Confidence-regression-based fatigue crack path recognition method for orthotropic steel decks

Due to their long, sparse, and curve-like characteristics, semantic segmentation–based methods struggle to preserve the geometric continuity and topological structure of crack propagation paths. This study proposes a crack-path recognition framework that formulates crack-path recognition as a confidence heatmap regression problem. Instead of delineating crack regions at the pixel level, the proposed approach directly models crack paths as continuous geometric entities, enabling extraction of crack paths under complex background. To support this formulation, a task-adapted crack-path dataset is constructed, in which crack annotations are represented as centerline-based confidence heatmaps rather than binary segmentation masks. Meanwhile, to address the requirements of multi-scale feature representation and path information preservation in this task, two feature enhancement modules are designed. These modules strengthen path-related responses while suppressing background interference during feature extraction and fusion. Experimental results demonstrate that, compared with segmentation-based crack path detection methods, the proposed method reduces path recognition error and inference time by approximately 50%. Multi-angle imaging further confirms the robustness of the proposed method under complex field conditions. When the data are integrated into a digital twin system, cracks can be automatically displayed on the virtual model.