Reconstructing building information models (BIMs) based on 2D drawings is an effective way to realize digital management of existing buildings. However, current image-based methods require much time as well as professional knowledge to manually design and extract features from drawing images. Moreover, the quality of generated BIM cannot be guaranteed when dealing with the drawings drawn under different design standards and drawing conventions. In this study, a novel hybrid method, integrating technologies like image processing, deep learning and optical character recognition (OCR), is proposed to extract the information of objects from the images of structural drawings (i.e. grids, columns and beams), and generate industry foundation classes (IFC) BIM for existing buildings. Experiments are carried out to verify the performance of the proposed method and the results demonstrate the feasibility and reliability of the method. Several accuracy-influencing factors are also analyzed and discussed in this paper.

基于2D图纸重建建筑物信息模型（BIM）是实现现有建筑物数字化管理的有效方法。但是，当前基于图像的方法需要大量时间和专业知识，才能从图纸图像中手动设计和提取特征。此外，在处理根据不同设计标准和图纸惯例绘制的图纸时，不能保证生成的BIM的质量。在这项研究中，提出了一种新颖的混合方法，该方法融合了图像处理，深度学习和光学字符识别（OCR）等技术，可以从结构图的图像（即网格，列和梁）中提取对象的信息，并生成现有建筑物的行业基础类（IFC）BIM。通过实验验证了该方法的有效性，结果证明了该方法的可行性和可靠性。本文还分析和讨论了几个精度影响因素。