Multi-Modal, Training-Free Crack Extraction via Generalized Frangi Graph
Year: 2026
Type: conference-paper
Abstract: Automatic crack extraction is essential for both infrastructure inspection and geosciences, but annotated crack data remain scarce and domain shifts are frequent. We address the problem of extracting a crack network from one or several scalar modalities, without training. To do so, we generalize the classical FRANGI vesselness filter from a pixel-wise operator to a pairwise spatial graph. While the standard filter is limited to independent pixel-wise geometry, our formulation enriches this description by explicitly encoding the orientation alignment between neighboring pixels, alongside elongation and contrast. The final crack network is then extracted through a graph pipeline combining minimum spanning tree reduction and similarity-weighted tree centrality. While supervised models perform best on clean, in-distribution data, our geometry-driven framework degrades more slowly than CrackSegDiff under the tested synthetic perturbations.
Source: HAL
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7202062500