n-tan/engineering-drawing-intelligence

Engineering Drawing Intelligence (EDI)

A high-performance Computer Vision pipeline designed to extract structured data from complex engineering drawings and blueprints. This system leverages RT-DETR for real-time precision object detection (targeting Tables and Notes) combined with a robust PaddleOCR / EasyOCR multi-method strategy for text extraction.

✨ Key Features

• —Precision Detection: Accurate localization of technical components using state-of-the-art RT-DETR architecture.
• —Hybrid OCR Engine: Modular extraction pipeline optimized for engineering fonts and technical notation.
• —Structured JSON Output: Automatic conversion of visual data into machine-readable formats.
• —Modern Workflow: Fully managed with uv for reproducible and fast environment setup.

🛠 Tech Stack

• —Detection: Ultralytics RT-DETR
• —OCR: PaddleOCR, EasyOCR
• —Processing: OpenCV, NumPy, PyTorch
• —Tooling: Python 3.10+, uv Package Manager

🚀 Quick Start

1. Environment Setup

This project uses uv for high-performance dependency management. If you don't have it, install it here.

# Sync dependencies and create virtual environment
uv sync

# Activate the environment
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

2. Model Weights

The core detection model relies on specialized weights.

• —Download: You can download the trained best.pt file from **Google Drive**.
• —Placement: Ensure the downloaded best.pt file is placed inside the weight/ directory.

Note: These files are typically excluded from Git via .gitignore. Verify their presence before running the pipeline.

3. Execution

To run the complete inference pipeline on input images:

python full_pipeline.py

Results are saved by default in the pipeline_results/ directory:

• —pipeline_output.json: The complete structured dataset.
• —table_*_crop.jpg: Individual crops for visual verification of table extraction.

📁 Project Structure

├── app/                 # Core processing logic
│   ├── detection.py     # Object Detection (RT-DETR)
│   ├── ocr.py           # Pre-processing & OCR (Paddle/EasyOCR)
│   └── config.py        # Configuration & Hyperparameters
├── weight/              # Directory for model weights (best.pt)
├── data/                # Input image directory
├── full_pipeline.py     # Main coordination script
├── pyproject.toml       # Dependency definitions
└── README.md            # Project documentation

📈 Model Training

To retrain the RT-DETR model on a new dataset:

from ultralytics import RTDETR

# Initialize model
model = RTDETR("rtdetr-l.pt")  # or load your existing checkpoint

# Start training
model.train(data="your_dataset/data.yaml", epochs=100, imgsz=640)

📝 Performance Notes

• —OpenCV & PyTorch: Ensure your system environment supports these libraries.
• —Hardware: While it runs on CPU, GPU acceleration (CUDA) is strongly recommended for real-time OCR and detection performance.