Dual-encoder Multi-Scale Refinement network for robust crack segmentation across diverse domains

Accurate crack segmentation is critical for infrastructure monitoring but remains challenging due to diverse crack morphologies, complex backgrounds, and domain shifts. This paper proposes DMSRCrack, a Dual-encoder Multi-Scale Refinement network for robust crack segmentation across diverse domains. The proposed architecture integrates a hybrid CNN–ViT encoder for local and global feature extraction, a Crack Detail Enhancement Module (CDEM) for preserving thin crack structures, a Boundary Refinement Head (BRH) for contour sharpening, and a Multi-Scale Fusion (MSF) module for scale-consistent representation. Across eight benchmark datasets, DMSRCrack achieves an average Dice of 0.7949 ± 0.0655 and IoU of 0.6639 ± 0.0917 in dataset-wise training. Under leave-one-dataset-out evaluation, it attains the highest IoU on DeepCrack (0.4056), Rissbilder (0.3540), and Crack500 (0.3008). Ablation and computational analyses further confirm the effectiveness and practical efficiency of the proposed contributions. • Dual-encoder network captures local texture and global crack structure. • CDEM, BRH, and MSF modules preserve thin cracks and sharpen boundaries. • Composite loss improves pixel accuracy and topological consistency. • Strong cross-dataset performance under leave-one-dataset-out testing. • Favorable accuracy-efficiency balance with 11.66M parameters.