A Crack Segmentation Network Integrating Multi-Scale Attention Residual and Context-Enhanced Transformer Block

Crack segmentation is crucial for infrastructure maintenance, yet it remains challenging due to false detection from crack-like textures and discontinuous predictions from inadequate continuity modeling. To address these issues, we propose a complementary synergistic fusion network (CSF-Net). Our approach features a dual-branch encoder: a local texture branch equipped with a novel multi-scale attention residual (MSAR) module to suppress texture interference, and a global structure branch incorporating a context-enhanced attention module (CEAM) to enhance the modeling of crack continuity. These complementary features are fused via a cross-branch fusion block (CFB). Extensive experiments on three public benchmarks demonstrate the superiority of CSF-Net. On the DeepCrack dataset, it achieves an mIoU of 72.79% and an F1-score of 86.67%, significantly outperforming the U-Net baseline by 7.08% and 7.49%, respectively. CSF-Net also exhibits state-of-the-art and robust performance on the CFD and Crack500 datasets, confirming its strong generalization capability.