AUTOMATIC RECOGNITION AND SEGMENTATION OF ARCHITECTURAL ELEMENTS FROM 2D DRAWINGS BY CONVOLUTIONAL NEURAL NETWORK - Abstract

Oct 24, 2026

Upload by Sarah

The paper addresses the time-intensive BIM modeling process by exploring deep-learning-based recognition and segmentation of architectural components from 2D CAD drawings. It proposes a two-part research workflow: dataset preparation using original CAD drawings with labeled architectural components for training, followed by network training and testing where a mature CNN model predicts component classes and boundaries. The study discusses potential utilization of the results, including optimization methods. Experiments indicate effective identification and segmentation that can reduce manpower requirements.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

1. Background, related work, and meaning

Building information modeling and need

Prior research and processing frameworks

Motivations and research meanings

2. Methodology

Deep learning and CNN foundations

Network structure and training process

AUTOMATIC RECOGNITION AND SEGMENTATION OF ARCHITECTURAL ELEMENTS FROM 2D DRAWINGS BY CONVOLUTIONAL NEURAL NETWORK YAHAN XIAO 1 , SEN CHEN2 , YASUSHI IKEDA3 and KENSUKE HOTTA4 1,2,3,4Keio University. Japan [1](1yahanxiao@gmail.com 2)[yahanxiao@gmail.com](1yahanxiao@gmail.com 2)[ 2](1yahanxiao@gmail.com 2),[3](3 {cs0715|yasushi}@sfc.keio.ac.jp)[ {cs0715|yasushi}@sfc.keio.ac.jp](3 {cs0715|yasushi}@sfc.keio.ac.jp) [4](4 hottakensuke@keio.jp)[ hottakensuke@keio.jp](4 hottakensuke@keio.jp) Abstract. The BIM modeling process is the most time-consuming aspect. This paper studies the possibility of applying the recognition and segmentation of architectural components by deep learning to assist automatic BIM modeling. The research has two parts: the first one is dataset preparing, that images with the labeled architectural components from an original CAD drawing are made for the network training, and second is training and testing, that a mature network which has been trained in hundreds of labeled images is used to make predictions. The utilization of the current study results is discussed and the optimization method as well. Keywords. BIM; CAD drawings; Recognition and Segmentation; Convolutional Neural Network; Computer vision. 1. Background, related work, and meaning: China’s economic development in the past few decades has been very fast which is particularly obvious in the construction that sustainable urban development is required. While a digital model that contains all the information of a building ora project or even a city is a useful tool to achieve this goal. Therefore, building information modeling is not only needed for big new projects in core cities as it enables better regional planning, fire prevention, and disaster relief designing and so on, but also for modification projects as it contains the information like building’s structure which is the main content in renovation mission. In general, 3D modeling is a manual process (Horna, 2009) (Sigalov, 2017) whether according to the digital 2D of a new building or the remaining paper document of an old building, that consuming lots of time and labor. Then in this research, we used semantic segmentation technology in deep learning to recognize the architectural components in 2D drawings to assist automatic modeling. Among the succeed automatic modeling research, lots of researchers have developed methods which are achieved by the algorithm for generating a 3D model from 2D drawings (including CAD and raster images) (Ren, 2018) . In most cases, algorithms have been proposed and implemented in stand-alone topotype tools. RE: Anthropocene, Proceedings of the 25 th International Conference of the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA) 2020 , Volume 1, 843-852 . © 2020 and published by the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA), Hong Kong. 844 Y. XIAO ET AL. Only a few of them are accomplished within existing software that people who work in the AEC industry may already be familiar with, which still requires the collaboration of multiple tools (Lim 2018) . However, those different methods have the similar framework of their processing structure and require some extra manual preparation work (Yin 2009), like transform the recognition target into a specific format, or clean up the 2D drawings and place the 2D geometry for different building elements onto separate layers (Lim 2018) . Moreover, the common recognition target of 3D modeling is the standard 2D drawing, while the raster images are mostly for pattern, shape and topology-based space detection, which means the designing process is finished before the 3D modeling process. The meanings ofthis research are: 1) It does not need to extract additional features such as grids lines and dimensions but be able to get good results directly through repeated data learning and image training. 2) it has a strong anti-interfe

FAQs

What problem does the paper target in BIM modeling?

It targets the fact that BIM modeling is a highly time-consuming task, especially the manual effort required for 3D modeling from 2D drawings or paper documents.

How is the proposed research carried out?

The work includes dataset preparation—creating labeled images of architectural components from original CAD drawings—followed by CNN training and testing to generate predictions using a model trained on hundreds of labeled images.

What role does the convolutional neural network play?

A CNN is used for image recognition, learning features from CAD drawing data and enabling semantic segmentation so important architectural elements can be distinguished from irrelevant information.

Oct 24, 2026

Upload by Sarah