Engineering Drawing Segmentation and Understanding For Building Construction Digital Twins

<p>Many legacy buildings lack Building Information Models (BIMs), even though they have detailed engineering drawings that contain essential mechanical, electrical, plumbing (MEP) components and structural components. Converting these drawings into BIM models remains a largely manual process, which is time-consuming, difficult to scale, and costly. This thesis presents a computer vision–based pipeline designed to automate two fundamental steps necessary for downstream BIM reconstruction from engineering drawings: legend extraction and engineering drawing preprocessing and noise removal. The legend extraction module combines histogram segmentation with Paddle OCR, enabling accurate extraction across heterogeneous MEP legend images. The engineering drawing preprocessing and noise removal module removes background structural components and employs a multi-stage approach for detecting and removing annotation arrows along with annotations, which commonly obscure MEP components. A sequence of morphological filtering and geometric analysis enables robust annotation arrows detection and removal, producing cleaner drawings for future analysis. Experimental results on engineering drawings of multiple real-world construction projects show strong legend segmentation accuracy, reliable annotation arrows detection, and consistent preprocessing performance across diverse engineering drawing layouts. The proposed pipeline establishes essential preprocessing and extraction capabilities that significantly reduce manual effort and provide a foundation for developing automated BIM models.</p>

Name thesis:degree_name

Master of Science (M.S.)

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Year dc:date.available

2025

Author dc:creator

• Chen, Suyang

Contributors dc:contributor

• Jin Chen
• Zhigang Zhu

• BIM Reconstruction
• Engineering Drawing
• Image Processing
• Computer Vision.
• Computer Sciences

Repository record dc:identifier

https://academicworks.cuny.edu/cc_etds_theses/1240

OAI identifier oai:identifier

oai:academicworks.cuny.edu:cc_etds_theses-2395