关键词: 神经网络, 服装压分布, 服装纸样结构, 虚拟试衣, 服装合体性

Abstract:

Objective Pattern structure adjustment in women's garment development mainly relies on empirical judgment and lacks objective quantitative criteria. This study aims to address these issues by establishing a pattern structure evaluation method based on garment pressure data obtained from virtual fitting. The objective is to clarify the relationships among pattern parameters, body parameters, and garment pressure distribution, and to examine the feasibility of using pressure indicators to predict pattern adjustment quantities for objective pattern modification.

Method Virtual fitting experiments were conducted using the CLO3D system with a standard female mannequin. Pattern structural parameters, including bust circumference and shoulder slope angle, as well as mannequin body parameters such as bust girth, shoulder slope, and neck root girth, were systematically varied to investigate their effects on garment pressure distribution. A series of control, single-variable, coordinated, and cross-variable experiments were designed to analyze the individual and combined influences of structural and body parameters. Garment pressure data were collected at seven predefined pressure measurement points: neck side point (P₁), shoulder endpoint (P₂), front armpit point (P₃), bust point (P₄), back neck point (P₅), back armpit point (P₆), and scapular prominence point (P₇). A total of 171 experimental datasets were obtained. Based on these data, a multilayer perceptron (MLP) was constructed to model the nonlinear relationships between pattern parameters, body measurements, and garment pressure. The trained model was further used to predict pattern adjustment quantities, and its performance was evaluated using error metrics.

Results The results showed that increasing pattern bust circumference led to a consistent decrease in garment pressure values across all seven measurement points. The correlation coefficients between pattern bust circumference and garment pressure were found to range from -0.50 to -0.21, indicating that bust adjustment produces directionally consistent pressure changes across multiple anatomical regions. Among the measurement points, pressure responses at the neck and shoulder-related locations showed higher sensitivity to bust variation than those at posterior torso regions. Feature importance analysis revealed that neck root girth had the strongest influence on the prediction of pattern bust adjustment, with the highest importance value (2.027). This result indicates that variations in neck root girth correspond more closely to required bust-related structural modification than other body parameters considered in this study. Shoulder slope mainly affected pressure redistribution in the upper torso and shoulder regions, while its influence on lower torso pressure distribution remained limited. The MLP model achieved stable predictive performance for pattern adjustment quantities. The mean absolute error for predicting pattern bust circumference adjustment was 0.15 cm, and the mean absolute error for shoulder slope angle adjustment was 0.14°. These error levels remained consistent across different experimental groups, including single-variable and coordinated adjustment experiments, indicating stable prediction performance under varying parameter combinations. Analysis of prediction residuals indicated no systematic bias related to pressure magnitude or measurement location. Further verification experiments were conducted to assess the reliability of the proposed approach. Subsequent physical garment pressure tests using an airbag-type contact pressure measurement system confirmed that, after structural modification, measured pressure values at all locations fell within corresponding comfort threshold ranges. The deviation between virtual and physical pressure measurements remained within an acceptable range.

Conclusion Results indicate that garment pressure data obtained through virtual fitting can quantitatively describe the relationship between women's garment pattern structure and pressure distribution. The research confirms that pattern structural parameters, particularly bust circumference and shoulder slope, are closely associated with pressure responses at key anatomical locations, while mannequin neck root girth plays an important role in determining required bust-related pattern adjustments. The MLP model provides stable predictions of pattern adjustment quantities within the experimental parameter range, and verification through physical pressure testing confirms consistency between virtual and real garment pressure distributions.

From a practical perspective, the proposed method offers an objective reference for pattern structure modification during virtual sample development, reducing reliance on empirical judgment. Future work may extend this approach to a wider range of garment types, body shapes, and dynamic postures, and further integrate pressure-based evaluation with digital pattern design systems, supporting data-driven garment development and structural optimization.

Key words: neural network, clothing pressure distribution, garment pattern structure, virtual fitting, garment fit