Virtual reality processing of fashion design, residual body scan data completion generation method

Abstract

The application provides a virtual reality processing incomplete human body scanning data completion generation method for fashion design, relates to the technical field of fashion design, and comprises the following steps: obtaining human body 3D scanning data required by fashion design, pre-processing the scanning data to obtain sparse incomplete human body 3D point cloud data; based on a trained generative adversarial network, the sparse incomplete human body 3D point cloud data is completed to obtain complete human body 3D point cloud data; the application adopts a generative adversarial network GAN and combines a fashion design scene to perform special adaptation optimization, proposes a GAN-C completion algorithm, solves the problem that existing generative AI completion methods are not adapted to the fashion design scene and the completion morphological distortion, the algorithm optimizes a loss function L1 and network parameters by specially constructing a training data set adapted to the fashion design, so that the key part morphology of the completed human body is more close to the real human body, and the core demand of garment production is accurately matched.

Description

Technical Field

[0001] This invention relates to the field of fashion design technology, and in particular to a method for generating complete data from virtual reality-processed human body scans in fashion design. Background Technology

[0002] In the field of fashion design, the completeness and accuracy of human body scanning data directly determine the suitability and rationality of fashion pattern design, serving as the core foundation for personalized fashion design, virtual try-on, and mass customization. With the popularization of virtual reality technology and low-cost scanning equipment, devices such as monocular depth cameras and mobile phones, leveraging their low cost and portability, are increasingly being applied to human body scanning scenarios in fashion design. These devices can quickly acquire 3D human body scanning data, providing convenient data support for fashion design. Simultaneously, the rapid development of generative AI technology has provided new technical pathways for incomplete data completion. Generative Adversarial Networks (GANs), due to their powerful generative capabilities, have been initially applied to the field of 3D point cloud data completion, generating point cloud data for missing areas by learning the distribution characteristics of real data. Furthermore, existing technologies include point cloud completion algorithms and size extraction methods that attempt to address the problem of incomplete scanning data. By combining virtual reality technology with scanning data to achieve linkage between scanning data and design scenarios, fashion design is being driven towards digitalization and intelligence.

[0003] However, existing technologies still have shortcomings in the virtual reality processing of incomplete human body scan data for fashion design, making it difficult to meet the high precision and adaptability requirements of fashion design: existing generative AI completion methods are not specifically adapted to fashion design scenarios, resulting in distorted shapes of key human body parts after completion, failing to meet the precise requirements of garment production; existing completion algorithms cannot effectively adapt to low-quality sparse point cloud data generated by low-cost devices such as monocular depth cameras and mobile phones, resulting in low completion accuracy, high noise, and difficulty in dealing with data defects caused by device performance; size extraction algorithms have poor robustness and are affected by errors in the point cloud data after completion. The existing technologies have several drawbacks. Firstly, the extracted garment dimensions have significant deviations, failing to meet the precision standards of fashion design. Secondly, they lack deep integration between incomplete data completion, dimension extraction, and the virtual reality system, preventing users from visually viewing the completion effect and size compatibility, resulting in low design efficiency. Thirdly, existing completion methods rely solely on single-modal point cloud data, neglecting to incorporate multimodal data such as texture and depth. This limits the completion effect to data type, hindering further accuracy improvement and failing to fully support the demands of high-precision fashion design. Therefore, this invention proposes a virtual reality processing method for completing and generating incomplete human body scan data for fashion design to address the problems in existing technologies. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a virtual reality-based method for completing and generating incomplete human body scan data for fashion design. This method employs a generative adversarial network (GAN) and incorporates specific adaptation and optimization tailored to fashion design scenarios. A GAN-C completion algorithm is proposed, resolving the problems of existing generative AI completion methods failing to adapt to fashion design scenarios and exhibiting distorted completion forms. This algorithm constructs a training dataset specifically adapted to fashion design and optimizes the L1 loss function and network parameters, ensuring that the completed key human body parts more closely resemble the real human body, accurately matching the core requirements of garment production and providing a reliable human body model foundation for fashion pattern design.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for generating data from incomplete human body scans using virtual reality processing in fashion design, comprising the following steps:

[0006] S1: Obtain the 3D human body scan data required for fashion design, preprocess the scan data to obtain sparse and incomplete 3D human body point cloud data;

[0007] S2: Based on the trained generative adversarial network, the sparse and incomplete human body 3D point cloud data is completed to obtain complete human body 3D point cloud data.

[0008] S3: A robust size extraction algorithm is used to process the complete human body 3D point cloud data to extract the human body size parameters required for clothing production;

[0009] S4: Import complete 3D point cloud data of the human body and extracted human body size parameters into the virtual reality system to achieve real-time visualization and interactive adjustment;

[0010] S5: Verify and correct human body size parameters, and output the final complete 3D point cloud data and human body size parameters that meet fashion design standards.

[0011] Further improvements are made in the following aspects: In S1, the preprocessing includes denoising, sampling and normalization processing, specifically: using a Gaussian filtering algorithm to remove noise points in the scanning data, retaining effective point clouds through uniform sampling, normalizing the point cloud coordinates to a preset range to adapt to the input requirements of the subsequent completion algorithm; the preprocessing adapts to the low-quality scanning data generated by the scanning device.

[0012] A further improvement lies in the following: In S2, the Generative Adversarial Network (GAN) completion algorithm is used, denoted as the GAN-C algorithm, where: G stands for Generator; A for Adversarial; N for Network; and C for Completion. This algorithm specifically includes the following steps:

[0013] A generative adversarial network is constructed, which includes a generator G and a discriminator D. The generator G is used to receive sparse and incomplete human 3D point cloud data X1 and generate complete human 3D point cloud data X1'. The discriminator D is used to receive X1' output by the generator G and real complete human 3D point cloud data Y1, and outputs a discrimination probability P1 to distinguish the authenticity of X1' and Y1.

[0014] Define a loss function L1, which includes adversarial loss L11 and reconstruction loss L12. L11 measures the adversarial performance of the generator G against the discriminator D, and L12 measures the similarity between the generated X1' and the real Y1. The loss function is expressed as L1 = α1 × L11 + β1 × L12, where α1 is the weight of the adversarial loss and β1 is the weight of the reconstruction loss. Both are preset adjustment coefficients, and α1 + β1 = 1. The value of α1 ranges from 0.6 to 0.8, and the value of β1 ranges from 0.2 to 0.4.

[0015] The generative adversarial network is trained using a human point cloud dataset adapted to fashion design scenarios. During training, an adaptive learning rate μ1 is used, with an initial value of 0.001. The learning rate μ1 decreases adaptively as the number of training iterations increases. The number of iterations K1 is preset to 8000~12000 times until the loss function L1 converges to a preset threshold δ1, where δ1 ranges from 0.005 to 0.01.

[0016] Input the preprocessed sparse and incomplete human 3D point cloud data X1 into the trained generator G, and output the completed human 3D point cloud data X1' to complete the incomplete point cloud data.

[0017] A further improvement is made in S2, which employs a low-quality sparse point cloud adaptation algorithm, denoted as the LSA algorithm, where L stands for Low-quality, S for Sparse, and A for Adaptation. This algorithm is used to optimize sparse and incomplete point cloud data generated by low-cost devices and adapt it to the input of the GAN-C algorithm. Specifically, it includes the following steps:

[0018] Neighborhood sampling is performed on the preprocessed sparse and incomplete human 3D point cloud data X1 to obtain the sampling point set X2, where the sampling interval T1 is preset to 0.01~0.03m to ensure that the sampled data retains the key contour features of the human body.

[0019] Calculate the neighborhood density D1 of each point in the sampling point set X2, where the number of m1 neighboring points is preset to 15~25, and D1 is the reciprocal of the average distance of the m1 neighboring points around each point, which is used to characterize the sparsity of the point cloud.

[0020] A density threshold T2 is set, with a value ranging from 0.04 to 0.06. Regions with a neighborhood density D1 < T2 are identified as extremely sparse regions. Linear interpolation is used to fill in the sparse areas in these regions to obtain the preliminary optimized point cloud X2'.

[0021] The coordinates of the initially optimized point cloud X2' are normalized and mapped to the interval [-1,1] to obtain the standardized point cloud X3, which is used as the input data for the GAN-C algorithm to improve the adaptability and accuracy of the completion algorithm.

[0022] A further improvement is made in S2, which employs a multimodal data fusion and completion optimization algorithm, denoted as the MFA algorithm, where M stands for Multi-modal, F for Fusion, and A for Optimization. This algorithm is used to fuse multimodal data to improve completion accuracy and specifically includes the following steps:

[0023] Acquire multimodal data from human body scanning, including sparse residual cloud data X1, human body texture data T1, and human body depth data D2. The texture data T1 is acquired by the image acquisition module of the scanning device, and the depth data D2 is acquired by the depth sensor.

[0024] Feature extraction was performed on each modal data to obtain point cloud features F1, texture features F2, and depth features F3, respectively. F1 was extracted using a PointNet network, F2 was extracted using a CNN network, and F3 was extracted using a convolutional neural network.

[0025] Attention mechanism A1 is used to assign weights to three modal features. The weight coefficients are ω1 point cloud weight, ω2 texture weight, and ω3 depth weight, respectively, and ω1+ω2+ω3=1. The value of ω1 ranges from 0.5 to 0.7, and the values ​​of ω2 and ω3 both range from 0.15 to 0.25.

[0026] The weighted features are fused to obtain fused features F4 = F1×ω1 + F2×ω2 + F3×ω3. The fused features F4 are then input into the generator G of the generative adversarial network to optimize the completion effect and obtain more accurate complete human 3D point cloud data X1''.

[0027] A further improvement is that, in S2, the training dataset of the generative adversarial network includes 10,000 to 15,000 sets of human 3D point cloud data, covering human data of different genders, ages, body types, and postures. Each set of data contains a complete human 3D point cloud and a corresponding sparse and incomplete human 3D point cloud, including various common incomplete types.

[0028] A further improvement lies in the following: In S3, a robust size extraction algorithm, denoted as the RSE algorithm, is used, where R stands for Robust, S for Size, and E for Extraction. This algorithm is used to accurately extract clothing manufacturing dimensions from the completed 3D point cloud data of the human body, and specifically includes the following steps:

[0029] Semantic segmentation is performed on the completed human body 3D point cloud data X1' to obtain the point cloud set Pi (i=1,2,...,n1) of key human body parts, where the number of n1 parts is preset to 8~12, and Pi corresponds to key parts for clothing production such as the top of the head, neck, shoulder, armpit, waist, hip, upper thigh, and middle of the calf.

[0030] Calculate the minimum bounding box Bi of the point cloud Pᵢ for each key part, and extract the length Li, width Wi, and height Hi of each bounding box Bi as the initial size parameters of that part.

[0031] An iterative elimination algorithm is used to remove outlier points O1 from each point set Pi. The number of iterations K2 is preset to 3 to 5 times, and the outlier judgment threshold δ2 is in the range of 0.02 to 0.04 m. After removing outlier points, the optimized key part point set Pi' is obtained.

[0032] Based on the optimized point cloud Pi', the minimum bounding box Bi' of each part is recalculated, and the corresponding size parameters Li', Wi', and Hi' are extracted. The average value of the extracted dimensions for the same part is calculated to obtain the final human body size parameter Si=(Li',Wi',Hi'), ensuring the robustness and accuracy of size extraction.

[0033] A further improvement lies in the following: In S4, the virtual reality real-time linkage feedback algorithm, denoted as the VRF algorithm, is used, where: V stands for Virtual Reality; R stands for Real-time; and F stands for Feedback. This algorithm is used to achieve linkage between point cloud completion, size extraction, and virtual reality fashion design scenes, specifically including the following steps:

[0034] The completed 3D point cloud data of the human body X1' and the extracted human body size parameters Si are imported into the virtual reality engine to construct a virtual human body model M1.

[0035] Real-time acquisition of user interaction commands I1 in the virtual reality system, including operation commands such as rotation, scaling, selection of key parts, and size adjustment of the virtual human body model, with the command response threshold T3 preset to 0.08~0.12s;

[0036] According to the interactive command I1, adjust the display state D1 of the virtual human body model M1, including the model posture, display ratio, and highlighting of key parts, and at the same time update the corresponding human body size parameters Si' to ensure the consistency between the model and the size parameters.

[0037] The adjusted virtual human body model M1' and the updated size parameters Si' are fed back to the user in real time, forming a closed-loop interactive process of "complete-display-interaction-update-feedback", allowing the user to intuitively view the completion effect and size adaptability.

[0038] A further improvement is made in S5, where the verification correction includes: comparing the actual human body manual measurement data Y2 with the extracted human body size parameters Si, calculating the size error Δ1, where Δ1 = |Si - Y2|; setting an error threshold θ1, where θ1 ranges from 0.3 to 0.5 cm; when Δ1 ≤ θ1, the size parameters are deemed qualified and output directly; when Δ1 > θ1, the process returns to S2 to re-complete the point cloud, or returns to S3 to re-extract the size, until the error meets the requirements.

[0039] A further improvement is that, in S5, the final output of complete human body 3D point cloud data and human body size parameters is exported in a format supported by mainstream fashion design software, achieving seamless data integration and direct use in fashion design-related operations.

[0040] The beneficial effects of this invention are as follows:

[0041] 1. This invention employs Generative Adversarial Networks (GANs) and combines them with fashion design scenarios for specialized adaptation and optimization, proposing a GAN-C completion algorithm. This addresses the problems of existing generative AI completion methods failing to adapt to fashion design scenarios and exhibiting distorted completion forms. By constructing a training dataset specifically adapted to fashion design and optimizing the L1 loss function and network parameters, the algorithm ensures that the completed key human body parts more closely resemble the real human body, accurately matching the core needs of garment production. This provides a reliable human body model foundation for fashion pattern design. Adversarial training enhances the realism and completeness of the completed data, thereby improving completion accuracy.

[0042] 2. This invention designs a low-quality sparse point cloud adaptation algorithm, LSA algorithm, specifically designed to address low-quality sparse point cloud data generated by low-cost scanning devices such as monocular depth cameras and mobile phones. It solves the problems of existing completion algorithms being unable to adapt to low-cost devices and having low completion accuracy. This algorithm effectively improves the adaptability of low-quality data by performing neighborhood density D1 analysis, sparse processing and normalization optimization, combined with reasonable settings of parameters T1, m1 and T2, and eliminates data noise. This enables the sparse point cloud X1 generated by low-cost devices to stably adapt to subsequent GAN-C algorithms, which not only reduces the equipment cost of human body scanning in fashion design, but also ensures the stability and reliability of the completed data, and expands the scope of application.

[0043] 3. This invention proposes a robust size extraction algorithm, RSE algorithm, which solves the problems of poor robustness, large influence of completion error, and inaccurate size extraction of existing size extraction algorithms. Through key part semantic segmentation, outlier point O1 iterative removal, and multiple size calibration steps, this algorithm can accurately extract the human body size parameters Si required for garment making from the completed point cloud data X1'. The size extraction error Δ1 is controlled within 0.5cm, which meets the high precision requirements of fashion design and effectively reduces the problems of unqualified fashion patterns and rework caused by size deviation, thereby improving the pass rate and efficiency of fashion design.

[0044] 4. This invention designs a Virtual Reality Real-Time Linkage Feedback (VRF) algorithm, which realizes deep linkage between incomplete data completion, size extraction, and virtual reality fashion design scenes. It solves the problems of existing technologies, such as lack of real-time linkage, low design efficiency, and inability to intuitively view the effects. This algorithm constructs a virtual human body model M1 and combines the setting of interactive instructions I1 and response threshold T3 to realize real-time visualization of the completed data X1' and size parameters Sᵢ. It supports user interaction and adjustment, forming a closed-loop feedback process. Fashion designers can intuitively view the completion effect and size adaptability, optimize the design scheme in real time, improve design efficiency, lower the design threshold, and improve the intelligence level of fashion design.

[0045] 5. This invention proposes a multimodal data fusion completion optimization algorithm (MFA), which integrates multimodal human body scanning data such as point cloud X1, texture T1, and depth D2. This solves the problem that existing completion methods rely on only a single modality of data and have insufficient completion accuracy. The algorithm uses an attention mechanism A1 to weight and fuse multimodal features F1, F2, and F3, making full use of the advantages of each modality of data. This further improves the completion accuracy and detail richness of the incomplete point cloud. The similarity between the completed human point cloud X1'' and the real human body reaches more than 95%, which can more accurately restore the human body contour features. This provides higher quality data support for high-end personalized fashion design and customized fashion production, and further broadens the application scenarios. Attached Figure Description

[0046] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0047] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0048] Example 1

[0049] according to Figure 1As shown, this embodiment proposes a method for generating incomplete human body scan data in virtual reality processing for fashion design, verifies the completion effect of the GAN-C algorithm, and adapts to the completion of incomplete human body point cloud data commonly found in fashion design scenarios, such as missing parts of the head and armpits. The specific steps are as follows:

[0050] Data preparation: Commonly used human body shape data in fashion design were selected, including 200 sets of 3D point cloud data of different genders (100 sets of males and 100 sets of females) and different body types (30 sets of thin, 140 sets of standard, and 30 sets of fat). Each set of data includes a real and complete human body point cloud Y1 and a sparse and incomplete point cloud X1 that simulates scanning with low-cost equipment (point clouds in the top of the head and underarms were manually removed to simulate the incompleteness in the actual scan). The resolution of the point cloud data was set to 1024×1024, and the number of points in the cloud was 50,000~80,000 per set.

[0051] Network Construction: A generative adversarial network is constructed. The generator G adopts a deconvolutional neural network structure, including an input layer, four convolutional layers, four deconvolutional layers, and an output layer. The input layer has a dimension of 3 (corresponding to the x, y, z three-dimensional coordinates of the point cloud), and the output layer has the same dimension as the input layer, which is used to output the complete point cloud X1' after completion. The discriminator D adopts a convolutional neural network structure, including an input layer, five convolutional layers, two fully connected layers, and an output layer. The output layer is a single channel and outputs the discrimination probability P1 (0≤P1≤1, the closer P1 is to 1, the closer the input point cloud is to the real complete point cloud Y1).

[0052] Algorithm parameter settings: Loss function L1 = α1 × L11 + β1 × L12, where α1 = 0.7 (adversarial loss weight), β1 = 0.3 (reconstruction loss weight); adversarial loss L11 uses the cross-entropy loss function, and reconstruction loss L12 uses the mean squared error loss function; the adaptive learning rate μ1 is initially 0.001, adopts an exponential decay strategy with a decay coefficient of 0.99, and decays once every 100 iterations; the number of iterations K1 = 10000, and the convergence threshold δ1 = 0.008; batch gradient descent is used during training, with a batch size of batch = 32.

[0053] Network training: 200 sets of point cloud data are divided into a training set (160 sets) and a test set (40 sets). The network is trained using the training set. During the training process, the change of the loss function L1 is monitored in real time. When the number of iterations reaches K1=10000 and L1≤δ1=0.008, the training is stopped, and the trained generative adversarial network is obtained.

[0054] Completeness test: Input the 40 sparse and incomplete point clouds X1 from the test set into the trained generator G one by one, and output the completed point cloud X1'. Compare the similarity between X1' and the real complete point cloud Y1, and the completion effect of the incomplete parts. The chamfer distance (CD) is used to measure the completion accuracy. The smaller the CD value, the higher the completion accuracy.

[0055] Test results: In 40 sets of test data, the similarity between the completed point cloud X1' and the real point cloud Y1 reached over 92%, with 32 sets achieving a similarity of over 95%; the chamfer distance CD value was 0.002~0.005m, far lower than the CD value of existing completion algorithms (0.008~0.012m); the deviation between the completed shape and the real shape of the top of the head and underarms was ≤0.3cm. The completed point cloud X1' can accurately restore the human body contour features, fully meeting the needs of fashion design.

[0056] Example 2

[0057] according to Figure 1 As shown, this embodiment proposes a virtual reality processing method for completing and generating incomplete human body scan data in fashion design, and verifies the adaptation effect of the LSA algorithm. A monocular depth camera and a mobile phone are used as scanning devices to acquire real human body scan data, and the algorithm's adaptability to low-quality sparse point clouds is verified. The specific steps are as follows:

[0058] Scanning equipment and data acquisition: Ten test subjects (5 males and 5 females, aged 20-45, with body types including thin, standard, and fat) were selected. Human body scans were performed using a monocular depth camera and a mobile phone, respectively, to acquire 3D human body scan data for the 10 test subjects, resulting in a total of 20 sets of scan data. Due to equipment cost limitations, the scan data exhibited significant sparsity and noise, with a point cloud density of 10,000-30,000 points per set, and varying degrees of point cloud missingness (top of the head, back, armpits, etc.). This scan data was preprocessed and used as sparse residual point cloud X1.

[0059] LSA algorithm parameter settings: sampling interval T1=0.02m, number of neighborhood points m1=20, density threshold T2=0.05, coordinate normalization range is [-1,1]; Gaussian filter parameter σ=0.01m (standard deviation) in preprocessing is used to remove noise points in the scan data.

[0060] Algorithm execution: The LSA algorithm is executed one by one on each of the 20 sets of low-quality scan data (X1 after preprocessing). The steps are as follows:

[0061] The Gaussian filtering algorithm is used to remove noise points in the scan data and eliminate isolated points (points without surrounding neighboring points) to obtain the denoised point cloud data (still X1).

[0062] Uniform sampling is performed at a sampling interval of T1=0.02m, and the effective point cloud is retained to obtain the sampling point set X2;

[0063] Calculate the neighborhood density D1 of each point in X2, determine the region where the neighborhood density D1 < T2 = 0.05 as an extremely sparse region, use linear interpolation to fill in the missing point cloud, and obtain the preliminary optimized point cloud X2'.

[0064] The coordinates of X2' are normalized to the interval [-1,1] to obtain the standardized point cloud X3, completing the adaptation process. X3 is then used as the input data for the subsequent GAN-C algorithm.

[0065] Adaptation effect verification: Input the standardized point cloud X3 after adaptation processing into the GAN-C algorithm trained in Example 1 to complete the point cloud and compare the difference in completion effect before and after adaptation; at the same time, compare the density, noise ratio and completion accuracy (CD value) of the point cloud before and after adaptation processing.

[0066] Verification results: After adaptation, the density of 20 sets of point cloud data increased to 40,000~60,000 points / set, and the noise ratio decreased from the original 15%~25% to 3%~8%. The completion accuracy (CD value) of the input GAN-C algorithm after adaptation was 0.003~0.006m, which is more than 60% higher than before adaptation (CD value 0.010~0.015m). The adapted point cloud X3 can stably adapt to the GAN-C algorithm, and the completed key human body parts are more complete and realistic. This shows that the LSA algorithm can effectively adapt to the low-quality sparse point cloud data X1 generated by low-cost equipment, and solves the problem of poor adaptability of existing algorithms.

[0067] Example 3

[0068] according to Figure 1 As shown, this embodiment proposes a virtual reality processing method for completing and generating incomplete human body scan data for fashion design. It verifies the size extraction accuracy and robustness of the RSE algorithm. Using the completed human body point cloud data X1' from Embodiment 1, eight key dimensions commonly used in fashion design (shoulder width, bust, waist, hip, sleeve length, garment length, thigh circumference, and calf circumference) are extracted and compared with manually measured dimensions. The specific steps are as follows:

[0069] Data preparation: 40 sets of complete human point clouds X1' from the test set in Example 1 were selected. At the same time, manual measurements were performed on the 40 corresponding test subjects to obtain manual measurement values ​​Y2 of 8 key dimensions (as the true values). The manual measurement accuracy was 0.1cm and was completed by professional measurement personnel.

[0070] RSE algorithm parameter settings: number of key parts n1=8, corresponding to the human body parts corresponding to shoulder width, chest circumference, waist circumference, hip circumference, sleeve length, garment length, thigh circumference, and calf circumference respectively; number of iterations for elimination K2=4 times, outlier judgment threshold δ2=0.03m; minimum bounding box calculation adopts axis-aligned bounding box (AABB) algorithm.

[0071] Algorithm execution: The RSE algorithm is executed on each of the 40 completed point cloud X1's, with the following steps:

[0072] The semantic segmentation network (PointNet++) is used to perform semantic segmentation on X1', resulting in point cloud sets Pi (i=1~8) of 8 key parts.

[0073] Calculate the minimum bounding box Bi for each Pi, and extract the length Li, width Wi, and height Hi of Bᵢ as initial size parameters;

[0074] An iterative elimination algorithm is used to eliminate outlier points O1 in Pᵢ that are more than 0.03m away from the center of the bounding box in each iteration. After K2=4 iterations, the optimized point cloud Pi' is obtained.

[0075] Recalculate the minimum bounding box Bi' of Pi', extract Li', Wi', and Hi', and calculate the average of the extracted dimensions for the same part 5 times to obtain the final human body size parameter Si=(Li', Wi', Hi'); where the annular dimensions such as shoulder width and chest circumference are taken from the perimeter of the bounding box, and the linear dimensions such as sleeve length and garment length are taken from the diagonal length of the bounding box.

[0076] Accuracy verification: Calculate the error Δ1=|Si-Y2| between each dimension parameter Si and the manually measured value Y2, count the number of samples with errors ≤0.3cm, 0.3cm<Δ1≤0.5cm, and Δ1>0.5cm, and calculate the average error and maximum error; at the same time, compare the error situation of existing dimension extraction algorithms and set the error threshold θ1=0.5cm.

[0077] Verification results: Among 40 samples, the average error of the 8 key dimensions was 0.25cm, and the maximum error was 0.48cm, all ≤θ1=0.5cm; among them, 36 samples had an error ≤0.3cm, accounting for 90%; 4 samples had an error between 0.3 and 0.5cm, accounting for 10%; no samples had an error >0.5cm. Compared with existing size extraction algorithms (average error 0.8~1.2cm, maximum error 1.5cm), the RSE algorithm in this embodiment improves the size extraction accuracy by more than 70% and has better error stability, indicating that the algorithm has strong robustness and can accurately extract the human body size parameters Si required for garment production, meeting the accuracy requirements of fashion design.

[0078] Example 4

[0079] according to Figure 1 As shown, this embodiment proposes a virtual reality processing method for completing and generating incomplete human body scan data in fashion design. It verifies the linkage effect and practicality of the VRF algorithm, builds a virtual reality fashion design platform, and realizes the linkage between the completed data X1', the size extraction Si, and the virtual reality design scene. The specific steps are as follows:

[0080] Virtual Reality Platform Setup: Hardware includes a VR headset (Model: Meta Quest 3), VR controllers, and a high-performance computer (CPU: Intel Core i9-13900K, GPU: NVIDIA RTX 4090, RAM: 32GB); Software uses Unity 2022.3 as the virtual reality engine, developing functional modules such as virtual try-on, size adjustment, and model interaction; A virtual fashion design scene is built, including a virtual human body display area, a size parameter display area, and an interactive operation area.

[0081] VRF algorithm parameter settings: interactive command response threshold T3=0.1s, virtual human model M1 display resolution is 2K (2560×1440), size parameter display accuracy is 0.1cm; model rotation angle step is 5°, scaling ratio step is 0.01.

[0082] Algorithm Execution and Linkage Testing: Ten sets of size parameters Si and corresponding completed point cloud data X1' extracted from Example 3 were selected, imported into the virtual reality platform, and the VRF algorithm was executed to conduct linkage testing. The steps are as follows:

[0083] Import the completed 3D point cloud data of the human body X1' and size parameters Sᵢ into the Unity engine to build a virtual human body model M1. The model material uses real human skin texture to ensure a realistic display effect.

[0084] Testers wear VR headsets and issue interactive commands I1 through VR controllers, including rotating the model (viewing the human body shape from different angles), zooming (enlarging to view details of key parts), selecting the underarm area (highlighting the area and the corresponding size parameter Si), and adjusting the waist size (manually entering the size value and updating the model synchronously).

[0085] According to the interaction command I1, the system adjusts the display state D1 of the virtual human body model M1 within T3=0.1s, including the model posture, display ratio, and highlighting of key parts, and at the same time updates the corresponding human body size parameter Si' to ensure the consistency between model M1' and size parameter Si'.

[0086] The tester checks the adjusted model M1' and dimensional parameters Si' to confirm whether they meet the design requirements. If they do not meet the requirements, the tester continues to issue adjustment instructions I1 until the requirements are met, forming a closed-loop interaction of "complete-display-interaction-update-feedback".

[0087] Practicality verification: Ten fashion designers were invited to participate in the test to evaluate the convenience, real-time performance, and display effect of virtual reality linkage, and to collect the designers' satisfaction data; at the same time, the design efficiency of using VRF algorithm linkage design was compared with that of traditional design methods (based on 2D size drawings).

[0088] Verification results: Over 90% of the 10 fashion designers were satisfied with the virtual reality linkage effect, believing that the linkage system allowed for intuitive viewing of the completion effect and size fit, and was easy to operate and responsive. The average efficiency of the VRF algorithm linkage design was 2.5 hours per fashion pattern, which is more than 55% higher than the traditional design method (5.5 hours per pattern). Designers were able to optimize pattern design in real time through the linkage system, reducing rework caused by size deviations. This indicates that the VRF algorithm effectively achieved deep linkage between completion, size extraction and virtual reality, improving the efficiency and convenience of fashion design.

[0089] Example 5

[0090] according to Figure 1 As shown, this embodiment proposes a virtual reality processing method for incomplete human body scan data in fashion design, verifies the optimization effect of the MFA algorithm, and integrates human body scan data from three modalities: point cloud X1, texture T1, and depth D2, to improve the completion accuracy and detail richness. The specific steps are as follows:

[0091] Multimodal data acquisition: Ten test subjects from Example 2 were selected. Human point cloud data X1 (sparse residual point cloud) and depth data D2 were acquired using a monocular depth camera (Kinect v2). Human texture data T1 (4K resolution, 3840×2160) was acquired using a mobile phone camera. A total of 10 sets of multimodal data were acquired. Each set of data includes sparse residual point cloud X1, corresponding texture data T1, depth data D2, and real complete human point cloud Y1 (as a comparison benchmark).

[0092] MFA algorithm parameter settings: attention weight coefficients ω1=0.6 (point cloud weight), ω2=0.2 (texture weight), ω3=0.2 (depth weight), and ω1+ω2+ω3=1; feature extraction network parameters: the kernel size of the PointNet network (extracting point cloud features F1) is 3×3, the kernel size of the CNN network (extracting texture features F2) is 5×5, and the depth feature F3 is extracted using a 3-layer convolutional neural network; the fused feature F4 is input into the generator G in Example 1 to optimize the completion effect and obtain X1''.

[0093] Algorithm execution: The MFA algorithm is executed one by one on the 10 sets of multimodal data. The steps are as follows:

[0094] Preprocessing of each modal data: The LSA algorithm in Example 2 is executed on the sparse residual point cloud X1 to obtain the standardized point cloud X3; the texture data T1 is denoised and normalized to remove image noise and normalize the pixel values ​​to the [0,1] interval; the depth data D2 is smoothed and the median filtering algorithm is used to remove depth noise to ensure the accuracy of the depth data.

[0095] Feature extraction: PointNet network is used to extract features from the standardized point cloud X3 to obtain point cloud features F1 (1024 dimensions); CNN network is used to extract features from the preprocessed texture data T1 to obtain texture features F2 (512 dimensions); a 3-layer convolutional neural network is used to extract features from the smoothed depth data D2 to obtain depth features F3 (512 dimensions).

[0096] Feature weight allocation: Attention mechanism A1 is used to allocate weights to F1, F2, and F3. Based on the importance of each modality feature, the point cloud feature F1 is given the highest weight ω1=0.6, while the texture feature F2 and depth feature F3 are given weights of ω2=0.2 and ω3=0.2 respectively, to ensure that the fused features can highlight the contour information of the point cloud while taking into account texture details and depth accuracy.

[0097] Feature fusion: The weighted features are fused using the formula F4 = F1 × ω1 + F2 × ω2 + F3 × ω3. The resulting fused feature F4 has a dimension of 1024, ensuring that the fused feature contains key information from each modality.

[0098] Optimization and completion: Input the fused feature F4 into the generative adversarial network generator G trained in Example 1 to generate optimized complete human 3D point cloud data X1'', thus completing the multimodal data fusion and completion.

[0099] Optimization effect verification: Comparing X1'' completed by the MFA algorithm with X1' completed by the GAN-C algorithm alone in Example 1, using the real complete human point cloud Y1 as the benchmark, the verification was carried out using three indicators: similarity, chamfer distance (CD), and deviation of key parts; at the same time, the completion efficiency was statistically analyzed and the time consumption of the two completion methods was compared.

[0100] Verification Results: In 10 sets of test data, the similarity between X1'' and the real point cloud Y1 after completion by the MFA algorithm reached 96%~98%, an improvement of 3%~5% compared to completion by the GAN-C algorithm alone (92%~95%); the chamfer distance CD value was 0.001~0.003m, a reduction of more than 40% compared to completion alone (0.002~0.005m); the morphological deviation of key parts such as the top of the head, armpits, and back was ≤0.2cm, an improvement of more than 33% in accuracy compared to completion alone (≤0.3cm); the completion time was 0.8~1.2s / set, basically the same as the completion time of completion alone (0.7~1.1s / set), without excessive increase in time due to multimodal fusion. This shows that the MFA algorithm can significantly improve the completion accuracy and detail richness of incomplete point clouds without reducing completion efficiency, further meeting the high-precision requirements of high-end fashion design.

[0101] Validation data:

[0102] This invention utilizes five core algorithms (GAN-C, LSA, RSE, VRF, and MFA) to collaboratively address the problem of completing and generating dimensions of sparse and incomplete 3D human body point cloud data generated by low-cost scanning equipment in fashion design scenarios. Verified through five embodiments, its performance indicators outperform existing technologies. Specific data summaries are as follows: Completion Accuracy: The GAN-C algorithm alone achieves completion accuracy (CD value 0.002~0.005m) that is more than 60% higher than existing completion algorithms. The MFA algorithm, after fusing multimodal data, further improves completion accuracy, reducing the CD value to 0.001~0.003m. The similarity between the completed point cloud and the real human body reaches 96%~98%, with key part morphological deviations ≤0.2cm, fully meeting the accuracy requirements of fashion design for human body models. Low-cost device compatibility: The LSA algorithm can effectively adapt to low-quality sparse point cloud data (point cloud density 10,000~30,000 points / group) generated by low-cost scanning devices such as monocular depth cameras and mobile phones. After adaptation, the point cloud density increases to 40,000~60,000 points / group, and the noise ratio decreases from 15%~25% to 3%~8%. The equipment cost is reduced by more than 80% compared to high-end 3D scanners, significantly reducing the hardware investment in fashion design. Size extraction accuracy: The RSE algorithm has an average size extraction error of 0.25cm and a maximum error of 0.48cm, both controlled within the error threshold θ1=0.5cm. This is more than 70% more accurate than existing size extraction algorithms. The size extraction is robust and less affected by completion errors, and can accurately extract 8~12 key size parameters required for garment production. Design Efficiency: The VRF algorithm achieves deep integration of pattern completion, size extraction, and the virtual reality system, improving fashion pattern design efficiency by over 55% compared to traditional methods. Average design time is reduced from 5.5 hours / set to 2.5 hours / set, with designer satisfaction exceeding 90%. This effectively lowers the design threshold and enhances design convenience. Overall Performance: The overall time consumption of this invention is controllable. Pattern completion takes 0.7~1.2 seconds per set, size extraction takes 1.0~1.5 seconds per set, and the virtual reality interaction response time is ≤0.1 seconds. All algorithms operate stably without lag or errors. The final output data can seamlessly integrate with mainstream fashion design software such as AutoCAD and Adobe Illustrator, enabling unified data flow. It is suitable for various application scenarios such as personalized fashion design, virtual try-on, and batch customization, possessing broad practicality and promotional value. In summary, this invention solves the core problems in the prior art, such as low completion accuracy, poor adaptability to low-cost equipment, insufficient robustness of size extraction, low design efficiency, and lack of fusion of multimodal data. It fills the technological gap in the fashion design scenario for generative AI to complete incomplete human body scan data and generate sizes, and promotes the development of fashion design towards digitalization, intelligence, and low cost.

[0103] This virtual reality processing method for completing incomplete human body scan data in fashion design employs a generative adversarial network (GAN) and is specifically adapted and optimized for fashion design scenarios. A GAN-C completion algorithm is proposed, addressing the issues of existing generative AI completion methods failing to adapt to fashion design scenarios and exhibiting distorted completion forms. This algorithm constructs a training dataset specifically adapted to fashion design, optimizes the L1 loss function and network parameters, making the completed key human body parts more closely resemble the real human body, accurately matching the core needs of garment production. This provides a reliable human body model foundation for fashion pattern design. Adversarial training enhances the realism and completeness of the completed data, improving completion accuracy. This invention designs a low-quality sparse point cloud adaptation algorithm, LSA algorithm, specifically designed to handle low-quality sparse point cloud data generated by low-cost scanning devices such as monocular depth cameras and mobile phones. It solves the problems of existing completion algorithms being unable to adapt to low-cost devices and having low completion accuracy. This algorithm effectively improves the adaptability of low-quality data by performing neighborhood density D1 analysis, sparse processing and normalization optimization, combined with reasonable settings of parameters T1, m1 and T2, and eliminates data noise. This enables the sparse point cloud X1 generated by low-cost devices to be stably adapted to subsequent GAN-C algorithms. This not only reduces the equipment cost of human body scanning in fashion design, but also ensures the stability and reliability of the completed data, expanding the scope of application. This invention proposes a robust size extraction algorithm, RSE, which solves the problems of poor robustness, large influence of completion errors, and inaccurate size extraction in existing size extraction algorithms. Through steps such as semantic segmentation of key parts, iterative removal of outliers O1, and multiple size calibrations, this algorithm can accurately extract the human body size parameters Si required for garment production from the completed point cloud data X1'. The size extraction error Δ1 is controlled within 0.5cm, meeting the high precision requirements of fashion design. It effectively reduces problems such as unqualified fashion patterns and rework caused by size deviations, and improves the pass rate and efficiency of fashion design. This invention designs a Virtual Reality Real-Time Linkage Feedback (VRF) algorithm, which realizes deep linkage between incomplete data completion, size extraction, and virtual reality fashion design scenes. It solves the problems of existing technologies, such as lack of real-time linkage, low design efficiency, and inability to intuitively view the effects. The algorithm constructs a virtual human body model M1 and combines the setting of interactive command I1 and response threshold T3 to realize real-time visualization of the completed data X1' and size parameter Sᵢ. It supports user interaction and adjustment, forming a closed-loop feedback process. Fashion designers can intuitively view the completion effect and size adaptability, optimize the design scheme in real time, improve design efficiency, lower the design threshold, and enhance the intelligence level of fashion design.This invention proposes a multimodal data fusion completion optimization algorithm (MFA), which integrates multimodal human body scanning data such as point cloud X1, texture T1, and depth D2. This addresses the problem of existing completion methods relying solely on single-modal data and lacking sufficient completion accuracy. The algorithm uses an attention mechanism A1 to weight and fuse multimodal features F1, F2, and F3, fully utilizing the advantages of each modality to further improve the completion accuracy and detail richness of incomplete point clouds. The completed human point cloud X1'' achieves a similarity of over 95% with the real human body, enabling more accurate reconstruction of human contour features. This provides higher-quality data support for high-end personalized fashion design and customized fashion production, further expanding application scenarios.

[0104] 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 illustrative of the 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 present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating data from incomplete human body scans using virtual reality processing in fashion design, characterized by: Includes the following steps: S1: Obtain the 3D human body scan data required for fashion design, preprocess the scan data to obtain sparse and incomplete 3D human body point cloud data; S2: Based on the trained generative adversarial network, the sparse and incomplete human body 3D point cloud data is completed to obtain complete human body 3D point cloud data. S3: A robust size extraction algorithm is used to process the complete human body 3D point cloud data to extract the human body size parameters required for clothing production; S4: Import complete 3D point cloud data of the human body and extracted human body size parameters into the virtual reality system to achieve real-time visualization and interactive adjustment; S5: Verify and correct human body size parameters, and output the final complete 3D point cloud data and human body size parameters that meet fashion design standards.

2. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 1, characterized in that, In S1, the preprocessing includes denoising, sampling and normalization processing, specifically: using a Gaussian filtering algorithm to remove noise points in the scan data, retaining effective point clouds through uniform sampling, and normalizing the point cloud coordinates to a preset range to adapt to the input requirements of the subsequent completion algorithm. Preprocessing low-quality scan data generated by the adapted scanning equipment.

3. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 1, characterized in that, In S2, the Generative Adversarial Network (GAN) completion algorithm used is denoted as the GAN-C algorithm, where: G stands for Generator; A stands for Adversarial; N stands for Network; and C stands for Completion. This algorithm specifically includes the following steps: A generative adversarial network is constructed, which includes a generator G and a discriminator D. The generator G is used to receive sparse and incomplete human 3D point cloud data X1 and generate complete human 3D point cloud data X1'. The discriminator D is used to receive X1' output by the generator G and real complete human 3D point cloud data Y1, and outputs a discrimination probability P1 to distinguish the authenticity of X1' and Y1. Define a loss function L1, which includes adversarial loss L11 and reconstruction loss L12. L11 measures the adversarial performance of the generator G against the discriminator D, and L12 measures the similarity between the generated X1' and the real Y1. The loss function is expressed as L1 = α1 × L11 + β1 × L12, where α1 is the weight of the adversarial loss and β1 is the weight of the reconstruction loss. Both are preset adjustment coefficients, and α1 + β1 = 1. The value of α1 ranges from 0.6 to 0.8, and the value of β1 ranges from 0.2 to 0.

4. The generative adversarial network is trained using a human point cloud dataset adapted to fashion design scenarios. During training, an adaptive learning rate μ1 is used, with an initial value of 0.

001. The learning rate μ1 decreases adaptively as the number of training iterations increases. The number of iterations K1 is preset to 8000~12000 times until the loss function L1 converges to a preset threshold δ1, where δ1 ranges from 0.005 to 0.

01. Input the preprocessed sparse and incomplete human 3D point cloud data X1 into the trained generator G, and output the completed human 3D point cloud data X1' to complete the incomplete point cloud data.

4. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 3, characterized in that, In step S2, a low-quality sparse point cloud adaptation algorithm, denoted as LSA algorithm, is also employed, where L stands for Low-quality, S for Sparse, and A for Adaptation. This algorithm is used to optimize sparse and incomplete point cloud data generated by low-cost devices and adapt it to the input of the GAN-C algorithm. Specifically, it includes the following steps: Neighborhood sampling is performed on the preprocessed sparse and incomplete human 3D point cloud data X1 to obtain the sampling point set X2, where the sampling interval T1 is preset to 0.01~0.03m to ensure that the sampled data retains the key contour features of the human body. Calculate the neighborhood density D1 of each point in the sampling point set X2, where the number of m1 neighboring points is preset to 15~25, and D1 is the reciprocal of the average distance of the m1 neighboring points around each point, which is used to characterize the sparsity of the point cloud. A density threshold T2 is set, with a value ranging from 0.04 to 0.

06. Regions with a neighborhood density D1 < T2 are identified as extremely sparse regions. Linear interpolation is used to fill in the sparse areas in these regions to obtain the preliminary optimized point cloud X2'. The coordinates of the initially optimized point cloud X2' are normalized and mapped to the interval [-1,1] to obtain the standardized point cloud X3, which is used as the input data for the GAN-C algorithm to improve the adaptability and accuracy of the completion algorithm.

5. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 4, characterized in that, In step S2, a multimodal data fusion completion optimization algorithm, denoted as the MFA algorithm, is also employed, where M stands for Multi-modal, F for Fusion, and A for Optimization. This algorithm is used to fuse multimodal data to improve completion accuracy and specifically includes the following steps: Acquire multimodal data from human body scanning, including sparse residual cloud data X1, human body texture data T1, and human body depth data D2. The texture data T1 is acquired by the image acquisition module of the scanning device, and the depth data D2 is acquired by the depth sensor. Feature extraction was performed on each modal data to obtain point cloud features F1, texture features F2, and depth features F3, respectively. F1 was extracted using a PointNet network, F2 was extracted using a CNN network, and F3 was extracted using a convolutional neural network. Attention mechanism A1 is used to assign weights to three modal features. The weight coefficients are ω1 point cloud weight, ω2 texture weight, and ω3 depth weight, respectively, and ω1+ω2+ω3=1. The value of ω1 ranges from 0.5 to 0.7, and the values ​​of ω2 and ω3 both range from 0.15 to 0.

25. The weighted features are fused to obtain fused features F4 = F1×ω1 + F2×ω2 + F3×ω3. The fused features F4 are then input into the generator G of the generative adversarial network to optimize the completion effect and obtain more accurate complete human 3D point cloud data X1''.

6. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 5, characterized in that, In S2, the training dataset of the generative adversarial network includes 10,000 to 15,000 sets of human 3D point cloud data, covering human data of different genders, ages, body types and postures. Each set of data contains complete human 3D point clouds and corresponding sparse and incomplete human 3D point clouds, including various common incomplete types.

7. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 6, characterized in that, In step S3, the robust size extraction algorithm used is denoted as the RSE algorithm, where: R stands for Robust; S stands for Size; and E stands for Extraction. This algorithm is used to accurately extract clothing manufacturing dimensions from the completed 3D human body point cloud data, and specifically includes the following steps: Semantic segmentation is performed on the completed human body 3D point cloud data X1' to obtain the point cloud set Pi (i=1,2,...,n1) of key human body parts, where the number of n1 parts is preset to 8~12, and Pi corresponds to key parts for clothing production such as the top of the head, neck, shoulder, armpit, waist, hip, upper thigh, and middle of the calf. Calculate the minimum bounding box Bi of the point cloud Pᵢ for each key part, and extract the length Li, width Wi, and height Hi of each bounding box Bi as the initial size parameters of that part. An iterative elimination algorithm is used to remove outlier points O1 from each point set Pi. The number of iterations K2 is preset to 3 to 5 times, and the outlier judgment threshold δ2 is in the range of 0.02 to 0.04 m. After removing outlier points, the optimized key part point set Pi' is obtained. Based on the optimized point cloud Pi', the minimum bounding box Bi' of each part is recalculated, and the corresponding size parameters Li', Wi', and Hi' are extracted. The average value of the extracted dimensions for the same part is calculated to obtain the final human body size parameter Si=(Li',Wi',Hi'), ensuring the robustness and accuracy of size extraction.

8. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 7, characterized in that, In step S4, the virtual reality real-time linkage feedback algorithm used is denoted as the VRF algorithm, where: V stands for Virtual Reality; R stands for Real-time; and F stands for Feedback. This algorithm is used to realize the linkage between point cloud completion, size extraction, and virtual reality fashion design scene, and specifically includes the following steps: The completed 3D point cloud data of the human body X1' and the extracted human body size parameters Si are imported into the virtual reality engine to construct a virtual human body model M1. Real-time acquisition of user interaction commands I1 in the virtual reality system, including operation commands such as rotation, scaling, selection of key parts, and size adjustment of the virtual human body model, with the command response threshold T3 preset to 0.08~0.12s; According to the interactive command I1, adjust the display state D1 of the virtual human body model M1, including the model posture, display ratio, and highlighting of key parts, and at the same time update the corresponding human body size parameters Si' to ensure the consistency between the model and the size parameters. The adjusted virtual human body model M1' and the updated size parameters Si' are fed back to the user in real time, forming a closed-loop interactive process of "complete-display-interaction-update-feedback", allowing the user to intuitively view the completion effect and size adaptability.

9. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 8, characterized in that, In step S5, the verification and correction include: comparing the actual human body manual measurement data Y2 with the extracted human body size parameters Si, calculating the size error Δ1, where Δ1 = |Si - Y2|; setting an error threshold θ1, where θ1 ranges from 0.3 to 0.5 cm; when Δ1 ≤ θ1, the size parameters are deemed qualified and output directly; when Δ1 > θ1, the process returns to S2 to re-complete the point cloud, or returns to S3 to re-extract the size, until the error meets the requirements.

10. The virtual reality processing method for completing and generating incomplete human body scan data in fashion design according to claim 1, characterized in that, In S5, the final output of complete human body 3D point cloud data and human body size parameters is exported in a format supported by mainstream fashion design software, achieving seamless data integration and direct use in fashion design-related operations.

Images