
Accurate crack segmentation is critical for structural health monitoring but is often hindered by the limited availability of annotated images, especially for complex crack patterns. Traditional data augmentation and GAN-based methods frequently introduce artifacts and fail to preserve fine structural details, reducing segmentation reliability. To address these challenges, this study employs diffusion-based generative models to produce high-fidelity synthetic crack images, enriching training datasets and improving robustness. The proposed approach enhances segmentation accuracy and structural continuity, providing an effective solution to data scarcity and quality limitations in crack detection tasks.
View this paper