패션 산업의 인공지능 전환을 위한 패션 이미지 라벨링 자동화 기법 연구

Abstract

In response to the growing emphasis on digital transformation (DX) and artificial intelligence (AI) transformation (AX) within the fashion industry, this study develops an AI-based fashion image–labeling system aimed at improving the efficiency and accuracy of large-scale data processing. Given the dominance of visual information in fashion, both the systematic extraction and structuring of meaningful attributes from image data are essential for trend forecasting and demand prediction. Therefore, this study proposes a two-step labeling framework based on the YOLO object-detection algorithm. First, a 1-depth AI model was trained on approximately 20,000 fashion images from the AI Hub to detect and classify garments into broad categories, specifically “top” and “bottom.” Second, a 2-depth model was constructed using 10,000 image–text pairs from the online platform Musinsa to assign fine-grained attribute values (e.g., “blouse,”, “pants”) to each detected item. The experimental results demonstrated a labeling success rate of 96.29%, indicating the strong potential of the system for high-performance, automated labeling at scale. This study contributes to the devel- opment of scalable fashion data–annotation methods, offering practical implications for online retailers aiming at efficient visual data structuring. It also lays a foundation for more accurate and real-time AI-driven fashion trend analytics, thereby supporting both academic research and industry innovation.