This study aims to address severe data scarcity in automated tunnel crack detection systems, particularly for micro-cracks in complex environments, which hinders urban infrastructure safety monitoring.
The authors propose a novel diffusion-based framework integrating three innovations: background-parametric diffusion module fusing multi-scale tunnel textures with controllable crack geometry; saliency-guided enhancement mechanism using binary masks to amplify micro-crack details; task-driven quality assessment system (TDQS) combining edge fidelity score (EFS) and semantic consistency score to bridge generation quality with detection performance.
Evaluated on 5,820 tunnel images, the proposed method achieves 19.2% reduction in FID versus state-of-the-art methods and 17.1% improvement in micro-crack detection F1-score, with enhanced sample realism validated by task-specific metrics (EFS: 0.83, TDQS: 0.82).
This work establishes the first framework enabling: parametric crack geometry control with background fusion; saliency-guided micro-crack enhancement; and TDQS linking synthetic data quality to detection utility, creating a new paradigm for structural health monitoring.
