CrackMambaNet: A road crack segmentation network based on dual-branch encoder ensemble with adaptive feature fusion

Automated crack detection is essential for extending the service life of roads and ensuring the operational safety of road structures. Numerous computer vision-based networks have been proposed for crack segmentation in recent years. However, existing networks frequently struggle to balance global context modeling with local detail preservation and lack sufficient robustness in complex scenarios. To address these issues, this study proposes CrackMambaNet for automated pavement crack detection. It is a U-shaped segmentation network with a dual-branch encoder incorporating an adaptive feature fusion module (FFM). The dual-branch encoder consists of global and local branches. The former stacks visual state space (VSS) blocks to capture multi-scale long-range dependencies with linear complexity. The latter is built upon an enhanced ResNet34 backbone integrated with visual attention convolution modules to reinforce edge-aware and fine-grained feature extraction. An adaptive FFM is thus introduced to suppress redundant features through dual SE-attention. Extensive experiments were conducted on various public crack datasets to validate the performance of CrackMambaNet. The results indicate that CrackMambaNet achieved the best performance across different datasets compared with seven existing networks. The mean intersection over union on the DeepCrack, EdmCrack600, and CFRL datasets reached 87.94 %, 76.33 %, and 75.58 %, respectively. These results demonstrate that CrackMambaNet can balance global context modeling and local detail preservation, exhibiting excellent robustness and accuracy in complex scenarios and fine crack segmentation tasks.