Unsupervised Domain Adaptation for Crack Segmentation via Cross-Domain Stylization and Dual Adversarial Feature Learning

Surface crack segmentation is critical for infrastructure inspection, yet deep-learning-based methods are hampered by their reliance on large annotated data sets and poor generalization across domains due to distribution shifts. While unsupervised domain adaptation (UDA) offers a promising solution, existing methods struggle with crack-specific challenges; since thin structures are highly sensitive to style and texture interference, feature-level disparities are overlooked, and critical edge information degrades during adaptation. To address these issues, we propose CDE-Crack, a novel unsupervised domain adaptation framework for robust crack segmentation, featuring three innovative modules: (1) a cross-domain stylization network (CSN) that aligns pixel-level distributions through style transfer, shape preservation, and frequency-domain amplitude fusion, effectively minimizing domain gaps while retaining structural fidelity; (2) a dual adversarial feature learning (DAFL) strategy that simultaneously aligns high-level semantic features and feature representations, significantly enhancing generalization; and (3) an edge-consistency refinement (ECR) module that leverages edge-aware supervision to correct boundary distortions and preserve crack integrity. Extensive experiments on five data sets, Crack500 (buildings), CrackTree 200 (pavement), CrackFCN (concrete), CrackForest (asphalt), and MVTec AD (marble tiles), demonstrate that CDE-Crack outperforms state-of-the-art methods, achieving remarkable improvements in cross-domain segmentation accuracy, including a 12.473% optimal image scale (OIS) improvement for Crack500→CT200, a 13.15% OIS improvement for CrackFCN→CFD, and a 10.038% OIS improvement for CrackFCN→MVTec AD, particularly in preserving fine crack details. These results validate the framework’s robustness to domain shifts and its substantial potential for real-world inspection scenarios, where data variability remains a persistent challenge.