Constructing a Traditional Hanbok Dataset for AI Training-Challenges in Digital Transformation; [AI 학습용 전통 한복 데이터 구축에 관한 연구-디지털 전환을 위한 도전과 과]

Abstract

This study examines the concepts and characteristics of traditional Hanbok, establishes a framework for collecting and classifying AI training data, and diagnoses issues arising during image collection to propose improvements. It assembled 1.1 million two-dimensional and three-dimensional images of garments and accessories in dressed and undressed states, capturing them at 10–20 degree intervals for full 360-degree coverage across the front, side, and back, under varied lighting, poses, and camera viewpoints. Master’s-and doctoral-level researchers designed and labeled a fifteen-field metadata schema,—covering type, subclass, period, usage, gender, age, status, item name, material, pattern, decorative technique, and color. Outfit combinations were collected in three stages by layering garments—adding outerwear over topwear and incorporating accessories—without duplication. Identified challenges include conflicts in period attribution owing to modern materials in replicas, ambiguity across transitional eras, unclear distinctions of use and social status, imbalances for underwear and socks, chroma-key backdrop failures for green hues, and color distortions in sheer weaves and gilded surfaces. The study proposes refined category boundaries, a controlled vocabulary and labeling guide, on-site expert verification, standardized capture and color criteria, and bias-mitigation measures. The resulting framework enables reliable training and supports academic, industrial, and policy applications. © 2026, The Korean Society of Clothing and Textiles. All rights reserved.