Clothing pattern intelligent generation method and system based on multi-modal visual identification

Abstract

The invention provides a clothing pattern intelligent generation method and system based on multi-modal visual identification, and relates to the technical field of visual identification. The method comprises the following steps: firstly, acquiring clothing image data and text description data, performing unified preprocessing, obtaining a style category label, an appearance attribute label and a local contour key point set through multi-modal visual identification, generating structure semantic elements and geometric anchor points by utilizing a semantic-to-structure mapping knowledge base, and constructing a plate type structure diagram; further analyzing the structural constraint and executing constraint solution to generate a candidate pattern; and in combination with manufacturability verification and constraint correction, closed-loop optimization is formed, a target plate type file meeting the structure and manufacturing requirements is output, and effective connection from semantic recognition to plate type structure generation is realized.

Description

Technical Field

[0001] This invention relates to the field of visual recognition technology, and in particular to a method and system for intelligent generation of clothing patterns based on multimodal visual recognition. Background Technology

[0002] In existing technologies, garment pattern generation mainly relies on garment pattern making experience and computer-aided design software to complete two-dimensional paper pattern design, and then uses three-dimensional virtual fitting and garment simulation for style preview and pattern iteration. Research on digital garments and virtual fitting is constantly increasing, and various synthetic and simulation data resources for three-dimensional human body and garment geometry have been made public. For example, the CLOTH3D dataset from ECCV 2020 is used for three-dimensional dressed human body sequence research. At the same time, three-dimensional garment data resources and generation pipelines containing sewing patterns have also emerged, providing a training foundation for learning-based methods.

[0003] With the development of deep learning and generative models, existing technologies are showing a trend of evolving from traditional rule-based pattern making to data-driven pattern making. Learning-based methods for recovering sewing patterns from everyday clothing images have been published. For example, SewFormer studies the recovery of clothing sewing patterns from single images and constructs corresponding datasets and network structures. At the same time, frameworks for generating sewing patterns from text descriptions have also emerged. For example, DressCode proposed SewingGPT to generate pattern sequences with text prompts. Furthermore, multimodal conditional pattern generation has been published. For example, GarmentDiffusion proposed generating centimeter-level accurate vectorized sewing patterns from multimodal inputs such as text and images.

[0004] However, the aforementioned disclosures mostly focus on directly regressing or generating pattern parameters from single-modal or weakly structured conditions, making it difficult to establish stable and consistent constraints between the recognition results and the manufacturable pattern structure. Multimodal visual recognition typically outputs semantic layer information such as style category, appearance attribute labels, local contours, or key points, while pattern generation requires geometric structure information such as structural line positions, panel geometric relationships, dimensional constraints, symmetry relationships, and stitching pairing relationships. The semantic layer output cannot directly constrain the pattern structure, causing the mapping from recognition results to pattern parameters to rely on empirical rules or simple regression, thus limiting stability and generalization ability. Related disclosures also point out that existing pattern generation methods have limitations such as single-modal dependence or generation efficiency, and that discriminative prediction from images to patterns tends to produce deterministic outputs, limiting design diversity. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for intelligent generation of clothing patterns based on multimodal visual recognition. By structuring the results of multimodal visual recognition into structural semantic elements and geometric anchor points that can directly constrain pattern generation, an effective connection is achieved from semantic recognition to pattern structure generation.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for intelligent generation of garment patterns based on multimodal visual recognition, comprising: The multimodal input data corresponding to the garment to be generated is obtained, and scale normalization and coordinate reference unification processing are performed on the multimodal input data to output a multimodal input dataset; the multimodal input data includes garment image data and garment text description data; Multimodal visual recognition is performed on the multimodal input dataset to generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from the clothing image data; A pre-built semantic-to-structure mapping knowledge base is invoked to convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, and a plate structure diagram is constructed based on the structural semantic element set and the geometric anchor point set; the plate structure diagram is used to describe the plate set, the structural line set, and the connection relationship between the plates; The plate structure diagram is parsed into a set of structural constraints, and the set of structural constraints is associated with the set of geometric anchor points to form constraint solution input data; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. An initial plate template is selected from the preset plate template library based on the set of structural semantic elements, and constraint solving is performed on the initial plate template according to the constraint solving input data to obtain a set of plate parameters that satisfy the set of structural constraints. Candidate plate types are generated based on the set of plate parameters. Perform manufacturability verification on the candidate board type to obtain the manufacturability verification result; If the manufacturability verification result fails, a constraint correction amount is generated. The structural constraint set is updated based on the constraint correction amount, and the constraint solution is re-executed. If the manufacturability verification result passes, the target plate shape file and the stitching relationship data corresponding to the target plate shape file are output.

[0007] Preferably, a pre-set semantic-to-structure mapping knowledge base is invoked to convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, including: The style category label is mapped to the target pattern category identifier, and the structural semantic element entries that match the target pattern category identifier are retrieved in the semantic-to-structure mapping knowledge base to generate the structural semantic element set; The appearance attribute tags are mapped to structural line type identifiers and plate type identifiers, and the structural line type identifiers and plate type identifiers are written into the structural semantic element set; The local contour key point set is grouped according to the preset key point topological relationship and fitted to obtain the contour segment set, so as to extract the anchor point coordinates used to constrain the position of the structure line from the contour segment set and generate the geometric anchor point set.

[0008] Preferably, constructing a plate-type structure diagram based on the structural semantic element set and the geometric anchor point set includes: Based on the structural semantic element set, generate a plate node set and a structural line node set, and write a plate type identifier and a structural line type identifier for the plate node set and the structural line node set respectively; Based on the set of structural semantic elements, determine the set of associated edges between the set of plate nodes and the set of structural line nodes, so as to obtain the structural graph connection relationship that represents the belonging relationship of the set of structural lines within the set of plate nodes; Based on the geometric anchor point set, anchor point coordinate attributes are written to the structural line node set so that the plate structure diagram can simultaneously represent structural semantic information and geometric positioning information.

[0009] Preferably, the plate structure diagram is parsed into a set of structural constraints, including: Based on the structural line type identifier and anchor point coordinate attributes in the plate structure diagram, structural line position constraints are generated; the structural line position constraints are used to limit the positioning relationship of the corresponding structural line on the corresponding plate. Based on the plate type identifier in the plate structure diagram and its connection relationship with the structure diagram, plate geometric relationship constraints are generated; these plate geometric relationship constraints are used to limit the relative positional and dimensional relationships between different plates. Based on the symmetry identifiers in the plate structure diagram, symmetry constraints are generated; these symmetry constraints are used to limit the geometric consistency of symmetrical plates or symmetrical structural lines. Based on the connection relationships in the plate structure diagram, stitching relationship constraints are generated; these constraints are used to limit the pairing relationship of the edges to be stitched and ensure consistency with the stitching direction.

[0010] Preferably, associating the set of structural constraints with the set of geometric anchor points to form constraint solution input data includes: Write a constraint action object identifier for each constraint in the set of structural constraints; the constraint action object identifier is used to indicate the plate node or structural line node corresponding to the constraint. The anchor point coordinates in the geometric anchor point set are bound to the constraint object identifier to form an anchor point location item for constraint solving. Write constraint weights for at least some of the constraints in the set of structural constraints to form the constraint solution input data, which includes constraint terms, anchor point location terms, and constraint weights.

[0011] Preferably, selecting an initial template from a preset template library based on the set of structural semantic elements includes: The pattern template library contains a pre-set pattern template index table for different style categories; the pattern template index table is used to declare the mapping relationship between style category labels and initial pattern templates; The target template identifier is retrieved from the template template index table based on the target template type identifier in the structural semantic element set. The plate parameter framework and structural line parameter framework corresponding to the target template identifier are called as the initial plate template.

[0012] Preferably, performing constraint solving on the initial plate template based on the constraint solving input data includes: The plate shape parameters in the initial plate shape template are used as the set of variables to be solved; the set of variables to be solved includes plate contour parameters and structural line parameters. The set of structural constraints is converted into a set of constraint equations for the set of variables to be solved, and the anchor point location terms corresponding to the set of geometric anchor points are written into the set of constraint equations. When the constraint equations satisfy the preset convergence conditions, the plate shape parameter set is output, and the candidate plate shape is generated based on the plate shape parameter set.

[0013] Preferably, manufacturability verification is performed on the candidate board type to obtain manufacturability verification results, including: A closure check is performed on each plate in the candidate plate type to generate a plate closure verification result; Perform a pairing consistency judgment on the edges to be stitched corresponding to the stitching relationship constraints in the candidate plate type to generate a stitching pairing consistency verification result; The length difference of the edge to be sewn in the candidate plate type is calculated, and the length difference is compared with a preset error threshold to generate a sewn edge length error verification result. The manufacturability verification result is generated based on the plate closure verification result, the stitch pairing consistency verification result, and the stitch edge length error verification result.

[0014] Preferably, when the manufacturability verification result fails, a constraint correction amount is generated, the structural constraint set is updated based on the constraint correction amount, and the constraint solution is re-executed, including: Based on the manufacturability verification results, determine the type of failure constraint, and locate the target constraint item in the structural constraint set that matches the type of failure constraint; The correction direction and correction magnitude are calculated for the target constraint to generate the constraint correction amount, which is used to adjust the anchor point positioning item or constraint weight corresponding to the target constraint. The constraint correction amount is written into the structural constraint set to obtain an updated structural constraint set, and the constraint solution is re-executed based on the updated structural constraint set until the manufacturability verification result passes.

[0015] A smart garment pattern generation system based on multimodal visual recognition includes: A multimodal input acquisition and preprocessing unit is used to acquire multimodal input data corresponding to the garment to be generated, and to perform scale normalization and coordinate reference unification processing on the multimodal input data to output a multimodal input dataset; the multimodal input data includes garment image data and garment text description data; A multimodal visual recognition unit is used to perform multimodal visual recognition on the multimodal input dataset and generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from the clothing image data; The semantic-to-structure mapping and plate structure construction unit is used to call a pre-set semantic-to-structure mapping knowledge base, convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, and construct a plate structure diagram based on the structural semantic element set and the geometric anchor point set; the plate structure diagram is used to describe the plate set, the structural line set, and the connection relationship between the plates; The structural constraint analysis and solution input construction unit is used to analyze the plate structure diagram into a set of structural constraints and associate the set of structural constraints with the set of geometric anchor points to form constraint solution input data; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. The constraint-solving plate type generation unit is used to select an initial plate type template from a preset plate type template library based on the set of structural semantic elements, and perform constraint solving on the initial plate type template according to the constraint solving input data to obtain a set of plate type parameters that satisfy the set of structural constraints, and generate candidate plate types based on the set of plate type parameters. A board type manufacturability verification unit is used to perform manufacturability verification on the candidate board type and obtain manufacturability verification results; The constraint correction and plate shape iteration output unit is used to generate a constraint correction amount when the manufacturability verification result fails, update the structural constraint set according to the constraint correction amount and re-execute the constraint solution, and output the target plate shape file and the stitching relationship data corresponding to the target plate shape file when the manufacturability verification result passes.

[0016] The present invention discloses the following technical effects: This invention addresses the problem in existing technologies where multimodal visual recognition results remain at the semantic level, such as style labels, appearance attributes, or local contours, making it difficult to directly constrain the garment pattern structure. By introducing a semantic-to-structure mapping knowledge base, the multimodal recognition results are explicitly converted into a set of structural semantic elements and a set of geometric anchor points. Furthermore, a pattern structure diagram is constructed, enabling the visual recognition results to participate in the subsequent pattern generation process in a structured form. This fundamentally breaks down the mapping path between semantic information and geometric structure, eliminating the disconnect between semantics and geometry that is prevalent in existing technologies.

[0017] This invention resolves the pattern structure diagram into a set of structural constraints, including structural line position constraints, panel geometric relationship constraints, symmetry constraints, and stitching relationship constraints. This transforms the pattern generation process from being driven by empirical rules or simple regression to a solution process driven by explicit structural constraints, thereby ensuring that the generated pattern has consistency and interpretability at the structural level. This avoids the structural instability problem caused by the difficulty in constraining the pattern structure due to the recognition results in the prior art.

[0018] This invention adjusts the initial pattern template using a constraint-solving method during the pattern generation process, so that the generation of the pattern parameter set is controlled by explicit structural constraints and geometric anchor points, rather than relying solely on end-to-end prediction results. This maintains the stability and generalization ability of the pattern generation process under different clothing styles and different input conditions, reduces reliance on human experience and hard-coded template rules, and improves the consistency and repeatability of pattern generation.

[0019] This invention introduces manufacturability verification after generating candidate plate types. When the verification result fails, the structural constraint set is updated based on the constraint correction amount and the constraint solution is re-executed, forming a closed-loop mechanism between plate type generation and structural verification. This ensures that the final output target plate type is structurally reasonable and has consistent stitching, avoiding the problem in the prior art where the generated results require a lot of manual rework before they can be used for actual plate manufacturing.

[0020] This invention organically combines multimodal visual recognition, structured semantic mapping, structural constraint solving, and manufacturability verification, transforming the garment pattern generation process from a fragmented process of "identification followed by manual adjustment" into a continuous process where the recognition results directly drive structural constraints and automatically generate usable patterns. This significantly improves the automation level and engineering practicality of intelligent garment pattern generation. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the publicly available clothing image resources provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0023] 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, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The purpose of this invention is to provide a method and system for intelligent generation of garment patterns based on multimodal visual recognition. By combining structural constraint solving with manufacturability verification, the generated garment patterns are guaranteed in terms of both structural consistency and manufacturing feasibility.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a method for intelligent generation of clothing patterns based on multimodal visual recognition, comprising: Step 100: Obtain the multimodal input data corresponding to the garment to be generated, and perform scale normalization and coordinate reference unification processing on the multimodal input data to output the multimodal input dataset; the multimodal input data includes garment image data and garment text description data; Step 200: Perform multimodal visual recognition on the multimodal input dataset to generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from clothing image data; Step 300: Call the pre-built semantic-to-structure mapping knowledge base to convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, and construct a plate structure diagram based on the structural semantic element set and the geometric anchor point set; the plate structure diagram is used to describe the plate set, the structural line set, and the connection relationship between the plates; Step 400: Parse the plate structure diagram into a set of structural constraints, and associate the set of structural constraints with the set of geometric anchor points to form the input data for constraint solution; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints; Step 500: Select an initial plate template from the preset plate template library based on the set of structural semantic elements, and perform constraint solving on the initial plate template according to the constraint solving input data to obtain a set of plate parameters that satisfy the set of structural constraints, and generate candidate plate types based on the set of plate parameters; Step 600: Perform manufacturability verification on the candidate board type and obtain the manufacturability verification results; Step 700: If the manufacturability verification result fails, generate constraint correction amount, update the structural constraint set according to the constraint correction amount and re-execute constraint solution, and output the target plate type file and the stitching relationship data corresponding to the target plate type file when the manufacturability verification result passes.

[0027] Specifically, in step 100 of this embodiment, the multimodal input data corresponding to the garment to be generated is first obtained. The multimodal input data includes at least garment image data and garment text description data. The garment image data is used to characterize the garment's appearance, structural form, and local details, while the garment text description data is used to characterize the garment's style intent, structural features, and design elements. In this embodiment, the garment image data can originate from garment design drawings, real-object photographs, or virtual garment renderings, with an image resolution of at least 256 x 256 pixels to ensure the recognizability of the garment's outline and key structural features. The garment text description data can be natural language descriptive text, preferably at least 10 Chinese characters in length, to cover the basic style information and main structural elements of the garment, thereby providing sufficient semantic information support for subsequent multimodal visual recognition. Exemplarily, in this embodiment, the garment image data can include visual samples from different sources, such as real-object photographs or design sketches. These visual samples can use example images from the DeepFashion-MultiModal dataset as examples (see...). Figure 2 (The numbers A through C are not distinguished). The image has clear style and structural features and key information. The corresponding clothing text description data can be natural language descriptions that match the clothing image, such as example text like "high-waisted jeans, slightly flared legs, five-pocket structure." The image and text are combined to form a multimodal input for subsequent multimodal visual recognition processing steps.

[0028] After acquiring the multimodal input data, this embodiment performs scale normalization processing on the multimodal input data. Scale normalization processing refers to mapping clothing image data from different sources and with different resolutions to a unified scale range to eliminate the impact of image size differences on subsequent recognition results. In this embodiment, the scale normalization processing includes proportionally scaling the long or short side of the clothing image data so that the scaled image size falls within a preset scale range, for example, normalizing the longest side of the image to within 512 pixels while maintaining the aspect ratio of the image. For clothing text description data, the text characters are uniformly encoded, for example, limiting the text encoding length to no more than 128 characters, to ensure the uniformity of the multimodal input at the scale level, thereby avoiding scale imbalance between different modal data during subsequent processing.

[0029] After completing the scale normalization process, this embodiment further performs coordinate reference unification processing on the multimodal input data, and outputs a multimodal input dataset based on this. The coordinate reference unification processing refers to establishing a unified reference coordinate system for the clothing image data, used to describe the relative spatial relationships of each pixel or key position in the image. In this embodiment, the coordinate reference unification processing includes setting the upper left corner of the clothing image as a fixed reference starting point and dividing the image plane into regular two-dimensional coordinate regions, so that any image position can be described by relative proportional coordinates, for example, uniformly mapping the horizontal and vertical coordinate ranges to the interval between 0 and 1. Through this coordinate reference unification processing, the local contour key points extracted from the clothing image data can be aligned and compared under the same coordinate reference, thereby forming a structurally consistent multimodal input dataset, providing a stable and reproducible input foundation for subsequent multimodal visual recognition steps.

[0030] Optionally, in step 200 of this embodiment, multimodal visual recognition is performed on the multimodal input dataset to generate a multimodal recognition result set. The multimodal visual recognition refers to performing feature analysis on clothing image data and clothing text description data separately in a unified recognition process, and performing consistent expression processing on the recognition results of different modalities at the semantic level. In this embodiment, for clothing image data, visual feature information that can characterize the overall appearance and local structural features of the clothing is extracted; for clothing text description data, semantic feature information that can characterize the style intent and appearance attributes of the clothing is analyzed. Through the multimodal visual recognition processing, the multimodal recognition result set generated in this embodiment includes at least style category labels and appearance attribute labels, wherein the style category label is used to characterize the overall style category to which the clothing belongs, and the appearance attribute label is used to characterize the attribute features of the clothing at the structural, detail, or decorative levels. To ensure the stability of the recognition results, in this embodiment, each style category label and appearance attribute label is associated with corresponding confidence information. The value range of the confidence information is limited to 0 to 1, and preferably only appearance attribute labels with a confidence level of not less than 0.50 are retained to avoid low-confidence recognition results interfering with subsequent processing.

[0031] In this embodiment, the multimodal recognition result set also includes a set of local contour key points extracted from the clothing image data. The local contour key point set refers to a set of spatial location points used to describe the outer contour and key structural positions of the clothing. These spatial location points are all described based on the image coordinate system after the coordinate reference unification processing in step 100. In this embodiment, the local contour key point set preferably contains no fewer than 30 key points, which cover the main contour area and areas of significant structural change in the clothing to ensure that the overall contour trend and local structural features of the clothing can be reflected. To improve the usability of the key points, this embodiment performs validity screening processing on the extracted local contour key points, preferably removing key points with a confidence level below 0.40, and performs a consistency check on the spatial distance between adjacent key points. When the distance between adjacent key points exceeds a preset proportional threshold, the corresponding key point is either positionally smoothed or marked as an anomaly. The local contour key point set obtained through the above processing, together with the style category label and appearance attribute label, constitutes the multimodal recognition result set, thereby providing a reliable data foundation for semantic-to-structure mapping and geometric anchor point generation in subsequent step 300.

[0032] Optionally, in step 300 of this embodiment, a pre-set semantic-to-structure mapping knowledge base is used to complete the feasible translation of semantic layer recognition results into panel structure expressions. The semantic-to-structure mapping knowledge base is stored in an entry-based manner, and each entry includes at least: a style category label field, a target panel category identifier field, an appearance attribute label field, a panel type identifier field, a structural line type identifier field, a structural line belonging panel field, a symmetry identifier field, a connection relationship field, an anchor point type field, and a mutual exclusion relationship field. Among the above fields, the style category label field indicates the style category to which the entry applies, the target pattern type identifier field indicates the corresponding basic pattern type category; the appearance attribute label field indicates the semantic details of the style, the pattern type identifier field declares the composition of the pattern set, the structural line type identifier field declares the composition of the structural line set; the structural line belonging to pattern field declares the belonging position of each structural line type identifier within the pattern set; the symmetry identifier field declares the pattern or structural line that needs to satisfy symmetry constraints; the connection relationship field declares the association edge rules that should be established between the pattern node set and the structural line node set; the anchor point type field declares the anchor point type that should be referenced by the structural line position constraints; and the mutual exclusion relationship field declares the exclusive combination between appearance attribute labels. In this embodiment, the semantic-to-structure mapping knowledge base preferably covers no less than 20 types of style category labels and no less than 60 types of appearance attribute labels, and configures at least 1 basic entry and no less than 3 refined entries for each type of style category label to ensure that the structural semantic element set can cover common clothing styles and be compatible with detail differences.

[0033] In this embodiment, the process of mapping style category tags to target pattern type category identifiers and retrieving structural semantic element entries to generate a structural semantic element set is as follows: This embodiment reads the style category tags and corresponding confidence scores from the multimodal recognition result set, and uses the style category tags as the search key to locate the target pattern type category identifier in the semantic-to-structure mapping knowledge base; when there are multiple candidate style category tags, this embodiment uses the principle of maximum confidence to determine the target style category tag. The confidence threshold is preferably 0.60. When the maximum confidence score is lower than 0.60, this embodiment uses a general pattern type category identifier to ensure that the generation of the structural semantic element set is not interrupted. Subsequently, this embodiment uses the target pattern type category identifier as the index to retrieve structural semantic element entries, and writes the plate type identifier field, structural line type identifier field, structural line belonging plate field, symmetry identifier field, and connection relationship field declared in the structural semantic element entry into the structural semantic element set, thereby forming a structural semantic element set that at least includes plate set definition, structural line set definition, belonging relationship definition, and symmetry relationship definition. To ensure the parsability of the structural semantic element set, this embodiment requires that the number of plate type identifiers written be no less than 2 and the number of structural line type identifiers be no less than 1, and that each structural line type identifier has a corresponding attribution record in the structural line attribution plate field; when any structural line type identifier is missing an attribution record, this embodiment marks the structural line type identifier as unattributed and calls the basic entry under the same target plate type category identifier to fill in the attribution record, so as to avoid missing connection relationships when constructing the plate type structure diagram in the future.

[0034] In this embodiment, the process of mapping appearance attribute labels to structural line type identifiers and plate type identifiers and writing them into the structural semantic element set is as follows: This embodiment reads the appearance attribute labels and corresponding confidence scores from the multimodal recognition result set, and retrieves the corresponding structural line type identifiers and plate type identifiers in the semantic-to-structure mapping knowledge base using the appearance attribute labels as search keys. To suppress structural noise introduced by weakly correlated attributes, this embodiment only writes the structural line type identifier and plate type identifier corresponding to the appearance attribute label into the structural semantic element set when the confidence score of the appearance attribute label is not lower than 0.50; when the number of appearance attribute labels exceeds a preset upper limit, this embodiment preferably retains only the top 10 appearance attribute labels in terms of confidence score to avoid excessive expansion of the structural semantic element set leading to increased complexity in subsequent constraint solving. For conflict resolution of mutually exclusive appearance attribute labels, this embodiment determines conflict pairs based on the mutual exclusion relationship field, preferably retains the entries corresponding to appearance attribute labels with higher confidence scores, and records the removed appearance attribute labels as conflict removal labels for subsequent traceability explanation. Through the above-mentioned writing and conflict resolution processes, the structural semantic element set can introduce detailed structural lines and detailed panel information while keeping the basic structure of the style unchanged, thereby providing complete semantic support for the generation of nodes and the determination of connection relationships in the panel structure diagram.

[0035] In this embodiment, the process of grouping the local contour keypoint set according to a preset keypoint topological relationship and fitting it to obtain a contour segment set, and then extracting anchor point coordinates from the contour segment set to generate a geometric anchor point set, is as follows: The preset keypoint topological relationship is a keypoint connection rule table, which is used to declare the connection order and grouping boundary between key points, so as to organize discrete key points into continuous contour segments that can be used for structural positioning. In this embodiment, the local contour keypoint set preferably contains no less than 30 key points, and the key point coordinates adopt the relative coordinates after the coordinate reference unification processing in step 100, with the coordinate range limited to 0 to 1. Based on the keypoint connection rule table, this embodiment divides the key points into several keypoint groups, each keypoint group corresponds to a contour segment candidate, and performs smooth fitting on each keypoint group to obtain a contour segment set; to ensure the stability of the contour segments, this embodiment requires that each keypoint group contains no less than 5 key points, and keypoint groups with fewer than 5 key points are supplemented by interpolation of the endpoints of adjacent contour segments. Subsequently, this embodiment extracts anchor point coordinates from the contour segment set to generate a geometric anchor point set. The anchor point coordinates include the coordinates of the contour segment endpoints, the coordinates of contour segment direction change points, and the coordinates of points with significant curvature changes in the contour segment. Direction change points are determined by a threshold value of the change amplitude in adjacent tangential directions, preferably 0.25. Points with significant curvature changes are determined by a threshold value of the curvature change amplitude, preferably 0.20. To avoid unstable structural line positioning due to excessively dense anchor points, this embodiment performs sparsification processing on the anchor point coordinates, preferably limiting the number of anchor points extracted from each contour segment to no more than 8, and prioritizing the retention of endpoint anchor points and anchor points with significant curvature changes. This ensures that the geometric anchor point set can stably constrain the structural line position and improve the robustness of subsequent structural constraint analysis.

[0036] In this embodiment, the process of constructing a plate structure diagram based on a set of structural semantic elements and a set of geometric anchor points is as follows: This embodiment generates a set of plate nodes based on the plate type identifiers in the set of structural semantic elements, and a set of structural line nodes based on the structural line type identifiers in the set of structural semantic elements. Plate type identifier attributes are written to the set of plate nodes, and structural line type identifier attributes are written to the set of structural line nodes, forming traceable node semantic attributes. Subsequently, this embodiment generates a set of associated edges based on the plate-to-slab field and connection relationship field of the structural lines in the set of structural semantic elements. The set of associated edges is used to characterize the belonging and connection relationships of the structural line set within the plate set. Specifically, each associated edge includes at least a starting node identifier, an ending node identifier, and an association type identifier, where the starting node identifier is a structural line node, the ending node identifier is a plate node, and the association type identifier is used to distinguish between belonging and connection relationships. Finally, in this embodiment, anchor point coordinate attributes are written to the structural line node set based on the geometric anchor point set. The anchor point coordinate attributes include at least the anchor point position coordinates and the anchor point type identifier. The anchor point type identifier corresponds to the anchor point category declared in the anchor point type field, so that the plate structure diagram can simultaneously represent structural semantic information and geometric positioning information. To ensure that the plate structure diagram can be used to parse into a set of structural constraints in the subsequent step 400, this embodiment performs an integrity check on the plate structure diagram, requiring that the number of nodes in the plate node set is not less than 2, the number of nodes in the structural line node set is not less than 1, and that each structural line node has at least one associated edge and is bound to at least one anchor point coordinate. When any structural line node does not meet the above constraints, this embodiment backtracks the structural semantic element set to complete the plate field to which the structural line belongs or backtracks the geometric anchor point set to complete the anchor point coordinates, thereby ensuring that the output of the plate structure diagram has parsability and consistency.

[0037] Further, in step 400 of this embodiment, the plate structure diagram obtained in step 300 is parsed into a set of structural constraints. The set of structural constraints refers to a group of constraint entries used to impose restrictions on the subsequent solution process of the plate parameter set. Each constraint entry records information such as constraint type, constraint object identifier, constraint parameters, and constraint weights using structured fields. The constraint object identifier indicates the plate node or structural line node to which the constraint applies, and the constraint parameters describe the geometric or relational objectives of the constraint. To facilitate unified solution in the subsequent process, this embodiment classifies the constraint types in the set of structural constraints into at least four categories: structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. Each type of constraint is assigned at least three fields to express the necessary information about the constraint, thereby ensuring that the set of structural constraints is computable at both the semantic and geometric levels.

[0038] In this embodiment, structural line position constraints are generated based on the structural line type identifier and anchor point coordinate attributes in the plate structure diagram. These structural line position constraints define the positioning relationship of the corresponding structural line on the corresponding plate, including at least the endpoint positioning and direction positioning of the structural line. In this embodiment, the structural line type identifier and anchor point coordinate attributes of the structural line node set are first read from the plate structure diagram, and the anchor point coordinates declared in the anchor point coordinate attributes are used as references for structural line endpoint positioning. When the same structural line node is bound to multiple anchor point coordinates, this embodiment preferably selects the pair of anchor point coordinates with the largest distance as the structural line endpoint anchor point to enhance the stability of the structural line coverage. Subsequently, this embodiment writes the structural line endpoint anchor point into the constraint parameter field of the structural line position constraint, and determines the corresponding plate based on the associated edge between the structural line node and the plate node in the plate structure diagram, thereby generating a structural line position constraint entry of "structural line node identifier - plate node identifier - endpoint anchor point coordinates". To avoid positioning distortion caused by abnormal anchor point coordinates, this embodiment performs a range check on the anchor point coordinates, requiring the anchor point coordinate values ​​to fall within the coordinate range of 0 to 1, and preferably removing anchor point coordinates that exceed the range; when the number of valid anchor points after removal is less than 2, this embodiment marks the structural line node as insufficient anchor points and backtracks to step 300 to complete the anchor point coordinate attributes.

[0039] In this embodiment, geometric constraints on the plates are generated based on the plate type identifiers and connection relationships in the plate structure diagram, and symmetry constraints are generated based on the symmetry identifiers in the plate structure diagram. These geometric constraints define the relative positional and dimensional relationships between different plates. The relative positional relationships may include the correspondence between adjacent boundaries or relative orientation relationships, while the relative dimensional relationships prevent unexpected drift in the size ratios of different plates. In this embodiment, the plate type identifiers in the plate node set are first read, and the adjacency relationships between plates are determined based on the associated edge types declared in the connection relationships of the structure diagram. These adjacency relationships are then converted into geometric constraint entries. The relative dimensional relationships are limited by a preset ratio threshold, preferably limiting the length ratio of corresponding edges of adjacent plates to fall within the range of 0.70 to 1.30 to avoid significant inconsistencies in plate dimensions. The symmetry constraints are used to limit the geometric consistency of symmetrical plates or symmetrical structural lines. In this embodiment, the symmetry identifier field is read, and the plate nodes or structural line nodes declared as symmetrical objects are paired to generate symmetry constraint entries. The symmetry axis direction identifier or symmetry reference datum identifier is written in the symmetry constraint entries to ensure that consistent geometric constraints can be applied to symmetrical objects in the subsequent solution process. When the number of symmetrical objects is 2, a single symmetry pair is formed. When the number of symmetrical objects exceeds 2, multiple symmetry constraint entries are generated according to the preset pairing order.

[0040] In this embodiment, stitching constraints are generated based on the connection relationships in the pattern structure diagram. These constraints define the pairing relationships and stitching direction consistency of the edges to be stitched. The edges to be stitched refer to the boundary segments in the pattern set that need to be stitched together to form a garment structure connection. In this embodiment, the stitching connection relationships declared in the connection relationship field of the pattern structure diagram are read, and mapped to a pairing entry of "first pattern node identifier—first edge to be stitched identifier—second pattern node identifier—second edge to be stitched identifier." Simultaneously, a stitching direction identifier is written into the stitching constraint to constrain stitching direction consistency. To ensure the stitching constraints are usable in subsequent solutions, this embodiment performs a consistency check on the stitching connection relationships. Preferably, each stitching connection relationship corresponds to two pattern nodes and only one edge to be stitched. When the same edge to be stitched is repeatedly paired, this embodiment retains the higher-priority pairing relationship according to the preset priority rules in the connection relationship field, and records the rejected pairing relationships as conflicting pairing relationships for traceability in subsequent verification.

[0041] In this embodiment, the set of structural constraints and the set of geometric anchor points are associated to form constraint solving input data. The constraint solving input data refers to a set of uniformly formatted data entries that can be directly referenced in the subsequent constraint solving process, including at least constraint terms, anchor point location terms, and constraint weights. In this embodiment, firstly, a constraint object identifier is written for each constraint in the set of structural constraints. The constraint object identifier includes at least a plate node identifier or a structural line node identifier and a corresponding boundary segment identifier to clarify the scope of the constraint's effect. Subsequently, the anchor point coordinates in the set of geometric anchor points are bound to the constraint object identifier to form an anchor point location term. The anchor point location term is used to establish a correspondence between structural line position constraints, plate geometric relationship constraints, or symmetry constraints and image spatial positions during the constraint solving process. Furthermore, in this embodiment, constraint weights are written for at least some of the constraints in the structural constraint set. These constraint weights characterize the importance of different constraints, and are preferably integers from 1 to 10. Specifically, the constraint weight for structural line position constraints is preferably no less than 7, and the constraint weight for stitching relationship constraints is preferably no less than 8, to ensure that structural positioning and stitching pairing are preferentially satisfied during the solution process. When constraint weights are not specified, this embodiment uses a default weight of 5. Through the above processing, constraint solution input data containing constraint terms, anchor point positioning terms, and constraint weights is obtained, enabling subsequent steps 500 to perform constraint solution under a unified data format and output a set of plate type parameters that satisfy the structural constraint set.

[0042] In this embodiment, step 500 is used to select an initial plate template from a preset plate template library based on a set of structural semantic elements, and to perform constraint solving on the initial plate template based on constraint solving input data, outputting a set of plate parameters that satisfy the set of structural constraints and generating candidate plate templates. The plate template library includes a set of plate templates organized by style category. Each plate template includes at least a plate parameter framework and a structural line parameter framework, used to describe the positioning of the plate outline and structural lines within the plate. The initial plate template is the starting template for subsequent constraint solving. In this embodiment, it is preferable to ensure that at least 3 candidate templates are configured for the same style category to cover common structural differences within the same category. The plate parameter set is used to drive the generation of candidate plate templates, and preferably includes no less than 20 adjustable parameters to ensure that the constraint solving has sufficient degrees of freedom and does not degenerate into a fixed template output.

[0043] In this embodiment, the initial plate type template is selected based on the structural semantic element set, preferably through a plate type template index table. The plate type template index table is used to declare the correspondence between "target plate type category identifier - target template identifier - applicable conditions," where the applicable conditions include at least plate type coverage conditions and structural line type coverage conditions. In this embodiment, when the index matches multiple target template identifiers, the matching degree of each candidate template is calculated, and the target template identifier with the highest matching degree is selected. The matching degree is preferably determined in the following manner: in, The panel type coverage rate represents the proportion of the panel type identifier set in the candidate template to the panel type identifier set in the structural semantic element set, with a value ranging from 0 to 1. The structure line type coverage rate represents the proportion of the structure line type identifier set within the candidate template that covers the structure line type identifier set within the structure semantic element set. Its value ranges from 0 to... Conflict rate, representing the proportion of entries in the candidate template that are inconsistent with the structural semantic element set in terms of mutually exclusive attributes or attribution relationships, ranging from 0 to 1. Parameter The matching score weight is used to adjust the relative importance of coverage and conflict. It is derived from the empirical preset of different style categories in this embodiment, and the preferred value is [value missing]. This embodiment is preferably used only when... The corresponding candidate template is selected as the initial board type template. If all candidate templates have a score below 0.70, the basic template under the same target board type category is selected as the initial board type template. For example, when a candidate template has a board coverage rate of 0.90, a structural line coverage rate of 0.80, and a conflict rate of 0.10, the matching degree can be obtained as 0.5 multiplied by 0.90 plus 0.4 multiplied by 0.80 minus 0.3 multiplied by 0.10, which is 0.74, thus meeting the threshold requirement.

[0044] In this embodiment, the plate parameter framework and structural line parameter framework corresponding to the target template identifier are used as the initial plate template. The plate parameter framework is used to parametrically describe the contour of each plate, preferably using several contour control points or boundary segment length parameters to characterize the shape of the plate boundary; the structural line parameter framework is used to parametrically describe the position and direction of each structural line within the corresponding plate, preferably using structural line endpoint coordinate parameters and direction control parameters to characterize the geometry of the structural line. In this embodiment, to ensure that the parameters are solvable and do not undergo non-physical drift, a preset allowable range of variation is set for each parameter, and it is written into the set of variables to be solved in the form of boundary constraints, preferably satisfying: in, For any parameter to be solved, and These are the lower and upper bounds of the parameter, derived from the basic parameters of the initial plate template and the preset proportional relaxation rules. In this embodiment, the allowable variation range is preferably set to no more than 0.30 times the basic parameter to balance structural adjustability and stability; for example, if the basic value of the length of a plate boundary segment is 100 mm, then the corresponding upper bound can be set to 130 mm and the lower bound can be set to 70 mm, thereby providing clear and implementable parameter constraint boundaries.

[0045] In this embodiment, constraint solving is performed on the initial plate template based on the constraint solving input data. First, a set of variables to be solved is constructed using the plate parameters in the initial plate template. This set of variables includes at least plate contour parameters and structural line parameters. Subsequently, this embodiment converts the set of structural constraints into a set of constraint equations for the set of variables to be solved, and writes the anchor point positioning terms corresponding to the set of geometric anchor points into the set of constraint equations. To handle multiple types of constraints within the same solution framework, this embodiment preferably constructs an objective function with constraint residuals as its core and solves it using a weighted minimization method. in, Let the parameter vector be the set of variables to be solved. To constrain the total number of entries, For the first The constraint weights of the constraints are derived from the constraint weight field written in step 400, and the preferred value range is 1 to 10; For the first The residual function corresponding to each constraint is used to quantify the degree of deviation from the constraint under the current parameters. For example, for structural line position constraints, the residual function can be defined as the Euclidean distance between the structural line endpoint position and the anchor point coordinates; for symmetry constraints, the residual function can be defined as the difference between symmetrical objects under a symmetry reference datum; for plate geometric relationship constraints, the residual function can be defined as the excess amount by which the length of the corresponding side of adjacent plates deviates from a preset proportional range. Through the above construction, this embodiment enables the anchor point positioning term to directly participate in the residual calculation, thereby explicitly placing the constraint of the visual geometric anchor point on the structural line positioning into a solvable mathematical expression.

[0046] in, For the first The parameter vector obtained from the next iteration. The threshold value for rose collection is preferably 0.01. This represents the upper limit for the number of iterations, preferably 50. If the convergence threshold is met, then output... The corresponding plate type parameter set is used as the final plate type parameter set. If the convergence threshold is not met even after reaching the upper limit of the number of iterations, the plate type parameter set corresponding to the minimum of the current objective function is output as an approximate solution and a flag indicating incomplete convergence is recorded so that subsequent manufacturability verification and constraint correction mechanisms can trigger further corrections. The process of generating candidate plate types is as follows: In this embodiment, the plate type parameter set is written into the plate parameter framework to obtain the geometric boundaries of each plate, and the plate type parameter set is written into the structural line parameter framework to obtain the geometry of each structural line, thereby obtaining candidate plate types containing the plate set and the structural line set, which are used as inputs for subsequent step 600.

[0047] Further, in step 600 of this embodiment, a manufacturability verification is performed on the candidate pattern generated in step 500 to obtain a manufacturability verification result. The manufacturability verification refers to checking the consistency and error determination of the geometric representation and stitching relationship representation of the pieces in the candidate pattern from the perspective of garment pattern making feasibility; the manufacturability verification result is a determination result used to indicate whether the candidate pattern meets the preset manufacturability conditions. This embodiment divides the manufacturability verification into three types of verification: piece closure judgment, stitching pairing consistency judgment, and stitching edge length error judgment, and generates piece closure verification results, stitching pairing consistency verification results, and stitching edge length error verification results respectively. Then, a manufacturability verification result is generated based on the three types of verification results, thereby providing a clear basis for the subsequent generation of constraint correction quantities. To improve the reproducibility of the verification, this embodiment preferably performs at least one closure judgment on each piece in the candidate pattern, and at least one pairing consistency judgment and one length error judgment on each edge to be stitched.

[0048] In this embodiment, a closure check is performed on each piece in the candidate plate type to generate a plate closure verification result. The closure check refers to checking the consistency of the first and last ends and the self-intersection consistency of the plate boundary curves or boundary segment sequences to confirm that the plate outline constitutes a closed region suitable for cutting. Preferably, the plate boundary is represented as a set of boundary segments arranged in sequence, and the distance between the first and last endpoints of the boundary segment set is judged. When the distance between the first and last endpoints is not greater than a preset closure threshold, it is determined to be closed; the preset closure threshold is preferably 2 mm to balance numerical accuracy and plate manufacturing tolerance. For plate boundaries with self-intersections, this embodiment determines that the closure fails, and the corresponding plate is marked as an abnormal plate for subsequent location of the failed constraint type. Through the above closure check, this embodiment obtains the plate closure verification result corresponding to each plate and summarizes the closure check conclusions for the candidate plate types.

[0049] In this embodiment, a pairing consistency judgment is performed on the edges to be stitched corresponding to the stitching relationship constraint, and the length difference of the edges to be stitched is calculated and compared with an error threshold to generate stitching pairing consistency verification results and stitching edge length error verification results, respectively. The edge to be stitched refers to the boundary segment of the plate in the candidate plate type that needs to be stitched according to the stitching relationship constraint declaration; the pairing consistency judgment is used to confirm that each edge to be stitched has one and only one paired edge, and the plate to which the paired edge belongs is consistent with the stitching relationship constraint declaration. In this embodiment, the pairing entries of the edges to be stitched corresponding to the stitching relationship constraint are traversed one by one in the candidate plate type. When an edge to be stitched is missing, repeatedly paired, or the pairing direction conflicts, the stitching pairing consistency is determined to fail, and the corresponding pairing entry identifier is recorded as the basis for failure. For the length error judgment, this embodiment calculates the boundary segment length difference for each edge to be stitched and compares the length difference with a preset error threshold, preferably 5 mm; when the length difference is greater than 5 mm, the stitching edge length error is determined to fail, and the corresponding edge to be stitched pairing entry is recorded as an error entry. Finally, this embodiment generates a manufacturability verification result based on the plate closure verification result, the stitching pairing consistency verification result, and the stitching edge length error verification result. When any verification result fails, the manufacturability verification result is determined to fail, thereby triggering the constraint correction amount generation in step 700.

[0050] Furthermore, in step 700 of this embodiment, when the manufacturability verification result fails, a constraint correction amount is generated. The structural constraint set is updated based on the constraint correction amount, and the constraint solution is re-executed until the manufacturability verification result passes. The failure constraint type refers to the constraint category that leads to failure, deduced from the manufacturability verification result. This embodiment includes at least the closure-related constraint type, the pairing consistency-related constraint type, and the length error-related constraint type. This embodiment locates the target constraint item that matches the failure constraint type in the structural constraint set based on the record information of the failure item. Subsequently, this embodiment calculates the correction direction and correction magnitude for the target constraint item to generate the constraint correction amount, wherein the correction direction is used to indicate the direction of the constraint parameter that should be relaxed or tightened, and the correction magnitude is used to indicate the size of the adjustment amount. For the length error-related constraint type, this embodiment preferably determines the correction magnitude according to the length difference, and the single correction magnitude does not exceed 0.20 of the allowable change range of the corresponding parameter of the target constraint item, so as to avoid structural oscillation caused by a one-time correction. For the pairing consistency-related constraint type, this embodiment preferably increases the priority of the pairing constraint by correcting the constraint weight of the target constraint item, and the single increment of the constraint weight is preferably 1 to 2. Finally, in this embodiment, the constraint correction amount is written into the structural constraint set to obtain the updated structural constraint set, and the constraint solution in step 500 is re-executed based on the updated structural constraint set; when the new candidate plate type obtained after re-solving passes the manufacturability verification, the target plate type file corresponding to the new candidate plate type and the stitching relationship data corresponding to the target plate type file are output, thereby forming a manufacturability-driven closed-loop solution mechanism.

[0051] Corresponding to the above methods, such as Figure 3 As shown, this embodiment also provides a garment pattern intelligent generation system based on multimodal visual recognition, including: The multimodal input acquisition and preprocessing unit is used to acquire the multimodal input data corresponding to the garment to be generated, and to perform scale normalization and coordinate reference unification processing on the multimodal input data to output the multimodal input dataset; the multimodal input data includes garment image data and garment text description data; The multimodal visual recognition unit is used to perform multimodal visual recognition on the multimodal input dataset and generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from clothing image data; The semantic-to-structure mapping and plate structure construction unit is used to call a pre-built semantic-to-structure mapping knowledge base, convert the multimodal recognition result set into a set of structural semantic elements and a set of geometric anchor points, and construct a plate structure diagram based on the set of structural semantic elements and the set of geometric anchor points; the plate structure diagram is used to describe the set of plates, the set of structural lines, and the connection relationships between plates; The structural constraint analysis and solution input building unit is used to analyze the plate structure diagram into a set of structural constraints and associate the set of structural constraints with the set of geometric anchor points to form constraint solution input data; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. The constraint-solving plate type generation unit is used to select an initial plate type template from a preset plate type template library based on a set of structural semantic elements, and to perform constraint solving on the initial plate type template according to the constraint solving input data to obtain a set of plate type parameters that satisfy the set of structural constraints, and to generate candidate plate types based on the set of plate type parameters. The board type manufacturability verification unit is used to perform manufacturability verification on candidate board types and obtain manufacturability verification results; The constraint correction and plate shape iteration output unit is used to generate constraint correction amounts when the manufacturability verification result fails, update the structural constraint set based on the constraint correction amounts and re-execute the constraint solution, and output the target plate shape file and the stitching relationship data corresponding to the target plate shape file when the manufacturability verification result passes.

[0052] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0053] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent generation of garment patterns based on multimodal visual recognition, characterized in that, include: The multimodal input data corresponding to the garment to be generated is obtained, and scale normalization and coordinate reference unification processing are performed on the multimodal input data to output a multimodal input dataset; the multimodal input data includes garment image data and garment text description data; Multimodal visual recognition is performed on the multimodal input dataset to generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from the clothing image data; The pre-set semantic-to-structure mapping knowledge base is invoked to convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, and a plate structure diagram is constructed based on the structural semantic element set and the geometric anchor point set. The plate structure diagram is used to describe the plate set, the structural line set, and the connection relationship between the plates; The plate structure diagram is parsed into a set of structural constraints, and the set of structural constraints is associated with the set of geometric anchor points to form constraint solution input data; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. An initial plate template is selected from the preset plate template library based on the set of structural semantic elements, and constraint solving is performed on the initial plate template according to the constraint solving input data to obtain a set of plate parameters that satisfy the set of structural constraints. Candidate plate types are generated based on the set of plate parameters. Perform manufacturability verification on the candidate board type to obtain the manufacturability verification result; If the manufacturability verification result fails, a constraint correction amount is generated. The structural constraint set is updated based on the constraint correction amount, and the constraint solution is re-executed. If the manufacturability verification result passes, the target plate shape file and the stitching relationship data corresponding to the target plate shape file are output.

2. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, The pre-built semantic-to-structure mapping knowledge base is invoked to convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, including: The style category label is mapped to the target pattern category identifier, and the structural semantic element entries that match the target pattern category identifier are retrieved in the semantic-to-structure mapping knowledge base to generate the structural semantic element set; The appearance attribute tags are mapped to structural line type identifiers and plate type identifiers, and the structural line type identifiers and plate type identifiers are written into the structural semantic element set; The local contour key point set is grouped according to the preset key point topological relationship and fitted to obtain the contour segment set, so as to extract the anchor point coordinates used to constrain the position of the structure line from the contour segment set and generate the geometric anchor point set.

3. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, Based on the set of structural semantic elements and the set of geometric anchor points, a plate-type structure diagram is constructed, including: Based on the structural semantic element set, generate a plate node set and a structural line node set, and write a plate type identifier and a structural line type identifier for the plate node set and the structural line node set respectively; Based on the set of structural semantic elements, determine the set of associated edges between the set of plate nodes and the set of structural line nodes, so as to obtain the structural graph connection relationship that represents the belonging relationship of the set of structural lines within the set of plate nodes; Based on the geometric anchor point set, anchor point coordinate attributes are written to the structural line node set so that the plate structure diagram can simultaneously represent structural semantic information and geometric positioning information.

4. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, The plate structure diagram is parsed into a set of structural constraints, including: Based on the structural line type identifier and anchor point coordinate attributes in the plate structure diagram, structural line position constraints are generated; the structural line position constraints are used to limit the positioning relationship of the corresponding structural line on the corresponding plate. Based on the plate type identifier in the plate structure diagram and its connection relationship with the structure diagram, plate geometric relationship constraints are generated; these plate geometric relationship constraints are used to limit the relative positional and dimensional relationships between different plates. Based on the symmetry identifiers in the plate structure diagram, symmetry constraints are generated; these symmetry constraints are used to limit the geometric consistency of symmetrical plates or symmetrical structural lines. Based on the connection relationships in the plate structure diagram, stitching relationship constraints are generated; these constraints are used to limit the pairing relationship of the edges to be stitched and ensure consistency with the stitching direction.

5. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, Associating the set of structural constraints with the set of geometric anchor points to form constraint solution input data includes: Write a constraint action object identifier for each constraint in the set of structural constraints; the constraint action object identifier is used to indicate the plate node or structural line node corresponding to the constraint. The anchor point coordinates in the geometric anchor point set are bound to the constraint object identifier to form an anchor point location item for constraint solving. Write constraint weights for at least some of the constraints in the set of structural constraints to form the constraint solution input data, which includes constraint terms, anchor point location terms, and constraint weights.

6. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, Selecting an initial template from a pre-defined template library based on the set of structural semantic elements includes: The pattern template library contains a pre-set pattern template index table for different style categories; the pattern template index table is used to declare the mapping relationship between style category labels and initial pattern templates; The target template identifier is retrieved from the template template index table based on the target template type identifier in the structural semantic element set. The plate parameter framework and structural line parameter framework corresponding to the target template identifier are called as the initial plate template.

7. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, Perform constraint solving on the initial plate template based on the constraint solving input data, including: The plate shape parameters in the initial plate shape template are used as the set of variables to be solved; the set of variables to be solved includes plate contour parameters and structural line parameters. The set of structural constraints is converted into a set of constraint equations for the set of variables to be solved, and the anchor point location terms corresponding to the set of geometric anchor points are written into the set of constraint equations. When the constraint equations satisfy the preset convergence conditions, the plate shape parameter set is output, and the candidate plate shape is generated based on the plate shape parameter set.

8. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, Perform manufacturability verification on the candidate board type to obtain manufacturability verification results, including: A closure check is performed on each plate in the candidate plate type to generate a plate closure verification result; Perform a pairing consistency judgment on the edges to be stitched corresponding to the stitching relationship constraints in the candidate plate type to generate a stitching pairing consistency verification result; The length difference of the edge to be sewn in the candidate plate type is calculated, and the length difference is compared with a preset error threshold to generate a sewn edge length error verification result. The manufacturability verification result is generated based on the plate closure verification result, the stitch pairing consistency verification result, and the stitch edge length error verification result.

9. The intelligent garment pattern generation method based on multimodal visual recognition according to claim 1, characterized in that, When the manufacturability verification result fails, a constraint correction amount is generated, the structural constraint set is updated based on the constraint correction amount, and the constraint solution is re-executed, including: Based on the manufacturability verification results, determine the type of failure constraint, and locate the target constraint item in the structural constraint set that matches the type of failure constraint; The correction direction and correction magnitude are calculated for the target constraint to generate the constraint correction amount, which is used to adjust the anchor point positioning item or constraint weight corresponding to the target constraint. The constraint correction amount is written into the structural constraint set to obtain an updated structural constraint set, and the constraint solution is re-executed based on the updated structural constraint set until the manufacturability verification result passes.

10. A smart garment pattern generation system based on multimodal visual recognition, characterized in that, include: A multimodal input acquisition and preprocessing unit is used to acquire multimodal input data corresponding to the garment to be generated, and to perform scale normalization and coordinate reference unification processing on the multimodal input data to output a multimodal input dataset; the multimodal input data includes garment image data and garment text description data; A multimodal visual recognition unit is used to perform multimodal visual recognition on the multimodal input dataset and generate a multimodal recognition result set; the multimodal recognition result set includes at least style category labels, appearance attribute labels, and a set of local contour key points extracted from the clothing image data; The semantic-to-structure mapping and plate structure construction unit is used to call a pre-set semantic-to-structure mapping knowledge base, convert the multimodal recognition result set into a structural semantic element set and a geometric anchor point set, and construct a plate structure diagram based on the structural semantic element set and the geometric anchor point set. The plate structure diagram is used to describe the plate set, the structural line set, and the connection relationship between the plates; The structural constraint analysis and solution input construction unit is used to analyze the plate structure diagram into a set of structural constraints and associate the set of structural constraints with the set of geometric anchor points to form constraint solution input data; the set of structural constraints includes at least structural line position constraints, plate geometric relationship constraints, symmetry constraints, and stitching relationship constraints. The constraint-solving plate type generation unit is used to select an initial plate type template from a preset plate type template library based on the set of structural semantic elements, and perform constraint solving on the initial plate type template according to the constraint solving input data to obtain a set of plate type parameters that satisfy the set of structural constraints, and generate candidate plate types based on the set of plate type parameters. A board type manufacturability verification unit is used to perform manufacturability verification on the candidate board type and obtain manufacturability verification results; The constraint correction and plate shape iteration output unit is used to generate a constraint correction amount when the manufacturability verification result fails, update the structural constraint set according to the constraint correction amount and re-execute the constraint solution, and output the target plate shape file and the stitching relationship data corresponding to the target plate shape file when the manufacturability verification result passes.

Images