건설장비 객체인식 모델 성능 향상을 위한 생성형 AI 기반 합성 이미지 데이터 활용

Enhancing Construction Equipment Detection with Generative AI-Based Synthetic Image Data

Dynamic and complex construction environments require effective monitoring to enhance productivity and ensure safety. While recent advancements in computer vision and deep learning have enabled the application of object detection models for this purpose, their performance is often constrained by the scarcity and diversity of training data. The process of collecting and annotating sufficient real-world construction images is labor-intensive and frequently limited by specific site conditions. This study addresses this challenge by investigating whether synthetic images, generated via a Text-to-Image model, can effectively supplement real datasets. For our experiments, we extracted roller and dozer classes from two publicly available datasets, ACID and MOCS, and generated corresponding synthetic images. We trained YOLOv11n models under three distinct scenarios: using synthetic images only, real images only, and a combination of both. Models trained with real data augmented by synthetic images showed consistent performance gains in both mAP50 and mAP50-95. This performance improvement was most significant with the smaller MOCS dataset, suggesting that synthetic data has a stronger supplementary effect when real-world data is limited.