metal‐dacl: Image‐Based Automated Damage Recognition for Steel Bridge Inspections

As infrastructure ages and the number of structures requiring inspection increases, effective monitoring of damage in built structures has become more crucial than ever. Staff shortages and budget constraints can make it difficult for authorities to conduct the necessary frequent inspections. To address these challenges, companies and research institutions are increasingly exploring digital approaches to building inspection. Digitalized inspection processes involve creating a digital shadow of the structure, which combines a BIM model with a record of classified, measured, localized and assessed defects. Key to this approach is the deployment of transformer‐based architectures for image‐based automated damage recognition. Accurate defect recognition is essential for evaluating the condition of specific areas and assessing the overall integrity of a structure. This paper presents a dataset extension for the dacl10k dataset tailored to steel defect recognition on bridges. We manually assigned polygonal annotations to 3,737 images of dacl10k that showed steel bridges or building parts. Despite the challenging nature of this segmentation dataset extension, our baseline model achieves a mean Intersection‐over‐Union of 17.37%. This result provides a valuable reference point for future models and highlights the complexity inherent in detecting fine‐grained steel defects. We conduct a detailed analysis to uncover the factors limiting model performance and suggest pathways for improvement.