Context- and refinement-driven convolutional architecture for robust crack segmentation under real-world and zero-shot conditions

Crack segmentation is essential for infrastructure monitoring but remains challenging due to complex textures, lighting variations, and the thin, fragmented nature of cracks. This paper presents a deep learning architecture, CrackRefineNet, which integrates three complementary modules, Context-Aware Modeling (CAM), Adaptive Feature Refinement (AFR), and Selective Feature Aggregation (SFA), to improve crack localization and boundary precision. Built on the ConvNeXt backbone, the method captures both fine-grained structures and global context by injecting spatial information into early layers and refining deep semantic features at later stages. Aggregated intermediate predictions guide final mask generation through a unified decoder. Comprehensive experiments on four benchmark datasets (RCFD, Crack500, CFD, and Sylvie) show that CrackRefineNet outperforms state-of-the-art methods by up to 3.5% in IoU and 2.8% in F1-score, achieving 42.91% F1-score and 63.03% mIoU in zero-shot testing, confirming its robustness and readiness for real-world applications. • Context- and refinement-driven CNN for reliable crack segmentation. • Improved feature quality to detect fine and discontinuous cracks. • Reliability assessed using RGA, RPBT, and Bayesian confidence metrics. • Attention maps provide clear and transparent model interpretability. • Strong generalization across real cases and unseen benchmark datasets.