|
Title |
Development of a Neural Network Model for Structural Surface Crack Detection Based on Data Augmentation Using DreamBooth
|
|
DOI |
https://doi.org/10.11112/jksmi.2025.29.6.70 |
|
Keywords |
균열 탐지; 드림부스; 스테이블 디퓨전; 데이터 증강; 딥러닝 Crack detection; DreamBooth; Stable diffusion; Data augmentation; Deep learning |
|
Abstract |
Cracks in aging concrete structures pose significant threats to structural stability and public safety. However, it is difficult to secure sufficient crack image datasets that reflect diverse deterioration conditions in real-world scenarios, which leads to performance degradation of crack detection neural networks when applied on-site. To address this data scarcity problem, this study employs a generative AI approach based on Stable Diffusion.
Specifically, an inpainting fine-tuning method using DreamBooth was applied to synthesize realistic cracks within masked regions, and additional training datasets were generated through combinations of different label images. For evaluation, 4,867 concrete crack images were divided into training, validation, and test sets with ratios of 15%, 15%, and 70%, respectively, and four segmentation models?DDRNet, RegSeg, SwinFormer(T), and PoolFormer(S24)?were assessed. The quality of the generated images was evaluated using the Frechet Inception Distance metric, confirming visually consistent and high-quality crack synthesis. In terms of detection performance, incorporating synthetic data improved the average F1-score by 5.73% compared with the baseline, with PoolFormer(S24) achieving the highest improvement of 11.96%. In conclusion, the proposed synthetic data-based augmentation method effectively enhances crack detection accuracy across diverse network architectures and environments, providing a practical and generalizable alternative for real-world applications.
|