Topology-informed deep learning for pavement crack detection: Preserving consistent crack structure and connectivity

Abstract

This paper addresses the challenge of crack detection, where incorrect connections often distort crack topology. By leveraging topology theory, which focuses on properties that remain invariant under continuous transformations, the goal is to preserve key geometric features like connectivity and loops. For future-oriented road maintenance, fine segmentation that preserves the topological integrity of crack structures is essential for efficient automated repairs and crack characterization. To this end, the research combines persistent homology (pH) with the U-Net architecture enhanced by the Vmamba model, forming TopoM-CrackNet. TopoM-CrackNet outperforms other topology-preserving methods, such as Topoloss, with a Betti number of 4.032. It also achieves a mean Intersection over Union (mIoU) of 0.727, surpassing traditional methods like nnUnet and Segformer, and is nearly twice as fast. Overall, the key contribution is its ability to significantly improve crack topology preservation during segmentation, offering technical support for crack detection and automatic repair.