Intelligent costume design method based on virtual fitting

Abstract

The invention discloses an intelligent costume design method based on virtual fitting, and belongs to the field of costume design, and the design method comprises the following specific steps: I, collecting basic body shape data and health parameters of a user, predicting the body shape change trend of the user, and constructing a dynamic body shape model of the user; iI, collecting emotional responses of the user under various visual stimuli when the user watches different garment styles, colors and materials, and establishing a user emotional preference label; according to the method, the body change trend and aesthetic preference of the user can be accurately captured, the diversified and customized requirements are met, manual drawing and multi-round sample drawing are not needed, the design efficiency is greatly improved, the development cost is reduced, a scientific basis is provided for design optimization, garment adjustment with user experience as the center is achieved, and environmental protection and garment full-life-cycle management are effectively promoted.

Description

Technical Field

[0001] The present invention relates to the field of clothing design, and in particular to an intelligent clothing design method based on virtual fitting. Background Art

[0002] With the rapid development of artificial intelligence, computer graphics, biomechanical modeling, and wearable devices, the apparel industry is undergoing a profound transformation, transitioning from traditional manufacturing to intelligent design and digital manufacturing. Consumers are no longer satisfied with standard sizes and popular styles; instead, they prioritize clothing's adaptability to individual body shapes, comfort, and emotional and aesthetic expression. Furthermore, environmental sustainability is becoming a core concern for global apparel brands and manufacturers. Meeting individual user needs while reducing resource waste and carbon emissions has become a key area of technological innovation. Current mainstream virtual fitting systems rely on static 3D modeling and manual parameter settings, making them incapable of responding to dynamic changes in user body shape, differing aesthetic preferences, and the diverse physical wearing environments. Furthermore, traditional pattern development processes still require extensive manual intervention and experimental modification, making them inefficient in supporting the fast-paced, highly complex market demands. Furthermore, resource allocation during the clothing design process (e.g., material usage and structural combinations) often lacks systematic optimization and sustainability assessment, resulting in inefficient design implementation and significant waste in the production process. Therefore, developing an intelligent clothing design method based on virtual fitting is crucial.

[0003] After searching, Chinese patent number CN117422896A discloses an intelligent design method and system for clothing process templates. Although this invention achieves efficient and accurate clothing process design, it cannot accurately capture the user's body changes and aesthetic preferences, and cannot meet the diverse and customized needs. In addition, the clothing design process requires manual drawing and multiple rounds of sampling, which reduces design efficiency and increases development costs. Therefore, we propose an intelligent clothing design method based on virtual fitting. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose an intelligent clothing design method based on virtual fitting.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: An intelligent clothing design method based on virtual fitting, the specific steps of the design method are as follows: Ⅰ. Collect the user's basic body shape data and health parameters, predict the user's body shape change trend, and build a dynamic body shape model for the user; II. Collect users’ emotional responses to visual stimulation of different clothing styles, colors, and materials, and establish user emotional preference labels; III. Based on the user's dynamic body model and emotional preference tags, intelligent pattern evolution is performed to generate user-personalized patterns while automatically adjusting cutting lines and structural points. IV. Convert the user's personalized pattern into a virtual clothing model and bind it to the user's 3D body model for a try-on simulation; V. During the virtual try-on process, the pressure distribution of the contact surface between the garment and the human body is modeled in real time, and the wearing comfort is evaluated; VI. Build a clothing life cycle carbon footprint database, conduct quantitative analysis of candidate materials and structures, and optimize material utilization and structural sustainability.

[0006] As a further solution of the present invention, the basic body shape data in step I is mainly obtained through 3D scanning, image recognition, manual measurement or user-initiated filling; the health parameters are obtained through wearable devices or health monitoring systems.

[0007] As a further solution of the present invention, the specific steps of constructing the user's dynamic body model in step I are as follows: S1.1: Collect health parameters from the user's wearable device or health monitoring system over a continuous period of time, fill missing values in the collected health parameters through linear interpolation, and then standardize each group of health parameters to unify them into pre-set ranges. Health parameters include weight, resting heart rate, activity intensity, and sleep quality; S1.2: Based on the processed health parameters, obtain the user's estimated daily energy expenditure and estimated weight change rate. A sliding window with a length of 7 days is then established, and a prediction sample is constructed using the established sliding window. The prediction sample consists of 7 daily feature vectors, each containing the daily average activity intensity, weight, resting heart rate, estimated daily energy expenditure, and estimated weight change rate. S1.3: Collect the user's historical health parameters and create training samples based on the prediction sample construction steps. Use the user's body shape changes over time as the true labels of each training sample to establish a change prediction model. Initialize the number of network layers, the number of hidden units in each layer, the activation function, the loss function, and the weight initialization method for the change prediction model. S1.4: The training samples are sequentially input into the change prediction model. The change prediction model uses the forward propagation algorithm to pass the training samples layer by layer. The LSTM layer calculates the state and updates the hidden state of the training samples at each moment. The temporal feature encoding vector is extracted from the final hidden state output by the LSTM layer. S1.5: Input the encoded vector into the fully connected regression layer to generate the user's body shape parameters for the next seven days. This is then output through the output layer. The deviation between the user's body shape parameters and the true label is calculated using the mean squared error function. The deviation is then propagated layer by layer from the output layer of the change prediction model through the backpropagation algorithm. The gradient of the deviation with respect to the parameters of each network layer of the change prediction model is also calculated. S1.6: Based on the obtained gradient values, the Adam optimizer is used to optimize the parameters of each network layer of the change prediction model. The trained change prediction model is then verified using training samples that did not participate in the training. The deviation value during the verification process is calculated. If the deviation value is higher than the preset threshold, the change prediction model is retrained and verified until the deviation value of the verification process converges to the preset threshold. S1.7: The latest prediction samples are input into the trained change prediction model. The prediction samples are forward propagated layer by layer in the change prediction model. The LSTM layer uses a gating mechanism to memorize and forget the body shape change trend and update the hidden state. The fully connected regression layer is then used to generate and output the predicted body shape parameters. The predicted body shape parameters for the next 30 to 90 days are continuously output, and a user dynamic body shape model for the corresponding time period is constructed. At the same time, the user dynamic body shape model is updated in real time based on the user's basic body shape data and health parameters.

[0008] As a further embodiment of the present invention, the specific calculation formula for the estimated daily energy expenditure in S1.2 is as follows: ; Where, Representative Estimated total energy expenditure for the day; Representative Daily user activity intensity; Representative User weight on the day; Represents metabolic conversion efficiency.

[0009] The specific calculation formula for the estimated rate of weight change described in S1.2 is as follows: ; Where, Representative Estimated user weight change over the day; Represents the basal metabolic expenditure required to maintain current weight; Represents the calories contained in 1kg of fat.

[0010] As a further solution of the present invention, the specific steps of the intelligent pattern evolution in step II are as follows: S2.1: The user's dynamic body model data and emotional preference labels are concatenated into a unified joint input code. The historical clothing patterns in the historical pattern database are uniformly encoded into corresponding latent space vectors through a convolutional autoencoder, and the encoding vectors of each historical clothing pattern are used as real samples. S2.2: Build a GAN network consisting of a generator network and a discriminator network. The joint input encoding is used as the starting input data and transmitted to the generator network. The generator network then transforms features layer by layer through fully connected units from shallow to deep layers and outputs the final joint features. S2.3: The generator network uses the control unit to guide the style of the generated results based on the sentiment preference label in the final joint feature, and generates intermediate features after conditional control. The output layer then compresses the intermediate feature map to the latent space dimension and outputs the latent vector of the corresponding version. S2.4: The real samples and the latent vectors generated by the generator network are used as input data and fed into the discriminator network. The discriminator network then performs multiple linear transformations and activation operations on the real samples and the latent vectors, extracting deep features of the input vectors. The output layer then outputs a confidence score for each set of input data, indicating whether it is a real sample, with 1 indicating a real sample and 0 indicating a false positive. S2.5: Based on the output of the generator network and the discriminator network, calculate the GAN network loss value using the adversarial loss function. Based on this loss value, update the generator network and the discriminator network using the backpropagation algorithm and the Adam optimizer. Repeatedly update the generator network and the discriminator network until the GAN network loss value converges to the preset range. S2.5: Input the latest user dynamic body model data and emotional preference labels into the GAN network, and after forward propagation through the generator network, generate a personalized pattern vector, and decode the generated personalized pattern vector to generate the corresponding personalized clothing pattern data.

[0011] As a further solution of the present invention, the specific steps of automatically adjusting the cutting line and the structure point in step III are as follows: S3.1: Based on the decoded personalized clothing pattern data, a corresponding two-dimensional plane structure graph is established. The two-dimensional plane structure graph is discretized using a topological grid, and the clothing pattern structure points are used as nodes and the cutting lines as edges to establish an initial topological graph; S3.2: Based on the aesthetic curvature smoothness, structural stitchability, and patch tension balance, a set of objective functions that combine multiple sub-energy terms is set, and constraint modeling is performed on the established objective functions. The positions of each node in the initial topology are then input into the objective functions, and the values to be optimized are output; S3.3: Calculate the gradient vector of the objective function for each node in the topology graph to obtain the gradient descent direction of the current node position, update the position of each node based on the gradient direction, and recalculate the objective function value of the updated node position. At the same time, compare the updated objective function value with the value to be optimized. Based on the comparison result, adaptively adjust the step size of each iteration; S3.4: Iterate and update the position of each node repeatedly until the change in the position of the structural point in multiple rounds of iterations converges to a preset range. Then stop the optimization, recombine the final node set and edge set, and generate a personalized clothing pattern structure diagram after topology optimization.

[0012] As a further solution of the present invention, the specific calculation formula of the objective function in S3.2 is as follows: ; Where, represents the total topological energy; represents the curvature smoothing energy; represents the contour fitting energy; represents the structural tension energy; 、 、 Represent the weight coefficient of each energy term; The specific calculation formula for the constraint modeling described in S3.2 is as follows: ; Where, Representative The edge and The angle between the edges, if the two sides are smoothly connected, then ; Represents the number of trimming sides; Representative Coordinates of structural points; Representative Coordinates of target body reference points; represents the total number of structural points; as well as Representing the The two structural points connected by an edge; Representative The reference length of the edge at initialization.

[0013] As a further solution of the present invention, the specific steps of performing real-time modeling of the pressure distribution on the contact surface between the garment and the human body and evaluating the wearing comfort in step V are as follows: S4.1: Based on the simulated fitting results, the Euclidean distance between each garment patch and the human body surface is calculated. If any Euclidean distance is less than a preset contact distance threshold, it indicates that the garment patch is in contact with the human body. Based on the judgment result, a contact area point set is selected; S4.3: Convert the normal deformation of each contact point in the contact area into a local contact pressure to establish a contact pressure distribution model, obtain the pressure values of all contact points, and interpolate and map each pressure value on the skin surface to form a continuous pressure field distribution map; S4.4: Simulate the friction between fabric and skin and calculate the charge density per unit area of the current garment based on the continuous pressure field distribution diagram and the corresponding relative motion coefficient. Based on the charge density distribution, solve the potential distribution near the skin surface and determine the area and intensity of static electricity accumulation. S4.5: Statistically analyze the contact pressure peak value, mean value, change gradient, and electrostatic potential amplitude to construct a comprehensive comfort scoring model and calculate the comfort score of the currently tried-on garment. If the comfort score is lower than the preset threshold, the material stiffness, friction coefficient, or local thickness of the contact area is automatically adjusted based on the analysis results, and corresponding correction suggestions are generated.

[0014] As a further solution of the present invention, the specific steps of optimizing material utilization and structural sustainability in step VI are as follows: S5.1: Discretize the design parameters of the virtual garment and encode them into multiple sets of individual chromosomes. Construct two sets of objective functions based on material utilization and structural sustainability, respectively. Based on the number of individual chromosomes generated in each set, generate an initial population using a random or semi-heuristic method. Then, calculate the effective material utilization and structural sustainability of each individual in the population using the two sets of objective functions. S5.2: Based on the effective material utilization rate and structural sustainability, all individuals in the population are stratified according to the Pareto dominance relationship. All non-dominated individuals are classified into the first-level non-dominated solution set. The next level of non-dominated solution sets are searched from the remaining individuals in turn until the stratification is completed and the non-dominated solution sets of multiple levels after stratification are obtained. S5.3: In each non-dominated layer, calculate the crowding distance between each group of individuals and their neighbors. From the lowest to the highest non-dominated layer, add individuals whose crowding distance exceeds the preset threshold to the new population until the preset population size is reached. If the population size exceeds the upper limit after individuals in the current non-dominated layer choose to join, fill the remaining vacancies based on descending crowding distance. S5.4: Randomly select a pair of parent chromosomes from the new population and generate two daughter chromosomes by simulating a binary crossover method. Then, mutate each design parameter in each daughter chromosome using Gaussian perturbation. Repeat the selection, crossover, and compilation operations multiple times to generate multiple sets of daughter chromosomes. S5.5: Decode each daughter chromosome to obtain the corresponding cutting plan, calculate the effective material utilization rate, and then comprehensively evaluate the structural sustainability of each daughter chromosome based on the disassembly index and reuse score of the structural components; S5.6: Based on the effective material utilization and structural sustainability of the generated daughter chromosomes, perform non-dominated sorting, stratify all the generated daughter chromosomes for the next round of iteration, repeat the selection, crossover, mutation and sorting process until the effective material utilization and structural sustainability, and the rate of change for multiple rounds of iterations is less than the threshold, and output the individual chromosome with the highest material utilization and structural sustainability, after which the designer performs the final virtual pattern making.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This intelligent clothing design method based on virtual fitting jointly encodes the user's dynamic body shape data and emotional labels, and uniformly encodes the historical clothing patterns in the historical pattern database into corresponding latent space vectors through a convolutional autoencoder. The joint encoding and latent space vectors are input into an anti-generation network to generate a personalized clothing pattern vector, and then the generated personalized clothing pattern vector is decoded to generate a two-dimensional pattern structure diagram. Subsequently, a topology optimization algorithm is used to adjust the position of the structural points. Then, through simulation of the user's try-on effect, the fabric contact pressure field and electric potential distribution are constructed to evaluate the wearing comfort. If the score is lower than the threshold, material or structural correction suggestions are proposed. Finally, the clothing design parameters are discretely encoded into genetic chromosomes, and a multi-objective genetic algorithm is combined to optimize material utilization and structural sustainability. The optimal design scheme is screened through non-dominated sorting and crossover mutation iteration. It can accurately capture the user's body change trends and aesthetic preferences, meet diverse and customized needs, and does not require manual drawing and multiple rounds of samples, greatly improving design efficiency and reducing development costs. It provides a scientific basis for design optimization, realizes clothing adjustment centered on user experience, and effectively promotes environmental protection and clothing life cycle management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0017] Figure 1 This is a flowchart of an intelligent clothing design method based on virtual fitting proposed by the present invention. DETAILED DESCRIPTION

[0018] Example 1

[0019] Reference Figure 1 , an intelligent clothing design method based on virtual fitting, the specific steps of the design method are as follows: Collect users' basic body data and health parameters, predict users' body shape change trends, and build users' dynamic body shape models.

[0020] Specifically, health parameters are collected from the user's wearable device or health monitoring system over a continuous period of time, and missing values in the collected health parameters are filled by linear interpolation. Each group of health parameters is then standardized to unify each health parameter into a preset range, where the health parameters include weight, resting heart rate, activity intensity, and sleep quality. The estimated daily energy consumption and weight change rate of the corresponding user are obtained based on the processed health parameters. A sliding window with a time length of 7 days is then established, and a prediction sample is constructed through the established sliding window. The prediction sample is composed of daily feature vectors for 7 days, and the daily feature vectors respectively contain the daily average Activity intensity, weight, resting heart rate, estimated daily energy consumption and estimated weight change rate of the day, collect user historical health parameters, and establish training samples based on the construction steps of prediction samples. The user's body shape changes at the corresponding time are used as the true labels of each training sample to establish a change prediction model, and initialize the number of network layers, the number of hidden units in each layer, the activation function, the loss function and the weight initialization method of the change prediction model. The training samples are input into the change prediction model in sequence. The change prediction model is based on the forward propagation algorithm, which transmits the training samples layer by layer and performs state calculation and hidden state update on the training samples at each moment through the LSTM layer. The temporal feature encoding vector is extracted from the final hidden state output by the TM layer, and the encoding vector is input into the fully connected regression layer to generate the user's body shape parameters for the next 7 days. The parameters are output through the output layer, and the deviation between the user's body shape parameters and the true label is calculated through the mean square error function. The deviation value is passed layer by layer from the output layer of the change prediction model through the back propagation algorithm. At the same time, the gradient value of the deviation value for the parameters of each network layer of the change prediction model is calculated. Based on the obtained gradient value, the Adam optimizer is used to optimize the parameters of each network layer of the change prediction model. After that, the training samples that did not participate in the training are used to verify the change prediction model after training, and the verification process is calculated. Deviation value. If the deviation value is higher than the preset threshold, the change prediction model is retrained and verified until the deviation value of the verification process converges to the preset threshold. The latest prediction sample is input into the trained change prediction model. The prediction sample is forward propagated layer by layer in the change prediction model. The LSTM layer uses a gating mechanism to memorize and forget the body shape change trend and update the hidden state. Then, the fully connected regression layer is used to generate the predicted body shape parameters and output them. The predicted body shape parameters for the next 30 to 90 days are continuously output, and the user dynamic body shape model for the corresponding time period is constructed. At the same time, the user dynamic body shape model is updated in real time based on the user's basic body shape data and health parameters.

[0021] In addition, it should be noted that the specific calculation formula for the estimated daily energy expenditure is as follows: ; Where, Representative Estimated total energy expenditure for the day; Representative Daily user activity intensity; Representative User weight on the day; Represents metabolic conversion efficiency.

[0022] The specific calculation formula for the estimated weight change rate is as follows: ; Where, Representative Estimated user weight change over the day; Represents the basal metabolic expenditure required to maintain current weight; Represents the calories contained in 1kg of fat.

[0023] It should be further explained that basic body shape data is mainly obtained through 3D scanning, image recognition, manual measurement or user-initiated filling; health parameters are obtained through wearable devices or health monitoring systems.

[0024] Collect users' emotional responses to visual stimuli of different clothing styles, colors and materials, and establish user emotional preference labels.

[0025] Specifically, the user's dynamic body model data and emotional preference labels are integrated into a unified joint input code through a splicing operation, and the historical clothing patterns in the historical pattern database are uniformly encoded into corresponding latent space vectors through a convolutional autoencoder. Each historical clothing pattern encoding vector is used as a real sample to establish a set of GAN networks. The GAN network includes a generator network and a discriminator network. The joint input code is used as the starting input data and transmitted to the generator network. After that, the generator network performs feature transformation layer by layer through fully connected units from shallow to deep layers, and outputs the final joint feature. Based on the emotional preference label in the final joint feature, the generator network guides the style direction of the generated result through the control unit and generates intermediate features after conditional control. After that, the output layer compresses the intermediate feature map to the latent space dimension and outputs the latent vector of the corresponding pattern. The real sample and the latent vector generated by the generator network are combined. As input data, it is input into the discriminator network. The discriminator network then performs multiple linear transformations and activation operations on the real samples and potential vectors, and extracts the deep features of the input vectors. The output layer then outputs a confidence score of whether each set of input data is a real sample, 1 for a real sample, otherwise it is not. According to the output results of the generator network and the discriminator network, the GAN network loss value is calculated through the adversarial loss function, and based on the loss value, the generator network and the discriminator network are updated based on the back propagation algorithm and the Adam optimizer. The generator network and the discriminator network are updated repeatedly until the GAN network loss value converges to the preset range. The latest user dynamic body model data and emotional preference labels are input into the GAN network, and after forward propagation processing through the generator network, a personalized pattern vector is generated, and the generated personalized pattern vector is decoded to generate the corresponding personalized clothing pattern data.

[0026] Based on the user's dynamic body model and emotional preference tags, intelligent pattern evolution is performed to generate user-personalized patterns, while automatically adjusting cutting lines and structural points.

[0027] Specifically, based on the personalized clothing pattern data generated by decoding, a corresponding two-dimensional plane structure graph is established, and the two-dimensional plane structure graph is discretized using a topological grid, and the clothing pattern structure points are used as nodes and the cutting lines are used as edges to establish an initial topological graph. According to the aesthetic curvature smoothness, structural sutureability and patch tension balance, a set of objective functions containing a combination of multiple sub-energy items is set, and the established objective function is constrained modeling. Then, the position of each node in the initial topological graph is input into the objective function, and the value to be optimized is output. The gradient vector of each node in the topological graph to the objective function is calculated to obtain the gradient descent direction of the current node position, and the position of each node is updated based on the gradient direction, and the objective function value of the updated node position is recalculated. At the same time, the updated objective function value is compared with the value to be optimized. Based on the comparison result, the step size of each round of iteration is adaptively adjusted, and the position of each node is repeatedly iterated until the change value of the structural point position in the update amount of multiple iterations converges to a preset range. Then the optimization is stopped, and the final node set and edge set are recombined to generate a personalized clothing pattern structure graph after topology optimization.

[0028] In this embodiment, the specific calculation formula of the indicator function is as follows: ; Where, represents the total topological energy; represents the curvature smoothing energy; represents the contour fitting energy; represents the structural tension energy; 、 、 Represent the weight coefficient of each energy term; The specific calculation formula for constraint modeling is as follows: ; Where, Representative The edge and The angle between the edges, if the two sides are smoothly connected, then ; Represents the number of trimming sides; Representative Coordinates of structural points; Representative Coordinates of target body reference points; represents the total number of structural points; as well as Representing the The two structural points connected by an edge; Representative The reference length of the edge at initialization.

[0029] Example 2 Reference Figure 1, an intelligent clothing design method based on virtual fitting, the specific steps of the design method are as follows: The user's personalized pattern is converted into a virtual clothing model and bound to the user's three-dimensional body model for try-on simulation.

[0030] During the virtual try-on process, the pressure distribution on the contact surface between the clothing and the human body is modeled in real time, and the wearing comfort is evaluated.

[0031] Specifically, based on the simulated fitting results, the Euclidean distance between each clothing patch and the human body surface is calculated. If there is a Euclidean distance less than a preset contact distance threshold, it means that the clothing patch is in contact with the human body. Based on the judgment result, a contact area point set is screened out, and the normal deformation of each contact point in the contact area point set is converted into local contact pressure to establish a contact pressure distribution model, and the pressure values of all contact points are obtained. Each pressure value is interpolated and mapped on the skin surface to form a continuous pressure field distribution map to simulate the friction between the fabric and the skin. Based on the continuous pressure field distribution map and the corresponding relative motion coefficient, the charge density per unit area of the current clothing is calculated. Based on the charge density distribution, the electric potential distribution near the skin surface is solved, and the existing static electricity accumulation area and intensity are determined. The contact pressure peak, mean, change gradient and electrostatic potential amplitude are statistically calculated to construct a comprehensive comfort scoring model, and the comfort score of the current tried-on clothing is calculated. If the comfort score is lower than the preset threshold, the material stiffness, friction coefficient or local thickness of the contact area is automatically adjusted according to the analysis results, and corresponding correction suggestions are generated.

[0032] Build a clothing life cycle carbon footprint database, conduct quantitative analysis of candidate materials and structures, and optimize material utilization and structural sustainability.

[0033] Specifically, the design parameters of the virtual clothing are discretized and encoded into multiple groups of individual chromosomes. Two sets of objective functions are constructed according to material utilization and structural sustainability. Based on the number of individual chromosomes generated in each group, the initial population is generated by random or semi-heuristic methods, and the effective material utilization and structural sustainability of each individual in the population are calculated by the two sets of objective functions. According to the effective material utilization and structural sustainability, all individuals in the population are stratified according to the Pareto dominance relationship, and all non-dominated individuals are classified into the first-level non-dominated solution set, and the next layer of non-dominated solution sets are searched from the remaining individuals in turn until the stratification is completed, and the non-dominated solution sets of multiple levels after stratification are obtained. In each non-dominated layer, the crowding distance between each group of individuals and their adjacent individuals is calculated. According to the non-dominated level from low to high, the individuals in each non-dominated level with a crowding distance higher than the preset threshold are added to the new population in turn until the preset population size is reached. If the current non-dominated layer individual chooses to join, the population size If the quantity exceeds the upper limit, the remaining places are filled in descending order of the crowding distance. The parent chromosome pair is randomly selected from the new population, and two daughter chromosomes are generated by simulating the binary crossover method. Each design parameter in each daughter chromosome is mutated through Gaussian perturbation. The selection, crossover and compilation operations are repeated many times to generate multiple groups of daughter chromosomes. The generated daughter chromosomes are decoded to obtain the corresponding cutting scheme, and the effective material utilization rate is calculated. Then, the structural sustainability of each daughter chromosome is comprehensively evaluated based on the detachable index and reuse score of the structural components. According to the effective material utilization and structural sustainability of the generated daughter chromosomes, non-dominated sorting is performed, and all generated daughter chromosomes are stratified for the next round of iteration. The selection, crossover, mutation and sorting process are repeated until the effective material utilization and structural sustainability are achieved. The rate of change for multiple rounds of iterations is less than the threshold, and the individual chromosome with the highest material utilization and structural sustainability is output. The designer then performs the final virtual plate making.

Claims

1. An intelligent clothing design method based on virtual fitting, characterized in that: The specific steps of this design method are as follows: Ⅰ. Collect the user's basic body shape data and health parameters, predict the user's body shape change trend, and build a dynamic body shape model for the user; II. Collect users’ emotional responses to visual stimulation of different clothing styles, colors, and materials, and establish user emotional preference labels; III. Based on the user's dynamic body model and emotional preference tags, intelligent pattern evolution is performed to generate user-personalized patterns while automatically adjusting cutting lines and structural points. IV. Convert the user's personalized pattern into a virtual clothing model and bind it to the user's 3D body model for a try-on simulation; V. During the virtual try-on process, the pressure distribution of the contact surface between the garment and the human body is modeled in real time, and the wearing comfort is evaluated; VI. Build a clothing life cycle carbon footprint database, conduct quantitative analysis of candidate materials and structures, and optimize material utilization and structural sustainability.

2. The intelligent clothing design method based on virtual fitting according to claim 1, characterized in that: The specific steps for constructing the user's dynamic body shape model described in step I are as follows: S1.1: Collect health parameters from the user's wearable device or health monitoring system over a continuous period of time, fill missing values in the collected health parameters through linear interpolation, and then standardize each group of health parameters to unify them into pre-set ranges. Health parameters include weight, resting heart rate, activity intensity, and sleep quality; S1.2: Based on the processed health parameters, obtain the user's estimated daily energy expenditure and estimated weight change rate. A sliding window with a length of 7 days is then established, and a prediction sample is constructed using the established sliding window. The prediction sample consists of 7 daily feature vectors, each containing the daily average activity intensity, weight, resting heart rate, estimated daily energy expenditure, and estimated weight change rate. S1.3: Collect the user's historical health parameters and create training samples based on the prediction sample construction steps. Use the user's body shape changes over time as the true labels of each training sample to establish a change prediction model. Initialize the number of network layers, the number of hidden units in each layer, the activation function, the loss function, and the weight initialization method for the change prediction model. S1.4: The training samples are sequentially input into the change prediction model. The change prediction model uses the forward propagation algorithm to pass the training samples layer by layer. The LSTM layer calculates the state and updates the hidden state of the training samples at each moment. The temporal feature encoding vector is extracted from the final hidden state output by the LSTM layer. S1.5: Input the encoded vector into the fully connected regression layer to generate the user's body shape parameters for the next seven days. This is then output through the output layer. The deviation between the user's body shape parameters and the true label is calculated using the mean squared error function. The deviation is then propagated layer by layer from the output layer of the change prediction model through the backpropagation algorithm. The gradient of the deviation with respect to the parameters of each network layer of the change prediction model is also calculated. S1.6: Based on the obtained gradient values, the Adam optimizer is used to optimize the parameters of each network layer of the change prediction model. The trained change prediction model is then verified using training samples that did not participate in the training. The deviation value during the verification process is calculated. If the deviation value is higher than the preset threshold, the change prediction model is retrained and verified until the deviation value of the verification process converges to the preset threshold. S1.7: The latest prediction samples are input into the trained change prediction model. The prediction samples are forward propagated layer by layer in the change prediction model. The LSTM layer uses a gating mechanism to memorize and forget the body shape change trend and update the hidden state. The fully connected regression layer is then used to generate and output the predicted body shape parameters. The predicted body shape parameters for the next 30 to 90 days are continuously output, and a user dynamic body shape model for the corresponding time period is constructed. At the same time, the user dynamic body shape model is updated in real time based on the user's basic body shape data and health parameters.

3. The intelligent clothing design method based on virtual fitting according to claim 2, characterized in that: The specific steps of the intelligent pattern evolution described in step II are as follows: S2.1: The user's dynamic body model data and emotional preference labels are concatenated into a unified joint input code. The historical clothing patterns in the historical pattern database are uniformly encoded into corresponding latent space vectors through a convolutional autoencoder, and the encoding vectors of each historical clothing pattern are used as real samples. S2.2: Build a GAN network consisting of a generator network and a discriminator network. The joint input encoding is used as the starting input data and transmitted to the generator network. The generator network then transforms features layer by layer through fully connected units from shallow to deep layers and outputs the final joint features. S2.3: The generator network uses the control unit to guide the style of the generated results based on the sentiment preference label in the final joint feature, and generates intermediate features after conditional control. The output layer then compresses the intermediate feature map to the latent space dimension and outputs the latent vector of the corresponding version. S2.4: The real samples and the latent vectors generated by the generator network are used as input data and fed into the discriminator network. The discriminator network then performs multiple linear transformations and activation operations on the real samples and the latent vectors, extracting deep features of the input vectors. The output layer then outputs a confidence score for each set of input data, indicating whether it is a real sample, with 1 indicating a real sample and 0 indicating a false positive. S2.5: Based on the output of the generator network and the discriminator network, calculate the GAN network loss value using the adversarial loss function. Based on this loss value, update the generator network and the discriminator network using the backpropagation algorithm and the Adam optimizer. Repeatedly update the generator network and the discriminator network until the GAN network loss value converges to the preset range. S2.5: Input the latest user dynamic body model data and emotional preference labels into the GAN network, and after forward propagation through the generator network, generate a personalized pattern vector, and decode the generated personalized pattern vector to generate the corresponding personalized clothing pattern data.

4. The intelligent clothing design method based on virtual fitting according to claim 3, characterized in that: The specific steps for automatically adjusting the cutting line and structure points in step III are as follows: S3.1: Based on the decoded personalized clothing pattern data, a corresponding two-dimensional plane structure graph is established. The two-dimensional plane structure graph is discretized using a topological grid, and the clothing pattern structure points are used as nodes and the cutting lines as edges to establish an initial topological graph; S3.2: Based on the aesthetic curvature smoothness, structural stitchability, and patch tension balance, a set of objective functions that combine multiple sub-energy terms is set, and constraint modeling is performed on the established objective functions. The positions of each node in the initial topology are then input into the objective functions, and the values to be optimized are output; S3.3: Calculate the gradient vector of the objective function for each node in the topology graph to obtain the gradient descent direction of the current node position, update the position of each node based on the gradient direction, and recalculate the objective function value of the updated node position. At the same time, compare the updated objective function value with the value to be optimized. Based on the comparison result, adaptively adjust the step size of each iteration; S3.4: Iterate and update the position of each node repeatedly until the change in the position of the structural point in multiple rounds of iterations converges to a preset range. Then stop the optimization, recombine the final node set and edge set, and generate a personalized clothing pattern structure diagram after topology optimization.

5. The intelligent clothing design method based on virtual fitting according to claim 4, characterized in that: The specific calculation formula of the objective function described in S3.2 is as follows: ; Where, represents the total topological energy; represents the curvature smoothing energy; represents the contour fitting energy; represents the structural tension energy; 、 、 Represent the weight coefficient of each energy term; The specific calculation formula for the constraint modeling described in S3.2 is as follows: ; Where, Representative The edge and The angle between the edges, if the two sides are smoothly connected, then ; Represents the number of trimming sides; Representative Coordinates of structural points; Representative Coordinates of target body reference points; represents the total number of structural points; as well as Representing the The two structural points connected by an edge; Representative The reference length of the edge at initialization.

6. The intelligent clothing design method based on virtual fitting according to claim 1, characterized in that: The specific steps for real-time modeling of the pressure distribution on the contact surface between the garment and the human body and evaluating wearing comfort described in step V are as follows: S4.1: Based on the simulated fitting results, the Euclidean distance between each garment patch and the human body surface is calculated. If any Euclidean distance is less than a preset contact distance threshold, it indicates that the garment patch is in contact with the human body. Based on the judgment result, a contact area point set is selected; S4.3: Convert the normal deformation of each contact point in the contact area into a local contact pressure to establish a contact pressure distribution model, obtain the pressure values of all contact points, and interpolate and map each pressure value on the skin surface to form a continuous pressure field distribution map; S4.4: Simulate the friction between fabric and skin and calculate the charge density per unit area of the current garment based on the continuous pressure field distribution diagram and the corresponding relative motion coefficient. Based on the charge density distribution, solve the potential distribution near the skin surface and determine the area and intensity of static electricity accumulation. S4.5: Statistically analyze the contact pressure peak value, mean value, change gradient, and electrostatic potential amplitude to construct a comprehensive comfort scoring model and calculate the comfort score of the currently tried-on garment. If the comfort score is lower than the preset threshold, the material stiffness, friction coefficient, or local thickness of the contact area is automatically adjusted based on the analysis results, and corresponding correction suggestions are generated.

7. The intelligent clothing design method based on virtual fitting according to claim 1, characterized in that: The specific steps for optimizing material utilization and structural sustainability in step VI are as follows: S5.1: Discretize the design parameters of the virtual garment and encode them into multiple sets of individual chromosomes. Construct two sets of objective functions based on material utilization and structural sustainability, respectively. Based on the number of individual chromosomes generated in each set, generate an initial population using a random or semi-heuristic method. Then, calculate the effective material utilization and structural sustainability of each individual in the population using the two sets of objective functions. S5.2: Based on the effective material utilization rate and structural sustainability, all individuals in the population are stratified according to the Pareto dominance relationship. All non-dominated individuals are classified into the first-level non-dominated solution set. The next level of non-dominated solution sets are searched from the remaining individuals in turn until the stratification is completed and the non-dominated solution sets of multiple levels after stratification are obtained. S5.3: In each non-dominated layer, calculate the crowding distance between each group of individuals and their neighbors. From the lowest to the highest non-dominated layer, add individuals whose crowding distance exceeds the preset threshold to the new population until the preset population size is reached. If the population size exceeds the upper limit after individuals in the current non-dominated layer choose to join, fill the remaining vacancies based on descending crowding distance. S5.4: Randomly select a pair of parent chromosomes from the new population and generate two daughter chromosomes by simulating a binary crossover method. Then, mutate each design parameter in each daughter chromosome using Gaussian perturbation. Repeat the selection, crossover, and compilation operations multiple times to generate multiple sets of daughter chromosomes. S5.5: Decode each daughter chromosome to obtain the corresponding cutting plan, calculate the effective material utilization rate, and then comprehensively evaluate the structural sustainability of each daughter chromosome based on the disassembly index and reuse score of the structural components; S5.6: Based on the effective material utilization and structural sustainability of the generated daughter chromosomes, perform non-dominated sorting, stratify all the generated daughter chromosomes for the next round of iteration, repeat the selection, crossover, mutation and sorting process until the effective material utilization and structural sustainability, and the rate of change for multiple rounds of iterations is less than the threshold, and output the individual chromosome with the highest material utilization and structural sustainability, after which the designer performs the final virtual pattern making.

Images