Abstract:

Objective An intelligent method is proposed for extracting clothing specification dimensions from two-dimensional (2-D) human body images, aiming to provide efficient and low-cost technical support for the intelligent production of apparel.

Method The proposed method uses 2D body images as input, and selects 19 key anthropometric parameters as the foundational data. In order to address the complexities of clothing sizing arising from human body curvature, imaging perspective, fabric properties, and style variations, a comprehensive multi-factor correction model is developed. This model systematically integrates three core correction mechanisms, which are the curvature compensation for converting chord lengths to chord lengths along body contours, perspective adjustment for compensating dimensional distortions caused by camera angles, and dynamic ease allocation that considers both garment style and specific body regions. Furthermore, a differentiated fitting strategy is employed to accurately quantify clothing dimensions across various body sections. For body areas with relatively uniform cross-sections, such as the waist and certain parts of the limbs, an elliptical fitting method is utilized. This approach models the body section as an ellipse, and the clothing dimension is derived by uniformly adding the design ease to its semi-major and semi-minor axes before calculating the adjusted perimeter. For irregular or non-uniform body sections, such as the armhole, chest, and hip areas, a feature-adaptive fitting method based on radial expansion is adopted.

Results This technique extends the original body contour points outward along radial vectors from a central point, with the extension magnitude modulated by a dynamic coefficient to allow for non-uniform ease distribution, thereby more accurately capturing the clothing's shape around complex anatomies. The experimental validation involved selecting representative daily clothing styles. The study focused on key measurement areas pertinent to each garment type: bust, waist, shoulder width, sleeve length, and garment length for tops; waist, hip, and length for bottoms; and a combination of these for dresses. In order to rigorously evaluate the method's performance, its results were compared against both physical measurements (treated as ground truth) and estimates made by ten experienced industry professionals (pattern makers, buyers, retail consultants) using a triple-blind assessment protocol. The primary metrics for comparison were the mean absolute percentage error (MAPE) and the mean relative error (MRE).Results demonstrated a significant enhancement in estimation accuracy compared to manual methods. The proposed method achieved an overall MAPE of 3.02% and an MRE of 2.23 cm across all tested clothing and measured dimensions. In contrast, the manual estimation by experts yielded a notably higher MAPE of 8.78% and an MRE of 6.28 cm. A detailed analysis revealed that the method performs exceptionally well for garments with regular, structured silhouettes, where errors for linear dimensions like garment length and sleeve length were often below 1 cm (approximately 1% error). Accuracy remained high for symmetrical lower-body garments in dimensions like hip circumference and inseam length. However, as anticipated, larger errors were observed for circumference measurements (e.g., waist, underbust) in loose-fitting garments like sweatshirts and relaxed dresses, primarily due to fabric drape, folds, and blurred contours in the 2-D images.

Conclusion A novel, image-based framework is prosposed for intelligent clothing size extraction that effectively marries the convenience of 2-D image processing with the incorporation of 3-D body morphological information through parametric modeling and correction factors. It successfully mitigates the subjectivity, inefficiency, and higher error rates inherent in manual estimation while avoiding the high cost, operational complexity, and computational demands associated with mainstream 3-D scanning solutions. The proposed method offers a practical, cost-effective tool for applications such as automated pattern making, personalized size recommendation, and smart manufacturing workflows, thereby contributing to enhanced precision, efficiency, and digital transformation within the apparel industry. Future work will involve expanding the clothing style dataset for training and integrating the framework with advanced deep learning models to further generalize and optimize its performance.

Key words: clothing dimension extraction, size quantification, parametric modeling, region-adaptive fitting, error-correction framework