HLGCrack: A Lightweight Hierarchical CNN–Mamba–Attention Network for Crack Segmentation

Accurate crack segmentation is important for intelligent infrastructure inspection and has potential applications in railway infrastructure monitoring. However, cracks usually exhibit thin, elongated, irregular, and spatially asymmetric morphologies, while weak boundaries and complex surface textures further increase segmentation difficulty. To address these challenges, this paper proposes HLGCrack, a lightweight hierarchical CNN-Mamba-attention network for efficient crack segmentation. The network adopts a stage-wise representation strategy in which shallow asymmetric convolution and large-kernel modeling enhance directional edge responses and local crack structures, while deeper visual state-space and local–global attention modules capture long-range dependencies and contextual information. A Lightweight Progressive Fusion decoder progressively aggregates multi-level features to recover fine crack boundaries with low computational cost, and BCE, Dice, and clDice losses are jointly employed to improve region overlap and structural continuity. HLGCrack is evaluated on CamCrack789 and four additional public pavement and structural crack datasets covering diverse crack morphologies, surface textures, and interference conditions. On CamCrack789, it achieves Dice and IoU scores of 0.860 and 0.755, respectively, with only 3.53 M parameters and 4.03 G FLOPs. The results demonstrate its effectiveness and robustness for general crack segmentation and provide a basis for subsequent adaptation to railway-specific inspection scenarios.