What question did this study set out to answer?

This research aims to compare various text recognition tools for extracting structured information from 2D bridge engineering drawings.

March 23, 2026Open Access

View Full Paper

Optimizing 2D Bridge Engineering Drawing Digitization: A Comparative Study of Text Recognition Tools and Development of Lightweight Post-Recognition Structured Information Extraction Methods

MPMengyan PengTechnische Universität Dresden HQHan QianChengdu Academy of Agriculture and Forestry Sciences SMSteffen MarxTechnische Universität Dresden

Key Points

This research aims to compare various text recognition tools for extracting structured information from 2D bridge engineering drawings.
Evaluation of pre-trained OCR tools and four VLMs for text recognition.
Testing with domain-specific datasets and controlled degradation factors.
Application of four metrics: Levenshtein similarity, Jaccard similarity, TF–IDF similarity, and WordNet similarity.
Development of lightweight, rule-based methods for structured information extraction.
All tools perform well on printed text, but VLMs show superior robustness across perturbations.
Handprint achieves the highest accuracy for handwritten text, particularly with rotations.
Developed methods enable extraction of structured content without prior segmentation.
The workflow supports digitization for integration into systems like BIM and digital twins.

Abstract

This paper investigates the application of text recognition technologies, spanning conventional Optical Character Recognition (OCR) and vision-language models (VLMs) in civil engineering, with a focus on the structured information extraction of textual information in two-dimensional (2D) bridge engineering drawings. Pre-trained OCR tools are evaluated for recognizing printed and handwritten text fragments extracted from bridge engineering drawings, and four VLMs are assessed as OCR-like recognizers. The evaluation employs domain-specific test datasets under controlled degradations (motion blur, partial occlusion, salt-and-pepper noise) and rotations (±45°, ±90°), and adopts four metrics: Levenshtein-based similarity, Jaccard similarity, Term Frequency–Inverse Document Frequency (TF–IDF) similarity, and WordNet-based similarity. The results demonstrate that while all tools perform well on printed text, VLMs achieve the strongest overall robustness across perturbations, and Handprint consistently achieves the highest accuracy among OCR tools, particularly for handwritten text and under rotation. To further enhance the interpretability and usability of raw recognition outputs, two lightweight, rule-based post-recognition structured information extraction methods are developed: a region-based method for extracting fixed-layout information such as title blocks, and a keyword-driven method for identifying domain-specific annotations like material specifications. These methods require no prior segmentation or additional trainable models and are integrated into a graphical user interface, TextBridge. The proposed workflow offers an effective solution for the semi-automated extraction of structured textual content from engineering drawings, with the potential to facilitate the digitization and integration of legacy infrastructure data into modern applications such as Building Information Modeling (BIM) and digital twin systems.

KI fragen

Bookmark

View Full Paper

Cite This Study

Peng et al. (Sun,) studied this question.

synapsesocial.com/papers/69c0de74fddb9876e79c1312 https://doi.org/https://doi.org/10.1016/j.rineng.2026.110186

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

KI fragen

Bookmark

View Full Paper