Edge-aware transformer-based damage segmentation framework for bridge inspection maps using synthetic data

Most bridge inspection records are kept in analog formats as inspection maps with cracks annotated by hand, which limits their usefulness in modern digital asset management systems. Although deep learning has achieved strong performance on crack detection in photographic imagery, these methods depend heavily on color and texture cues that are absent in inspection maps. To address this gap, this paper presents a deep-learning framework designed to segment cracks directly from binary inspection maps. A synthetic dataset generation pipeline was developed using Auto LISP scripts to simulate cracks in CAD-based bridge elevations, reducing the need for manual annotations. Building on advances in general-purpose segmentation architectures, we design a customized model that incorporates an edge-sensitive decoding module, a structure-aware loss combining geometric and pixel-level accuracy, and a sliding-window inference strategy for processing large, high-resolution drawings. The model is trained on synthetic data and evaluated on real inspection maps to test its generalization ability. Results show that the proposed method consistently outperforms widely used segmentation baselines in both quantitative accuracy and visual clarity. Ablation studies further confirm the contribution of each architectural component. Beyond static segmentation, the framework enables time-series visualization of crack evolution, supporting condition tracking across historical records. This approach provides a scalable and practical solution for digitizing analog inspection data, making it compatible with Building Information Modeling (BIM) and digital twin systems. By transforming long-term inspection archives into actionable digital resources, the proposed method enhances efficiency, continuity, and data-driven decision-making in bridge maintenance workflows.