Photo by ThisisEngineering on Unsplash

CadIntel: Deriving Insights From Engineering Drawings

Problem Statement

Engineering drawings are crucial for manufacturing and quality assurance, as they contain critical information such as part specifications, dimensions, and material details. However, extracting this information in a structured, machine-readable format remains a major challenge due to:

• Overlapping of textual and graphical elements (e.g., dimensions, notes, and symbols)
• Variability in formats across drawings, fonts, and styles
• Poor performance of traditional OCR tools on noisy or complex images

This manual data extraction process is both time-consuming and prone to errors, resulting in inefficiencies in workflows and increased operational costs.

The manufacturing industry faces a staggering reality: manual engineering drawing processing is quietly draining resources at an unprecedented scale. Research indicates that over 75% of manufacturing organizations still rely on manual methods to extract GD&T information from drawings, creating a bottleneck that reverberates across entire operations. [1]

Manual data extraction from a single complex engineering drawing can take multiple days for intricate parts, with quality-related costs reaching up to 20% of total revenue for manufacturers. When multiplied across thousands of drawings processed annually, this represents millions of dollars in hidden operational expenses. The 2021 Level Research Payables Insight Report found that manual processing costs average $15.97 per document, meaning a manufacturer processing 12,000 drawings annually burns through nearly $192,000 in manual processing costs alone [2]

The inefficiencies extend far beyond direct processing costs. Manual GD&T extraction creates a cascade of operational challenges that impact every department:

1. Manufacturing Delays and Quality Issues : When dimensional data is incorrectly extracted or misinterpreted, it leads to costly rework, scrap parts, and production delays. Studies show that design-related errors can account for up to 5% of total project costs in manufacturing, while bottlenecks in production can be responsible for up to 30% of throughput losses.[3]
2. Procurement inefficiencies : Sourcing teams struggle to make informed decisions without structured data from engineering drawings. This results in longer quote generation times, missed cost optimization opportunities, and delayed supplier negotiations. One client case study revealed a 90% reduction in quote generation time after implementing automated drawing data extraction [4]
3. Limitations for sales teams : Sales engineers often need to manually interpret complex drawings to provide accurate quotes and technical specifications to customers. This manual process can extend sales cycles significantly, with some organizations reporting 15% faster time-to-market after implementing automated CAD data processing

Many organizations attempt to solve this challenge with traditional OCR solutions, only to discover that engineering drawings present unique challenges that generic OCR cannot address. Traditional OCR tools struggle with:

Complex Visual Layouts: Engineering drawings contain overlapping text, symbols, and graphical elements that confuse standard OCR algorithms. The fragmented text and complex data formats commonly found in GD&T specifications require contextual understanding that traditional OCR lacks

Technical Symbol Recognition: GD&T symbols, tolerances, and engineering-specific notations require specialized recognition capabilities that go beyond simple character recognition. Research shows that traditional OCR has an error rate of around 1% for standard documents, but this rate increases significantly when processing complex technical drawings.

Format Variability: The vast variability in drawing formats, fonts, and styles across different CAD systems and historical documents creates additional complexity that generic OCR solutions cannot handle effectively

Generated using AI

AI to the Rescue

At YlogX we applied our experience and expertise in having incorporated AI based solutions to solve business problems across domains built a solution accelerator — CadIntel that automates this process and improves data reliability. The end-users include engineering teams, quality analysts, and manufacturing units that rely heavily on accurate interpretation of CAD documentation.

Our Approach

To tackle the complexities of extracting structured information from engineering drawings, we designed a modular three-pipeline system using a blend of image processing, OCR, and AI techniques:

Text Extraction Pipeline

• Applied pre-processing techniques like noise removal, contrast enhancement, and line elimination.
• Used Tesseract OCR for baseline text recognition.
• Integrated a local Large Language Model (LLM) from Ollama to intelligently structure and refine extracted text.
• Extracted key metadata such as drawing number, title, engineer name, approval dates, and notes.

Table Extraction Pipeline

• Employed image-based algorithms to detect and isolate tabular regions.
• Extracted table bounding boxes and passed them downstream for exclusion from other pipelines.
• Handled complex structured data that conventional OCR fails to capture accurately.

Engineering Drawing Extraction Pipeline

• Focused on isolating core technical illustrations from text and tables.
• Used contour detection and shape analysis to clean and segment graphical components.
• Removed detected tables based on bounding box info from the table pipeline.
• Applied geometric pattern recognition to detect symbols, annotations, and GD&T elements.

Generated using AI

Tech Stack

• Python — Core programming language for backend pipelines
• OpenCV — For image pre-processing, contour detection, and shape analysis
• Paddle OCR — Used for baseline text extraction from drawings
• Ollama LLM — Local large language model for refining and structuring extracted text
• img2table — For accurate table detection and structured table extraction
• NumPy / Pandas — Data handling and post-processing
• FastAPI — For serving the pipeline as an API
• Postgressql — Database for search and retrieval
• Streamlit — Frontend application

Implementation Highlights

During development, several technical innovations and design strategies were implemented to ensure high accuracy and robustness in extracting structured data from complex engineering drawings. Key highlights include:

Three-Stage Modular Architecture

Independent pipelines for text, tables, and drawing content to ensure specialized processing and reduce cross-contamination of features.

Hybrid OCR + LLM Pipeline

Combined Paddle OCR with a local LLM (Ollama) to not only extract text but also structure and label it meaningfully — boosting accuracy for fields like “dwg no”, “eng”, “rvwr”.

Robust Table Detection with img2table

Img2table for precise detection and extraction of complex tabular structures, outperforming standard OCR methods on engineering layout tables.

Contour-Based Drawing Isolation

OpenCV for contour detection and shape filtering to isolate only the graphical content (symbols, designs, annotations) while excluding noise like text and tables.

Bounding Box Inter-Pipeline Communication

Smart coordination between table and drawing pipelines allowed for spatial exclusion of tables during graphical extraction, ensuring cleaner output.

Noise-Resilient Preprocessing

Applied contrast adjustment, noise removal, and line elimination to enhance image clarity before OCR and shape detection.

Results

• Achieved over 90% field-level accuracy on key metadata extraction (e.g., drawing number, engineer, revision details) using the hybrid OCR + LLM pipeline.
• Successfully extracted structured tables from the input drawings, even in cases where traditional OCR tools failed.
• Improved drawing content segmentation by removing noise from tables and text, leading to cleaner visual outputs for downstream CAD processing.
• Reduced manual annotation time, significantly accelerating the document digitization workflow.

Screenshots

What’s Next

We plan to extend CadIntel with a multimodal embedding-based search system that enables intelligent retrieval of engineering drawings using both text and image queries.

Build a hybrid vector search system by combining:

Textual embeddings (e.g., product specs, annotations) from LLMs.

Visual embeddings from engineering drawings using vision models.

Enable users to perform:

Text-based searches (e.g., “Find all drawings with specific material)

Image-based searches (e.g., upload a drawing to find similar designs)

This upgrade will significantly enhance design reuse and streamline the retrieval process unlocking more value from engineering archives.

Team Credits

Kirtisikka

Akshaya

Vinaya Vijayan

Tags

AI #ComputerVision #OCR #LLM #EngineeringDrawings #CAD #Automation #ImageProcessing #MachineLearning #POC #CaseStudy #Ylogx #Innovation #DocumentDigitization

References:

[1] https://www.qualitymag.com/articles/96128-can-gdt-extraction-really-be-automated

[2] https://icaptur.ai/cad-data-extraction-automation-for-higher-efficiency/

[3] https://www.sciencedirect.com/science/article/pii/S0360835222007896

[4] https://www.infrrd.ai/solutions/engineering-drawing-data-extraction-ai-tool