Answers are grounded in this paper and cite their sources.
Accurate concrete crack segmentation is critical for reliable structural assessment, yet it remains challenging due to irregular morphology of cracks and the ambiguity of their boundaries. This leads to a fundamental limitation in existing methods: while engineering assessment requires reliable predictive uncertainty estimation, most existing segmentation models produce a single deterministic mask, resulting in unstable geometric measurements and limited interpretability. Recent diffusion-based approaches introduce probabilistic segmentation but often lack explicit structural guidance during the denoising process. To address these limitations, this study proposes ProbCrack-Diff, a structure-aware probabilistic framework for concrete crack segmentation, quantification, and uncertainty estimation. The framework integrates a Lightweight Local–Global Encoder (LLGE) to capture multi-scale structural representations and a Structure-aware Discriminator (SA-Discriminator) to provide geometry-aware supervision through structural distillation. By incorporating crack-specific structural priors into the diffusion process, ProbCrack-Diff promotes the preservation of crack continuity, boundary integrity, and geometric fidelity. Extensive experiments demonstrate that ProbCrack-Diff achieves superior segmentation accuracy and geometric fidelity compared with existing discriminative and diffusion-based approaches. Furthermore, uncertainty–reliability analyses indicate that higher predictive uncertainty estimates are generally associated with larger segmentation and crack-geometry estimation errors, suggesting that the estimated predictive uncertainty may provide useful supplementary information regarding prediction reliability in automated crack assessment.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: