[0063] Furthermore, the steps for identifying the current fitting scenario and optimizing the personalized, fine-tuned processing plan in a contextualized manner include:

[0064] By analyzing the scenario description information in user requests, the type of try-on scenario can be identified;

[0065] Based on the identified scene type, select a suitable template group from the template library. The template group contains multiple clothing matching templates suitable for that scene.

[0066] Read the personalized, fine-tuned processing solution and analyze its suitability for the current scenario;

[0067] Parameter optimization of the processing solution based on scenario characteristics;

[0068] Establish a scenario-based parameter adjustment rule library and define parameter correction rules for each scenario type;

[0069] Generate the final contextualized optimization solution.

[0070] This application provides a clothing try-on system based on a 3D human body model, which is used to implement a clothing try-on method based on a 3D human body model, including: an intent mapping module, a solution output module, a data acquisition module, a template construction module, a template matching module, a matching processing module, a personal profile creation module, a profile fine-tuning module, a scene optimization module, and an effect generation module;

[0071] The intent mapping module is used to receive and parse the abstract clothing intent description input by the user, and map the intent description into visual targets of body modification, style emotion and scene fit.

[0072] The solution output module is used to generate a target multi-clothing combination solution that satisfies all visual constraints, using the visual target as multiple constraints, and output the optimized target try-on solution.

[0073] The data acquisition module is used to acquire the multi-person clothing combination and human body three-dimensional model data selected by the user;

[0074] The template building module is used to build a clothing combination template library to form a core template set;

[0075] The template matching module is used to perform template matching from the clothing combination template library based on the user-selected multi-person clothing combination, and generate a processing solution suitable for the current combination.

[0076] The matching processing module is used to process the matching of each garment with the human body 3D model in sequence according to the stacking order rules defined in the processing scheme generated by template matching;

[0077] The personal profile creation module is used to obtain the user's fitting history, adjustment preferences and evaluation feedback, and to create a personal preference profile;

[0078] The file fine-tuning module is used to personalize the processing scheme generated by template matching by combining personal preference files;

[0079] The scenario optimization module is used to identify the current fitting scenario and perform contextual optimization on the personalized fine-tuning processing solution.

[0080] The effect generation module is used to generate multiple clothing try-on effects based on the contextualized optimized processing scheme.

[0081] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0082] By constructing a clothing combination template library and employing a template matching mechanism, the layering order of multiple garments can be effectively handled, ensuring the integrity of the matching foundation. This overcomes the limitations of single-item matching logic in handling complex layering, making the try-on process more realistic in real-life scenarios. Furthermore, when handling the layering order, each garment is processed sequentially according to a rule sequence, with collision detection and adaptive parameter adjustments. This avoids penetration or distortion issues that occur when matching inner and outer garments, improving the accuracy and visual realism of the try-on effect. Going further, while ensuring the realism of the try-on, personalized fine-tuning and contextual optimization are performed by incorporating personal preference profiles. This allows for flexible adaptation to individual user differences and different scenario needs, enhancing the practicality and user satisfaction in clothing matching evaluation. Simultaneously, by introducing a reverse reasoning mechanism based on wearing intentions, professional matching schemes can be proactively generated based on the user's abstract goals. This overcomes the limitations of traditional try-on methods that rely on users selecting clothing themselves, achieving a shift from passive display to proactive intelligent design, and improving the personalization and practicality of matching schemes. Attached Figure Description

[0083] Figure 1 A flowchart of a clothing try-on method based on a three-dimensional human body model is provided for an embodiment of this application;

[0084] Figure 2 This is a schematic diagram of the structure of a clothing try-on system based on a three-dimensional human body model, provided in an embodiment of this application. Detailed Implementation

[0085] This application provides a clothing try-on method and system based on a 3D human body model, which solves the problem of matching distortion caused by the inability to handle the stacking order of multiple garments in the prior art. By constructing a clothing combination template library containing stacking rules, a three-level matching mechanism of precise matching, similarity matching and rule deduction is adopted to generate a processing scheme. Clothing matching and collision detection are performed sequentially according to the stacking order, realizing a realistic and natural stacking effect and accurate spatial relationship presentation when trying on multiple layers of clothing.

[0086] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0087] like Figure 1 The diagram shows a flowchart of a clothing try-on method based on a three-dimensional human body model provided in this application embodiment. The method is applied to a clothing try-on system based on a three-dimensional human body model. The method includes the following steps: receiving and parsing an abstract description of the wearing intention input by the user, and mapping the intention description into quantifiable visual targets of body modification, style emotion and scene fit through a natural language processing model.

[0088] Using the visual target as multiple constraints, in the pre-constructed knowledge graph of clothing attributes, matching rules and aesthetic feedback, graph neural networks are used for reverse reasoning and multi-objective optimization to generate a target multi-clothing combination scheme that satisfies all visual constraints. The clothing in the target multi-clothing combination scheme is dynamically fitted and layered with the user's three-dimensional human body model for physical simulation. Through joint optimization algorithms, the clothing selection, size and layering order are iteratively adjusted to output an optimized target try-on scheme with excellent visual performance and physical rationality.

[0089] Acquire user-selected multi-person clothing combinations and 3D human body model data;

[0090] A clothing combination template library is constructed, which integrates clothing matching expert knowledge, fashion rules and dressing standards to form a core template set that defines clothing type combinations, stacking order rules, matching priorities and transformation parameters. It adopts a hierarchical storage structure to classify and manage the templates according to usage frequency, applicable scenarios and user evaluations.

[0091] Based on the user's selected multi-person clothing combination, template matching is performed from the clothing combination template library, which is divided into three levels: exact matching, similarity matching, and rule inference, to generate a processing solution suitable for the current combination.

[0092] According to the stacking order rules defined in the processing scheme generated by template matching, the matching of each garment with the human body 3D model is processed in turn, and the matching parameters are adaptively adjusted to avoid stacking interference.

[0093] Collect users' fitting history, adjustment preferences, and evaluation feedback to build personal preference profiles;

[0094] By combining personal preference profiles, the processing scheme generated by template matching is fine-tuned to ensure that the matching parameters match user preferences;

[0095] Identify the current fitting scenario and select the appropriate template group accordingly, then optimize the personalized, fine-tuned processing solution in a contextualized manner.

[0096] Based on the contextualized and optimized processing scheme, multiple clothing try-on effects are generated;

[0097] By collecting user trial behavior data and effect evaluations, the template parameters and metadata in the clothing combination template library are dynamically updated.

[0098] Furthermore, the steps of receiving and parsing the abstract clothing intention description input by the user, and mapping the intention description to visual targets of body modification, style emotion, and scene fit include:

[0099] Receive text describing the user's intended clothing style submitted through a text input interface;

[0100] A pre-trained natural language processing model is used to perform word segmentation and semantic parsing on the descriptive text to extract semantic units from the descriptive text;

[0101] The semantic units are mapped and matched with a predefined visual target concept library, which stores quantitative parameter labels related to body modification, style and emotion, and scene fit.

[0102] Natural language processing (NLP) models typically employ Transformer-based architectures, such as BERT or GPT series. Their construction begins with large-scale corpus collection, drawing from publicly available text datasets and encompassing diverse texts including clothing descriptions, fashion reviews, and scene narratives. The corpus then undergoes preprocessing, including text cleaning, word segmentation, and vocabulary construction. Pre-training is performed through self-supervised learning tasks, such as masked language modeling, where parts of the input sentence are randomly masked, and the model is trained to predict the masked words based on context, thereby learning deep language representations and semantic relationships. After pre-training, the model possesses the ability to perform semantic understanding, entity recognition, and relation extraction from natural language.

[0103] The construction of a predefined visual target concept library is a knowledge engineering process. First, core concepts and attributes are extracted from structured knowledge sources in fields such as fashion design, ergonomics, and color psychology to form an initial concept set. These concepts are categorized into body shaping (e.g., "waist-cinching," "shoulder-expanding"), style and mood (e.g., "minimalist," "elegant"), and scenario (e.g., "business meeting," "beach vacation"). Each concept is assigned a unique identifier and bound to a set of quantifiable parameters or parameter ranges. For example, the concept of "waist-cinching" is associated with the quantifiable parameter of "waistline curve contraction intensity"; the "elegant" style may be mapped to a set of quantifiable values ​​across multiple dimensions such as "fabric gloss," "color saturation," and "contouring softness." Finally, the concept library is stored in a structured database or ontology, establishing a mapping table between natural language vocabulary and quantifiable visual parameters.

[0104] For body contouring, the mapping process generates quantified contouring instructions for specific body parts, including part identifiers and expected contouring levels. The mapping process first identifies contouring action keywords (e.g., "slimming," "upright") and body part keywords (e.g., "waist," "shoulder") from the parsed core semantic units. Then, it queries the visual target concept library for body contouring categories to find the predefined contouring instruction template that best matches these keywords. This template specifies the affected body part identifier (e.g., "waist_line") and a base contouring strength value. Combining this with degree adverbs in the descriptive text (e.g., "very," "slightly"), the base contouring strength value is adjusted using a preset adverb strength coefficient, ultimately generating a contouring instruction containing the specific body part and the final quantified strength value.

[0105] For style sentiment, a style vector describing clothing style attributes is generated through mapping. The coordinates of this vector in the multidimensional style space represent the intensity of a specific style sentiment. The mapping of style sentiment combines style-related words (such as "retro" and "casual") from the core semantic units. These style words are input into a style encoder, which has been trained on a massive amount of clothing image-text description pairs and can convert discrete style words into a fixed-dimensional continuous style vector. The coordinate values ​​of each dimension of this vector in the multidimensional style space represent the quantified intensity of the combined style on a certain abstract style attribute (such as "formalism," "complexity," and "livelyness"). This numerical style vector is then output.

[0106] For scene fit, a scene type identifier and a list of constraints related to scene fit are generated. For the scene fit mapping, named entity recognition or keyword matching techniques are used to extract scene entities (such as "office" or "wedding") from the descriptive text. Next, precise or fuzzy matching is performed in the "scene" category of the visual target concept library to obtain the corresponding scene type identifier (such as "SCENE_BUSINESS_FORMAL"). Simultaneously, a list of constraints pre-associated with this scene identifier is extracted. These constraints may include allowance or prohibition rules for clothing category, color range, degree of nudity, and fabric type. This identifier is output along with the constraint list.

[0107] The quantization modification instructions, style vectors, scene type identifiers, and constraint list are integrated and encapsulated into a structured visual target data object, completing the transformation process from abstract language description to structured quantized visual targets.

[0108] Furthermore, the steps of generating target multi-clothing combination schemes that satisfy all visual constraints, and outputting the optimized target try-on scheme, using the visual target as multiple constraints, include:

[0109] The visual target data object is parsed into multiple independent constraints, and a pre-built knowledge graph is loaded. The nodes of the knowledge graph represent clothing items, clothing attributes, matching rules or aesthetic tags, and the edges represent the relationships or association strength between the nodes.

[0110] Using the style vector and scene constraints in the visual target data as initial query conditions, a reverse reasoning process is initiated in the graph neural network. This process starts from potential clothing combination nodes that satisfy high-level constraints, searches backward along the relation edges of the knowledge graph, and gradually adds clothing nodes that satisfy low-level constraints such as body modification and material matching to form multiple initial clothing combination candidate schemes.

[0111] For each initial candidate solution, based on the historical aesthetic feedback data and association rules stored in the knowledge graph, a multi-objective optimization algorithm is used to calculate its satisfaction score on each visual constraint dimension, and the set of solutions with a comprehensive score exceeding the preset score threshold is selected as the target multi-clothing combination solution.

[0112] Satisfaction rating The calculation formula is:

[0113] ;

[0114] In the formula, The overall satisfaction score of the solution is given by a dimensionless scalar value. This represents the total number of visual constraint dimensions. For the first One visual constraint dimension For the first The weight coefficients of each dimension satisfy... This weight is derived from the importance distribution of each dimension learned from historical feedback data in the knowledge graph. For the current candidate clothing combination schemes in the 1st The actual attribute values ​​in each dimension are obtained by querying the attribute edges connected to the clothing combination node in the knowledge graph, or by aggregating the attribute values ​​of individual items within the combination (such as average or maximum / minimum values). The first one obtained from visual target analysis The target value or target value range for each dimension. For the first A satisfaction rating function for each dimension, whose output is a dimensionless rating between 0 and 1. The specific form of the function depends on the dimension type:

[0115] For those with clear target values The dimensions (such as "waist contraction strength") are typically calculated using the following function:

[0116] ;

[0117] in This is the length of the reasonable value range for this dimension attribute, used for normalization;

[0118] For those with a target range The dimensions (such as the color saturation range) can be calculated using the following function:

[0119] ;

[0120] For Boolean constraints (such as "must include a suit jacket"), It is either 1 (satisfied) or 0 (not satisfied).

[0121] For each target multi-clothing combination scheme, a physical simulation is performed. The three-dimensional model of each garment in the scheme is initially aligned with the user's three-dimensional human body model. Then, dynamic deformation calculation based on physical rules is initiated to simulate the shape and layering relationship of the clothing under gravity, human body collision and interaction between clothing.

[0122] The initial alignment step begins with feature point matching. A predefined set of corresponding anatomical or structural feature points (e.g., the shoulder and neck points on the human body corresponding to the midpoint of the clothing collar) are extracted from the mesh vertices of both the human and clothing 3D models. Next, a least-squares rigid transformation algorithm is used to derive an optimal rotation and translation matrix, transforming the clothing model to the vicinity of the human body model. This minimizes the spatial distance error between these corresponding feature points, completing the initial localization.

[0123] Then, dynamic deformation calculations based on physical rules are initiated. The triangular mesh of the clothing 3D model is discretized into a mass-spring system, where each mesh vertex is considered a mass and the mesh edges are considered springs with tensile and bending stiffness. Simultaneously, the surface of the human 3D model is defined as an impenetrable collision body. During the simulation, the system performs the following calculations at each time step: First, the internal forces of the springs between the mass points are calculated according to Hooke's Law; second, a vertically downward gravitational force is applied; then, the distance between the clothing mass points and the human collision body is detected, and if it is less than a preset distance threshold, a repulsive force is calculated and applied to prevent penetration; if it is a multi-layered garment, the collision forces between the garment layers are simultaneously detected and calculated. Next, based on the net force acting on the mass points, the velocity and position of the mass points are updated using numerical integration methods (such as explicit Euler's method or Verlet integral). This process iterates for multiple time steps until the overall kinetic energy of the system stabilizes or the preset maximum number of simulation steps is reached. At this point, the clothing model reaches an equilibrium state that fits the human body perfectly and has no penetration between layers, under the physical rules.

[0124] In this process, the joint optimization algorithm simultaneously uses the satisfaction of visual objectives and the stability of physical simulation as optimization objectives, and adjusts the constituent items of the clothing combination scheme, the size scaling factor of the items, and the stacking order of multiple garments through multiple rounds of iteration.

[0125] After each iteration, the visual performance and physical rationality of the solution are re-evaluated until the preset iteration termination condition is met, and finally the optimized target fitting solution that meets both visual and physical constraints is output.

[0126] The preset iteration termination condition is usually composed of a set of logical judgments. When any condition is met, the iterative optimization process stops.

[0127] The first condition is reaching the maximum number of iterations. The system sets a positive integer N as the upper limit for iterations before optimization begins. After each iteration, a counter is incremented by 1. When the counter value equals N, the iteration is forcibly terminated regardless of the quality of the current solution.

[0128] The second condition is that the scheme performance converges. The system records the overall satisfaction score of the optimized target try-on scheme generated by a series of consecutive iterations (e.g., M times). Calculate the variance of these M ratings. If the variance is less than a pre-defined, very small positive threshold close to zero... If the performance of the solution has stabilized and there is no longer any significant improvement, the iteration will terminate.

[0129] The third condition is the satisfaction of key constraints. The system sets a Boolean "hard" satisfaction flag for the core constraints in the visual objective (such as the clothing category that must be met or the style conflict that must not occur). After each iteration, it checks whether the current solution satisfies all hard constraints. Once all hard constraints are satisfied, and the overall score is... An acceptable baseline level has been reached. If the iteration is terminated prematurely, then the iteration will be terminated early.

[0130] Furthermore, the steps to build a clothing combination template library and form a core template set include:

[0131] The system extracts clothing type combination data from the clothing matching expert knowledge base, layering order rule data from the fashion rule database, matching priority data from the dressing specification archive, and deformation parameter data from the deformation parameter set. Among these, clothing type combination data specifies the combination method of top and bottom garments, layering order rule data specifies the order in which underwear is layered under outer garments, matching priority data sorts the matching of inner garments before matching of outer garments, and deformation parameter data defines the range of fabric stretch coefficients.

[0132] The clothing styling expert knowledge base extracts clothing type combination rules by collecting professional literature, fashion magazines, and designer interviews in the field of clothing design and using natural language processing technology. For example, it parses matching patterns such as "shirt paired with suit pants" from text and stores them as a structured data table, where each record contains a clothing type identifier and a combination rule string.

[0133] The fashion rules database extracts layering order rules by analyzing historical fashion trend reports and social media data using data mining algorithms. For example, it identifies clothing layering relationships (such as underwear under outerwear) from image data and encodes them as rule sequences, storing them in the database.

[0134] The dress code archive collects dress codes through questionnaires and industry standard documents, and uses a rule engine to define matching priorities. For example, it extracts rules such as "prioritize matching inner garments" from corporate dress codes and converts them into a list of numerical priorities.

[0135] Deformation parameter set: Deformation parameters are obtained by measuring the physical properties of clothing fabrics in the laboratory and using a tensile testing machine. For example, tensile tests are performed on cotton fabric samples, and the maximum tensile coefficient is recorded as 0.2 and the minimum as 0.1, forming a parameter threshold range, which is then stored in the parameter library.

[0136] The extracted clothing type combination data, stacking order rule data, matching priority data, and deformation parameter data are integrated into a core template set. Each template stores clothing type combination identifiers, stacking order rule sequences, matching priority value lists, and deformation parameter threshold ranges in a structured format.

[0137] A hierarchical storage structure is adopted, with template subsets arranged in descending order of usage frequency, template subsets categorized by applicable scenarios, and template subsets filtered by user rating score thresholds.

[0138] The template set subset refers to the subset of the core template set divided according to the classification criteria. The subset includes template groups arranged in descending order of usage frequency (such as high-frequency use templates), template groups classified by applicable scenarios (such as leisure scenario templates), and template groups filtered by user rating score thresholds (such as templates with a rating higher than 4.0).

[0139] Generate a metadata record for each template, which includes the applicable body size range boundary values, seasonal feature codes, and occasion label strings;

[0140] The template retrieval process is achieved by linking the core template set with metadata records through a database indexing mechanism.

[0141] Initialize the template library storage space and ensure that all data is persistently saved in JSON format;

[0142] Verify the integrity of the integrated data by checking and calculating to confirm that there are no missing fields.

[0143] Verifying the integrity of the integrated data involves calculating a checksum using a cyclic redundancy check algorithm. The specific process is as follows:

[0144] Calculate the checksum: Take the entire data block (including all field values) as the input byte stream and apply the CRC32 algorithm to generate a 32-bit checksum value. The CRC32 algorithm has a polynomial of 0xEDB88320 and an initial value of 0xFFFFFFFF. Each byte is XORed and shifted.

[0145] Parameter retrieval: Data blocks are read from a stored JSON format file and converted into byte sequences.

[0146] Suppose a data block contains the field value "shirt_pants"["underwear", "outerwear"][1,2][0.1,0.2], which, after conversion to a byte sequence, is 100 bytes long. The calculated CRC32 checksum is 0x5A5A5A5A. During verification, the checksum is recalculated. If the results match (both are 0x5A5A5A5A), the data is complete; otherwise, it indicates a missing field.

[0147] Furthermore, the steps of matching templates from the clothing combination template library to generate a processing solution suitable for the current combination include:

[0148] Read the user-selected multi-person clothing combination data, which includes a list of clothing type identifiers;

[0149] The system queries the clothing combination template library, performs precise matching at different levels, and achieves a complete string match by comparing the user's clothing type combination identifier with the template's clothing type combination identifier.

[0150] The user clothing type combination identifier is a list of clothing type identifiers from the data of multiple clothing combinations selected by the user. For example, if the user selected "shirt" and "pants", the identifier might be ["shirt", "pants"]. This list is used to compare with the identifiers in the template.

[0151] When an exact match fails, a similarity matching hierarchy is initiated to obtain the cosine similarity score between the user's clothing type combination identifier and the template's clothing type combination identifier, and the template with the highest similarity score is selected as the candidate template.

[0152] The formula for calculating the cosine similarity score is:

[0153] ;

[0154] In the formula, It is the i-th component of the user clothing type combination vector. It is the i-th component of the template clothing type combination vector, where n is the vector dimension.

[0155] When all similarity matching scores are below the preset similarity threshold, rule derivation hierarchy is performed to extract the stacked order rule sequence and the matching priority value list from the core template set;

[0156] A new stacked order rule sequence is generated by combining logical reasoning algorithms with extracted rule sequences;

[0157] The new processing scheme is formed by merging the new stacking order rule sequence, the user clothing type combination identifier, and the deformation parameter threshold range, and then outputting the scheme, which includes the stacking order rule sequence, the matching priority value list, and the deformation parameter threshold range.

[0158] Furthermore, following the stacking order rules defined in the template matching processing scheme, the steps for matching each garment with the 3D human body model sequentially include:

[0159] Extract the stacking order rule sequence from the processing scheme generated according to template matching, which specifies the order of clothing matching processing;

[0160] Starting from the first element of the stacking order sequence, select the first garment data, including the garment mesh model and size parameters;

[0161] The selected clothing mesh model is mapped onto the surface of the human body 3D model, and the position of the clothing mesh model is adjusted using a vertex coordinate transformation algorithm;

[0162] Collision detection between clothing mesh model and human 3D model is calculated, and bounding box algorithm is used to identify overlapping areas;

[0163] Collision detection uses a bounding box overlap determination formula:

[0164] For each axis (x, y, z), check the following conditions: and , where A and B are the bounding boxes of the clothing mesh and the human body mesh, respectively.

[0165] In the formula, and These are the minimum and maximum coordinates of the clothing bounding box, extracted from the grid vertex coordinates, for example, by traversing the vertices to calculate the minimum and maximum values ​​for each axis. and These are the minimum and maximum coordinates of the human bounding box, extracted from the grid vertex coordinates, for example, by traversing the vertices to calculate the minimum and maximum values ​​for each axis.

[0166] For example, if the x-axis range of the clothing bounding box is [1,3] and the x-axis range of the human body bounding box is [2,4], then the condition 1≤4 and 2≤3 holds, indicating overlap.

[0167] Adaptively adjust matching parameters to avoid overlapping areas penetrating by scaling the size parameters of the clothing mesh model;

[0168] The adaptive adjustment of matching parameters uses a scaling algorithm: based on the collision detection results, the penetration depth is determined, and the clothing mesh size parameters are adjusted proportionally. For example, if penetration is detected, the scaling factor of the clothing mesh is multiplied by 0.9 (reduced by 10%), and the detection is repeated until there is no penetration.

[0169] Repeat the above selection, mapping, collision detection and adjustment process to process the next garment data according to the stacking order rule sequence;

[0170] When processing multiple garments, the matching results of the previous garment are accumulated as the matching base of the current garment to ensure that the outer garment mesh model covers the inner garment mesh model.

[0171] A complete matching model is generated, which integrates all clothing mesh models and human 3D models. The effectiveness of matching parameter adjustments is verified, and the absence of overlapping interference areas is confirmed through distance calculation.

[0172] The distance calculation formula is:

[0173] ;

[0174] in, and These are the vertex coordinates of the two grids.

[0175] For example, given vertex A(0,0,0) and vertex B(1,1,1), the distance = = ≈1.732. If the threshold is set to 0.5, then d>0.5 indicates no interference.

[0176] Furthermore, the steps to obtain users' fitting history, adjust preferences, and provide feedback to build a personal preference profile include:

[0177] Extract fitting history data from user interaction logs. This data records the identifiers and matching parameter values ​​of clothing combinations tried on in the past. At the same time, parse adjustment preference data, which specifies the user's modification of matching parameters, such as the adjustment range of tightness, and collect evaluation feedback data, which includes the user's score assignment and text description of the fitting effect.

[0178] The historical data of fitting, the data of adjusting preferences and the data of evaluation feedback are classified and stored. The historical data of fitting is sorted by timestamp, the data of adjusting preferences is grouped by parameter type, and the data of evaluation feedback is stratified by score level.

[0179] Obtain preference statistics, aggregate the average adjustment magnitude from the preference adjustment data, and calculate the weighted average score from the evaluation feedback data;

[0180] Preference statistics include the average adjustment magnitude and the weighted average score:

[0181] Average adjustment range: ;

[0182] in Let n be the magnitude of the i-th adjustment, and n be the number of adjustments.

[0183] Weighted average score: ;

[0184] in Let i be the score for the i-th evaluation. As weights (such as time weights).

[0185] Construct a personal preference profile structure, which organizes the history of try-on data, adjusts the preference data mapping, and evaluate feedback data vectors in XML format.

[0186] Furthermore, the steps for personalizing the template matching processing scheme by incorporating individual preference profiles include:

[0187] Read the processing scheme generated by template matching, and parse the stacking order rules, matching priority settings and deformation parameter values;

[0188] Extract user preference parameters from personal preference profiles, including preferred values ​​for layer spacing, clothing fit requirements, and acceptable range for deformation effects;

[0189] The parameters in the processing scheme generated by template matching are compared and analyzed with the user preference parameters to identify the parameter items that need to be adjusted;

[0190] The steps to identify the parameters that need adjustment include:

[0191] The comparative analysis uses a threshold comparison method to determine the absolute difference between the template parameters and the user preference parameters. If the difference exceeds a preset threshold (e.g., the difference in layer spacing > 0.1), the parameter is marked as an item that needs adjustment.

[0192] Based on the comparative analysis results, the processing scheme was adjusted item by item, including modifying the spacing settings in the stacking order rules, adjusting the accuracy requirements in the matching priority, and optimizing the deformation coefficient in the deformation parameters.

[0193] During the adjustment process, the overall structure of the processing solution remains unchanged, and only the specific parameter values ​​are optimized individually.

[0194] Establish an adjustment constraint mechanism to ensure that the adjusted parameters are within a reasonable range and will not lead to matching failures or display anomalies. Perform a consistency check on the adjusted processing scheme to ensure the coordination and compatibility between the parameters.

[0195] The fine-tuned processing solution is stored as a new solution version, associated with the corresponding user ID and preference profile version number.

[0196] Furthermore, the steps for identifying the current fitting scenario and optimizing the personalized, fine-tuned processing plan in a contextualized manner include:

[0197] By analyzing the scenario description information in user requests, the types of try-on scenarios can be identified, including everyday casual wear, business formal wear, sports and fitness, and social occasions.

[0198] Based on the identified scene type, select a suitable template group from the template library. The template group contains multiple clothing matching templates suitable for that scene.

[0199] Read the personalized, fine-tuned processing solution and analyze its suitability for the current scenario;

[0200] The steps for analyzing the degree of fit include:

[0201] The method extracts features of the current scenario from a scenario rule base using rule matching (e.g., "business formal attire" requires a formality score > 0.8), and then calculates the matching degree between the clothing features and scenario features in the processing solution. For example, it checks whether the solution contains the "suit" type, and uses binary logic (1 indicates a match, 0 indicates a mismatch) to calculate the overall matching ratio; the higher the ratio, the better the fit. In existing technologies, this is achieved by querying the scenario rule table and performing simple arithmetic operations.

[0202] The processing scheme is optimized based on the characteristics of the scene, including adjusting the tightness of the stacking order, modifying the matching accuracy requirements, and optimizing the deformation parameter settings.

[0203] For formal occasions, the requirements for layering regularity are increased, and the matching accuracy standard is improved.

[0204] For sports scenarios, increase the spacing between layers to allow for more movement in clothing;

[0205] Establish a scenario-based parameter adjustment rule library and define parameter correction rules for each scenario type;

[0206] Record the scene optimization process, including scene recognition results, template group identifiers used, and parameter optimization content;

[0207] The final scenario-based optimization solution is generated, which includes both personalized fine-tuning features and reflects scenario-based requirements.

[0208] like Figure 2 The diagram shown is a structural schematic of a clothing try-on system based on a three-dimensional human body model provided in this application embodiment. The clothing try-on system based on a three-dimensional human body model provided in this application embodiment includes: an intent mapping module, a solution output module, a data acquisition module, a template construction module, a template matching module, a matching processing module, a personal profile creation module, a profile fine-tuning module, a scene optimization module, and an effect generation module.

[0209] The intent mapping module is used to receive and parse the abstract clothing intent description input by the user, and map the intent description into visual targets of body modification, style emotion and scene fit.

[0210] The solution output module is used to generate a target multi-clothing combination solution that satisfies all visual constraints, using the visual target as multiple constraints, and output the optimized target try-on solution.

[0211] The data acquisition module is used to acquire the multi-person clothing combination and human body three-dimensional model data selected by the user;

[0212] The template building module is used to build a clothing combination template library to form a core template set;

[0213] The template matching module is used to perform template matching from the clothing combination template library based on the user-selected multi-person clothing combination, and generate a processing solution suitable for the current combination.

[0214] The matching processing module is used to process the matching of each garment with the human body 3D model in sequence according to the stacking order rules defined in the processing scheme generated by template matching;

[0215] The personal profile creation module is used to obtain the user's fitting history, adjustment preferences and evaluation feedback, and to create a personal preference profile;

[0216] The file fine-tuning module is used to personalize the processing scheme generated by template matching by combining personal preference files;

[0217] The scenario optimization module is used to identify the current fitting scenario and perform contextual optimization on the personalized fine-tuning processing solution.

[0218] The effect generation module is used to generate multiple clothing try-on effects based on the contextualized optimized processing scheme.

[0219] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0220] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0221] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0222] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0223] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0224] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of trying on clothes based on a three-dimensional model of a human body, characterized by, The method comprises the following steps: receiving and analyzing the abstract dressing intention description input by the user, mapping the intention description into visual targets of body modification, style mood and scene fit; generating a target multi-outfit combination scheme satisfying all visual constraints with the visual targets as multiple constraints, and outputting an optimized target try-on scheme; obtaining multi-person outfit combinations and human body three-dimensional model data selected by the user; constructing an outfit combination template library to form a core template set; based on the multi-person outfit combinations selected by the user, performing template matching in the outfit combination template library to generate a processing scheme suitable for the current combination; processing the matching of each outfit and the human body three-dimensional model in turn according to the layering order rules defined in the processing scheme generated by the template matching; obtaining the try-on history, adjustment preferences and evaluation feedback of the user to establish a personal preference profile; combining the personal preference profile, the processing scheme generated by the template matching is fine-tuned; identifying the current try-on scene, and optimizing the fine-tuned processing scheme; generating a multi-outfit try-on effect based on the scene-optimized processing scheme.

2. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of the user. The step of receiving and analyzing the abstract dressing intention description input by the user, mapping the intention description into visual targets of body modification, style mood and scene fit comprises: receiving the dressing intention description text submitted by the user through a text input interface; performing word segmentation processing and semantic analysis on the description text by using a pre-trained natural language processing model to extract semantic units in the description text; mapping the semantic units with a pre-defined visual target concept library; for body modification, mapping generates quantified modification instructions about the contour change direction of specific parts of the human body; for style mood, mapping generates a style vector describing the style attributes of the outfit; for scene fit, mapping generates a scene type identifier and a list of constraint conditions related to scene fit; integrating and packaging the quantified modification instructions, style vector, scene type identifier and constraint condition list into a visual target data object.

3. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of the user. The step of generating a target multi-outfit combination scheme satisfying all visual constraints with the visual targets as multiple constraints, and outputting an optimized target try-on scheme comprises: analyzing the visual target data object into multiple independent constraint conditions, and loading a pre-constructed knowledge graph; taking the style vector and scene constraints in the visual target data as initial query conditions to form multiple initial outfit combination candidate schemes; for each initial candidate scheme, calculating its satisfaction score in each visual constraint dimension, and selecting a scheme set with a comprehensive score exceeding a preset score threshold as the target multi-outfit combination scheme; performing physical simulation on each target multi-outfit combination scheme to simulate the shape and layering relationship of the outfit under the interaction of gravity, human body collision and outfits; outputting an optimized target try-on scheme satisfying both visual and physical constraints.

4. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of the user. The step of constructing an outfit combination template library to form a core template set comprises: extracting outfit type combination data, layering order rule data, matching priority data and deformation parameter data; integrating the extracted outfit type combination data, layering order rule data, matching priority data and deformation parameter data into the core template set; The hierarchical storage structure is adopted, the template set subsets are arranged in descending order of use frequency, the template set subsets are stored according to applicable scenarios, and the template set subsets are filtered according to user evaluation score thresholds; Metadata records are generated for each template; The core template set is linked with the metadata records, and a template retrieval process is implemented; The template library storage space is initialized, and the integrity of integrated data is verified.

5. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of the user. The steps of template matching from the garment combination template library to generate a processing scheme suitable for the current combination include: Reading user-selected multi-person garment combination data; Querying the garment combination template library for accurate matching levels; When accurate matching fails, starting a similarity matching level to obtain a cosine similarity score of the user garment type combination identifier and the template garment type combination identifier, and selecting the template with the highest similarity score as the candidate template; When all similarity matching scores are lower than a preset similarity threshold, performing a rule derivation level to extract a layering order rule sequence and a matching priority value list from the core template set; Generating a new layering order rule sequence through a logical reasoning algorithm combined with the extracted rule sequence; Merging the new layering order rule sequence, the user garment type combination identifier, and the deformation parameter threshold interval to form a new processing scheme, and outputting the new processing scheme.

6. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of a specific individual. The steps of processing each garment and the human body three-dimensional model in turn according to the layering order rules defined in the processing scheme generated by template matching include: Extracting a layering order rule sequence from the processing scheme generated by template matching; Starting from the first position of the layering order rule sequence, selecting the first garment data, including a garment mesh model and size parameters; Mapping the selected garment mesh model to the surface of the human body three-dimensional model and adjusting the position of the garment mesh model; Calculating the collision detection of the garment mesh model and the human body three-dimensional model to identify overlapping areas; When processing multiple garments, accumulating the matching results of the previous garments as the matching base for the current garment; Generating a complete matching model to verify the effectiveness of the matching parameter adjustment and confirming the non-overlapping interference area through distance calculation.

7. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of a specific individual. The steps of obtaining the user's try-on history, adjustment preferences, and evaluation feedback to establish a personal preference profile include: Extracting try-on history data, analyzing adjustment preference data, and collecting evaluation feedback data; Storing the try-on history data, adjustment preference data, and evaluation feedback data, sorting the try-on history data by timestamp, grouping the adjustment preference data by parameter type, and layering the evaluation feedback data by score level; Obtaining preference statistical indicators, aggregating the average adjustment amplitude from the adjustment preference data, and calculating the weighted average score from the evaluation feedback data; Building a personal preference profile structure.

8. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of a specific individual. The steps of combining the personal preference profile to individually fine-tune the processing scheme generated by template matching include: Reading the processing scheme generated by template matching, analyzing the layering order rules, matching priority settings, and deformation parameter values in the processing scheme; Extracting the user's preference parameters from the personal preference profile, including layering interval preference values, garment fit requirements, and deformation effect acceptance ranges; Comparing and analyzing the parameters in the processing scheme generated by template matching with the user's preference parameters to identify parameters that need to be adjusted; Adjusting the processing scheme item by item based on the comparison and analysis results; Establish an adjustment constraint mechanism, and perform consistency checking on the adjusted processing scheme; Store the fine-tuned processing scheme as a new scheme version, and associate the corresponding user identifier and preference profile version number.

9. The method of claim 1, wherein the three-dimensional model of the human body is a three-dimensional model of a human body of a specific individual. The steps of identifying the current fitting scene and performing scenario-based optimization on the personalized fine-tuned processing scheme include: Identify the fitting scene type by analyzing the scene description information in the user request; According to the identified scene type, select an adaptive template group from the template library, which contains multiple clothing combination templates suitable for the scene; Read the personalized fine-tuned processing scheme and analyze its adaptation to the current scene; Optimize the processing scheme based on the characteristics of the scene; Establish a scenario-based parameter adjustment rule library and define parameter correction rules for each scene type; Generate the final scenario-based optimized processing scheme.

10. A system for fitting clothes based on a three-dimensional model of a human body, for implementing the method for fitting clothes based on a three-dimensional model of a human body according to any one of claims 1-9, characterized in that, Comprise: Intention mapping module, scheme output module, data acquisition module, template construction module, template matching module, matching processing module, personal archive establishment module, archive fine-tuning module, scenario optimization module, effect generation module; The intention mapping module is used to receive and analyze the abstract dressing intention description input by the user, and map the intention description to the visual target of body modification, style emotion and scene fit; The scheme output module is used to generate a target multi-clothing combination scheme that satisfies all visual constraints based on the visual target as multiple constraint conditions, and output the optimized target fitting scheme; The data acquisition module is used to acquire multi-person clothing combination and human body three-dimensional model data selected by the user; The template construction module is used to construct a clothing combination template library and form a core template set; The template matching module is used to perform template matching from the clothing combination template library based on the multi-person clothing combination selected by the user, and generate a processing scheme suitable for the current combination; The matching processing module is used to process the matching of each clothing and the human body three-dimensional model in turn according to the layering order rules defined in the processing scheme generated by the template matching; The personal archive establishment module is used to acquire the fitting history, adjustment preferences and evaluation feedback of the user, and establish a personal preference archive; The archive fine-tuning module is used to fine-tune the processing scheme generated by the template matching in combination with the personal preference archive; The scenario optimization module is used to identify the current fitting scene and perform scenario-based optimization on the personalized fine-tuned processing scheme; The effect generation module is used to generate a multi-clothing fitting effect based on the scenario-based optimized processing scheme.

Images