A lightweight crack segmentation network based on the importance-enhanced Mamba model

In the maintenance of transportation infrastructure, crack segmentation is critical for ensuring road safety and prolonging the service life of bridges and other facilities. Existing methods struggle with complex background interference and intricate crack morphologies (e.g., mesh-like or tree-like morphologies). Meanwhile, high-precision models frequently suffer from excessive computational costs. To address these limitations, this study proposes, a lightweight crack segmentation network based on an Importance-Enhanced Mamba model. Built upon the U-Net architecture, innovatively features a dual-branch design that integrates CNN and Mamba modules for synergistic feature extraction. Within the Mamba branch, we designed an importance-enhanced dynamic scanning module, which adaptively adjusts scanning paths according to actual crack geometries, thereby significantly enhancing the perception of global key crack features. Concurrently, the CNN branch specializes in capturing fine-grained local features such as edges and textures. These complementary features are fused via an attention-guided module, which assigns adaptive weights to enable pixel-wise integration of local and global information, thus preserving both microstructural details and macroscopic relationships of crack. Comprehensive experiments conducted on three public datasets (Crack500, CrackTree260, and CrackForest) demonstrate that outperforms other advanced methods in segmentation accuracy while achieving significant reductions in model parameters and computational complexity.