Intelligent auxiliary garment pattern making method and system based on multi-modal model

Abstract

The application discloses an intelligent auxiliary garment pattern making method and system based on a multi-modal model, takes a garment design sketch as input, carries out semantic deconstruction on the sketch through a multi-modal visual language model, identifies a garment category and a multi-level structure component, and realizes component semantic standardized mapping in combination with a garment component knowledge base; on the basis, introduces a design attribute prediction mechanism and a garment category self-adaptive dynamic constraint, converts a modeling intention in the sketch into a calculable parameterized design attribute; further combines the design attribute with human body key size data, completes parameterized modeling of a garment pattern in a two-dimensional coordinate system, constructs a sewing topological relationship between patterns, carries out physical rationality checking and correction on a generated result, and finally outputs an editable pattern file conforming to an industrial pattern making software specification. The application realizes automatic auxiliary conversion from a design sketch to engineering available pattern data, and significantly improves garment pattern making efficiency, consistency and availability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of clothing design and artificial intelligence, and in particular relates to an intelligent assisted clothing pattern making method and system based on a multimodal model. Background Technology

[0002] Pattern making is a crucial technical link between garment design and industrial production. Its main task is to transform the designer's creative ideas into two-dimensional pattern data that meets ergonomic and production process requirements. In the current garment production system, this process heavily relies on experienced pattern makers. Typically, it requires understanding the design sketches and combining human body dimensions, pattern rules, and process requirements to refine and revise the garment structure multiple times. This process demands a high level of expertise, and the efficiency and quality of pattern making largely depend on accumulated personal experience.

[0003] With the development of digitalization and intelligentization in the apparel industry, computer-aided pattern making software has been widely used in industrial production, improving pattern making efficiency through parametric modeling and rule-driven methods. However, most existing pattern making software still relies on manual input of structural parameters and key points, and cannot directly understand the design intentions contained in the design sketches. The hand-drawn sketches provided by designers are usually highly abstract and subjective, often lacking clearly marked size information, dart structures, seam line positions, and curved surface contour definitions. Pattern makers still need to manually interpret and judge these sketches based on experience to convert them into production-ready pattern data, and the overall process has not changed substantially.

[0004] In recent years, artificial intelligence technology, especially generative and multimodal models based on deep learning, has made significant progress in image understanding and content generation. Some studies have attempted to introduce these technologies into garment pattern making, enabling the automatic generation of garment patterns from text descriptions or sketches. However, existing technologies primarily focus on generating images of garment appearance or rough pattern outlines. The outputs are typically bitmaps or non-editable geometric forms, making them difficult to use directly for subsequent pattern adjustments, grading, and process design. Furthermore, such generated results often lack constraints on the rationality of garment structure and human body fit, easily leading to problems such as unreasonable dimensions, unbalanced component proportions, or inconsistent stitching relationships, making it difficult to meet the precision and controllability requirements of actual production.

[0005] On the other hand, garment pattern making is essentially a highly rule-bound engineering activity. Different garment categories exhibit significant differences in structural composition, design attributes, and parameter value ranges. Existing general-purpose image recognition or generation models typically lack dedicated decoupling mechanisms for garment pattern making rules, making it difficult to stably output structural parameters that match the target garment type in complex design scenarios. Furthermore, current technologies lack a technical system that can effectively connect sketch understanding, structural parameterization, geometric construction, and industrial software adaptation. This results in the application of artificial intelligence in garment pattern making remaining largely at the visual display level, failing to truly integrate into the pattern-making production process.

[0006] In summary, existing garment pattern-making technologies still have significant shortcomings in terms of sketch semantic understanding, structural information integrity, pattern physical rationality, and the editability of generated results. They are unable to simultaneously address design intent expression, engineering constraints, and industrial usability, and also limit the possibility of developing the pattern-making process towards intelligence, standardization, and efficiency. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an intelligent assisted garment pattern making method and system based on a multimodal model, the specific technical solution of which is as follows:

[0008] A smart assisted garment pattern making method based on a multimodal model includes the following steps:

[0009] S1. Input the clothing design sketch and prompts into the multimodal visual language model to obtain the recognition results; the recognition results include the clothing category corresponding to the sketch and the clothing components it contains;

[0010] S2. Based on the recognition results, a retrieval and recall is performed in a pre-built clothing component knowledge base based on semantic similarity to obtain standardized component information corresponding to the identified clothing component, so as to generate a parameterized configuration of the clothing component.

[0011] S3. Perform design attribute prediction on the clothing design sketch to obtain design attribute parameters that reflect the clothing shape characteristics and structural proportions, and write them into the corresponding clothing component parameterization configuration.

[0012] S4. Combine the parameterized configuration of the garment components with key human body size data, and perform parameterized garment pattern generation in a two-dimensional coordinate system to construct the pattern geometry structure corresponding to each garment component.

[0013] S5. After completing the generation of the pattern geometry, construct the stitching topology between each garment pattern and perform physical rationality verification on the generated garment pattern. Based on the verification results, correct the relevant parameters.

[0014] S6. Convert the verified garment pattern data into a standardized file format supported by the pattern-making software and output it for use in garment pattern editing and production.

[0015] Furthermore, the prompts mentioned in S1 include system prompts and user prompts; the system prompts are used to set the task role of the multimodal visual language model, constrain the model output format, and limit the recognition range, so that the model only analyzes content related to garment pattern making; the user prompts are used to attach sketches and clarify task objectives, instructing the model to perform garment category recognition and garment component step-by-step parsing on the input sketches.

[0016] Furthermore, the clothing component knowledge base mentioned in S2 is used to store the structured description information of each clothing component, including the standard name of the component, the component level to which it belongs, the parent component identifier, the applicable clothing category range, and the parameter description information related to the component. The parameter description information is used to characterize the adjustable attributes of the component in the pattern making process. The adjustable attributes include length ratio, angle range, curvature characteristics, dart parameters, or mapping relationships related to key human body dimensions.

[0017] The clothing component knowledge base is also used to store semantic vectors corresponding to the structured description information of each clothing component. The semantic vectors are obtained by mapping the structured description information text of the clothing components through a semantic embedding model, and are used to reflect the similarity relationship between different clothing components at the semantic level.

[0018] Further, in S2: First, the text description of the recognition result is mapped to a corresponding semantic vector through a semantic embedding model. Then, a retrieval and recall operation is performed in the clothing component knowledge base based on the semantic vector similarity. The semantic vector of the recognition result is used as the query vector, and the vector similarity is used as the matching criterion to retrieve the standard component information that is semantically closest to it from the knowledge base. The retrieved standard component information is then subjected to structured post-processing to generate a preliminary component parameterized configuration. In the component parameterized configuration, each clothing component corresponds to a set of structured parameter description information. This parameter description information includes at least the name of the adjustable parameter item of the component, the explanation of the parameter meaning, and the initial value range of the parameter.

[0019] Furthermore, in S3, the design sketch and corresponding text prompts are input into the attribute predictor to obtain the corresponding design attribute parameter values. The attribute predictor includes a basic multimodal feature extraction module and an attribute prediction module. The basic multimodal feature extraction module is used to perform joint feature extraction on the input design sketch and corresponding text prompts to obtain multimodal features that can simultaneously represent visual structural information and semantic information. The attribute prediction module is used to map the multimodal features into parameter values ​​of each design attribute in the design attribute space and output them.

[0020] Furthermore, the attribute prediction module adopts a multilayer perceptron structure, and its output layer is used to output a prediction vector containing multiple design attribute parameter values.

[0021] The predicted vector is multiplied by the corresponding attribute mask matrix of the clothing category, retaining only the design attribute parameter values ​​related to that clothing category. Different clothing categories have corresponding attribute mask matrices, which are vectors with the same dimension as the predicted vector. Each vector element is used to indicate whether the corresponding design attribute is a valid attribute under the current clothing category. When a design attribute is not applicable to the current clothing category, its corresponding vector element value is set to a suppressed state, where 0 indicates suppression and 1 indicates retention.

[0022] Furthermore, the parameterized garment pattern generation described in S4 includes the following steps:

[0023] By combining parametric configuration of garment components with key human body dimensions, a baseline coordinate framework for the garment pattern is established in a two-dimensional plane coordinate system, and the key human body dimensions are standardized in terms of units and scale. Based on the standardized key human body dimensions, design attribute parameters, and preset pattern-making rules, the basic outline key points of the garment pattern are determined sequentially in the two-dimensional coordinate system. After determining the basic outline key points, the dart structure is generated based on the dart amount, dart depth, and dart inclination angle parameters, and the dart outline is constructed in a linear or curved form. The curved boundary of the garment pattern is modeled to generate a continuous pattern outline line. The outline data of each garment pattern is recorded in a structured manner to form pattern data containing vertex coordinate sequences, boundary connection relationships, and boundary type information.

[0024] Furthermore, S5 includes the following steps:

[0025] For each garment pattern piece, a sewing interface is set for the boundary segment or connection position involved in sewing. The sewing interface includes at least the interface name, interface type and corresponding geometric boundary information, and engineering parameters for describing the sewing adjustment rules are configured for the sewing interface.

[0026] Based on the preset garment structure rules, the sewing relationship between garment pieces is established with the sewing interface as the basic unit. The pairing relationship between corresponding interfaces is established for left and right symmetrical parts. For parts with different shapes but need to be connected to each other, the sewing rules are established based on the principle of geometric compatibility. The length analysis of the interface boundary involved in sewing is performed. When there is a difference in the length of the interface boundary, the difference is allocated and adjusted according to the preset rules or the interface boundary is corrected by curve reparameterization so that the interface involved in sewing remains consistent within the effective length range.

[0027] After obtaining the stitching rules between each pattern, the stitching rules are integrated into a stitching topology graph, wherein the stitching interface is used as the topology node and the stitching relationship between the interfaces is used as the topology connection edge, and a consistency check is performed on the stitching topology graph.

[0028] After completing the construction of the stitching topology, the generated garment pattern is subjected to physical rationality verification. The physical rationality verification includes at least key circumference dimension verification, garment component proportion verification, stitching interface boundary length matching degree verification, and pattern geometric anomaly detection. When the physical rationality verification result does not meet the preset conditions, the relevant pattern parameters are adjusted according to the preset correction rules, and the physical rationality verification is re-executed after adjustment until the verification result meets the requirements.

[0029] Furthermore, S6 includes the following steps:

[0030] The garment pattern data is parsed and processed. The garment pattern data includes at least the identification information of the garment pattern, vertex coordinate sequence, boundary connection relationship, boundary type information, curve control parameters, and stitching topology relationship between the patterns. The garment pattern data is then standardized using a uniform size unit.

[0031] The parsed and standardized 2D plate vertex coordinates are mapped to vertex elements in the target plate-making software file format, and each vertex is assigned a unique identifier to establish the correspondence between vertex coordinates and file structure.

[0032] The plate boundaries are transformed according to the boundary type. Line segment elements are generated for straight boundaries, and curve boundaries are reconstructed into curve or spline curve representations supported by the target plate-making software based on their parameter forms, so as to maintain the geometric consistency of the plate outline.

[0033] According to the structural specifications of the target pattern-making software file format, the vertex, boundary and curve elements of the garment pattern are organized to build a corresponding file structure unit for each garment pattern.

[0034] The stitching topology is mapped to the file structure, and the boundary segments of the pattern pieces involved in stitching and their corresponding relationships are recorded, so that the pattern-making software can identify the stitching relationship between different garment pattern pieces when loading the file.

[0035] An intelligent assisted garment pattern making system based on the above method, the system comprising:

[0036] The sketch input and preprocessing module is used to receive clothing design sketches and perform background separation and noise suppression processing on them to obtain normalized sketch images;

[0037] The multimodal semantic parsing module is used to input the normalized sketch image and prompts into the multimodal model, identify the clothing category corresponding to the sketch and the clothing parts contained in the sketch, and generate clothing part recognition results;

[0038] The clothing component knowledge management module is used to perform a retrieval and recall in a pre-built clothing component knowledge base based on semantic similarity for the clothing component identification results, obtain standardized component information corresponding to the identified clothing component, and generate parameterized configuration of the clothing component.

[0039] The design attribute prediction and constraint module is used to predict the design attribute parameters that reflect the shape features and structural proportions of the clothing based on the visual semantic features of the sketch, and write them into the corresponding parametric configuration of the clothing component.

[0040] The parametric pattern generation module is used to combine the parametric configuration of the garment components with key human body size data, perform parametric garment pattern generation in a two-dimensional coordinate system, and construct the pattern geometry structure corresponding to each garment component.

[0041] The stitching topology and verification module is used to construct the stitching topology relationship between each garment pattern after the pattern geometry is generated, and to perform physical rationality verification and correction on the generated garment patterns.

[0042] The pattern conversion and output module is used to convert the verified garment pattern data into a standardized file format supported by the pattern making software and output it for garment pattern editing and production.

[0043] Compared with the prior art, the present invention has at least the following beneficial effects:

[0044] This invention starts with clothing design sketches, uses a multimodal model to semantically deconstruct the sketches, and introduces a collaborative mechanism of clothing component knowledge base, design attribute prediction, and parametric pattern-making rules. This bridges the long-standing technical gap between understanding design intent and generating engineered patterns, effectively changing the traditional pattern-making process's reliance on human experience, low efficiency, and poor consistency. Unlike existing intelligent generation technologies that can only generate images of clothing appearance or uneditable geometric contours, this invention, through layered component recognition, semantic similarity retrieval, and parametric configuration, systematically and standardizedly transforms the structural information implicit in the sketches into calculable pattern-making parameters. This fundamentally solves the problems of abstract, incomplete, and unsuitable-for-direct-use industrial pattern-making information in sketches.

[0045] Furthermore, this invention introduces adaptive dynamic mask constraints for clothing categories during the design attribute prediction stage. This ensures that only relevant design attributes from different clothing categories participate in the inference calculation, thereby avoiding the risk of irrelevant or conflicting parameters generated by the general model in complex pattern-making scenarios and significantly improving the stability and rationality of pattern generation. Simultaneously, by mathematically modeling design attributes and key human body size data in a two-dimensional coordinate system, and combining this with precise construction rules for darts, curves, and contours, the generated pattern possesses clear constraints in both geometric structure and human body adaptability, avoiding common problems in existing generative methods such as size imbalance, proportion discrepancy, and structural unavailability.

[0046] Furthermore, this invention explicitly models the assembly relationships between pattern pieces by constructing stitching interfaces, stitching rules, and a global stitching topology diagram. It also introduces a systematic physical rationality verification and parameter correction mechanism during the generation stage, ensuring that the pattern pieces meet actual production requirements in terms of circumference, proportion, and stitching matching. This guarantees that the generated results are not only visually appealing but also usable. Finally, by converting the parameterized pattern piece data into a standardized file format conforming to industrial pattern-making software specifications, seamless integration with existing pattern-making software environments is achieved, allowing pattern makers to further edit and refine the initial pattern pieces using familiar tools. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0048] Figure 1 This is a flowchart illustrating the intelligent assisted garment pattern making method provided in an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the attribute predictor provided in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the framework of the intelligent assisted garment pattern making system provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0052] Example 1

[0053] This embodiment provides an intelligent assisted garment pattern making method based on a multimodal model, which is used to convert garment design sketches into physically plausible and editable two-dimensional garment pattern data to assist pattern makers in completing the pattern making work.

[0054] like Figure 1 As shown, the method mainly includes the following steps:

[0055] I. Building a Knowledge Base for Clothing Components

[0056] To achieve standardized identification and subsequent parametric processing of component information in garment design sketches, it is necessary to pre-build a garment component knowledge base to store component structural information related to garment pattern making, and to serve as an intermediate bridge between sketch semantic recognition results and pattern making parameter system.

[0057] In this embodiment, the clothing components in the clothing component knowledge base are organized according to a preset hierarchical system, including at least first-level components and multi-level sub-components corresponding to the first-level components. First-level components describe the main structural units of the clothing, including at least the body, sleeves, collar, pockets, front placket, skirt panels, and trouser panels. Second- or third-level sub-components describe the detailed structures under each first-level component, such as the sleeve cap, cuff, large sleeve panel, and small sleeve panel under the sleeve component; the collar depth, collar width, collar stand, and lapel panel under the collar component; and the pocket opening, pocket flap, and pocket body under the pocket component.

[0058] During the actual construction process, for each garment component, its corresponding structured description information is pre-organized. This structured description information includes at least: the component's standard name, its component hierarchy, its parent component identifier, the applicable range of garment categories, and parameter description information related to that component. The parameter description information is used to characterize the adjustable properties of the component during the pattern-making process. These properties may include length proportions, angle ranges, curvature characteristics, dart-related parameters, or mapping relationships related to key human body dimensions.

[0059] The structured descriptions of the aforementioned garment components are organized in a unified data format, such as structured records in key-value pair (JSON) form. Different fields correspond to the component name, hierarchy label, parameter name, and parameter meaning description. This approach ensures that garment components from different sources and with different naming conventions have a unified standard representation in the knowledge base.

[0060] After organizing the structured descriptions of clothing components, semantic vectorization is performed. Specifically, a pre-selected semantic embedding model (such as Qwen3-Embedding-4B) is used to map the structured description text of each clothing component into a corresponding high-dimensional semantic vector, which represents the component's position in the semantic space. These semantic vectors reflect the semantic similarity relationships between different clothing components, enabling matching based on semantic distance even if component names differ.

[0061] Subsequently, the generated high-dimensional semantic vectors and the corresponding structured descriptions of clothing components are stored in a vector database, thus forming a clothing component knowledge base. This vector database supports semantic similarity-based retrieval and can limit the retrieval scope by combining component-level tags, achieving efficient and accurate component information retrieval.

[0062] II. Identifying Garment Components in Design Sketches

[0063] To effectively understand and analyze the garment structure information contained in garment design sketches, a prompt is constructed for input sketches, along with the sketch itself. Figure 1 The data is then input into a multimodal visual language model for sketch semantic deconstruction processing, thereby obtaining structured and hierarchical clothing component recognition results.

[0064] Specifically, this embodiment constructs different prompts for single-view and multi-view sketches to guide the multimodal visual language model to focus on the overall structure of the clothing and key component features in the sketches during inference. A single-view sketch contains at least frontal structural information of the clothing, while a multi-view sketch includes at least a front view and a back view. Multi-view sketches are input into the multimodal visual language model simultaneously as multiple images, while single-view sketches only require a single image.

[0065] The prompts include at least system prompts and user prompts. The system prompts define the task role of the multimodal visual language model, constrain the model's output format, and limit the recognition scope, ensuring the model analyzes only content related to garment pattern making. The user prompts include a sketch and clarify the current task objective, instructing the model to perform garment category recognition and step-by-step analysis of garment components from the input sketch. For multi-view sketches, the user prompts further specify the viewpoint type corresponding to each input image, enabling the model to integrate structural feature information from different viewpoints during analysis.

[0066] For example, here is a prompt that includes a two-view sketch:

[0067] [{"role": "system", "content":"content":"Role setting and overall requirements"},{"role": "user", "content": [{"type": "text", "text": detailed description of specific task requirements}, {"type": "image_url", "image_url": {"url": f"data:image / jpeg;base64,{base64_image1}"}}, {"type": "image_url", "image_url": {"url": f"data:image / jpeg;base64,{base64_image2}"}}]}].

[0068] After constructing the prompts, the sketch image and corresponding prompts are input into the multimodal visual language model for sketch semantic deconstruction processing. The sketch can be a hand-drawn scanned image, a digital tablet drawing, or an image file exported from design software. To ensure the accuracy and stability of the multimodal visual language model's recognition of clothing structural features, preprocessing operations such as background separation and noise suppression are performed on the acquired sketch image before inputting it into the model. This eliminates background interference and irrelevant details to highlight the outline and structural features of the clothing, providing high-quality, structurally clear, and standardized input data for the model's feature extraction.

[0069] The multimodal visual language model first performs overall semantic analysis on the input sketch to identify the clothing category corresponding to the sketch, thereby determining the clothing type to which the sketch belongs, such as T-shirt, dress, suit, etc. After completing the clothing category identification, the identified clothing category is used as a priori condition to guide the multimodal visual language model to perform component-level analysis and identification of the clothing structure in the sketch. Specifically, the model first performs core component identification, identifying whether the sketch contains major structural components of clothing such as the body, sleeves, collar, pockets, and front placket. Then, for each identified core component, it further completes the in-depth identification of sub-components. For example, after identifying the sleeve component, it further identifies sub-components such as sleeve cap, cuff, large sleeve piece, and small sleeve piece; after identifying the collar component, it further identifies sub-components such as collar depth, collar width, lapel piece, and collar stand. Finally, the model outputs the identification results in structured data form (JSON), which includes at least the clothing category identifier, the core components, and a list of their corresponding sub-components. Each garment component includes its name, hierarchical information, and its relationship within the garment structure.

[0070] In the aforementioned recognition process, the multimodal visual language model comprehensively judges the garment structure based on the line distribution, contour shape, and structural connection relationships in the sketch. For multi-view sketches, the model comprehensively utilizes complementary information from different perspectives when recognizing components to compensate for potential structural occlusion or information loss in a single perspective, thereby improving the completeness of component recognition. In this embodiment, the multimodal visual language model selected is a visual language model with image and text joint reasoning capabilities, such as Qwen-VL-Max. This model supports joint reasoning with multiple image inputs and text commands, enabling a comprehensive understanding of the overall garment structure, local components, and their hierarchical relationships in the sketch. Considering the differences in line expression, symbolic darts / dividing lines, variations in pattern makers' drawing methods, and industry standardization requirements for component naming systems compared to general visual corpora, existing visual language models can be lightweightly transferred and adapted before application to obtain stable and consistent component hierarchy output and a low false recognition rate.

[0071] III. Component Information Recall Based on Semantic Similarity

[0072] To transform the component recognition results output by the multimodal visual language model into standardized structural data that can be used for subsequent pattern making and modeling, it is necessary to obtain information such as corresponding component parameter descriptions from a pre-built garment component knowledge base.

[0073] Since the component labels and other information in the recognition results are generated by the model based on joint visual and linguistic reasoning, their naming methods may differ in expression from the predefined standard component names in the clothing component knowledge base. Therefore, accurate mapping and matching cannot be achieved directly through string matching. To address this, this embodiment first performs semantic vectorization processing on the recognition results, that is, it uses a semantic embedding model (Qwen3-Embedding-4B) to map the text description of the recognition results into corresponding high-dimensional semantic vectors. Then, based on the semantic vector similarity, a retrieval and recall operation is performed in the clothing component knowledge base. That is, the semantic vector of the recognition results is used as the query vector, and the vector similarity is used as the matching criterion to retrieve the standard component information that is semantically closest to it from the knowledge base.

[0074] To improve retrieval accuracy and reduce false recall, this embodiment introduces a tag-level filtering mechanism during the retrieval process. Specifically, based on the component hierarchy information output by the multimodal model, the parent component category to which the current component belongs is first determined, and the retrieval scope is limited to the candidate component set corresponding to that parent tag. For example, when the currently identified component belongs to the parent component "pocket," semantic similarity retrieval is performed only in the set of child components belonging to the "pocket" category, without matching in other irrelevant component sets, thereby avoiding false matching between different component categories. When a standard component that highly matches the semantics of the current component tag exists within the limited retrieval scope, the standard component with the highest semantic similarity is selected as the recall result; when no standard component with a completely matching semantics does not exist under the corresponding parent tag, the component information with the highest semantic similarity is returned as a substitute component to ensure the completeness of component configuration.

[0075] After retrieving and recalling all identified components one by one, the recalled standard component information is post-processed in a structured manner to generate a preliminary component parameter configuration. In the component parameter configuration, each garment component corresponds to a set of structured parameter description information, which includes at least the name of the adjustable parameter item of the component, a description of the parameter meaning, and the initial value range of the parameter.

[0076] IV. The design predictor adjusts the design attributes of components.

[0077] After obtaining the structured parametric configuration of garment components, in order to further extract the design intent from the design sketch and refine and fit the garment pattern outline, a design attribute prediction step is performed to generate design attribute parameters related to the appearance and structure of the garment, and the design attribute parameters are written into the corresponding component parameter configuration.

[0078] Design attributes refer to parametric information that reflects the shape characteristics and structural tendencies of clothing. They differ from direct absolute size data, primarily existing in the form of relative proportions, angular relationships, and morphological parameters. Examples include the proportional relationship between the length of each component and its corresponding human body dimensions (e.g., a sleeve length of 0.3 refers to 0.3 * arm length), the angle values ​​of structural lines (e.g., a collar of 95 refers to a collar angle of 95 degrees, i.e., a standard V-neck), the degree of opening and closing of local structures, and geometric control parameters affecting the silhouette. Through these design attributes, the design style depicted in the sketch can be parametrically expressed while maintaining the overall proportional rationality of the garment.

[0079] To achieve the aforementioned design attribute prediction, this embodiment constructs a design attribute predictor, which consists of a basic multimodal feature extraction module and an attribute prediction module. The basic multimodal feature extraction module performs joint feature extraction on the input sketch image and its corresponding text description (used to describe semantic dimensions related to design attribute prediction, such as prompting the model to pay attention to the overall proportions of the clothing, the degree of opening and closing of local structures, line direction, or angular relationships, thereby guiding the model to strengthen the semantic expression related to design attributes in the joint feature space) to obtain a multimodal feature representation that can simultaneously characterize visual structural information and semantic information. The attribute prediction module maps the multimodal features to numerical outputs in a predefined design attribute space. Specifically, as shown... Figure 2 As shown, the basic multimodal feature extraction module in this embodiment adopts Qwen2.5-VL-3B-Instruct, while the attribute prediction module adopts a multilayer perceptron (MLP) structure. It performs layer-by-layer nonlinear transformation on the input features and completes the mapping from the multimodal feature space to the design attribute space in the output layer, thereby generating a prediction result vector containing multiple design attributes.

[0080] To ensure that the structural differences between different clothing categories are reasonably reflected and to avoid irrelevant attributes from interfering with the prediction results, this embodiment introduces a clothing category adaptive dynamic masking mechanism in the design attribute prediction process.

[0081] Specifically, attribute mask matrices are pre-constructed for different clothing categories. These matrix vectors have the same dimension as the predicted design attribute results (e.g., 1*126, corresponding to the 126 selected general clothing component attributes). Each element indicates whether the corresponding design attribute is valid within the current clothing category. When a design attribute is not applicable to the current clothing category, its corresponding mask value is set to a suppressed state; a mask of 0 indicates suppression, and 1 indicates retention. During the predictor inference phase, based on the clothing category identified in the previous steps, an attribute mask matrix corresponding to that category is selected. This mask matrix is ​​then applied to the output of the attribute prediction module. That is, by performing element-wise multiplication on the predicted output vector and the attribute mask matrix, design attributes unrelated to the current clothing category are suppressed, and only the predicted values ​​of design attributes relevant to that category are retained, thus obtaining the category-constrained design attribute prediction results.

[0082] To enable the design attribute predictor to better adapt to the mapping relationship between garment sketch semantics and pattern-making parameters, transfer adaptation training can be performed on the predictor. While keeping the main parameters of the basic multimodal feature extraction module frozen, the attribute prediction module and a small number of related trainable parameters are fine-tuned to make the model output more consistent with the distribution characteristics of design attributes in the garment pattern-making field. In one feasible implementation, the transfer adaptation training employs a parameter-efficient fine-tuning method (Weight-Decomposed Low-Rank Adaptation, DoRA), which decomposes the weights of the basic model and updates only the magnitude or low-rank parameters related to the design attribute prediction task, thereby reducing training costs while improving the model's prediction stability in garment sketch scenarios.

[0083] V. Parametric Garment Pattern Generation

[0084] After completing the structured configuration and design attribute prediction of garment components, the updated component parameter configuration is combined with key human body dimension data to perform parametric garment pattern generation in a two-dimensional coordinate system, thereby obtaining garment pattern data with clear geometric definitions and structural relationships. Key human body dimension data serves as the basic input for pattern generation, and its sources can include manual measurement data, 3D human body scan data, or standard size chart data, including waist circumference, hip circumference, waist-hip vertical distance, height, leg length, back width, and front waist length.

[0085] Taking the pre-production version as an example, the main steps are as follows:

[0086] 1. Calculation of key parameters

[0087] A two-dimensional coordinate system is established with the neck hollow as the origin, the front midline as the y-axis (positive downwards), and the chest circumference line as the x-axis (positive to the left). Key parameters are calculated based on anthropometric data, including:

[0088] The formula for calculating chest width is: ,

[0089] in, Bust and W back These are measurements of chest circumference and back width, respectively. K ease This is the loosening coefficient.

[0090] The formula for calculating the shoulder slope angle is: ,

[0091] in, H shoulder This represents the shoulder height difference, which is the vertical distance from the shoulder neck point to the shoulder endpoint. W shoulder Shoulder width.

[0092] The formula for calculating the front waist circumference is: ,

[0093] in, Waist This is the waist circumference measurement. W waist_back For back width, K waist Provided in the component properties.

[0094] The formula for calculating garment length is: ,

[0095] in, L design To determine the garment length, multiply the proportion from the component properties by the waistline length. , Hip Hip circumference, H waist_hip Waist-hip height is the distance from the waistline to the hipline.

[0096] 2. Generation of basic outline and provincial highway structure

[0097] Locate the key points of the front piece's basic outline in the defined coordinate system, including:

[0098] Calculation of shoulder and neck point coordinates: ,

[0099] in, Neck Neck circumference, D neck_front The front of the neck is deep. This is the compensation value for shoulder tilt.

[0100] Calculation of chest height coordinates: ,

[0101] in, The provincial highway is deviated from its course. H bust It indicates chest height.

[0102] Waist dart parameters include:

[0103] Save ,

[0104] Provincial Deep ,

[0105] in, The inclination angle of the provincial highway. H waist The waist length.

[0106] Side seam width reduction ,

[0107] Provincial apex coordinates: ,

[0108] The opening curve of the provincial highway achieves a smooth transition through control points located on both sides of the center of the provincial highway, with an offset of half the width of the province.

[0109] 3. Precise modeling of curved surface edges and determination of key point coordinates

[0110] The armhole curve is constructed using a piecewise cubic Bézier curve. Key calculation points are as follows:

[0111] Anterior axillary point:

[0112] in, D armhole For the depth of the armhole, H armhole This indicates the depth of the armpit, which is the distance from the side of the neck to the armpit.

[0113] The armhole control points are as follows: ,

[0114] .

[0115] The coordinates of the shoulder endpoint are calculated using the shoulder slope angle.

[0116] Neckline curve includes the front neckline point ,

[0117] Deeper point ,

[0118] Neckline curve radius .

[0119] 4. Plate integration and parameter standardization

[0120] Calculate the shoulder line alignment point and mark the incision point at 1 / 5 of the shoulder length from the neck point:

[0121] ,

[0122] Side seam hip measurement corresponding point: , K ease_hip Allowance for hip circumference.

[0123] After completing the modeling of all contour lines, channels, and surfaces, the spatial positions and edge connections of all contour points will be systematically recorded in a counterclockwise direction, with the vertex coordinate sequence as the core, to describe the boundary contour of the plate.

[0124] VI. Construction of Topological Relationships and Verification of Physical Rationality

[0125] After completing the geometric contour modeling of each garment component and obtaining independent two-dimensional pattern data, the stitching topology between garment patterns is further constructed to clarify the connection method and connection position between different patterns, thereby forming a complete garment structure assembly relationship.

[0126] First, an interface identifier for sewing is set for each generated garment pattern. The interface represents the boundary segment or key connection position involved in sewing within the pattern. Each interface includes at least an interface name, an interface type, and its corresponding geometric boundary information. The interface name describes the sewing role of the interface in the garment structure, such as a side seam interface on the front bodice, a shoulder seam interface on the back bodice, or a sleeve cap interface; the interface type distinguishes whether the interface corresponds to a straight boundary or a curved boundary; the geometric boundary information describes the start and end points of the contour segment corresponding to the interface in the pattern and its curve parameters. In this embodiment, the interface can also be attached with sewing-related engineering parameters to describe the adjustment rules of the interface during the sewing process. For example, a relaxation coefficient parameter is set for the interface to indicate whether the interface boundary length needs to be scaled or allocated before sewing; an edge direction marker is set to ensure the geometric consistency of the two interfaces involved in sewing.

[0127] After setting up the interfaces, the stitching relationships between the garment pieces are established based on the structural principles of clothing. These stitching relationships use interfaces as the basic unit to describe the connection pairing rules between two or more interfaces. For symmetrical garment parts, the connection relationships between corresponding interfaces are established according to preset symmetry rules. For example, the side seam interfaces of the left front bodice and the right front bodice are set as a pair of stitching interfaces to ensure consistency in the assembly relationship between the left and right structures. For garment parts with different shapes but that must be connected, such as the connection between the sleeve and the bodice, stitching rules are established based on the principle of geometric compatibility.

[0128] Specifically, the sleeve cap interface used for connection in the sleeve pattern is designated as an auxiliary interface, while the armhole interfaces in the front and back bodice pieces are designated as the main interfaces. The geometric boundaries of the auxiliary interfaces are projected or reconstructed to correspond with the main interfaces in terms of length and curvature characteristics. During the geometric compatibility processing, length analysis is performed on the interface boundaries involved in sewing: when there is a difference in the original boundary lengths of the two interfaces, the difference is allocated and adjusted according to preset garment structure rules. For example, the length difference is proportionally distributed to the front or back interface, or the interface boundaries are fine-tuned through local curve reparameterization to ensure that the final pair of interfaces involved in sewing remain consistent within the effective length range.

[0129] Using the above method, a corresponding list of seam interface pairs is generated for each garment component, forming seam rules that describe the connection relationships between the garment pieces. These seam rules are then integrated into a global seam topology diagram.

[0130] Specifically, each interface is treated as a topological node, and the stitching relationships between interfaces are considered topological edges, thus constructing a topological network with interfaces as nodes and stitching rules as connecting edges. During the construction of the topological relationship graph, a globally unique identifier is assigned to each stitching relationship, and the corresponding interface reference information and its associated engineering parameters are recorded, including stitching type, relaxation coefficient, and length adjustment method. Subsequently, a consistency check is performed on the topological relationship graph to ensure that all interfaces declared as requiring stitching have valid pairing relationships and that there are no duplicate or missing connections. This converts the component configuration into digital patch data of geometric information and topological relationships.

[0131] Next, the physical rationality of the generated pattern needs to be verified to ensure that the generated pattern meets the actual wearing and production requirements in terms of structure and size. This mainly includes: 1) Verification of key circumference dimensions. Based on key human body circumference data, calculate key channel dimensions such as neckline, chest, waist, hip, and armhole circumference, and determine whether they meet the minimum space requirements for wearing. For example, verify the neckline based on head circumference, and the armhole depth based on arm circumference. 2) Structural proportion check: Analyze the proportional coordination between components. For example, the width and depth of the armhole should maintain a reasonable proportional range with the shoulder width. 3) Check the matching degree of seam length. For example, calculate whether the difference between the sleeve cap arc and the armhole arc length is within a reasonable range. 4) Inspect the pattern for excessively sharp corners or pieces with excessively small areas.

[0132] When physical inconsistencies are detected during the above verification process, relevant parameters are adjusted according to preset correction rules. For example, for insufficient circumference, the allowance parameter for the corresponding part is appropriately increased; for mismatched boundary lengths, the interface length balance calculation is re-executed; and local geometric anomalies are corrected by adjusting the positions of control points or key points. After correction, the board data can be re-verified until all verification items meet the preset conditions. Finally, a parameterized and adjustable set of board data is output.

[0133] VII. Plate Data Format Conversion and Output

[0134] To ensure that the generated pattern data can be directly loaded, edited, and reused by existing pattern-making software, the parameterized pattern data needs to be converted into a standardized file format supported by the target pattern-making software. In this embodiment, the target pattern-making software is the open-source 2D pattern-making software Seamless2D, whose corresponding pattern file format is .sm2d. This file uses an XML-based structured data format to describe the geometric elements, pattern organization relationships, and related metadata of the garment pattern.

[0135] Specifically, the conversion steps mainly include:

[0136] 1. Parsing and standardization of parametric plate data

[0137] The parameterized pattern data generated in the previous steps is parsed and processed. The parameterized pattern data includes at least: a unique identifier for each garment pattern, a sequence of vertex coordinates for each pattern, boundary connection relationships between vertices, boundary type information (straight line or curve), control parameters for the curve, and a description of the stitching topology between patterns.

[0138] During the parsing process, all geometric data undergoes uniform unit standardization. For example, the length units used in internal calculations are uniformly converted to the units of measurement required by the Seamless2D file format, ensuring consistency of plate size when loaded across different systems and software environments.

[0139] 2. Mapping from 2D coordinates to SM2D vertex elements

[0140] The coordinates of the 2D vertices in each tile are mapped to point elements in the SM2D file. Specifically, a unique vertex identifier is assigned to each vertex in each tile, and a corresponding point node is created for it in the XML structure. The point node contains at least a coordinate attribute representing the vertex's position, used to record the vertex's geometric position in the 2D plane. During the conversion process, a mapping relationship is established between the original vertex index and the XML node identifier to ensure that subsequent boundary and curve elements can correctly reference the corresponding vertices.

[0141] 3. Reconstruction of boundary and curve geometric elements

[0142] Refer to the corresponding vertex ID based on the vertex index.

[0143] For a straight boundary formed by directly connecting two vertices, a corresponding line element is created in the SM2D file, and the start and end points of the straight boundary are specified through vertex identifiers. The line element has the `firstPoint` and `secondPoint` properties. For curved boundaries, based on the curve type and its parameter format, it is converted into a spline curve representation supported by the SM2D file. Specifically, for circular arc curves, the radius parameter is extracted and converted into the path point angle control of the spline; for quadratic Bézier curves, a single set of control point parameters is parsed; for cubic Bézier curves, two sets of control point parameters are parsed, all converted into the `angle1`, `angle2`, `length1`, and `length2` control parameters of the spline element.

[0144] 4. XML organization and assembly of the page structure

[0145] After converting vertex and boundary elements, a complete XML document structure is constructed according to the SM2D file format specification. Specifically, the `pattern` root element of the pattern file is initialized, and metadata information related to the pattern, such as version, unit, and description, is set. Then, `draftBlock` is created as the main container in the document, and geometric elements are systematically organized under the `calculation` node, inserting the vertices, lines, and curves corresponding to each piece in sequence. Based on the stitching relationship definition, a `piece` element is created for each piece under the `pieces` node, and geometric elements are assembled in edge order using the `nodes` child elements.

[0146] Follow the steps above to generate a .sm2d file that conforms to the specified format, and store it as the output in the preset path. The generated pattern file is in an editable format, allowing pattern makers to review, correct, and add process details (such as seam allowances, cuts, and texture direction) to the initial pattern in the Seamless2D software environment for further redesign, ultimately creating a standard pattern that can be used for grading, layout, and production.

[0147] Example 2

[0148] Based on the above method, this embodiment provides an intelligent assisted garment pattern making system based on a multimodal model. This system achieves semantic understanding, structural analysis, and engineered pattern output from garment design sketches through hardware and software collaboration. Overall, it assists pattern makers in completing the conversion process from design intent to editable garment patterns. The system is deployed on general-purpose computing devices, is compatible with existing garment pattern making software environments, and achieves intelligent assistance in the pattern making process without changing traditional pattern making work habits.

[0149] like Figure 3As shown, the system generally includes a sketch input and preprocessing module, a multimodal semantic parsing module, a garment component knowledge management module, a design attribute prediction and constraint module, a parametric pattern generation module, a seam topology and verification module, and a pattern format conversion and output module. These modules are connected through a unified data interface, forming a continuous pattern-making process. During system operation, the modules can work automatically and collaboratively according to preset processes, or they can be manually intervened at key nodes to meet the actual needs of pattern-making tasks of varying complexity.

[0150] During system operation, the sketch input and preprocessing module first receives clothing design sketches and performs processing operations such as background separation, noise suppression, scale normalization, and orientation correction on the sketches to generate standardized sketch images with clear structure and uniform format. Subsequently, the multimodal semantic parsing module receives the standardized sketch image and its corresponding text information, performs joint semantic reasoning on the sketches through a multimodal model, identifies the clothing category corresponding to the sketch, and parses the overall structure of the clothing and multi-level clothing components, outputting component recognition results.

[0151] After obtaining the component identification results, the garment component knowledge management module performs semantic alignment processing on the identification results. It retrieves standardized component description information that matches the identified component from a pre-built garment component knowledge base through semantic similarity retrieval, and generates the corresponding component parameterized configuration accordingly. This module acts as a semantic bridge in the system, enabling the non-standardized semantic results output by the multimodal model to be mapped into an engineering data structure that conforms to pattern-making rules, providing a unified foundation for subsequent parameter calculations.

[0152] Based on this, the design attribute prediction and constraint module predicts design attributes reflecting the garment's shape features and structural proportions according to the visual semantic features of the sketch and the component parameter configuration. It then applies adaptive constraints based on the identified garment category to ensure that only design attributes relevant to the current garment category participate in subsequent pattern-making calculations, thus avoiding interference from irrelevant parameters. The predicted design attribute parameters are subsequently written into the corresponding component parameter configuration as important input for pattern geometric modeling.

[0153] The parametric pattern generation module combines component parameter configurations with key human body dimension data to perform geometric modeling of garment patterns in a two-dimensional coordinate system. Based on preset pattern-making rules, this module calculates the positions of key points, outlines, and curve boundaries of each pattern piece, and generates two-dimensional pattern geometric structures for different garment components such as the body, sleeves, and collar, transforming the design intent in the sketch into pattern data with clear mathematical definitions.

[0154] After the pattern geometry is generated, the stitching topology and verification module further models the connection relationships between the various garment patterns. By defining stitching interfaces, establishing stitching rules, and constructing global stitching topology relationships, it explicitly describes the assembly relationships of the overall garment structure. Simultaneously, this module performs physical rationality verification on the generated patterns, checking circumference dimensions, component proportions, and stitching boundary matching. When anomalies are detected, relevant parameters are corrected to ensure that the patterns meet actual wearing and production requirements in terms of structure and dimensions.

[0155] Finally, the plate format conversion and output module converts the verified parametric plate data into a standardized file format supported by the plate-making software and outputs it as an editable plate file, enabling plate makers to directly load, view, and further adjust, encode, and refine the generated initial plate in existing plate-making software.

[0156] Through the collaborative work of the above modules, the system realizes a complete auxiliary pattern-making process from clothing design sketches to industrially usable clothing pattern data. While ensuring the usability of the project, it significantly reduces the reliance on human experience and improves pattern-making efficiency and result consistency.

[0157] The above system can execute the intelligent assisted garment pattern making method based on a multimodal model as described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to the intelligent assisted garment pattern making method provided in Embodiment 1 of this invention.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A smart assisted garment pattern making method based on a multimodal model, characterized in that, Including the following steps: S1. Input the clothing design sketch and prompts into the multimodal visual language model to obtain the recognition results; the recognition results include the clothing category corresponding to the sketch and the clothing components it contains; S2. Based on the recognition results, a retrieval and recall is performed in a pre-built clothing component knowledge base based on semantic similarity to obtain standardized component information corresponding to the identified clothing component, so as to generate a parameterized configuration of the clothing component. S3. Perform design attribute prediction on the garment design sketch to obtain design attribute parameters that reflect the garment's shape characteristics and structural proportions, and write them into the corresponding garment component parameterization configuration: The design sketch and corresponding text prompts are input into the attribute predictor to obtain the corresponding design attribute parameter values. The attribute predictor includes a basic multimodal feature extraction module and an attribute prediction module. The basic multimodal feature extraction module is used to perform joint feature extraction on the input design sketch and corresponding text prompts to obtain multimodal features that can simultaneously represent visual structural information and semantic information. The attribute prediction module is used to map the multimodal features to parameter values ​​of each design attribute in the design attribute space and output them. The attribute prediction module adopts a multilayer perceptron structure, and its output layer is used to output a prediction vector containing multiple design attribute parameter values. The prediction vector is multiplied by the attribute mask matrix of the clothing category, and only the design attribute parameter values ​​related to the clothing category are retained. Different clothing categories have corresponding attribute mask matrices, which are vectors with the same dimension as the prediction vector. Each vector element is used to indicate whether the corresponding design attribute is a valid attribute under the current clothing category. When a design attribute is not applicable to the current clothing category, its corresponding vector element value is set to a suppressed state, where 0 indicates suppression and 1 indicates retention. S4. Combine the parameterized configuration of the garment components with key human body size data, and perform parameterized garment pattern generation in a two-dimensional coordinate system to construct the pattern geometry corresponding to each garment component: By combining the parametric configuration of clothing components with key human body size data, a reference coordinate framework for the pattern is established in a two-dimensional plane coordinate system, and the units and scales of the key human body size data are standardized. Based on standardized human body key dimension data, design attribute parameters, and preset pattern-making rules, the basic outline key points of the garment pattern are determined sequentially in the two-dimensional coordinate system. After determining the basic outline key points, the dart structure is generated according to the dart amount, dart depth, and dart inclination angle parameters, and the dart outline is constructed in a linear or curved form. The curved boundary of the garment pattern is modeled to generate continuous pattern outline lines. The outline data of each garment pattern is recorded in a structured manner to form pattern data containing vertex coordinate sequences, boundary connection relationships, and boundary type information. S5. After completing the generation of the pattern geometry, construct the stitching topology between each garment pattern and perform physical rationality verification on the generated garment pattern. Based on the verification results, correct the relevant parameters. S6. Convert the verified garment pattern data into a standardized file format supported by the pattern-making software and output it for use in garment pattern editing and production.

2. The intelligent assisted garment pattern making method as described in claim 1, characterized in that, The prompts mentioned in S1 include system prompts and user prompts. System prompts are used to set the task role of the multimodal visual language model, constrain the output format of the model, and limit the recognition range, so that the model only analyzes content related to clothing pattern making. User prompts are used to attach sketches and clarify task objectives, instructing the model to perform clothing category recognition and step-by-step parsing of clothing parts on the input sketches.

3. The intelligent assisted garment pattern making method as described in claim 1, characterized in that, The garment component knowledge base described in S2 is used to store the structured description information of each garment component, including the standard name of the component, the component level to which it belongs, the parent component identifier, the applicable garment category range, and the parameter description information related to the component. The parameter description information is used to characterize the adjustable attributes of the component in the pattern making process. The adjustable attributes include length ratio, angle range, curvature characteristics, dart parameters, or mapping relationships related to key human body dimensions. The clothing component knowledge base is also used to store semantic vectors corresponding to the structured description information of each clothing component. The semantic vectors are obtained by mapping the structured description information text of the clothing components through a semantic embedding model, and are used to reflect the similarity relationship between different clothing components at the semantic level.

4. The intelligent assisted garment pattern making method as described in claim 3, characterized in that, In S2: First, the text description of the recognition result is mapped to a corresponding semantic vector using a semantic embedding model. Then, a retrieval and recall operation is performed in the clothing component knowledge base based on the semantic vector similarity. The semantic vector of the recognition result is used as the query vector, and the vector similarity is used as the matching criterion to retrieve the standard component information that is semantically closest to it from the knowledge base. The retrieved standard component information is then subjected to structured post-processing to generate a preliminary component parameterized configuration. In the component parameterized configuration, each clothing component corresponds to a set of structured parameter description information. This parameter description information includes at least the name of the adjustable parameter item of the component, the explanation of the parameter meaning, and the initial value range of the parameter.

5. The intelligent assisted garment pattern making method as described in claim 1, characterized in that, S5 includes the following steps: For each garment pattern piece, a sewing interface is set for the boundary segment or connection position involved in sewing. The sewing interface includes at least the interface name, interface type and corresponding geometric boundary information, and engineering parameters for describing the sewing adjustment rules are configured for the sewing interface. Based on the preset garment structure rules, the sewing relationship between garment pieces is established with the sewing interface as the basic unit. The pairing relationship between corresponding interfaces is established for left and right symmetrical parts. For parts with different shapes but need to be connected to each other, the sewing rules are established based on the principle of geometric compatibility. The length analysis of the interface boundary involved in sewing is performed. When there is a difference in the length of the interface boundary, the difference is allocated and adjusted according to the preset rules or the interface boundary is corrected by curve reparameterization so that the interface involved in sewing remains consistent within the effective length range. After obtaining the stitching rules between each pattern, the stitching rules are integrated into a stitching topology graph, wherein the stitching interface is used as the topology node and the stitching relationship between the interfaces is used as the topology connection edge, and a consistency check is performed on the stitching topology graph. After completing the construction of the stitching topology, the generated garment pattern is subjected to physical rationality verification. The physical rationality verification includes at least key circumference dimension verification, garment component proportion verification, stitching interface boundary length matching degree verification, and pattern geometric anomaly detection. When the physical rationality verification result does not meet the preset conditions, the relevant pattern parameters are adjusted according to the preset correction rules, and the physical rationality verification is re-executed after adjustment until the verification result meets the requirements.

6. The intelligent assisted garment pattern making method as described in claim 1, characterized in that, S6 includes the following steps: The garment pattern data is parsed and processed. The garment pattern data includes at least the identification information of the garment pattern, vertex coordinate sequence, boundary connection relationship, boundary type information, curve control parameters, and stitching topology relationship between the patterns. The garment pattern data is then standardized using a uniform size unit. The parsed and standardized 2D plate vertex coordinates are mapped to vertex elements in the target plate-making software file format, and each vertex is assigned a unique identifier to establish the correspondence between vertex coordinates and file structure. The plate boundaries are transformed according to the boundary type. Line segment elements are generated for straight boundaries, and curve boundaries are reconstructed into curve or spline curve representations supported by the target plate-making software based on their parameter forms, so as to maintain the geometric consistency of the plate outline. According to the structural specifications of the target pattern-making software file format, the vertex, boundary and curve elements of the garment pattern are organized to build a corresponding file structure unit for each garment pattern. The stitching topology is mapped to the file structure, and the boundary segments of the pattern pieces involved in stitching and their corresponding relationships are recorded, so that the pattern-making software can identify the stitching relationship between different garment pattern pieces when loading the file.

7. An intelligent assisted garment pattern making system based on the method of any one of claims 1 to 6, characterized in that, include: The sketch input and preprocessing module is used to receive clothing design sketches and perform background separation and noise suppression processing on them to obtain normalized sketch images; The multimodal semantic parsing module is used to input the normalized sketch image and prompts into the multimodal model, identify the clothing category corresponding to the sketch and the clothing parts contained in the sketch, and generate clothing part recognition results; The clothing component knowledge management module is used to perform a retrieval and recall in a pre-built clothing component knowledge base based on semantic similarity for the clothing component identification results, obtain standardized component information corresponding to the identified clothing component, and generate parameterized configuration of the clothing component. The design attribute prediction and constraint module is used to predict the design attribute parameters that reflect the shape features and structural proportions of the clothing based on the visual semantic features of the sketch, and write them into the corresponding parametric configuration of the clothing component. The parametric pattern generation module is used to combine the parametric configuration of the garment components with key human body size data, perform parametric garment pattern generation in a two-dimensional coordinate system, and construct the pattern geometry structure corresponding to each garment component. The stitching topology and verification module is used to construct the stitching topology relationship between each garment pattern after the pattern geometry is generated, and to perform physical rationality verification and correction on the generated garment patterns. The pattern conversion and output module is used to convert the verified garment pattern data into a standardized file format supported by the pattern making software and output it for garment pattern editing and production.

Images