Digital system and method for suit pants design

Abstract

The invention discloses a digital system and method for suit pants design, and relates to the field of suit pants design, and the method comprises the steps: setting a man-machine interaction port, receiving suit pants design parameters inputted by a user, and displaying a suit pants three-dimensional contour map; constructing a standardized three-dimensional standard template of the suit pants, and constructing a graph data structure of the suit pants; based on the conditional graph neural network model, generating a three-dimensional contour map of the suit pants; setting an evaluation interface, receiving user evaluation information, extracting evaluation features through semantic recognition, and establishing a mapping relationship between the evaluation features and design parameters; according to historical user evaluation information, the satisfaction degree, accuracy and scheme iteration efficiency are counted and calculated, and the whole suit pants design system is evaluated. The method has the advantages that convenient operation of a non-professional user and high-precision and personalized requirements of suit pants design are considered, intelligent and efficient digital design is realized through a full closed loop of interaction-generation-feedback-evaluation, the professional threshold is greatly reduced, and the customization efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of men's trousers design, and more specifically to a digital system and method for men's trousers design. Background Technology

[0002] With the explosive growth of personalized demands in the clothing customization industry and the iteration of intelligent manufacturing technologies, dress trousers, as a core item in the clothing category with strong structural standardization and high fit requirements, are undergoing a transformation in their design model from traditional hand-made patterns to digital ones. Traditional dress trouser design relies on the experience of senior pattern makers, completing the pattern design by hand-drawing two-dimensional patterns and repeatedly trying them on and modifying them. This approach suffers from problems such as long cycles, high costs, and difficulty in personalizing the fit, and can no longer meet the needs of modern consumers for rapid customization and precise fit. Against this backdrop, digital design technologies are gradually being applied to the field of dress trouser design, with core technologies including parametric CAD systems, 3D modeling software, and basic neural network generation technology. Currently, most mainstream digital solutions are based on parametric CAD systems, which drive the generation of two-dimensional pattern pieces by pre-setting key trouser dimensions (such as waist and hip circumference). Some advanced solutions introduce 3D modeling technology to achieve a visual preview of the pattern. Meanwhile, with the development of artificial intelligence technology, some solutions attempt to use neural network models to generate trouser silhouettes, but these mostly focus on generating single styles or fixed patterns. Furthermore, the demand for "user-participatory design" in the clothing customization industry is increasingly prominent; ordinary consumers want to actively participate in customizing the style and details of their trousers, rather than passively accepting standardized products.

[0003] Mainstream parametric CAD systems are essentially a collection of "if-then" rules. Their design logic is rigid and they cannot learn the complex, non-linear pattern changes in the data. For example, the system may preset that "for every 2 cm increase in waist circumference, the waistband circumference increases linearly by 6.28 cm." However, in reality, waist circumference changes may be accompanied by coordinated changes in waistband shape and dart treatment. This complex effect of multi-parameter coupling is difficult for preset rules to exhaust. Secondly, the human-computer interaction interface of existing systems is mostly based on professional terminology and numerical sliders, which is extremely unfriendly to non-professional users. Users find it difficult to convert subjective feelings into specific parameters. More importantly, the system is completely unable to understand subsequent user feedback and cannot form an interactive closed loop of "design-feedback-optimization." This leads users to fall into the predicament of blindly adjusting parameters when faced with unsatisfactory initial results, resulting in a poor experience and low efficiency. Summary of the Invention

[0004] To address the aforementioned technical problems, a digital system and method for trouser design is provided. This technical solution solves the problem mentioned in the background that mainstream parametric CAD systems have rigid design logic and cannot learn complex and nonlinear pattern variation patterns in the data. Furthermore, the human-computer interaction interface of existing systems is mostly based on professional terminology and numerical sliders, which is extremely unfriendly to non-professional users.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A digital approach to trouser design includes: Set up a human-computer interaction port, provide selectable labels, receive user input of trouser design parameters, and display the trouser's 3D outline through a synchronous preview area; Construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and build the graph data structure of the trousers. Based on the graph data structure of the trousers, and using a conditional graph neural network model, combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated. Set up an evaluation interface to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; Based on historical user reviews, we statistically analyzed and calculated satisfaction, accuracy, and solution iteration efficiency to evaluate the entire trousers design system.

[0006] Preferably, the step of constructing a standardized 3D reference template for trousers, extracting the 3D feature points of the template to construct graph nodes and edges, and constructing the graph data structure of the trousers specifically includes: Based on the principles of 3D garment CAD pattern making, a standardized 3D baseline template library for trousers is generated for different target styles, serving as the baseline template for all designs. The baseline template is broken down into several key structural parts according to garment structure, including at least the following parts: waistband, front trouser piece, back trouser piece, left trouser leg, right trouser leg, front dart / pleat, and back dart. For each structural part, based on its two-dimensional unfolded plate outline and key internal structural lines, three-dimensional reference feature points are manually defined. The set of all feature points constitutes the basic dataset of the reference template shape. Using a right-handed Cartesian coordinate system, the origin of the coordinate system is defined as 5cm directly above the navel in a standard human standing posture. The positive X-axis is horizontal to the left along the human body, the positive Y-axis is vertically downward, and the positive Z-axis is horizontal to the back. All defined 3D reference feature points are mapped to the coordinate system according to their geometric positions on the reference template to obtain their corresponding initial standard values ​​of 3D coordinates; Each 3D reference feature point and its coordinates are treated as a node in the graph, and each node is assigned an attribute, which includes at least its coordinates and a label to identify its structural part and semantic role. Based on the spatial adjacency and technological logic of clothing structure, connection relationships are defined in the node set as edges of the graph. The definition rules include: Contour edge: An edge is established between two adjacent feature point nodes on a continuous contour line of the same structural part; Symmetric edge: Establish an edge between a pair of feature point nodes that are symmetrical about the YZ plane of the human body; Seam edge: Establish an edge between corresponding matching feature point nodes of two different structural parts that need to be seamed; Spatial Adjacency Edge: An edge is established between feature point nodes in three-dimensional space whose Euclidean distance is less than a preset threshold and belong to the same surface region; Construct the graph data structure for the trousers based on the set of graph nodes and the set of graph edges.

[0007] Preferably, the step of generating a 3D outline of the trousers based on the graph data structure of the trousers, using a conditional graph neural network model, and combining the trousers design parameters input by the user specifically includes: Obtain the design parameters of the trousers input by the user and perform standardized preprocessing on them; The preprocessing includes: normalizing the continuous numerical parameters, converting the discrete option parameters into one-hot encoded vectors, and concatenating all the preprocessed parameters into a unified global design feature vector. Construct a conditional graph neural network model that takes the global design feature vector as the conditional input and the graph topology and initial node feature inputs obtained from the benchmark template library. In each layer of the conditional graph neural network, the global design feature vector is processed by a conditional coding network to generate affine transformation parameters for the dynamic scaling parameters and bias parameters corresponding to that layer. By using the generated scaling and bias parameters, the fixed weight matrix of the current graph convolutional layer is dynamically modulated to obtain the conditionally dependent effective weight matrix. After multiple layers of conditional graph convolution, the network output layer generates a predicted 3D coordinate offset value for each graph node; Add the original 3D coordinates of each node to its corresponding predicted offset to obtain the 3D coordinates of the node under the user-input design parameters; All adjusted feature point nodes are sorted according to the predefined contour edge connection relationship in the graph data structure, and then connected sequentially using the B-spline curve algorithm to generate the three-dimensional contour edges of each structural part. For the area enclosed by the contour lines, a triangular patch filling algorithm is used to construct a three-dimensional surface, and the surfaces of each part are spliced ​​together to form a complete three-dimensional mesh model of the trousers. Based on the surface continuity constraints implied by the predefined spatial adjacency edges in the graph data structure, the generated 3D mesh model is subjected to local Laplacian smoothing. The optimized 3D contour model is converted into a universal 3D format and rendered in real time to the synchronous preview area of ​​the human-computer interaction port.

[0008] Preferably, the step of setting an evaluation interface, receiving user evaluation information, extracting evaluation features through semantic recognition, and establishing a mapping relationship between evaluation features and design parameters specifically includes: Next to the preview area of ​​the 3D outline of the trousers on the human-computer interaction port, set up a user evaluation interface; The evaluation interface receives user evaluation information for the current design result in the form of text input boxes, structured rating forms, and sentiment selection buttons. The collected unstructured text evaluations are cleaned, including removing irrelevant characters, correcting spelling errors, and word segmentation. Construct a fine-grained semantic recognition model for the field of apparel design to extract structured evaluation features from preprocessed user reviews; The semantic recognition model adopts a fine-tuned architecture based on a pre-trained language model, and its output layer is adapted for multi-label classification tasks. The recognition results of the semantic recognition model are integrated into a structured evaluation feature vector; Construct an evaluation-parameter mapping learning module. This module takes the structured evaluation feature vector and the design parameter vector corresponding to this evaluation as input, and aims to output a set of recommended design parameter adjustments. The evaluation-parameter mapping learning module uses a multilayer perceptron regression model, and its training data comes from historical interaction logs. Each historical interaction log contains the initial design parameters, user evaluation features, and the design parameters finally accepted by the user. The training objective of the evaluation-parameter mapping learning module is set to minimize the mean square error between the parameter adjustment of its output and the ideal adjustment implied by real user behavior. Through training, the evaluation-parameter mapping learning module learns the mapping relationship from abstract evaluation to specific quantitative parameter adjustment suggestions, and outputs the recommended adjustment amount and optimized design parameter candidates. The optimized design parameter candidates are input into the conditional graph neural network model to generate the corresponding 3D outline of the trousers. The generated optimized 3D silhouette of the trousers will be presented to the user in the preview area in a side-by-side comparison or carousel format, and the user's selection results will be recorded together with the evaluation information.

[0009] Furthermore, this solution proposes a digital system for designing trousers, used to implement the aforementioned digital method for designing trousers, including: The human-computer interaction module is used to set up the human-computer interaction port, provide selectable labels, receive the trouser design parameters input by the user, and display the 3D outline of the trousers through a synchronous preview area. The design generation module is used to construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers; based on the graph data structure of the trousers, and using a conditional graph neural network model combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated; an evaluation interface is set up to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; The system evaluation module is used to evaluate the entire trousers design system by statistically analyzing and calculating satisfaction, accuracy, and solution iteration efficiency based on historical user evaluation information.

[0010] Preferably, the design generation module includes: The graph data unit is used to construct a standardized three-dimensional reference template for trousers, extract the three-dimensional feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers. The design generation unit is used to generate a three-dimensional outline of the trousers based on the graphical data structure of the trousers, a conditional graphical neural network model, and user-input trouser design parameters. The evaluation optimization unit is used to set up an evaluation interface, receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a digital system and method for designing trousers. It collects user-inputted trouser design requirements in a structured manner through a tagged design of the human-computer interaction interface. Combining a standardized 3D baseline template library for trousers with embedded garment process constraints in a graph data structure, it utilizes a conditional graph neural network to accurately generate personalized 3D outline diagrams of trousers. Simultaneously, a semantic recognition module analyzes user evaluations and constructs a mapping relationship between evaluation features and design parameters. Furthermore, a system evaluation mechanism is used to quantitatively optimize the entire design process. This solution effectively lowers the operational threshold for non-professional users, improves the personalization and pattern compliance of trouser designs, and thus effectively enhances the efficiency and user satisfaction of digital trouser design. Attached Figure Description

[0012] Figure 1 This is a flowchart of a digital method for designing trousers according to the present invention; Figure 2 This is a flowchart illustrating the graph data structure for constructing trousers according to the present invention; Figure 3 The flowchart of generating a 3D outline of trousers based on the conditional graph neural network model of the present invention, combined with the user-input trousers design parameters; Figure 4 The flowchart below illustrates the process of extracting evaluation features through semantic recognition and establishing a mapping relationship between evaluation features and design parameters according to the present invention. Detailed Implementation

[0013] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0014] Reference Figure 1 As shown, a digital method is used to design trousers, including: Set up a human-computer interaction port, provide selectable labels, receive user input of trouser design parameters, and display the trouser's 3D outline through a synchronous preview area; Construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and build the graph data structure of the trousers. Based on the graph data structure of the trousers, and using a conditional graph neural network model, combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated. Set up an evaluation interface to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; Based on historical user reviews, we statistically analyzed and calculated satisfaction, accuracy, and solution iteration efficiency to evaluate the entire trousers design system.

[0015] To achieve efficient, precise, personalized, and user-friendly trouser design, it's essential to first clarify the core boundaries of the design. Therefore, the first step is to establish a human-computer interaction tag system, outlining all design parameter types and corresponding input methods required for the fit and personalization of trousers. Based on the core principles of user-friendliness and machine-recognizable design, essential basic tags and optional personalized tags are categorized, creating a standardized parameter input list. This serves as the foundation for all subsequent design processes, preventing issues such as missing parameters, input confusion, or user difficulties. After clarifying the input basis, the compliance and structural rationality of the generated pattern need to be addressed. Therefore, a standardized 3D base for trousers is constructed. The system uses a quasi-template library and disassembles key structural parts, extracting 3D feature points to construct graph nodes and edges. It embeds garment manufacturing constraints (such as symmetry, stitching, and surface smoothness) to build a graph data structure. This ensures the basic rationality of the pattern through a baseline template and transforms manufacturing knowledge into computable topological relationships through the graph data structure. This breaks away from the traditional single-mode digital design approach of "no manufacturing constraints, resulting in difficult-to-produce patterns," balancing design flexibility with manufacturing compliance. To transform abstract design parameters and graph data structures into personalized 3D silhouettes of trousers, a conditional graph neural network model is constructed. This model serves as the core of design generation, integrating preprocessed global feature vectors from user input parameters with... The topological information of the graph data structure is used to dynamically modulate the convolutional weights through a conditional coding network, generating precise feature point coordinate offsets. This enables intelligent deformation from a baseline template to a personalized design, addressing the industry pain points of traditional parametric design, such as "uniform style and unnatural deformation." Secondly, after determining the core generation logic, the connection between user feedback and parameter adjustment needs to be addressed. Therefore, an evaluation feedback loop is built, multiple evaluation interfaces are set up to collect user feedback, and a domain-adapted fine-grained semantic recognition model is used to extract structured evaluation features. An evaluation-parameter mapping learning module is constructed to transform abstract evaluations into quantitative adjustment quantities, simultaneously generating optimized solutions and comparing and displaying them, thus completing the process from "design generation." From the implementation of "user feedback optimization" to the final solution, to address the pain points of difficulty in quantifying system performance and unclear optimization directions, a full-process evaluation system was constructed. Based on historical design session logs, automated evaluation tasks were set up to quantitatively calculate satisfaction, accuracy, and solution iteration efficiency. The results were displayed and stored through a visualization platform. At the same time, the correlation logic between evaluation results and system optimization was constructed to accurately locate and guide optimization of issues such as human-computer interaction, model accuracy, and efficiency. This formed a complete closed loop of "parameter input - intelligent generation - evaluation optimization - quantitative evaluation - system iteration," which significantly improved the stability and practicality of digital design of trousers and ensured the efficiency, accuracy, and user-friendliness of the entire process.

[0016] The setup of the human-computer interaction port, which provides selectable labels, receives user-inputted design parameters for the trousers, and displays a 3D outline of the trousers through a synchronized preview area, specifically includes: The human-computer interaction port adopts a UI interface, which is presented in the form of a graphical user interface and has selectable labels deployed. The selectable tags include: basic mandatory tags and personalized optional tags; The basic required labels include: gender, size, waist circumference, hip circumference, pant length, thigh circumference, leg opening width, front crotch length, and back crotch length; The personalized optional tags include: style, color, pant type, pleat type, waistband, pocket style, hem treatment, and fabric; The pre-design parameters of the trousers can be set through the human-computer interaction port by means of radio buttons, drop-down menus, slider controls, direct input boxes and visual icons. Based on the selectable tag types, each tag is bound to a fixed parameter key name and standardized value through manual predefinition, thus constructing a front-end data mapping table; Users select tags through the interface, and the front-end data mapping table is used to obtain the parameter key name and standardized value of the selected tag, which are then stored in key-value pairs or structured JSON format. On the human-computer interaction interface, there is a synchronous preview area that displays a 3D outline of the trousers.

[0017] Explained in this context, the core of digital design for dress trousers is transforming traditional, experience-based pattern making into a process driven by a set of precisely defined parameters. To achieve high-precision and personalized design, the system needs to quantify parameters in two dimensions: Fit: These parameters directly determine the degree of fit between the dress trousers and the wearer's body, forming the foundation of the pattern structure; they are typically continuous values. Aesthetic and Functional: These parameters determine the dress trousers' appearance, details, and functions, and are key to meeting personalized needs; they are typically discrete options. Based on this principle, this solution achieves personalized digital design for dress trousers by setting the following tag system. In a specific implementation case, the basic mandatory tags specifically include: Size: Offers standard Chinese size options (sizes 30, 31, 32...40), and also supports custom size input (sizes 28-42) to accommodate special body types; Waist circumference: In centimeters, it offers gradient options such as 65cm, 68cm, 70cm...90cm, with gradient intervals of 2cm. It also supports custom input accurate to 1cm (input range 60cm-100cm). Pants length: There are options for straight-waisted and bent-waisted styles. The length range for straight-waisted pants is 85cm-110cm, and the length range for bent-waisted pants is 80cm-105cm, with a gradient interval of 3cm. It supports custom adjustment to a precision of 1cm (custom range for straight-waisted pants is 80cm-115cm, and for bent-waisted pants it is 75cm-110cm). Hip circumference: In centimeters, it offers gradient options such as 85cm, 88cm, 90cm...110cm, with gradient intervals of 2cm. It supports custom input accurate to 1cm (input range 80cm-120cm) to adapt to the circumference needs of different body types. Thigh circumference: In centimeters, it offers gradient options such as 45cm, 47cm, 49cm...60cm, with gradient intervals of 2cm. It also supports custom input accurate to 1cm (input range 40cm-65cm) to ensure a comfortable and unrestricted feel in the thigh area. Foot opening width: In centimeters, it offers gradient options such as 18cm, 19cm, 20cm...25cm, with gradient intervals of 1cm. It supports custom input accurate to 0.5cm (input range 16cm-28cm) to adapt to different shoe styles and matching needs. Front crotch length: In centimeters, it offers gradient options such as 22cm, 23cm, 24cm...30cm, with gradient intervals of 1cm. It supports custom input accurate to 0.5cm (input range 20cm-32cm) to avoid discomfort caused by the front crotch being too short or too long. Back crotch length: In centimeters, it offers gradient options such as 28cm, 29cm, 30cm...36cm, with gradient intervals of 1cm. It supports custom input accurate to 0.5cm (input range 26cm-38cm) to ensure a good fit to the buttocks and comfortable movement. Personalized optional tags provide style and detail customization parameters, specifically including: Style: Single selection available, including business, casual, retro, and commuter styles; Pants style: Single selection, including straight leg pants, tapered pants, loose pants, and wide-leg pants. Straight leg pants refer to pants where the difference between the thigh circumference and the leg opening width is ≤5cm, with straight lines. Tapered pants refer to pants where the difference between the thigh circumference and the leg opening width is 6-10cm, with a wider top and narrower bottom. Loose pants refer to pants where the thigh circumference is 3-5cm larger than that of straight leg pants of the same size. Wide-leg pants refer to pants where the leg opening width is ≥26cm, and the difference between the thigh circumference and the leg opening width is ≤3cm. Pleats: Single selection, including no pleats, single pleats, and double pleats. No pleats means the front of the pants is flat and without pleats. Single pleats means there is one pleat on one side of the front of the pants, extending from the waistband to the hem. Double pleats means there are one pleat symmetrically distributed on each side of the front of the pants. Waistband: Available in single-choice options, including no waistband, single-choice waistband, double-choice waistband, and elastic waistband. Single-choice waistbands are secured with one metal buckle on one side, double-choice waistbands are secured with two metal buckles on both sides, and elastic waistbands have a built-in 3-5cm elastic band that can stretch the waist by 1-3cm for comfortable wear. Pocket styles: Multiple options are available, including front slanted pockets, front straight pockets, back patch pockets, back concealed pockets, no back pockets, pocket flap designs, and zippered pockets; Trouser hem treatment: Single option, including regular overlock, rolled hem, blind seam, and vent; Fabric: Single selection, including selvedge cotton, Tencel blend, stretch cotton, polyester blend, and wool blend; The human-computer interaction port uses one or more of the following methods to implement user operations: Radio buttons: Used for mutually exclusive options, such as gender (male / female) and pleat type (no pleat / single pleat / double pleat). Drop-down selection menu: Used for lists with many options, such as fabric library or specific colors; Slider control: Used to adjust values ​​continuously or in segments within a certain range, such as waist circumference or pant length; Direct input box: Used for entering precise numerical values; Visual icon selection: Used for intuitive selection, such as clicking the outline icon of different pants styles.

[0018] Reference Figure 2 As shown, the graph data structure for constructing the trousers specifically includes: Based on the principles of 3D garment CAD pattern making, a standardized 3D baseline template library for trousers is generated for different target styles, serving as the baseline template for all designs. The baseline template is broken down into several key structural parts according to garment structure, including at least the following parts: waistband, front trouser piece, back trouser piece, left trouser leg, right trouser leg, front dart / pleat, and back dart. For each structural part, based on its two-dimensional unfolded plate outline and key internal structural lines, three-dimensional reference feature points are manually defined. The set of all feature points constitutes the basic dataset of the reference template shape. Using a right-handed Cartesian coordinate system, the origin of the coordinate system is defined as 5cm directly above the navel in a standard human standing posture. The positive X-axis is horizontal to the left along the human body, the positive Y-axis is vertically downward, and the positive Z-axis is horizontal to the back. All defined 3D reference feature points are mapped to the coordinate system according to their geometric positions on the reference template to obtain their corresponding initial standard values ​​of 3D coordinates; Each 3D reference feature point and its coordinates are treated as a node in the graph, and each node is assigned an attribute, which includes at least its coordinates and a label to identify its structural part and semantic role. Based on the spatial adjacency and technological logic of clothing structure, connection relationships are defined in the node set as edges of the graph. The definition rules include: Contour edge: An edge is established between two adjacent feature point nodes on a continuous contour line of the same structural part; Symmetric edge: Establish an edge between a pair of feature point nodes that are symmetrical about the YZ plane of the human body; Seam edge: Establish an edge between corresponding matching feature point nodes of two different structural parts that need to be seamed; Spatial Adjacency Edge: An edge is established between feature point nodes in three-dimensional space whose Euclidean distance is less than a preset threshold and belong to the same surface region; Construct the graph data structure for the trousers based on the set of graph nodes and the set of graph edges.

[0019] This can be explained by the fact that, as a highly structured garment, the design of dress trousers can be decoupled from the basic pattern structure and personalized parameter variations. This solution uses a pre-built 3D baseline template library for dress trousers, with templates covering mainstream styles (such as classic straight leg, slim tapered, etc.). Each template in the library corresponds to a mainstream style (such as classic straight leg, slim tapered, etc.) and its style attributes have been marked through template metadata. This creates an efficient digital design starting point. By reading the design parameters submitted by the user through the human-computer interaction port, the parameter key names of the selected tags in the style are extracted. Based on the parameter key names, the template library is quickly searched for templates whose style attributes match the base template. The quasi-template is an efficient key-value query process that loads the matched baseline template as the geometric base of the current design. All subsequent refined parameters input by the user (such as specific dimensions, pleats, pocket styles, etc.) are based on this high starting point, driving the template to perform relatively minor and precise geometric adjustments and structural reconstructions. This avoids the need for complex structural modeling for each trouser design from scratch, significantly improving design efficiency and response speed while ensuring pattern rationality and style accuracy. The preset threshold is determined as follows: for each feature point, calculate its distance to all other feature points within its structural region, and take the Kth distance from these distances. The minimum value is used as the local adjacency radius of the point, where K is a preset integer (e.g., K=3 or 5). For any two feature points, if the Euclidean distance between them is less than the arithmetic mean of their respective local adjacency radii, then a spatial adjacency edge is established between them. This allows the threshold to automatically decrease and the connection to be more refined in areas with dense feature points (e.g., curved hips); and to automatically increase and the connection to be more relaxed in sparse areas (e.g., the middle of a flat trouser leg). The manually defined three-dimensional reference feature points are determined based on the two-dimensional unfolded pattern outline and key internal structural lines of the trousers, such as the curves and darts on the cutting diagram. Endpoints and the intersections of two lines are the "key points" that determine the shape of the trousers. Mapping these points on the two-dimensional cutting diagram to a three-dimensional template creates three-dimensional feature points. When all the points are put together, the basic shape of the trousers is determined. Abstracting feature points into graph nodes prepares for subsequent applications of graph neural networks and other algorithms. The definition of edges is the soul of graph data structures. Contour edges ensure the continuity and smoothness of the contour lines in the same area; symmetrical edges force the model to maintain the left-right symmetry of the garment during learning or prediction; and seam edges ensure that the corresponding points of different pieces at the seam can change in tandem during adjustments, avoiding cracks or overlaps.Contour edges define the continuity of linear paths, symmetrical edges and seam edges define macroscopic structural correspondences, while spatial adjacent edges construct local smooth constraints on the surface of the 3D model, forcing adjacent feature points to maintain a natural transition and flatness of the surface during deformation, preventing unreasonable local bulges, depressions, or sharp edges, and effectively compensating for unnatural distortions that may still occur on the model surface in areas outside the contour lines. Through this structured relational expression, garment manufacturing knowledge is embedded into the data-driven model, resulting in a structured representation containing geometric (coordinates), semantic (labels), and topological (connections) information, laying the data foundation for intelligent and structured processing of 3D models of trousers; where the same surface region refers to the division according to the disassembled structural parts, and feature points within the same structural part are considered the same surface region; The trouser diagram data structure constructed in this scheme provides an ideal data foundation for subsequent data-driven intelligent design. This structure can be directly used as input to a graph neural network (GNN). By learning the complex relationships between nodes and edges in the graph, the model can achieve a precise mapping from design parameters to three-dimensional shapes that conforms to process constraints. Various edges in the graph naturally define the constraints in model training, such as symmetry loss, stitch consistency loss, and surface smoothness loss, thereby ensuring the rationality and usability of the generated results.

[0020] Reference Figure 3 As shown, the process of generating a 3D outline of the trousers based on the conditional graph neural network model and combined with the user-input trouser design parameters specifically includes: Obtain the design parameters of the trousers input by the user and perform standardized preprocessing on them; The preprocessing includes: normalizing the continuous numerical parameters, converting the discrete option parameters into one-hot encoded vectors, and concatenating all the preprocessed parameters into a unified global design feature vector. Construct a conditional graph neural network model that takes the global design feature vector as the conditional input and the graph topology and initial node feature inputs obtained from the benchmark template library. In each layer of the conditional graph neural network, the global design feature vector is processed by a conditional coding network to generate affine transformation parameters for the dynamic scaling parameters and bias parameters corresponding to that layer. By using the generated scaling and bias parameters, the fixed weight matrix of the current graph convolutional layer is dynamically modulated to obtain the conditionally dependent effective weight matrix. After multiple layers of conditional graph convolution, the network output layer generates a predicted 3D coordinate offset value for each graph node; Add the original 3D coordinates of each node to its corresponding predicted offset to obtain the 3D coordinates of the node under the user-input design parameters; All adjusted feature point nodes are sorted according to the predefined contour edge connection relationship in the graph data structure, and then connected sequentially using the B-spline curve algorithm to generate the three-dimensional contour edges of each structural part. For the area enclosed by the contour lines, a triangular patch filling algorithm is used to construct a three-dimensional surface, and the surfaces of each part are spliced ​​together to form a complete three-dimensional mesh model of the trousers. Based on the surface continuity constraints implied by the predefined spatial adjacency edges in the graph data structure, the generated 3D mesh model is subjected to local Laplacian smoothing. The optimized 3D contour model is converted into a universal 3D format and rendered in real time to the synchronous preview area of ​​the human-computer interaction port.

[0021] This solution explains that by introducing a conditional graph neural network, it deeply integrates and maps user-personalized design parameters with a standardized graph data structure. Through a conditional encoding network, the abstract, high-dimensional global design feature vector of the design intent is compiled into a set of low-dimensional dynamic scaling and bias parameters that control the graph convolution operations. These parameters act like "control knobs," dynamically changing the rules of information transmission (i.e., the weight matrix) at each layer of the network. This allows the same baseline graph data structure to generate 3D node coordinate offsets that conform to the expected style and size under different design conditions (global design feature vectors). This eliminates the need for the model to store independent weights for each parameter combination; instead, it shares a powerful basic transformation capability and allows for precise fine-tuning through conditional inputs, achieving efficient and flexible personalized design generation. Specifically, the preprocessing includes: normalizing the continuous numerical parameters (waist circumference, hip circumference, pant length, etc.) input by the user according to the formula "circumference scaling factor = user value / reference value, length offset = user value - reference value"; and converting the discrete optional parameters (pant style, pleat style, waistband, etc.) input by the user into one-hot encoded vectors (e.g., "double pleat" is encoded as [0,1,0], "single pleat" as [1,0,0], and "no pleat" as [0,0,1]). The conditional coding network is constructed using a lightweight architecture based on a multilayer perceptron. Its technical implementation includes: an input layer that receives a global design feature vector with a feature dimension of d; a feature transformation layer consisting of two fully connected layers. The first fully connected layer receives the input feature vector with dimension d and maps its dimension from d to d / 2. The ReLU activation function is used for dimensionality reduction and nonlinear feature extraction of the input features (where d is a positive integer and a multiple of 2 to ensure the output dimension is an integer); and a second fully connected layer that receives the d / 2-dimensional feature vector output from the first layer and maps it to a 2m-dimensional output vector (where m is the feature dimension of the hidden layer in the graph neural network). The 2m-dimensional vector output from the second layer is uniformly divided into two parts, which serve as the scaling parameter and bias parameter for the affine transformation, respectively. The mathematical principle behind the dynamic modulation of the fixed weight matrix of the current graph convolutional layer is as follows: In a graph neural network, a graph convolutional layer is given a base weight matrix. Simultaneously, a conditional coding network calculates a set of dynamically adjustable parameters for that specific layer: a scaling parameter vector and a bias parameter vector. During calculation, a broadcast numerical operation mechanism automatically expands the original single scaling and bias parameter vectors into a matrix with dimensions compatible with the base weight matrix. Specifically, this broadcast numerical operation mechanism combines the scaling parameter vector with a column vector containing all 1s to form a scaling coefficient matrix along the column direction; and combines the bias parameter vector with a column vector containing all 1s to form a bias matrix. The modulation process is performed in two steps: First, each element in the base weight matrix is ​​multiplied by the result of adding the corresponding row scaling coefficient to the number 1. Second, based on the scaled matrix obtained in the first step, the corresponding row bias coefficient is added to each element. After these two steps, a conditionally dynamically adjusted effective weight matrix is ​​obtained, effectively ensuring that the transformation behavior of the graph convolutional layer can flexibly respond to the user's design input, while the entire operation process is mathematically stable and controllable. The training of the graph neural network model uses paired data of "different design parameters and corresponding 3D models" as supervision signals. The training objective is to minimize the Euclidean distance loss between the model's predicted node coordinates and the actual coordinates. Through this process, the model automatically learns the complex, nonlinear mapping relationship from design parameter conditions to geometric deformation. This mapping implicitly includes all reasonable constraints in garment manufacturing regarding scaling, symmetry, seam alignment, and surface smoothness. The paired data can be constructed in the following way: using a parametric garment CAD system, automatically batch-generate 3D trouser models covering the design space (different sizes, styles, and detail combinations) and their corresponding parameters to form a large-scale, high-quality supervision dataset.

[0022] Reference Figure 4 As shown, the step of extracting evaluation features through semantic recognition and establishing a mapping relationship between evaluation features and design parameters specifically includes: Next to the preview area of ​​the 3D outline of the trousers on the human-computer interaction port, set up a user evaluation interface; The evaluation interface receives user evaluation information for the current design result in the form of text input boxes, structured rating forms, and sentiment selection buttons. The collected unstructured text evaluations are cleaned, including removing irrelevant characters, correcting spelling errors, and word segmentation. Construct a fine-grained semantic recognition model for the field of apparel design to extract structured evaluation features from preprocessed user reviews; The semantic recognition model adopts a fine-tuned architecture based on a pre-trained language model, and its output layer is adapted for multi-label classification tasks. The recognition results of the semantic recognition model are integrated into a structured evaluation feature vector; Construct an evaluation-parameter mapping learning module. This module takes the structured evaluation feature vector and the design parameter vector corresponding to this evaluation as input, and aims to output a set of recommended design parameter adjustments. The evaluation-parameter mapping learning module uses a multilayer perceptron regression model, and its training data comes from historical interaction logs. Each historical interaction log contains the initial design parameters, user evaluation features, and the design parameters finally accepted by the user. The training objective of the evaluation-parameter mapping learning module is set to minimize the mean square error between the parameter adjustment of its output and the ideal adjustment implied by real user behavior. Through training, the evaluation-parameter mapping learning module learns the mapping relationship from abstract evaluation to specific quantitative parameter adjustment suggestions, and outputs the recommended adjustment amount and optimized design parameter candidates. The optimized design parameter candidates are input into the conditional graph neural network model to generate the corresponding 3D outline of the trousers. The generated optimized 3D silhouette of the trousers will be presented to the user in the preview area in a side-by-side comparison or carousel format, and the user's selection results will be recorded together with the evaluation information.

[0023] This can be explained by the fact that a conditional graph neural network model, combined with user-inputted trouser design parameters, generates a 3D outline of the trousers. However, this generated 3D outline may not meet the user's needs. Simply adjusting the trouser design parameters would be extremely complex and difficult for non-professionals to obtain the ideal trouser design. Therefore, this solution, based on the conditional graph neural network model, incorporates user feedback and uses semantic recognition and evaluation-parameter mapping to dynamically quantify and adjust the trouser design parameters, thereby achieving a complete closed loop of "design generation → evaluation feedback → parameter optimization → iterative upgrade." The optimized candidate design parameter value is equal to the sum of the product of the currently used design parameter value, the step size coefficient, and the recommended design parameter adjustment amount. The step size coefficient is dynamically determined by the modification intensity branch result output by the semantic recognition model. When a slight adjustment is identified, the step size coefficient is 0.5 for small-scale exploratory optimization. When a strong modification is identified, the step size coefficient is 1.0, and the model's adjustment suggestion is fully adopted. When a strong modification is identified and the sentiment polarity is strongly negative, the step size coefficient is 1.5 for aggressive exploration beyond the suggestion to deal with situations where users are extremely dissatisfied. Specific implementation examples of the fine-grained semantic recognition model for the field of apparel design include: Input layer: Receives cleaned user review text, with a single text length limit of 128 characters, and passes it to subsequent modules in character-level input format; Preprocessing layer: Word segmentation is performed based on the dictionary of clothing design domain, with jieba segmentation as the preferred method and supplemented with domain vocabulary to generate a word index sequence. At the same time, positional encoding is added to adapt to the input requirements of the pre-trained encoding layer. Pre-trained encoding layer: RoBERTa-Base is selected as the basic encoding module. The multi-layer Transformer structure captures the contextual semantic relationships of the text and outputs word-level semantic vectors with a dimension of 768. Domain Feature Enhancement Layer: A new fully connected layer (input dimension 768, output dimension 512, activation function using ReLU) is added to enhance the encoding vector by combining part-of-speech features (such as part nouns and evaluative adjectives) from the clothing design domain. Multi-dimensional output layer: Four parallel classification branches are designed, each corresponding to one of the four major structured features: Evaluation object branches: Softmax activation is used, and the output labels include 12 categories of key parts of trousers such as waistband, front trousers, and back trousers; Evaluation Dimensions Branch: Using Softmax activation, outputs four dimensions: size, length, tightness, and appearance style; Emotional polarity branch: Using Sigmoid activation, it outputs three polarities: positive, neutral, and negative. Modify the intensity branch: Use Sigmoid activation, and output three intensity categories: strongly recommended to modify, slightly adjust, or keep as is. The quality of training data directly determines the model's extraction accuracy. Therefore, it is necessary to construct a labeled dataset that fits the design scenario of trousers. Specific data sources include: Historical user review logs (including genuine reviews of the trousers design, with priority given to reviews with edit history); simulated data (based on key parts and design parameters of the trousers, manually constructed review texts of different dimensions, such as "The front crotch length is too short, it is strongly recommended to increase it by 2cm"). Publicly available clothing review data (reviews related to dress pants were filtered and further annotated); The final training sample dataset is formed, and the training set, validation set, and test set are divided according to a 7:2:1 ratio. The "entity-attribute-sentiment-intensity" four-tuple annotation method is adopted to clearly define the labels of each dimension in the training sample dataset: Evaluation subject: Only the core structural parts and attributes of the trousers (such as waistband, front panel, thigh circumference, trouser length, etc.) are marked. Evaluation dimensions: Strictly distinguish between four categories: size (too big / too small), length (too long / too short), tightness (too tight / too loose), and appearance style (incompatible / fitting); Emotional polarity: negative (expressing dissatisfaction, needing revision), neutral (objective description with no attitude), positive (expressing satisfaction, no revision needed); Modification intensity: Strongly recommend modification (including words such as "strongly" and "must"), make minor adjustments (including words such as "slightly" and "a little"), keep as is (positive evaluation or no modification required); Training objective: Employ a multi-task joint loss function to weighted sum the cross-entropy losses of the four output branches (with weights allocated according to importance: evaluation object 0.3, evaluation dimension 0.25, sentiment polarity 0.25, and modification intensity 0.2) to ensure that the model simultaneously optimizes the accuracy of multi-dimensional feature extraction. Key parameters: Batch size is set to 32, learning rate is dynamically adjusted (initial learning rate 2e-5, decaying by 0.9 every 2 epochs), training epochs are set to 10 (early stopping strategy is adopted, if the validation set loss does not decrease for 3 consecutive epochs, training is stopped), and the optimizer is AdamW (weight decay of 0.01 to avoid overfitting). Regularization optimization: Add a Dropout layer (probability 0.1) between the encoding layer and the domain augmentation layer, and at the same time use data augmentation techniques (synonym replacement, such as replacing "too tight" with "tight"; sentence word order adjustment) to improve the model's generalization ability.

[0024] The evaluation of the entire trousers design system based on historical user feedback information, including statistical calculations of satisfaction, accuracy, and solution iteration efficiency, specifically includes: Set up a periodically executed automated evaluation task that retrieves all complete design session logs generated during the evaluation cycle from the system database; Based on the user's final behavior in a single design session, if the user adopts the system-generated solution and completes the acceptance, the session is recorded as satisfactory; if the user abandons or explicitly rejects all solutions midway, it is recorded as unsatisfactory. Based on the output of the evaluation-parameter mapping learning module, determine the consistency between the parameter adjustment direction of the result and the parameter adjustment direction finally adopted by the user; Based on historical user evaluation information, we statistically calculate satisfaction, accuracy, and solution iteration efficiency, and display and store the calculation results through a visualization platform.

[0025] It can be explained that this solution achieves an objective and comprehensive evaluation of the digital design system for trousers by quantitatively evaluating all complete design session logs generated during the evaluation period. The core purpose is to verify the practical application effect of the "conditional graph neural network and evaluation feedback closed loop". It should be noted that each session log entry must contain five key data categories: Initial data: Initial design parameters input by the user, and initial 3D contour plot parameters generated by the system; Interactive data: users' past evaluation texts and structured evaluation features output by the semantic recognition model; Model output data: Evaluation - Adjustment amount recommended by the parameter mapping module, parameters of previous optimization schemes generated by the system; User behavior data: Records of users' choices of optimization solutions over the years, and markers of their final adoption / abandonment behavior; Efficiency data: total duration of a single session, number of iterations, and time taken to generate each optimization solution; It should be further explained that the quantitative logic and design significance of the aforementioned satisfaction, accuracy, and solution iteration efficiency are as follows: The quantification logic of the satisfaction is: Satisfaction = Number of Satisfactory Sessions / Total Number of Design Sessions, which is used to measure the user's final acceptance of the generated solution and reflect the system's comprehensive ability to meet user needs; it can be further refined into "Initial Solution Satisfaction = Number of Sessions in which the Initial Solution is Adopted / Total Number of Design Sessions", which is specifically used to evaluate the initial prediction accuracy of the conditional graph neural network. If the initial solution satisfaction is high, it means that the model does not need to rely on subsequent evaluation and optimization to accurately match the user's initial needs, and the core model performance is better; The accuracy quantification logic is as follows: a comprehensive evaluation is conducted using both direction and magnitude. The core is to measure the degree of match between the adjustment suggestions output by the evaluation-parameter mapping learning module and the user's final adoption of the parameter adjustment direction and magnitude. This reflects the model's accuracy in understanding user evaluations (especially fuzzy evaluations). The consistency of the parameter adjustment direction refers to the fact that in the results output by the evaluation-parameter mapping learning module, a positive sign for the corresponding adjustment value indicates a suggested increase, a negative sign indicates a suggested decrease, and 0 indicates no change. If the difference between the user's final adopted design parameter and the initial parameter triggered in this evaluation has the same sign, then the parameter is considered to be adjusted in the correct direction. The adjustment direction is consistent; the quantification logic of the direction accuracy is: accuracy = number of parameters with consistent adjustment direction / total number of parameters involved in adjustment; the quantification logic of the amplitude accuracy is: amplitude deviation rate = |model recommended adjustment amount - user actual adjustment amount| / user actual adjustment amount (if the user adjustment amount is not 0, the amplitude deviation rate is 0); the final comprehensive accuracy = direction consistency ratio (number of parameters with consistent adjustment direction / total number of parameters involved in adjustment) × 0.6 + (1 - average amplitude deviation rate of all parameters) × 0.4, which ensures that the adjustment direction does not deviate from the user's needs and controls the amplitude deviation within a reasonable range; The quantitative logic for the iteration efficiency of the proposed solution is as follows: It employs a dual evaluation metric of iteration count and time. The core objective is to measure the iteration speed from the "initial solution" to the "final user-adopted solution," reflecting the efficiency of the evaluation feedback loop. Specifically, this includes: Average iteration count = Total optimization rounds / Total number of sessions triggering optimization (only sessions triggering optimization are counted, as sessions without optimization do not require iteration). The fewer the iteration counts, the more accurate the adjustment solution recommended by the model and the lower the user interaction cost. Average iteration time = Total iteration time for all sessions / Total optimization rounds. The shorter the time, the faster the conditional graph neural network generates the optimization solution and the semantic recognition model processes the evaluation, resulting in higher system response efficiency. The combination of these two metrics comprehensively reflects the quality of iteration efficiency. It is necessary to further clarify the rules for determining edge scenarios during the evaluation process: If the user does not adjust any parameters (such as directly adopting the initial solution), this session will not be counted in the "total number of parameters involved in the adjustment" to avoid diluting the accuracy assessment results; If the same parameter is adjusted multiple times in multiple rounds of evaluation and optimization, the "final adjustment direction / magnitude by the user" is compared with the model's "last recommended direction / magnitude for this parameter" to ensure that the evaluation focuses on the adjustment logic that ultimately meets the user's needs. After the evaluation is completed, the calculated satisfaction, accuracy, and solution iteration efficiency results will be displayed and stored through a visualization platform. The visualization content includes indicator trend charts during the evaluation period (intuitively presenting changes in system performance), indicator comparison charts for different styles of trousers (straight leg / tapered / wide leg) and user groups (identifying high-performance and low-performance scenarios), and a list of abnormal conversations with low satisfaction / high iteration counts (facilitating tracing the root cause of problems). The stored evaluation results will be directly used for system optimization: if the satisfaction level is low, prioritize optimizing the human-computer interaction interface (such as supplementing evaluation guidance prompts) or expanding the benchmark template library to cover more styles; if the accuracy is low, supplement the training data for the corresponding evaluation scenarios (such as evaluations of "pant leg tightness" and "waistband height"), and fine-tune the parameters of the semantic recognition model and the evaluation-parameter mapping module; if the iteration efficiency is low, optimize the inference speed of the conditional graph neural network (such as using model distillation technology) or adjust the step size coefficient value strategy to reduce invalid iterations.

[0026] Furthermore, based on the same inventive concept as the aforementioned digital method for designing trousers, this solution proposes a digital system for designing trousers, comprising: The human-computer interaction module is used to set up the human-computer interaction port, provide selectable labels, receive the trouser design parameters input by the user, and display the 3D outline of the trousers through a synchronous preview area. The design generation module is used to construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers; based on the graph data structure of the trousers, and using a conditional graph neural network model combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated; an evaluation interface is set up to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; The system evaluation module is used to evaluate the entire trousers design system by statistically analyzing and calculating satisfaction, accuracy, and solution iteration efficiency based on historical user evaluation information. The design generation module includes: The graph data unit is used to construct a standardized three-dimensional reference template for trousers, extract the three-dimensional feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers. The design generation unit is used to generate a three-dimensional outline of the trousers based on the graphical data structure of the trousers, a conditional graphical neural network model, and user-input trouser design parameters. The evaluation optimization unit is used to set up an evaluation interface, receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters.

[0027] In summary, the advantages of this invention are: it takes into account both the ease of operation for non-professional users and the high precision and personalization needs of trouser design; it achieves intelligent and efficient digital design through a closed loop of "interaction-generation-feedback-evaluation", which significantly reduces the professional threshold and improves customization efficiency and accuracy.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A digital method for designing trousers, characterized in that, include: Set up a human-computer interaction port, provide selectable labels, receive user input of trouser design parameters, and display the trouser's 3D outline through a synchronous preview area; Construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and build the graph data structure of the trousers. Based on the graph data structure of the trousers, and using a conditional graph neural network model, combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated. Set up an evaluation interface to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; Based on historical user reviews, we statistically analyzed and calculated satisfaction, accuracy, and solution iteration efficiency to evaluate the entire trousers design system.

2. The digital method for designing trousers according to claim 1, characterized in that, The setup of the human-computer interaction port, which provides selectable labels, receives user-inputted design parameters for the trousers, and displays a 3D outline of the trousers through a synchronized preview area, specifically includes: The human-computer interaction port adopts a UI interface, which is presented in the form of a graphical user interface and has selectable labels deployed. The selectable tags include: basic mandatory tags and personalized optional tags; The basic required labels include: gender, size, waist circumference, hip circumference, pant length, thigh circumference, leg opening width, front crotch length, and back crotch length; The personalized optional tags include: style, color, pant type, pleat type, waistband, pocket style, hem treatment, and fabric; The pre-design parameters of the trousers can be set through the human-computer interaction port by means of radio buttons, drop-down menus, slider controls, direct input boxes and visual icons. Based on the selectable tag types, each tag is bound to a fixed parameter key name and standardized value through manual predefinition, thus constructing a front-end data mapping table; Users select tags through the interface, and the front-end data mapping table is used to obtain the parameter key name and standardized value of the selected tag, which are then stored in key-value pairs or structured JSON format. On the human-computer interaction interface, there is a synchronous preview area that displays a 3D outline of the trousers.

3. The digital method for designing trousers according to claim 2, characterized in that, The process of constructing a standardized 3D baseline template for trousers, extracting the 3D feature points of the template to construct graph nodes and edges, and building the graph data structure of the trousers specifically includes: Based on the principles of 3D garment CAD pattern making, a standardized 3D baseline template library for trousers is generated for different target styles, serving as the baseline template for all designs. The baseline template is broken down into several key structural parts according to garment structure, including at least: waistband, front trouser piece, back trouser piece, left trouser leg, right trouser leg, front dart / pleat, and back dart; For each structural part, based on its two-dimensional unfolded plate outline and key internal structural lines, three-dimensional reference feature points are manually defined. The set of all feature points constitutes the basic dataset of the reference template shape. Using a right-handed Cartesian coordinate system, the origin of the coordinate system is defined as 5cm directly above the navel in a standard human standing posture. The positive X-axis is horizontal to the left along the human body, the positive Y-axis is vertically downward, and the positive Z-axis is horizontal to the back. All defined 3D reference feature points are mapped to the coordinate system according to their geometric positions on the reference template to obtain their corresponding initial standard values ​​of 3D coordinates; Each 3D reference feature point and its coordinates are treated as a node in the graph, and each node is assigned an attribute, which includes at least its coordinates and a label to identify its structural part and semantic role. Based on the spatial adjacency and technological logic of clothing structure, connection relationships are defined in the node set as edges of the graph. The definition rules include: Contour edge: An edge is established between two adjacent feature point nodes on a continuous contour line of the same structural part; Symmetric edge: Establish an edge between a pair of feature point nodes that are symmetrical about the YZ plane of the human body; Seam edge: Establish an edge between corresponding matching feature point nodes of two different structural parts that need to be seamed; Spatial Adjacency Edge: An edge is established between feature point nodes in three-dimensional space whose Euclidean distance is less than a preset threshold and belong to the same surface region; Construct the graph data structure for the trousers based on the set of graph nodes and the set of graph edges.

4. The digital method for designing trousers according to claim 3, characterized in that, The process of generating a 3D outline of the trousers based on the graph data structure of the trousers, using a conditional graph neural network model, and combining user-input trouser design parameters, specifically includes: Obtain the design parameters of the trousers input by the user and perform standardized preprocessing on them; The preprocessing includes: normalizing the continuous numerical parameters, converting the discrete option parameters into one-hot encoded vectors, and concatenating all the preprocessed parameters into a unified global design feature vector. Construct a conditional graph neural network model that takes the global design feature vector as the conditional input and the graph topology and initial node feature inputs obtained from the benchmark template library. In each layer of the conditional graph neural network, the global design feature vector is processed by a conditional coding network to generate affine transformation parameters for the dynamic scaling parameters and bias parameters corresponding to that layer. By using the generated scaling and bias parameters, the fixed weight matrix of the current graph convolutional layer is dynamically modulated to obtain the conditionally dependent effective weight matrix. After multiple layers of conditional graph convolution, the network output layer generates a predicted 3D coordinate offset value for each graph node; Add the original 3D coordinates of each node to its corresponding predicted offset to obtain the 3D coordinates of the node under the user-input design parameters; All adjusted feature point nodes are sorted according to the predefined contour edge connection relationship in the graph data structure, and then connected sequentially using the B-spline curve algorithm to generate the three-dimensional contour edges of each structural part. For the area enclosed by the contour lines, a triangular patch filling algorithm is used to construct a three-dimensional surface, and the surfaces of each part are spliced ​​together to form a complete three-dimensional mesh model of the trousers. Based on the surface continuity constraints implied by the predefined spatial adjacency edges in the graph data structure, the generated 3D mesh model is subjected to local Laplacian smoothing. The optimized 3D contour model is converted into a universal 3D format and rendered in real time to the synchronous preview area of ​​the human-computer interaction port.

5. The digital method for designing trousers according to claim 4, characterized in that, The setting of the evaluation interface, which receives user evaluation information, extracts evaluation features through semantic recognition, and establishes a mapping relationship between evaluation features and design parameters, specifically includes: Next to the preview area of ​​the 3D outline of the trousers on the human-computer interaction port, set up a user evaluation interface; The evaluation interface receives user evaluation information for the current design result in the form of text input boxes, structured rating forms, and sentiment selection buttons. The collected unstructured text evaluations are cleaned, including removing irrelevant characters, correcting spelling errors, and word segmentation. Construct a fine-grained semantic recognition model for the field of apparel design to extract structured evaluation features from preprocessed user reviews; The semantic recognition model adopts a fine-tuned architecture based on a pre-trained language model, and its output layer is adapted for multi-label classification tasks. The recognition results of the semantic recognition model are integrated into a structured evaluation feature vector; Construct an evaluation-parameter mapping learning module. This module takes the structured evaluation feature vector and the design parameter vector corresponding to this evaluation as input, and aims to output a set of recommended design parameter adjustments. The evaluation-parameter mapping learning module uses a multilayer perceptron regression model, and its training data comes from historical interaction logs. Each historical interaction log contains the initial design parameters, user evaluation features, and the design parameters finally accepted by the user. The training objective of the evaluation-parameter mapping learning module is set to minimize the mean square error between the parameter adjustment of its output and the ideal adjustment implied by real user behavior. Through training, the evaluation-parameter mapping learning module learns the mapping relationship from abstract evaluation to specific quantitative parameter adjustment suggestions, and outputs the recommended adjustment amount and optimized design parameter candidates. The optimized design parameter candidates are input into the conditional graph neural network model to generate the corresponding 3D outline of the trousers. The generated optimized 3D silhouette of the trousers will be presented to the user in the preview area in a side-by-side comparison or carousel format, and the user's selection results will be recorded together with the evaluation information.

6. The digital method for designing trousers according to claim 5, characterized in that, The evaluation of the entire trousers design system based on historical user feedback information, including statistical calculations of satisfaction, accuracy, and solution iteration efficiency, specifically includes: Set up a periodically executed automated evaluation task that retrieves all complete design session logs generated during the evaluation cycle from the system database; Based on the user's final behavior in a single design session, if the user adopts the system-generated solution and completes the acceptance, the session is recorded as satisfactory; if the user abandons or explicitly rejects all solutions midway, it is recorded as unsatisfactory. Based on the output of the evaluation-parameter mapping learning module, determine the consistency between the parameter adjustment direction of the result and the parameter adjustment direction finally adopted by the user; Based on historical user evaluation information, we statistically calculate satisfaction, accuracy, and solution iteration efficiency, and display and store the calculation results through a visualization platform.

7. A digital system for designing trousers, characterized in that, A digital method for designing trousers as described in any one of claims 1-6 includes: The human-computer interaction module is used to set up the human-computer interaction port, provide selectable labels, receive the trouser design parameters input by the user, and display the 3D outline of the trousers through a synchronous preview area. The design generation module is used to construct a standardized 3D baseline template for trousers, extract the 3D feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers; based on the graph data structure of the trousers, and using a conditional graph neural network model combined with the trousers design parameters input by the user, a 3D outline of the trousers is generated; an evaluation interface is set up to receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters; The system evaluation module is used to evaluate the entire trousers design system by statistically analyzing and calculating satisfaction, accuracy, and solution iteration efficiency based on historical user evaluation information.

8. The digital system for designing trousers according to claim 7, characterized in that, The design generation module includes: The graph data unit is used to construct a standardized three-dimensional reference template for trousers, extract the three-dimensional feature points of the template to construct graph nodes and edges, and construct the graph data structure of the trousers. The design generation unit is used to generate a three-dimensional outline of the trousers based on the graphical data structure of the trousers, a conditional graphical neural network model, and user-input trouser design parameters. The evaluation optimization unit is used to set up an evaluation interface, receive user evaluation information, extract evaluation features through semantic recognition, and establish a mapping relationship between evaluation features and design parameters.

Images