Identifier Extraction from Engineering Drawings via Object Detection and OCR - May 5, 2026

Oct 28, 2026

Engineering drawings often contain critical sensor and cable identifiers only as raster graphics inside PDF files, making manual lookup and validation slow, costly, and error-prone. This thesis develops an end-to-end pipeline that extracts those identifiers automatically using object detection and OCR. AYOLOv5m localises small symbols and applies class thresholds with non-maximum suppression, while a sufficiency analysis evaluates added annotations. Several OCR engines and bounding-box correction strategies are benchmarked, achieving high detection precision and recall on a held-out test set and enabling structured identifier storage for downstream search and traceability.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages55

Reading Time2 h 19 min

AccessFree to download

Introduction

Background and Related Work

Research Questions

Methods

Data Collection and Labelling

Tiling for Small Object Detection

Evaluation Metrics

Post-Processing

Data Sufficiency Analysis

Baseline Model

Hyperparameter Optimisation

Optical Character Recognition

OCR Problem Definition and Text Fields

Cropping and Detector-Based Preprocessing

OCR engines

Results

Object Detection Results

OCR Results

Automatic Bounding Box Quality Assessment for OCR Optimisation

Final Pipeline Results

Master’s thesis Master’s Programme in Data Science Identifier Extraction from Engineering Drawings via Object Detection and OCR Mia Lindblad May 5, 2026 University of Helsinki Faculty of Science P. O. Box 68 (Pietari Kalmin katu 5) 00014 University of Helsinki HELSINGIN YLIOPISTO—HELSINGFORS UNIVERSITET—UNIVERSITY OF HELSINKI | Tiedekunta—Fakultet—Faculty
Faculty of Science | | Koulutusohjelma—Utbildningsprogram—Degree programme
Master’s Programme in Data Science | | | --- | --- | --- | --- | | Tekijä—Författare—Author
Mia Lindblad | | | | | Työn nimi—Arbetets titel—Title
Identifier Extraction from Engineering Drawings via Object Detection and OCR | | | | | Työn laji—Arbetets art—Level
Master’s thesis | Aika—Datum—Month and year
May 5, 2026 | | Sivumäärä—Sidantal—Number of pages
49 | | Tiivistelmä—Referat—Abstract
Engineering drawings include critical sensor and cable identifiers that frequently exist only as raster graphics within PDF documents, making lookups and validations slow and error-prone. This thesis investigates whether sensor and cable information can be automatically extracted from electrical engineering drawings by combining object detection and Optical Character Recognition (OCR) to produce structured identifiers reusable in downstream systems.
The proposed solution is an end-to-end pipeline developed for internal Wärtsilä drawings. AYOLOv5m-based object detection model localises small sensor and cable symbols on high-resolution pages, and post-processing applies class-specific confidence thresholds and non-maximum suppression. A data sufficiency analysis examines how additional annotations impact class-wise precision and recall. For text extraction, several OCR engines are benchmarked on short alphanumeric IDs cropped from detected symbols: PyTesseract, Amazon Rekognition, and a multimodal GPT-based model. In addition, automatic bounding-box correction strategies are analysed. On a held-out test set, the final detector achieves a precision of 0.99 and a recall of 1.00 for both cables and sensors after post-processing, and Amazon Rekognition recovers approximately 97–98% of sensor and cable IDs from detector crops. The integrated pipeline stores extracted classes and IDs in a database table for search and traceability within internal tools.
This work is limited to data from a single company and a small test set with two imbalanced symbol classes and relies on one detector family and a single commercial OCR service without custom postcorrection. As such, reported performance is specific to this context, and transferring the pipeline to other contexts requires additional data and validation. Results show that automatic extraction of sensor and cable IDs from engineering drawings is feasible with current tools and can reduce manual effort when used as an assistive system.
ACM Computing Classification System (CCS):
Applied computing → Document management and text processing → Document capture → Optical character recognition
Computing methodologies → Artificial intelligence → Computer vision → Computer vision problems → Object detection | | | | | Avainsanat—Nyckelord—Keywords
engineering drawings, object detection, optical character recognition, information extraction | | | | | Säilytyspaikka—Förvaringsställe—Where deposited | | | | | Muita tietoja—Övriga uppgifter—Additional information | | | | Contents Acknowledgements 1 1 Introduction 3 2 Background and Related Work 5 2.1 Modern Approaches for Symbol and Text Extraction .......... 5 3 Research Questions 9 4 Methods 11 4.1 Data Collection and Labelling ....................... 12 4.2 Tiling for Small Object Detection ..................... 13 4.3 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 4.4 Post-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 4.5 Data Sufficiency Analysis . . . . . . . . . . . .

FAQs

What problem does the thesis address in electrical engineering drawings?

The thesis addresses that sensor and cable identifiers are frequently stored only as raster graphics in PDF drawings, which makes manual searching and validation slow and error-prone.

How does the proposed pipeline extract identifiers from drawings?

It first uses an AYOLOv5m object detection model to locate sensor and cable symbols, then applies OCR on cropped text areas, supported by post-processing and bounding-box correction strategies.

What performance results are reported for detection and OCR?

On a held-out test set, the final detector achieves precision 0.99 and recall 1.00 for both cable and sensor symbols after post-processing, and Amazon Rekognition recovers about 97–98% of sensor and cable IDs from detector crops.

Oct 28, 2026

Upload by Riley West