Digitalization of 2D Bridge Drawings Using Deep Learning Models

Technical drawings are a resource to create the geometric digital twin (DT) of existing bridges. A bridge DT demonstrates the current geometric-semantic information of the structure and supports the operation and maintenance process of bridges. Despite the significant advantages of a bridge DT, creating its 3D model from drawings is costly and labor-intensive. This paper presents a method to digitalize the technical drawing of bridges by deep learning models such that the required data for geometric modeling can be extracted more straightforwardly. The parametric model of bridge elements is created and used to generate a synthetic dataset. This dataset is combined with the actual drawings and a deep learning model is trained to detect bridge elements. Dimensions are also extracted using a pre-trained model and digitalized through optical character recognition (OCR). The results of the paper show that the model can detect different elements in drawings with a mean average precision (mAP) of 89.15%.