Label-efficient semi-supervised learning with hard negative samples for crack segmentation in concrete and asphalt structures

Crack segmentation is critical for maintaining structural safety, but fully supervised deep learning approaches require extensive pixel-level annotations, making large-scale deployment costly. This study adapts semisupervised learning framework for crack segmentation that extends weak-to-strong consistency regularization with domain-aware hard negative sample (HNS) integration to minimize annotation needs while improving robustness. The framework utilizes a dual-stream architecture with weak-to-strong image- and feature-level perturbation-based consistency regularization, adapted from general semantic segmentation to the domain-specific challenges of structural crack inspection. HNS, which are visually deceptive non-crack patterns such as shadows, stains, and surface textures, were integrated into the unlabeled training stream, enabling the model to better distinguish cracks from background noise and substantially reduce false positives. The proposed method achieved performance comparable to that of a fully supervised model using only 20% of the labeled data, effectively reducing the labeling costs by 80%. Experiments on concrete and asphalt structures demonstrate that HNS integration improves precision, especially in visually complex conditions, while a mixed-domain model trained on both surfaces achieves strong generalization. Furthermore, the proposed method was compared with other state-of-the-art semi-supervised methods and consistently achieved improved performance in both concrete and asphalt datasets. These findings indicate that integrating SSL with HNS enhances crack segmentation, providing a scalable solution for real-world applications where background complexity remains challenging.