MambaFuse: Cross-scale state space fusion for crack segmentation

The detection of cracks on the surface of infrastructure structures is a critical component of structural health monitoring. Addressing the core challenge of insufficient long-range dependency modeling in low-resolution crack detection, this paper proposes MambaFuse, a novel multilevel encoder–decoder model. This framework innovatively integrates the local feature extraction capability of CNNs, the global modeling strength of Transformers, and the long-sequence processing characteristics of Mamba. Field tests based on an autonomous mobile detection platform confirm the model’s exceptional ability to maintain crack topological continuity in real-time detection, with its selective state space mechanism successfully resolving fracture issues commonly encountered in dynamic mobile imaging. To advance research in this field, we constructed the CrackBench benchmark dataset containing 1,000 annotated images from multiple scenarios, and developed a geometry-based crack quantification method that enables direct conversion from pixel-level detection to engineering-applicable quantitative metrics. Experimental results demonstrate state-of-the-art performance in multiple benchmark datasets: 90.04% mIoU on DeepCrack, 79. 58% mIoU on Crack500 and 86. 17% mIoU on CrackBench, validating its superior segmentation accuracy and cross-scenario robustness. • Proposed MambaFuse, a multi-modal fusion architecture integrating CNN, Transformer, and Mamba to address crack misclassification and poor continuity modeling in current detection methods. • Built CrackBench dataset containing 1000 RGB images with high-quality labels across multiple scenarios. • Developed a geometry-based quantification method converting pixel-level detection into engineering-applicable metrics.