GeoMamba: Geometric-Prior-Infused Multi-Scale Deformable Visual Mamba for Crack Semantic Segmentation

Accurate pavement crack segmentation is critical for road infrastructure assessment, yet it remains challenging due to complex background noise and highly variable crack topologies. While emerging visual Mamba models excel in long-range contextual modeling, their inherent 1D sequence flattening process compromises local 2D spatial continuity. To address this limitation, we propose GeoMamba, a geometric-prior-infused multi-scale deformable visual Mamba network for road crack semantic segmentation. First, we design a Multi-Scale Deformable Visual State Space (MDVSS) module to extract multi-scale contextual features and dynamically adapt to tortuous crack paths through a novel deformable scanning mechanism. Second, a Geometric-Topology Prior Injection (GTPI) module is introduced to mitigate serialization artifacts. By leveraging deterministic, parameter-free analytical operators (i.e., Sobel and Laplace), the GTPI module explicitly extracts and adaptively infuses multi-scale structural priors into the Mamba decoder via gated skip connections, intrinsically reconstructing crack typologies while suppressing pseudo-structural noise. Comprehensive experiments on DeepCrack and Concrete3K datasets demonstrate that GeoMamba outperforms nine state-of-the-art methods. Specifically, it achieves peak performance on the DeepCrack dataset with an mIoU of 83.79% and an F1 score of 89.27%, demonstrating exceptional semantic segmentation performance, superior topological continuity, and robust generalization across diverse pavement materials.