Data Augmentation of Engineering Drawings for Data-driven Component Segmentation

Oct 23, 2026

Upload by Finn

Data-driven component segmentation of raster engineering drawings is targeted through an automatic pipeline that improves inspection efficiency. The approach vectorizes B&W raster drawings and uses a synthesis method to construct a large labelled dataset from existing DXF examples. It trains a model to predict the type of each vectorized component and deploys on raster inputs. A key requirement is generating arbitrarily large synthetic drawings while enforcing validity of engineering technical rules, avoiding invalid data under domain constraints.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages26

Reading Time1 h 6 min

AccessFree to download

Overview

Ultimate goal and automatic system

Motivation

Industrial part quotation systems

Notation and time-consuming steps

Literature Review: Engineering Drawing Analysis

Requirements

Data synthesis from labeled examples

Validity under technical rules

Related Work

General and simulated data augmentation in CV

Algorithm Overview

From vector drawings to training and deployment

Approach Overview

Four major modules: parse, synthesize, vectorize, predict

Data Augmentation of Engineering Drawings for Data-driven Component Segmentation Wentai Zhang, Quan Chen, Can Koz, Louise Xie, Amit Regmi, Soji Yamakawa, Tomotake Furuhata, Kenji Shimada, Levent Burak Kara Carnegie Mellon University 1 Overview Ultimate goal: Data-driven component segmentation of raster drawings An automatic system B&W raster drawing Vectocontozres,ddsegmentaimensionionandamontextsg 2 Motivation: Industrial Part Quotation Systems Drawing Format When an order is placed 16% 84% A survey on the project and issues in Japan’s manufacturing industry, 2017 Average time for a part quote: 7 days 3 Motivation: Industrial Part Quotation Systems Notation Part drawings Metadata This step is time-consuming and dull. We aim at building an automatic system to aid the inspector. 4 Literature Review: Engineering Drawing Analysis Prior works mainly focus on pattern recognition, shape identiﬁcation and drawing retrieval. Diagram recognition in ﬂoor plans, ﬂow charts and electric circuit diagrams and vibratory mechanical systems [Delalandre [et. al](et. al), 2010, [Kara et. al. 2008](Kara et. al. 2008), [Schafer et. al. 2021](Schafer et. al. 2021)] Shape identiﬁcation for part drawings with sampled points, histograms or shape descriptors. [[Liu et. al. 2009](Liu et. al. 2009), [Huet et. al. 2001](Huet et. al. 2001)] Drawing retrieval or matching using lines, pixel blocks or patches.[[Mednonogov et. al. 2000](Mednonogov et. al. 2000), [Jiao et. al](Jiao et. al). 2009, [Sousa et. al. 2010](Sousa et. al. 2010)] Only achieve a partial interpretation of the drawings for a speciﬁc scenario. We aim at developing a data-driven system to analyze all the components for general analysis. 5 Overview Ultimate goal: Data-driven component segmentation of raster drawings An automatic system with: 1. Avectorization preprocessing 2. A synthesis method to automatically construct a large labelled dataset 3. A data-driven model that predicts the type of each vectorized component raster drawing B&W Vectorized segmentation 6 Requirements A data synthesis method that: (1) Utilize the information stored in existing labelled examples (2) Generate an arbitrarily large set of synthetic drawings to train a data-driven model for binary component segmentation (contour shape/dimension set) (3) The generated drawings are subjected to validity of technical rules 7 Related work General Data Augmentation Methods in CV Simulated Data Augmentation Basic geometric transformation, ﬁltering, random erasing and mixing. [[Kang et. al. 2017](Kang et. al. 2017), [Zhong et. al. 2020](Zhong et. al. 2020), Chatﬁ[eld et. al. 2014](eld et. al. 2014), Inoue 2018] Carla[[Alexey et. al. 2017](Alexey et. al. 2017)], Udacity[molyakov [et. al. 2018](et. al. 2018)], Kuka[Lukaˇ[c et. al. 2018](c et. al. 2018)] Most of these manipulations will result in invalid image data in the context of engineering drawings The simulated data can be generated with ﬂexible experiment conditions in a reasonably short time We aim to create a parametric drawing generator that can synthesize a pool of new drawings in a simulated manner with a handful of existing drawing examples 8 Algorithm Overview The labeling for such drawings requires humans with technical training. But vector drawings can serve to create a dataset for training a model that deploys on raster images. Data-driven Model DXF drawings Synthesize and Convert to raster images Training Raster images Vectorized segmentation Deployment 9 Approach Overview Four major modules to parse existing data, generate new data, vectorize the drawing, and predict the component type Avector drawing (DXF) JSON Metadata Synthesis Image Vectorized Features Type Prediction (To prepare training raster drawings) 10

FAQs

What is the ultimate goal of the proposed system?

The system aims at data-driven component segmentation of raster engineering drawings, enabling automated analysis of all components rather than partial interpretation for a specific scenario.

How does the pipeline prepare training data?

It uses vectorized drawings (DXF) as a source, synthesizes additional drawings, converts them to raster images, and then trains a model to perform vectorized segmentation and component type prediction.

Why is data synthesis required to follow technical validity rules?

Because many generic image augmentations would produce invalid engineering-drawing outputs, the generated synthetic drawings must be checked against technical rules to ensure they remain valid for domain learning.

Oct 23, 2026

Upload by Finn