Proposal of a deep learning-based automatic segmentation method for raster diagrams of existing bridges

Abstract

In recent years, there has been an increasing demand for efficient and quantitative analysis models in the maintenance and management of existing bridges. For bridges where design diagrams are only available in raster formats, high-precision automatic segmentation techniques are essential for generating finite element method (FEM) models. This study proposes a method for automatically segmenting and recognizing components in raster diagrams, focusing on side, plan, and cross-sectional views of existing bridges. Using supervised learning with the DeepLabv3+ deep learning model, structural elements were extracted from simplified diagrams, and segmentation accuracy was further enhanced using a region-based refinement algorithm. Experimental results demonstrated an average segmentation accuracy of 97.0% with DeepLabv3+ and 97.9% after applying the refinement algorithm, confirming the effectiveness of the approach. The findings of this study are expected to contribute to the three-dimensional modeling and efficient management of existing bridges.