CrackSnakeNet: Multi-directional snake scanning with Mamba for topology-preserving crack segmentation

Thin cracks, typically only 1–3 pixels wide, remain the most challenging surface defects to segment in automated structural inspection. This paper presents CrackSnakeNet, a hybrid CNN-Mamba encoder-decoder designed to enhance both detection sensitivity and structural consistency. Its core module, SnakeMamba2D, scans feature maps along eight boustrophedon trajectories aligned with principal crack directions, effectively preserving spatial adjacency. A geometric analysis shows that the four axis-aligned scans preserve only 66.4% of the adjacencies along real crack centerlines, a deficit that the diagonal trajectories remove entirely. In addition, a centerline recall loss enforces prediction coverage on crack medial axes without requiring differentiable skeletonization. Experimental results show that CrackSnakeNet performs comparably to state-of-the-art methods on wide-crack datasets such as DeepCrack, while demonstrating clear advantages as crack width decreases. On CrackTree260, where cracks average 1 pixel wide, it achieves 87.23% T sens , the topology-sensitivity metric that measures the fraction of ground-truth centerline pixels covered by the prediction, outperforming DBCNet (85.26%) and U-Net (74.49%). On CRACK500, it achieves the best T sens of 82.39%. This trend indicates that the proposed method is particularly effective for extremely thin cracks. In cross-domain tests on CFD, CrackSnakeNet consistently achieves the highest clDice across all source domains, reaching 69.16% when trained on DeepCrack, compared to 58.62% for the second-best method. All results obtained with only 9.14 M parameters, a count that is less than half the 24.44 M of U-Net and roughly one-fifth the 50.57 M of DBCNet, and without ImageNet pretraining.