DuEx-Crack: A reliability-aware dual-expert framework for cross-domain crack segmentation

Automated crack segmentation for civil infrastructure must deliver both label efficiency and cross-domain robustness. Recent pipelines couple a domain-specific detector with a frozen vision foundation model such as the Segment Anything Model (SAM) in a cascade, and therefore inherit the failure modes of the foundation model: both the regime in which it returns no foreground at all, and the regime in which it returns noisy false positives. We propose DuEx-Crack, a rule-based framework with three core contributions: (i) the silent rate , a training-free, label-free, dataset-level probe that quantifies how often a frozen foundation model returns an empty mask on the target domain before fusion; (ii) a deterministic adaptive rule driven by the silent rate that selects between connectivity-driven and conservative-union fusion without learned gating or per-dataset tuning; and (iii) a parallel dual-expert architecture that replaces the cascade design, providing structural graceful degradation when the foundation model is silent and bounding rather than propagating the contamination from noisy predictions, as quantified against a true cascade under identical weights. Instantiated with YOLO26-seg-s (11.4 M parameters, bounding-box supervision) and frozen SAM3, DuEx-Crack is evaluated on the Roboflow and CrackVision12 K benchmarks, surpassing the supervised state of the art on Roboflow by IoU with only of the trainable parameters, and outperforming it by and IoU under cross-domain transfer. As an end-to-end engineering integration, a mask-grounded vision-language module converts the fused masks into structured JSON diagnostic reports, demonstrating a practical inspection use case built on the framework.