Requirements for Parametric Design of Physics-Based Synthetic Data Generation for Learning and Inference of Defect Conditions

The advancement of Artificial Intelligence (AI)-driven defect detection has already demonstrated promises to boost quality assurance and control, as well as condition assessment in the built environment. However, training defect detection models requires hefty amounts of reality capture data, and labeling is considered expensive. In most cases, such data may not cover all situations of defects. Synthetic data, most recently made with Building Information Models (BIM), is turbocharging model development for learning defect features. Nevertheless, few studies focused on characterizing defects to classify their severity, which is crucial to the condition assessment. To that end, this study explores the requirements for generating synthetic data. Parametric physics-based modeling approaches are carefully examined. Using the underlying geometric properties of such data, the condition of each defect can be determined. The feasibility of synthetic defect data is validated with a case study of crack segmentation using the transformer-based model, SegFormer. Examples of how different scenarios can be generated photo-realistically with the use of physics-based rendering for creating varying geometrical characteristics, appearance, and viewpoints of defects are presented. The generated synthetic crack datasets can successfully be used to train the SegFormer model and reach promising predictions on real crack images.