Directional multi-scale CNN-transformer hybrid network for robust crack segmentation

Crack segmentation is a critical task in structural health monitoring for infrastructure safety assessment. However, existing methods often struggle with incomplete detection of fine cracks and imprecise boundary delineation under complex backgrounds. To address these issues, this study proposes a directional multi-scale CNN-Transformer hybrid network (DMSF-Net) within an encoder-decoder architecture for robust crack segmentation. The proposed method incorporates three key components. First, an adaptive multi-scale spatial attention module (AMSAM) is designed to enable effective local-global feature fusion, allowing the model to simultaneously preserve fine crack details and structural continuity. Second, a crack-aware fusion module (CAFM) is introduced to explicitly encode the directional and anisotropic characteristics of cracks through multi-directional convolutions and channel attention, thereby improving boundary localization and reducing false detections. Third, a deep supervision strategy is employed to enhance gradient propagation and accelerate convergence. Extensive experiments on four public datasets (DeepCrack, CrackTree260, ConcreteCrack, and BCL2.0) and a self-constructed dataset demonstrate that DMSF-Net consistently outperforms state-of-the-art methods, achieving IoU scores of 82.41% on DeepCrack and 83.20% on BCL2.0. The proposed model exhibits strong robustness in fine crack detection and complex background scenarios, highlighting its potential for practical engineering applications.