Defect Augmented Heritage-BIM For Masonry Structures Leveraging Deep Learning-Based Segmentation

Defects on masonry built heritage require regular monitoring and effective management to reduce structural risks. Deep learning techniques such as CNN have been employed for automated defect identification; however, previous models underperform in complex defect scenarios due to inadequate sample size, unbalanced categories, and unclear classification standards of the training dataset. Moreover, previous mapping of defect information onto H-BIMs is superficial and not deeply integrated with specific components, which is not conducive to subsequent numerical simulation. To address these challenges, this study establishes quantifiable classification rules and introduces data augmentation and k-fold cross-validation in dataset preparation and model training to improve identification accuracy. A multi-layer masonry H-BIM is proposed, where the masonry arrangement is extracted automatically, individual masonry unit components with realistic geometries are generated parametrically, and defect information is registered with and integrated into corresponding components. A case study was conducted on an ancient city wall in Suzhou, China. The trained CNN model accurately segmented various defects, and the generated multi-layer masonry H-BIM incorporates defect visualisation, lightweighting, and semantic enrichment, laying a foundation for subsequent numerical simulation.