ScholarWorks@성균관대학교

초록

Structural damage identification is crucial in civil engineering for ensuring infrastructure safety and durability. While machine learning offers the potential for automating this process, limited and inaccessible real-world data pose significant challenges. To address this, synthetic data generation has emerged as a promising solution to expand datasets and enhance model performance. This study introduces a novel approach using a 3D engine environment to generate diverse synthetic crack images through randomization of lighting, scale, and background. The synthetic dataset was meticulously designed to match the quantity of real data for a fair comparison. Experimental results show that models trained on synthetic data perform better in both accuracy and generalization than those trained only on real data. Using mean average precision (mAP) as a performance metric, we achieved an impressive 95.9% accuracy. These findings underscore the potential of synthetic data for improving crack detection and emphasize the value of simulation-based techniques for creating high-quality synthetic datasets. This research contributes to advancing classification models in data-scarce environments, paving the way for safer and more resilient infrastructure.

키워드