Intelligent costume design method fusing AI and CAD technologies

Abstract

The invention provides an intelligent costume design method fusing AI and CAD technologies, and belongs to the technical field of intelligent design, and the method comprises the steps: building a parameterized CAD engine based on NURBS curve modeling and a topological correlation algorithm; performing intelligent classification and component segmentation on the input sketch, and obtaining a matched parameterized model based on a parameterized CAD engine to generate a reusable component; matching and recommending a plurality of reusable components based on an AI technology to obtain a similar version, analyzing the sketch and the similar version, and recommending an optimal component combination scheme; a distributed collaborative design mechanism is constructed, when the same part of the combined sample clothes of the optimal part combination scheme modified by multiple persons is received, conflict points are automatically prompted through operation log comparison, version combination suggestions are provided, and a 3D sample clothes effect picture is output for a user to judge. The problems that traditional design is low in efficiency, prone to making mistakes and difficult to cooperate are solved, and technical support is provided for intelligent design innovation.

Description

Technical Field

[0001] The present invention relates to the field of intelligent design technology, and in particular to an intelligent clothing design method that integrates AI and CAD technologies. Background Art

[0002] The traditional clothing design process has long been constrained by multi-dimensional technical bottlenecks: first, pattern design is highly dependent on the designer's experience, and style adjustments require manual and repeated modification of patterns. Optimizing a single pattern alone takes 3-5 days, seriously restricting the efficiency of design iteration; second, there are structural defects in data management. Data such as style renderings, pattern parameters, fabric properties and process specifications are stored in different systems. Cross-departmental collaboration often leads to the loss of key information due to data barriers, and the reuse rate of historical design resources is less than 15%; third, the existing CAD system has limited intelligence and lacks intelligent recommendation mechanisms based on historical cases and automatic component combination capabilities, making it difficult to support the 2-3 design iteration requirements of the fast fashion field per week; fourth, there is a gap in data connectivity in the industrial chain. The parametric model on the design end and the process standards on the production end are not unified in data format, resulting in a process parameter conversion error rate of more than 20%, which in turn leads to a long clothing design cycle and difficulty in quickly responding to design needs.

[0003] Therefore, the present invention proposes an intelligent clothing design method that integrates AI and CAD technologies. Summary of the Invention

[0004] The present invention provides an intelligent clothing design method that integrates AI and CAD technologies to solve the above-mentioned technical problems.

[0005] The present invention proposes an intelligent clothing design method integrating AI and CAD technologies, comprising: Step 1: Build a parametric CAD engine based on NURBS curve modeling and topological association algorithms. This engine contains a structured pattern database of several historical styles, which is stored in a three-level index by style, category, and component type. When any pattern parameter is adjusted, the topological association rules automatically trigger the linkage modification of related parameters. Step 2: Intelligently classify and segment the input sketch, and generate reusable components based on the matching parametric model obtained by the parametric CAD engine; Step 3: Based on AI technology, several reusable components are matched and recommended to obtain similar patterns, and the sketch and similar patterns are analyzed to recommend the optimal component combination solution; Step 4: Build a distributed collaborative design mechanism. When receiving multiple people modifying the same component of the combined sample of the optimal component combination solution, the conflict points are automatically prompted through operation log comparison and version merging suggestions are provided, and a 3D sample rendering is output for users to decide.

[0006] Preferably, the sketch and similar templates are analyzed to recommend the optimal component combination solution, including: Extracting global style features of the sketch and local features of each reusable component to construct a multidimensional feature vector including color features, texture features, and contour features; Determine the layout features of each similar layout to obtain the layout complexity coefficient of the corresponding similar layout, and splice the reusable components in the order of the coefficients according to the component placement positions in the sketch to obtain a complete layout; Determining a line connection diagram between the sketch and the complete layout based on the spatial relationship between the parts; Extracting three-dimensional geometric parameters of each line connection point in the line connection diagram to determine the geometric feature distribution; Extracting style features of the sketch and the line connection diagram and performing feature alignment processing to obtain a difference feature vector; Determine the designer's potential preference type and obtain pattern parameters of historical designs of similar patterns, and construct an embedding vector of the potential preference type; Determine the preference association between the current design of the sketch and each embedding vector, and construct a pattern knowledge graph based on the style similarity and geometric matching between the sketch and similar patterns, wherein the nodes of the pattern knowledge graph are patterns and the edges are preference association, style similarity, and geometric matching; Modifying the pattern knowledge graph based on the geometric feature distribution and the difference feature vector, and searching the design database to obtain a complete candidate set; Transferring the multi-dimensional feature vector of the sketch to each complete candidate template to obtain a preliminary template; Analyze the functional compatibility and aesthetic coordination between the components of the preliminary pattern and output an optimal component combination scheme, wherein the functional compatibility is related to the engineering pattern parameters, and the aesthetic coordination is related to color theory and style trends.

[0007] Preferably, the input sketch is intelligently classified and segmented into components, and a matching parametric model is obtained based on a parametric CAD engine to generate reusable components, including: A deep learning model combination architecture based on deep reinforcement learning dynamic regulation performs feature extraction and semantic segmentation on sketches; When using the U-Net network to perform semantic segmentation on sketches, its encoder-decoder structure fuses multi-scale features through skip connections, and introduces hierarchical energy constraints in the segmentation process. The energy function expression is expanded to : ; in, is the hierarchical correlation potential energy function; is a global semantic label; It is a local component label; is the hierarchical constraint coefficient; For sketches; is the segmentation label; is the set of adjacent pixels of the i-th pixel in the sketch; is the unit potential energy function, used to calculate a single segmentation label element The potential energy between it and the input x; is a pairwise potential energy function used to calculate two adjacent segmentation label elements , yj and the potential energy between input x; When extracting features from the segmented part contours, construct a spatiotemporal attention-enhanced sketch vector ,in, is the rate of change of curvature based on the time dimension, and , is the contour tangent vector of point s; is the scale factor; Encode textures; Generate the index address Dz of the sketch feature vector V1, including: Calculate the sketch feature vector V1 and component type The product of ; Calculate styles separately , Category , component type Adaptive hash value, and perform hash value fusion to generate dynamic index ,in, is a hash function; Based on style , Category , component type , the index function of the sketch feature vector V1; ⊕ is the exclusive OR symbol; is the Hadamard product symbol; Indexing the index address Dz in a structured pattern database based on the parametric CAD engine, and screening reference vectors with a matching degree greater than a preset degree to obtain a parametric model; The sketch feature VI is adaptively adjusted based on the parameterized model, and a reusable component is generated through topological association rules.

[0008] Preferably, the model combination architecture consists of a ResNet-152 network and a MaskR-CNN network.

[0009] Preferably, determining a line connection diagram between the sketch and the complete layout based on the spatial relationship between the parts includes: Extracting the three-dimensional bounding boxes and two-dimensional projection contours of the sketch and the complete layout to construct a component topology relationship diagram; Constructing a line constraint optimization model based on the topological relationship graph, wherein the line control points of the sketch and the line control points of the complete layout are used as optimization variables to define an objective function and constraint conditions; The maximum flow-minimum cut algorithm is used to solve the line connection scheme of the line constraint optimization model to generate a line connection graph.

[0010] Preferably, it also includes: After outputting the 3D sample rendering for the user to decide, a QR code is generated and sent to the supplier; After the supplier scans the QR code received by the supply end, the supplier can view the fabric drape effect of the 3D sample garment and confirm it.

[0011] Preferably, conflict points are automatically indicated through operation log comparison and version merge suggestions are provided, including: Determine the operation permissions of each designer based on the data access control list; Based on the operation log, the modification time and parameters of each designer on the reusable component are recorded. When it is detected that the same component is modified by multiple people at the same time, the conflict vector set at each moment is constructed. , where n1 represents the number of people involved in the conflict; represents the conflict vector of the j1th conflicting person, They represent the modification time t, trajectory change information, and operation permissions respectively; Continuously obtain conflict vector sets at N0 consecutive moments, and map the trajectory change information to a blank coordinate system to lock the changed position point at the first moment. Then, construct a sub-point set for the same changed position point at the first moment based on the operation authority, and extract the trajectory points at each consecutive moment from the mapped coordinate system to obtain a change trajectory set for each changed position point. Each trajectory line in the change trajectory set is related to the operation authority and modification time. Inputting the change trajectory set into a trajectory analysis model to determine a recommended line corresponding to the change position point; Based on the recommended lines of all the first independent change locations, a version merge suggestion is generated and sent to the user for approval.

[0012] Preferably, it also includes: Build a design modification history database, timestamp and hash each design change to form an unalterable version traceability chain. When a dispute over process parameter conversion occurs, the complete design modification trajectory can be queried to provide a reminder.

[0013] Compared with the prior art, the present invention has the following advantages: By building a parametric engine foundation, intelligent sketch conversion, AI innovation recommendation, and collaborative process guarantee, a full-process digital system for clothing design has been built. This not only solves the pain points of traditional design such as low efficiency, prone to errors, and difficult collaboration, but also provides technical support for intelligent design innovation.

[0014] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of an intelligent clothing design method that integrates AI and CAD technologies in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0018] The present invention proposes an intelligent clothing design method that integrates AI and CAD technology. Figure 1 As shown, including: Step 1: Build a parametric CAD engine based on NURBS curve modeling and topological association algorithms. This engine contains a structured pattern database of several historical styles, which is stored in a three-level index by style, category, and component type. When any pattern parameter is adjusted, the topological association rules automatically trigger the linkage modification of related parameters. Step 2: Intelligently classify and segment the input sketch, and generate reusable components based on the matching parametric model obtained by the parametric CAD engine; Step 3: Based on AI technology, several reusable components are matched and recommended to obtain similar patterns, and the sketch and similar patterns are analyzed to recommend the optimal component combination solution; Step 4: Build a distributed collaborative design mechanism. When receiving multiple people modifying the same component of the combined sample of the optimal component combination solution, the conflict points are automatically prompted through operation log comparison and version merging suggestions are provided, and a 3D sample rendering is output for users to decide.

[0019] In this embodiment, NURBS curve modeling is based on non-uniform rational B-splines (Non-Uniform Rational B-Splines), a mathematical model that accurately describes complex curves / surfaces and can flexibly express straight lines, circular arcs, and free-form curves. Specifically, NURBS curve construction is implemented using an open source library (such as OpenCASCADE), and three core parameters are defined: control vertex Pi, node vector U, and degree p. For example, when drawing the armhole arc of a dropped shoulder sleeve, armhole curves of different curvatures can be quickly generated by adjusting the coordinates of control vertex P3 and combining the node vector U = [0, 0, 0, 0.2, 0.5, 1, 1, 1].

[0020] The topological association algorithm establishes association rules between pattern parameters (such as increasing garment length → synchronously adjusting the hem curvature and dart position) to ensure the consistency of parameter modifications. Specifically, it builds an association rule library and uses a directed acyclic graph (DAG) to store parameter dependencies, such as garment length → hem curvature → seam width. For example, when a designer modifies the garment length parameter (from 60cm to 65cm), the algorithm automatically triggers the DAG traversal and updates 12 associated parameters according to the rules, including the hem curvature (from 1.2rad to 1.5rad) and the dart position (offset by 0.3cm).

[0021] In this embodiment, the structured pattern database stores historical patterns according to the three-level index of style (such as minimalist / retro) - category (such as shirt / jacket) - component type (such as collar type / sleeve type) to achieve fast retrieval. Specifically, MongoDB is used to build a database and design a composite index Index=(Style, Category, PartType). For example, when searching for a retro-style-dress-square-neck pattern, the storage path DB / retro-style / dress / square-neck / param.json is directly located through the index, and the time taken is <10ms.

[0022] In this embodiment, intelligent classification and component segmentation uses a deep learning model to identify sketch styles and segment components (collar type, sleeve type, etc.). Specifically, the training data set is: 50,000+ hand-drawn sketches are labeled (including style labels and component outlines), and the model is optimized using transfer learning. For example, a sketch of ink-and-wash style Hanfu is input, and the model outputs the classification result (style = ink-and-wash style, category = Hanfu), and the cross collar, wide sleeves, and skirt are segmented.

[0023] Parametric model matching is to match the segmented component contour with the NURBS parametric model in the structured database to generate reusable digital components. Specifically, the cosine similarity algorithm is used to match contour features (such as curvature and proportion), and the threshold is set to 0.85 (a match is determined if it is higher than the threshold). For example, the segmented wide sleeve contour has a similarity of 0.92 with the NURBS model feature of the wide sleeves of Hanfu in the database. The model is automatically matched to generate a reusable wide sleeve parametric component (including parameters such as sleeve length, sleeve width, and cuff curvature).

[0024] In this embodiment, AI matching recommendation uses collaborative filtering and graph neural network (GNN) to recommend similar styles and calculate the compatibility of components (such as the adaptability of square collar + pipa sleeves). Specifically, a GNN model is constructed, with nodes as components (collar type, sleeve type) and edges as compatibility scores (trained based on historical design matching data). For example, if a cross collar + wide sleeve component is input, the GNN calculates that the compatibility scores of the ruqun and the mamian skirt are 0.91 and 0.78 respectively, and recommends a similar style of cross collar + wide sleeve + ruqun.

[0025] In this embodiment, the optimal component combination analysis comprehensively considers style consistency, structural rationality, and fashion trends to screen the optimal component combination. Specifically, a multi-objective optimization algorithm (NSGA-II) is used with the optimization objectives of style matching (>0.8), structural conflict number (<2), and fashion index (>0.7). For example, for components such as a cross collar, wide sleeves, a ruqun, and a horse-faced skirt, the algorithm selects a cross collar + wide sleeves + ruqun (style matching 0.92, no structural conflict, and fashion index 0.85) as the optimal combination.

[0026] In this embodiment, the distributed collaborative design mechanism is that multiple people can modify the design online at the same time, and achieve collaboration through version control and permission management. Specifically, Git distributed version control is used in combination with the RBAC permission model (designers can modify patterns, technicians can modify parameters, and suppliers can only read). For example: Designer A modifies the length of the skirt, and technician B adjusts the seam width. The system automatically merges non-conflicting modifications and marks conflicting items for adjudication.

[0027] In this embodiment, operation logs and conflict resolution record modification times and parameter changes, and use conflict resolution algorithms (such as three-way merge) to generate version recommendations. The log format is: Log = (User, Time, Param, ΔValue), which is stored as blockchain evidence (tamper-proof). For example, designer A (10:00) changes the length of a skirt from 80 to 85 cm, and designer B (10:02) changes it to 82 cm. The algorithm compares the logs and recommends adopting A's modification (first modification takes priority), or a compromise of 83 cm (requires manual judgment), and generates a 3D rendering for comparison.

[0028] The beneficial effects of the above technical solution are: through the parametric engine foundation, intelligent conversion of sketches, AI innovation recommendation, and collaborative process guarantee, a full-process digital system for clothing design is constructed, which not only solves the pain points of traditional design such as low efficiency, easy errors, and difficulty in collaboration, but also provides technical support for intelligent design innovation.

[0029] The present invention proposes an intelligent clothing design method that integrates AI and CAD technologies, analyzes the sketch and similar patterns, and recommends the optimal component combination scheme, including: Extracting global style features of the sketch and local features of each reusable component to construct a multidimensional feature vector including color features, texture features, and contour features; Determine the layout features of each similar layout to obtain the layout complexity coefficient of the corresponding similar layout, and splice the reusable components in the order of the coefficients according to the component placement positions in the sketch to obtain a complete layout; Determining a line connection diagram between the sketch and the complete layout based on the spatial relationship between the parts; Extracting three-dimensional geometric parameters of each line connection point in the line connection diagram to determine the geometric feature distribution; Extracting style features of the sketch and the line connection diagram and performing feature alignment processing to obtain a difference feature vector; Determine the designer's potential preference type and obtain pattern parameters of historical designs of similar patterns, and construct an embedding vector of the potential preference type; Determine the preference association between the current design of the sketch and each embedding vector, and construct a pattern knowledge graph based on the style similarity and geometric matching between the sketch and similar patterns, wherein the nodes of the pattern knowledge graph are patterns and the edges are preference association, style similarity, and geometric matching; Modifying the pattern knowledge graph based on the geometric feature distribution and the difference feature vector, and searching the design database to obtain a complete candidate set; Transferring the multi-dimensional feature vector of the sketch to each complete candidate template to obtain a preliminary template; Analyze the functional compatibility and aesthetic coordination between the components of the preliminary pattern and output an optimal component combination scheme, wherein the functional compatibility is related to the engineering pattern parameters, and the aesthetic coordination is related to color theory and style trends.

[0030] In this embodiment, global style features are abstract descriptions summarizing the overall design style of a sketch (e.g., "retro" or "minimalist"), encompassing color tones, texture patterns, and silhouette trends. Local component features are detailed features of each reusable component (collar, sleeve, or hem), such as "collar angle curvature of a square collar" and "pleat density of lantern sleeves." For example, a retro dress sketch has global style features: low-saturation colors (primarily camel), jacquard texture, and A-line silhouette. Local component features include: square collar (120° collar angle), lantern sleeves (pleat density of 5 / 10cm), and pleated skirt (pleat width of 3cm). Specifically, for global style, a pre-trained StyleNet algorithm extracts image style vectors from the sketch input to generate 256-dimensional style features. For local components, Mask R-CNN is used to segment the components, and keypoint detection (e.g., a modified version of OpenPose) is used to extract parameters such as contour corners and texture periodicity, generating 128-dimensional component features.

[0031] In this example, global and local features are fused to construct a high-dimensional vector containing color (LAB space values), texture (LBP texture encoding), and contour (curvature, inflection point coordinates). For example, the feature vector of a vintage dress includes: color: camel (L=60, A=10, B=40), jacquard texture LBP encoding (binary sequence). Contour: the coordinates of the three corner points of the square collar and the 10 curvature values ​​of the hem. Specifically, the Python OpenCV library is used to extract LAB color values, the Skimage library is used to calculate the LBP texture, and OpenCASCADE NURBS curve curvature analysis is combined to form a 512-dimensional feature vector.

[0032] In this embodiment, pattern features are parameters that describe the complexity of the pattern structure (such as the number of darts, the number of dividing line layers, and the proportion of asymmetric designs). The pattern complexity coefficient is a complexity score calculated using the entropy method (a higher value indicates a more complex structure). For example, similar pattern screening (vintage dress): Version A (complexity coefficient 0.8): 3 darts, 2 dividing lines, symmetrical design.

[0033] Version B (complexity coefficient 0.6): 1 dart, 1 dividing line, symmetrical design.

[0034] Select pattern A in descending order of coefficients, and assemble a square collar, lantern sleeves, and pleated skirt. Specifically, use Graph Theory to analyze the pattern dividing line network, and calculate the ratio of the number of network nodes (darts, pockets, etc.) to the number of edges (dividing lines) as the complexity coefficient. In this embodiment, the complete pattern is formed by assembling the parametric models of reusable components into a complete pattern according to the placement of the sketch components (e.g., collar on top, sleeve on the side), and generating a 2D / 3D rendering, for example: 1. Vintage Dress Stitching: Square collar model (parameters: collar width 20cm, collar depth 8cm) → Place on top of the body.

[0035] Lantern sleeve model (parameters: sleeve length 35cm, cuff circumference 40cm) → spliced ​​to the shoulder of the body.

[0036] Pleated skirt model (parameters: pleat width 3cm, skirt circumference 120cm) → spliced ​​to the bottom of the body.

[0037] Specifically: Based on a parametric CAD engine (such as the NURBS engine in step 1), components are automatically spliced ​​together using topological association rules, and a 3D preview is generated using Three.js.

[0038] In this embodiment, the line connection diagram is a schematic diagram that describes the line connection relationship between components, such as the connection curve between the collar bottom edge and the upper edge of the body, and the stitching relationship between the armhole line and the sleeve cap line. For example, the line connection of a vintage dress: the square collar bottom edge curve (NURBS parameters: control points (0,20), (10,18), (20,20)) → seamlessly connects with the upper edge curve (control points (0,18), (10,16), (20,18)).

[0039] In this embodiment, the three-dimensional parameters of the connection between three-dimensional geometric parameter lines (such as the curvature radius of the connection curve and the normal vector angle) are statistically analyzed by the distribution law (mean and variance) of these parameters. For example, the parameters of the connection of a vintage dress are: Collar-body connection curve: curvature radius 15cm (mean), variance 2cm².

[0040] Armhole-sleeve connection curve: normal vector angle 10° (mean), variance 1°².

[0041] In this embodiment, the style features (such as color distribution and texture direction) of the sketch and the line connection map are aligned, and the difference vector of the misaligned part (such as color deviation ΔL=2, ΔA=1) is calculated. For example, the style difference of a vintage dress is: Sketch jacquard texture direction (45°) → connection pattern texture direction (60°), difference vector Δθ = 15°.

[0042] Specifically, Procrustes is used to analyze the aligned style features and calculate the residual as the difference vector. The formula is: Difference vector = sketch features − aligned connection graph features.

[0043] In this embodiment, a clustering algorithm (such as LDA) is used to mine the potential preferences of designers' historical designs (such as "prefer high collar design" and "prefer asymmetrical skirt"), and the preference type is encoded into a low-dimensional embedding vector. For example, the designer's historical design clustering: Preference type 1 (coded [0.8, 0.2]): 80% of designs have high collars and 20% have asymmetrical hems.

[0044] Preference type 2 (coded [0.3, 0.7]): 30% of designs have high collars and 70% have asymmetrical hems.

[0045] Specifically: Use Python's Gensim library to train the Word2Vec model and encode preference labels (such as "high collar" and "asymmetric") into 2D embedding vectors.

[0046] In this embodiment, preference association is to calculate the association between the current design and the historical preference type (such as cosine similarity), and combine style similarity and geometric matching to construct a knowledge graph with "pattern" as a node and "association / similarity" as an edge. For example, the association between vintage dresses and preference type 1 is: Cosine similarity = 0.9 (high collar design match), style similarity = 0.85 (retro style match), geometric matching = 0.92 (line connection match).

[0047] In this embodiment, the knowledge graph is modified by using geometric feature distribution and difference feature vectors to modify the edge weights of the knowledge graph (for example, reducing the weights of patterns with large style differences). Highly matching patterns are retrieved from the database as candidate sets. For example, the candidate set for vintage dresses is screened: After correction, the weight of version A (original weight 0.8) is reduced to 0.7 due to the style difference Δθ=15°; the weight of version B (original weight 0.7) is increased to 0.8 due to the high geometric matching degree.

[0048] In this embodiment, the multi-dimensional feature vector (color, texture, outline) of the sketch is transferred to the candidate pattern to generate a preliminary pattern that retains the candidate pattern structure and integrates the sketch details, for example: candidate pattern B (straight skirt) + sketch features (jacquard texture, square collar) → preliminary pattern: straight skirt structure + jacquard texture + square collar.

[0049] In this embodiment, functional compatibility refers to the matching degree of engineering parameters between components (such as whether the sleeve movement meets ergonomic requirements), where compatibility = 1 − standard value | parameter − standard value |.

[0050] Aesthetic harmony: whether the combination of color, texture, and silhouette conforms to popular trends (such as retro style + Morandi color scheme). Harmony = SoftMax(trend model(eigenvector)).

[0051] Vintage Dress Evaluation: Function: The lantern sleeves allow for ample movement (armhole depth 18cm) to accommodate hand-lifting needs (compatibility score 0.9).

[0052] Aesthetics: Camel + jacquard + A-line skirt is in line with the 2025 retro trend (coordination score 0.85).

[0053] The beneficial effects of the above technical solution are: through feature quantification, knowledge graph, intelligent migration, and two-dimensional evaluation, an AI-driven closed loop for clothing design is constructed, providing quantifiable and traceable technical support for innovative design.

[0054] The present invention proposes an intelligent clothing design method that integrates AI and CAD technologies. The method intelligently classifies and segments input sketches, and generates reusable components based on a matching parametric model obtained by a parametric CAD engine. The method includes: A deep learning model combination architecture based on deep reinforcement learning dynamic regulation performs feature extraction and semantic segmentation on sketches; When using the U-Net network to perform semantic segmentation on sketches, its encoder-decoder structure fuses multi-scale features through skip connections, and introduces hierarchical energy constraints in the segmentation process. The energy function expression is expanded to :

[0055] in, is the hierarchical correlation potential energy function; is a global semantic label; It is a local component label; is the hierarchical constraint coefficient; For sketches; is the segmentation label; is the set of adjacent pixels of the i-th pixel in the sketch; is the unit potential energy function, used to calculate a single segmentation label element The potential energy between it and the input x; is a pairwise potential energy function used to calculate two adjacent segmentation label elements , yj and the potential energy between input x; When extracting features from the segmented part contours, construct a spatiotemporal attention-enhanced sketch vector ,in, is the rate of change of curvature based on the time dimension, and , is the contour tangent vector of point s; is the scale factor; Encode textures; Generate the index address Dz of the sketch feature vector V1, including: Calculate the sketch feature vector V1 and component type The product of ; Calculate styles separately , Category , component type Adaptive hash value, and perform hash value fusion to generate dynamic index ,in, is a hash function; Based on style , Category , component type , the index function of the sketch feature vector V1; ⊕ is the exclusive OR symbol; is the Hadamard product symbol; Indexing the index address Dz in a structured pattern database based on the parametric CAD engine, and screening reference vectors with a matching degree greater than a preset degree to obtain a parametric model; The sketch feature VI is adaptively adjusted based on the parameterized model, and a reusable component is generated through topological association rules.

[0056] Preferably, the model combination architecture consists of a ResNet-152 network and a MaskR-CNN network.

[0057] In this embodiment, reinforcement learning (such as DQN) is used to dynamically adjust the collaborative weights of multiple deep learning models, allowing different models (ResNet-152 to extract global features and MaskR-CNN to segment components) to divide the work as needed. For example, when processing the "sketch of a Hanfu cross-collared top", reinforcement learning discovered that the "cross-collar outline is blurred" and automatically increased the weight of MaskR-CNN (from 0.5→0.7) to enhance segmentation accuracy.

[0058] In this embodiment, It is to capture the dynamic changes of the designer's brush strokes (such as "quickly sketching the corners of the collar"), It is the ratio of the component size to the standard version (e.g. "the width of the cross collar is 1.2 times that of the standard Hanfu collar"), It is the LBP (local binary pattern) encoding of the sketch texture (such as "texture cycle of Hanfu jacquard").

[0059] In this embodiment, the dynamic index Dz and the adaptive hash are the fusion of style (St), category (Ca), component type (Pa) and sketch vector V1 to generate a dynamic index, which solves the problem that traditional static indexes cannot adapt to diverse sketches. For example: Pa⊙V1 (Hadamard product): the "cross collar type" and "collar width 1.2 times" features are fused, and the dynamic index Dz=H(St)⊕H(Ca)⊕H(Pa⊙V1) (XOR fusion hash value).

[0060] In this embodiment, the dynamic index Dz is used to search the structured pattern database (such as the "retro style-Hanfu-cross collar" database) to screen the parametric models with a matching degree > 0.85 (such as the NURBS model of "cross collar width 16 cm, collar depth 8 cm").

[0061] In this embodiment, reusable component generation is to adaptively adjust the retrieved parametric model (e.g., changing the collar width from 17cm to 18cm) and generate a reusable digital component (e.g., "cross collar component V2.0") based on topological association rules (e.g., "collar width adjustment requires simultaneous modification of collar depth and collar curve"). For example, when adjusting the collar width of the cross collar model from 17cm to 18cm, the topological rules are automatically modified: The collar depth changes from 8cm to 8.5cm (maintaining the collar proportions), and the NURBS control points of the collar curve change from (0,10), (10,8), (20,10) to (0,10), (10,9), (20,10).

[0062] The beneficial effects of the above technical solution are: through dynamic model collaboration, spatiotemporal feature enhancement, intelligent index matching, and topological self-consistent adjustment, a complete technical chain from sketches to reusable components is constructed, which not only solves the pain points of inaccurate segmentation, inefficient retrieval, and difficult reuse in traditional design, but also provides accurate and efficient underlying support for intelligent clothing design.

[0063] The present invention proposes an intelligent clothing design method that integrates AI and CAD technologies, and determines a line connection diagram between the sketch and the complete layout based on the spatial relationship between parts, including: Extracting the three-dimensional bounding boxes and two-dimensional projection contours of the sketch and the complete layout to construct a component topology relationship diagram; Constructing a line constraint optimization model based on the topological relationship graph, wherein the line control points of the sketch and the line control points of the complete layout are used as optimization variables to define an objective function and constraint conditions; The maximum flow-minimum cut algorithm is used to solve the line connection scheme of the line constraint optimization model to generate a line connection graph.

[0064] In this embodiment, the 3D bounding box is the smallest 3D rectangular box that encloses a garment component. It is used to describe the position and range of the component in 3D space. The parameters include minimum and maximum coordinates. For example, taking the standard collar of a "commuter shirt" as an example, the 3D bounding box parameters are: x∈[10,30]cm,y∈[0,10]cm,z∈[50,60]cm, represents the left and right, front and back, top and bottom ranges of the collar type at the top of the body.

[0065] In this embodiment, the two-dimensional projection contour is the projection contour of the three-dimensional component onto a two-dimensional plane (such as the XY plane), which is usually a closed curve and is used to express the plane shape of the component (such as the top view contour of a collar). For example, the two-dimensional projection contour of a standard collar is a "trapezoid that is wide at the top and narrow at the bottom". After being projected onto the XY plane, the contour curve consists of four NURBS curves, and the control point coordinates are (10,0), (30,0), (15,8), and (25,8).

[0066] In this embodiment, the component topology relationship diagram is a directed graph with components as nodes and spatial relationships between components as edges, describing the connection methods of the components (such as "collar type and body fitting" and "sleeve type and body stitching"). For example, in the shirt component topology diagram, nodes are: standard collar, regular sleeve, body; edges are: "standard collar-body" edge is marked "fit" (z-axis overlap >90%), and "regular sleeve-body" edge is marked "stitching" (armhole arc and sleeve cap arc match >85%).

[0067] In this embodiment, the line control points are key coordinate points of the defined NURBS curve shape. By adjusting the control points, the curve profile (such as the shape of the hem arc) can be changed. For example, the NURBS curve of the hem arc is defined by five control points: (0,0), (10,−5), (20,−8), (30,−5), and (40,0). Adjusting the middle control points can change the hem arc.

[0068] In this embodiment, the objective function is a mathematical expression used to measure the quality of line connections. It usually includes minimizing geometric differences (such as the line distance between the sketch and the complete layout) and energy (such as the curve bending energy). The objective function is defined as: , where D is the Hausdorff distance, Ebend is the sum of squares of the second-order derivative of the curve, λ1=0.7,λ2=0.3, and is the weight coefficient.

[0069] In this embodiment, the constraints are restrictions to ensure reasonable line connection, such as "the coordinates of the lower edge of the collar and the upper edge of the body are consistent" and "the tangent lines of the armhole arc and the sleeve cap arc are continuous". For example, the constraints for the collar-body connection are: Geometric constraints: z-coordinate of the collar bottom edge zcollar,bottom = z-coordinate of the body top edge zbody,top; Process constraints: seam width ≥ 0.8cm.

[0070] In this embodiment, the maximum flow-minimum cut algorithm is a classic algorithm in graph theory. It finds the minimum cut by calculating the maximum flow of the network and is used to solve the problem of "optimal segmentation of line connection schemes". Here, the association of line control points is regarded as network flow, and the minimum cut corresponds to the optimal connection point. For example, the association of the control points of the armhole arc and the sleeve cap arc is regarded as a network edge, and the edge weight is the "adjustment cost". The maximum flow algorithm finds the cut set with the minimum cost and determines the optimal connection point position (such as aligning the third control point of the armhole arc with the fifth control point of the sleeve cap arc).

[0071] In this embodiment, the line connection diagram is a schematic diagram showing the line connection relationship between components, including the coordinates of the connection points, curve parameters, and process instructions (such as "sewing line type" and "seam width"). For example, the line connection diagram of the shirt sleeve-body shows: The armhole arc (control points A1-A5) and the sleeve cap arc (control points B1-B5) connect at point B3, with the coordinates of the connection point being (20,30); Marked with "rolled seam" process, seam width 1.2cm.

[0072] The beneficial effect of the above technical solution is: through the technical chain of spatial modeling, constraint optimization, and intelligent solution, accurate data can be connected from sketch to production.

[0073] The present invention proposes an intelligent clothing design method that integrates AI and CAD technologies, further comprising: After outputting the 3D sample rendering for the user to decide, a QR code is generated and sent to the supplier; After the supplier scans the QR code received by the supply end, the supplier can view the fabric drape effect of the 3D sample garment and confirm it.

[0074] The beneficial effect of the above technical solution is that it is convenient for suppliers to check and confirm by sending it to the supply side.

[0075] This paper proposes an intelligent clothing design method that integrates AI and CAD technologies. It automatically prompts conflict points and provides version merging suggestions through operation log comparison, including: Determine the operation permissions of each designer based on the data access control list; Based on the operation log, the modification time and parameters of each designer on the reusable component are recorded. When it is detected that the same component is modified by multiple people at the same time, the conflict vector set at each moment is constructed. , where n1 represents the number of people involved in the conflict; represents the conflict vector of the j1th conflicting person, They represent the modification time t, trajectory change information, and operation permissions respectively; Continuously obtain conflict vector sets at N0 consecutive moments, and map the trajectory change information to a blank coordinate system to lock the changed position point at the first moment. Then, construct a sub-point set for the same changed position point at the first moment based on the operation authority, and extract the trajectory points at each consecutive moment from the mapped coordinate system to obtain a change trajectory set for each changed position point. Each trajectory line in the change trajectory set is related to the operation authority and modification time. Inputting the change trajectory set into a trajectory analysis model to determine a recommended line corresponding to the change position point; Based on the recommended lines of all the first independent change locations, a version merge suggestion is generated and sent to the user for approval.

[0076] In this embodiment, the data access control list is a fine-grained permission management mechanism. By associating a list of allowed / denied operations for each user or role, it controls their access to data resources. Specifically, in the clothing design scenario, the ACL can define: designers have the right to edit pattern parameters (such as modifying clothing length and sleeve type), technicians can only adjust sewing parameters (such as stitch length and stitch type), and suppliers can only view fabric demand data (such as material and quantity). The RBAC (Role-Based Access Control) model is combined with the ACL, and permission rules are stored in JSON format. For example: json { "Role":"Designer", "Permissions": ["Modify pattern parameters", "Submit design version", "View process instructions"]}.

[0077] In this embodiment, the operation day records user operations on system resources, including information such as time, object, and parameter changes, for traceability and conflict detection. A conflict vector set is formed when multiple people modify the same component simultaneously. Each user's modification time, parameter change trajectory, and operation permissions are encapsulated into a vector, forming a conflict vector set {V01, V02, …, V0n}. Specifically, if Designer A (10:00:00) changes "sleeve length" from 60cm to 65cm (operation permission level 3), and Designer B (10:00:05) changes "sleeve length" to 62cm (operation permission level 3), then the conflict vector set is: VA = (t = 10:00:00, Δ parameter = (60→65), permission = 3), VB = (t = 10:00:05, Δ parameter = (60→62), permission = 3).

[0078] Specifically: Blockchain technology is used to record operation logs to ensure that they cannot be tampered with. The log format is: json { "User ID":"designerA", "Timestamp": "2025-06-29 10:00:00", "Operation object": "sleeve length", "Original parameter": 60, "New Parameters": 65, "Authorization Level": 3}.

[0079] In this embodiment, conflict detection compares modification logs of the same component through a scheduled task (such as scanning every second), and triggers the construction of a conflict vector set when the time difference is less than a threshold (such as 5 seconds).

[0080] In this embodiment, the sub-point set refers to dividing the modification track points with the same operation authority in the same change position point into a subset when handling the conflict of multiple people modifying the same component at the same time.

[0081] In this embodiment, the change trajectory set is to map the parameter change trajectory in the conflict vector at consecutive moments to the coordinate system, forming a trajectory line for each modification operation, reflecting the trend of parameter changes over time, and the change position point: the coordinate point of the parameter change in the coordinate system, such as "sleeve length = 65cm" corresponds to the point (10:00:00, 65) in the coordinate system. For example, the sleeve length modification record within 10 seconds is continuously obtained. The modification trajectory of designer A is: (10:00:00, 65) → (10:00:10, 66), and the trajectory of designer B is: (10:00:05, 62) → (10:00:10, 63). Mapped to a coordinate system with time as the x-axis and parameter value as the y-axis, two trajectory lines are formed. Specifically: a two-dimensional Cartesian coordinate system is used, the x-axis is time (accurate to seconds), and the y-axis is the parameter value (such as sleeve length, width of clothes). Use the D3.js library to visualize the trajectory, and process it through the following steps: Convert trajectory change information into a set of coordinate points {(t1,v1),(t2,v2),…}; Group points at the same location by operating permissions, such as the point sets of high-authority users are aggregated first; Use Bezier curve to fit the trajectory points and generate a smooth trajectory line.

[0082] In this embodiment, the trajectory analysis model is based on a machine learning algorithm to analyze the trend, conflict points and rationality of the parameter change trajectory, and generate a recommended merged parameter value (recommendation line), and the recommendation line is: the parameter merging suggestion curve output by the model, such as the trend line of "sleeve length should be adjusted to 63cm". For example, the sleeve length modification trajectories of designers A and B are input, and the model analysis finds that the trends of both are "lengthening", but the amplitudes are different. The generated recommendation line is the weighted average of the two trajectories (such as 63cm), taking into account the modification intentions of both parties.

[0083] In this embodiment, an LSTM neural network is used to construct a trajectory analysis model, with the input being a sequence of trajectory points and the output being recommended parameter values, ensuring that the recommended line conforms to the best practices of historical conflict resolution.

[0084] In this embodiment, the version merging suggestion is based on the recommended line of all conflicting position points to generate a final parameter merging solution, such as "the sleeve length adopts 63cm, and the hem curvature remains the modification of designer A". For example, in response to the conflict between the sleeve length and the hem curvature, after the model generates the recommended line, the merging suggestion is: "The sleeve length takes the recommended value of 63cm, the hem curvature is based on the modification of designer A (because of the same authority and modification first), and the seam width is adjusted according to the technician's suggestion". Specifically, a greedy algorithm is used to preferentially merge non-conflicting parameters, and conflicting parameters are suggested according to the recommended line. The user is prompted through a pop-up window, which includes: a comparison table of conflicting points (original parameters, various modified versions, recommended values); a comparison of 3D sample renderings (before modification, after modification, and recommended solution); and an explanation of the reason for the merger (such as "based on the time priority principle, combined with the modification trends of both parties").

[0085] The beneficial effects of the above technical solution are: the detection and suggestion generation of multi-person collaborative modification conflicts effectively reduces waiting costs, the generation of version merge suggestions avoids version errors caused by conflicts, ACL-based permission control clarifies the operation scope of different roles, improves the efficiency of cross-departmental collaboration, and the visualization of trajectory analysis and recommendation lines makes the conflict resolution process explainable.

[0086] The present invention proposes an intelligent clothing design method that integrates AI and CAD technologies, further comprising: Build a design modification history database, timestamp and hash each design change to form an unalterable version traceability chain. When a dispute over process parameter conversion occurs, the complete design modification trajectory can be queried to provide a reminder.

[0087] In this embodiment, the design modification history database is used to store modification records for all versions throughout the entire garment design process. This structured database, which includes design parameters, modification personnel, time, and operation content, is the core carrier for tracing design changes. For example, the modification history of a scenario such as a "women's shirt design" is stored: Designer A modified the length from 60cm to 65cm at 10:00 on June 29, 2025, and Technician B adjusted the seam width from 1cm to 1.2cm at 10:30. These records are stored in the database, specifically using MongoDB. The design document structure is as follows: json { "versionId": "V001", "designer": "designerA", "timestamp": "2025-06-29T10:00:00Z", "modification": { "part": "length", "oldValue": 60, "newValue": 65, "reason": "Adjust the proportions of the template" }, "hash": "a1b2c3..." / / hash value}.

[0088] In this embodiment, the timestamp tag adds an accurate time identifier for each design change, usually in UTC time format, to ensure the indisputability of the modification order. For example, when a designer modifies the "sleeve type" parameter, the system automatically generates a timestamp "2025-06-29T14:30:22Z" accurate to the second level for subsequent version sorting. Specifically, the timestamp is generated through a blockchain consensus mechanism (such as PoS), or the server NTP (Network Time Protocol) is called to obtain the precise time to ensure time consistency in a distributed environment.

[0089] In this embodiment, the version traceability chain is a chain structure of the previous version hash value + the current record, which connects all design modification records into an unalterable chain, similar to the block link mechanism of the blockchain.

[0090] In this embodiment, dispute query and trace reminders are used when a dispute arises during process parameter conversion (e.g., when production discovers a discrepancy between a parameter and the design). By querying the historical record database, the complete design modification history is restored, automatically notifying the source of the disputed modification. For example, if production discovers a discrepancy between the "Seam Allowance" parameter and the design document, the historical record may show that Designer A set it to 1cm at 10:00 on June 29, 2025, and Technician B changed it to 1.2cm at 10:30 (without syncing with production). The traceability chain displays the final version as 1.2cm, reminding the technologist that the production document has not been updated.

[0091] The beneficial effect of the above technical solution is that the version traceability chain requires modifications to be traced, which improves the clarity of responsibility definition in cross-departmental collaboration.

[0092] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. An intelligent clothing design method integrating AI and CAD technology, characterized in that: include: Step 1: Build a parametric CAD engine based on NURBS curve modeling and topological association algorithms. This engine contains a structured pattern database of several historical styles, which is stored in a three-level index by style, category, and component type. When any pattern parameter is adjusted, the topological association rules automatically trigger the linkage modification of related parameters. Step 2: Intelligently classify and segment the input sketch, and generate reusable components based on the matching parametric model obtained by the parametric CAD engine; Step 3: Based on AI technology, several reusable components are matched and recommended to obtain similar patterns, and the sketch and similar patterns are analyzed to recommend the optimal component combination solution; Step 4: Build a distributed collaborative design mechanism. When receiving multiple people modifying the same component of the combined sample of the optimal component combination solution, the conflict points are automatically prompted through operation log comparison and version merging suggestions are provided, and a 3D sample rendering is output for users to decide.

2. The intelligent clothing design method according to claim 1, characterized in that: Analyze the sketch and similar patterns and recommend the best component combination, including: Extracting global style features of the sketch and local features of each reusable component to construct a multidimensional feature vector including color features, texture features, and contour features; Determine the layout features of each similar layout to obtain the layout complexity coefficient of the corresponding similar layout, and splice the reusable components in the order of the coefficients according to the component placement positions in the sketch to obtain a complete layout; Determining a line connection diagram between the sketch and the complete layout based on the spatial relationship between the parts; Extracting three-dimensional geometric parameters of each line connection point in the line connection diagram to determine the geometric feature distribution; Extracting style features of the sketch and the line connection diagram and performing feature alignment processing to obtain a difference feature vector; Determine the designer's potential preference type and obtain pattern parameters of historical designs of similar patterns, and construct an embedding vector of the potential preference type; Determine the preference association between the current design of the sketch and each embedding vector, and construct a pattern knowledge graph based on the style similarity and geometric matching between the sketch and similar patterns, wherein the nodes of the pattern knowledge graph are patterns and the edges are preference association, style similarity, and geometric matching; Modifying the pattern knowledge graph based on the geometric feature distribution and the difference feature vector, and searching the design database to obtain a complete candidate set; Transferring the multi-dimensional feature vector of the sketch to each complete candidate template to obtain a preliminary template; Analyze the functional compatibility and aesthetic coordination between the components of the preliminary pattern and output an optimal component combination scheme, wherein the functional compatibility is related to the engineering pattern parameters, and the aesthetic coordination is related to color theory and style trends.

3. The intelligent clothing design method according to claim 1, characterized in that: Intelligently classify and segment the input sketches, and generate reusable components based on the matching parametric models obtained by the parametric CAD engine, including: A deep learning model combination architecture based on deep reinforcement learning dynamic regulation performs feature extraction and semantic segmentation on sketches; When using the U-Net network to perform semantic segmentation on sketches, its encoder-decoder structure fuses multi-scale features through skip connections, and introduces hierarchical energy constraints in the segmentation process. The energy function expression is expanded to : ； in, is the hierarchical correlation potential energy function; is a global semantic label; It is a local component label; is the hierarchical constraint coefficient; For sketches; is the segmentation label; is the set of adjacent pixels of the i-th pixel in the sketch; is the unit potential energy function, used to calculate a single segmentation label element The potential energy between it and the input x; is a pairwise potential energy function used to calculate two adjacent segmentation label elements , yj and the potential energy between input x; When extracting features from the segmented part contours, construct a spatiotemporal attention-enhanced sketch vector ,in, is the rate of change of curvature based on the time dimension, and , is the contour tangent vector of point s; is the scale factor; Encode textures; Generate the index address Dz of the sketch feature vector V1, including: Calculate the sketch feature vector V1 and component type The product of ; Calculate styles separately , Category , component type Adaptive hash value, and perform hash value fusion to generate dynamic index ;in, is a hash function; Based on style , Category , component type , the index function of the sketch feature vector V1; ⊕ is the exclusive OR symbol; is the Hadamard product symbol; Indexing the index address Dz in a structured pattern database based on the parametric CAD engine, and screening reference vectors with a matching degree greater than a preset degree to obtain a parametric model; The sketch feature VI is adaptively adjusted based on the parameterized model, and a reusable component is generated through topological association rules.

4. The intelligent clothing design method according to claim 3, characterized in that: The model combination architecture consists of a ResNet-152 network and a Mask R-CNN network.

5. The intelligent clothing design method according to claim 2, characterized in that: Determining a line connection diagram between the sketch and the complete layout based on the spatial relationship between the parts includes: Extracting the three-dimensional bounding boxes and two-dimensional projection contours of the sketch and the complete layout to construct a component topology relationship diagram; Constructing a line constraint optimization model based on the topological relationship graph, wherein the line control points of the sketch and the line control points of the complete layout are used as optimization variables to define an objective function and constraint conditions; The maximum flow-minimum cut algorithm is used to solve the line connection scheme of the line constraint optimization model to generate a line connection graph.

6. The intelligent clothing design method according to claim 1, characterized in that: Also includes: After outputting the 3D sample rendering for the user to decide, a QR code is generated and sent to the supplier; After the supplier scans the QR code received by the supply end, the supplier can view the fabric drape effect of the 3D sample garment and confirm it.

7. The intelligent clothing design method according to claim 1, characterized in that: Automatically prompt conflict points through operation log comparison and provide version merge suggestions, including: Determine the operation permissions of each designer based on the data access control list; Based on the operation log, the modification time and change parameters of each designer on the reusable component are recorded. When it is detected that the same component is modified by multiple people at the same time, the conflict vector set at each moment is constructed. , where n1 represents the number of people involved in the conflict; represents the conflict vector of the j1th conflicting person, They represent the modification time t, trajectory change information, and operation permissions respectively; Continuously obtain conflict vector sets at N0 consecutive moments, and map the trajectory change information to the blank coordinate system to lock the changed position point at the first moment. Then, construct a sub-point set for the same changed position point at the first moment based on the operation authority, and extract the trajectory points at each consecutive moment from the mapped coordinate system to obtain a change trajectory set for each changed position point, where each trajectory line in the change trajectory set is related to the operation authority and modification time; Inputting the change trajectory set into a trajectory analysis model to determine a recommended line corresponding to the change position point; Based on the recommended lines of all the first independent change locations, a version merge suggestion is generated and sent to the user for approval.

8. The intelligent clothing design method according to claim 1, characterized in that: Also includes: Build a design modification history database, timestamp and hash each design change to form an unalterable version traceability chain. When a dispute over process parameter conversion occurs, the complete design modification trajectory can be queried to provide a reminder.

Images