Connectivity-preserving Crack Extraction from Geometrically Aligned Tunnel Face Images in Sequential Excavation

Accurate extraction of rock mass cracks from tunnel face images is essential for assessing rock conditions and supporting construction safety. However, on-site tunnel face images often suffer from perspective distortion, complex textures, and severe background interference, which make consistent spatial localization and comparison difficult and reduce the reliability of crack extraction. Moreover, cracks typically exhibit thin, elongated, and highly connected structures; preserving their continuity is difficult with conventional pixel-wise segmentation methods. This paper proposes an automated connectivity-preserving framework for tunnel face crack extraction. Stage 1 maps each tunnel-face image to a standardized two-dimensional image layout. In the tunnel-face workflow, Stage 2 applies a Transformer-based segmentation model to patches extracted from the normalized view. During training, clDice supervision encourages crack continuity within individual patches. Experimental results show that the proposed method outperforms representative baselines in both region-level accuracy and structural connectivity. The full-face and sequential examples illustrate the use of the resulting maps for fracture documentation and qualitative comparison. Overall, the proposed framework offers a practical solution for automated tunnel face crack analysis and strong potential for integration into intelligent tunnel construction and monitoring systems.