Synthetic OCR Dataset Generation for Mechanical Floor Plans

Overview

This repository provides an automated pipeline for generating a high-volume synthetic dataset tailored for optical character recognition (OCR) in mechanical floor plan contexts. The dataset includes:

• Alphanumeric tag strings in various styles.
• Rotated (sideways) text to mimic text alignment in plans.
• Underlined text to address model biases.
• Text on visually noisy backgrounds matching floor plan diagrams.
• Text within simple icons to simulate equipment tagging.

Sample Outputs

Below are example images (reduced-size for illustration) depicting the kind of synthetic data generated:

Each image has an accompanying annotation file containing text ground truth and bounding box coordinates.

These synthetic images are accompanied by JSON annotations containing:

1. Image path
2. Bounding box coordinates
3. OCR text (ground truth)
4. Rotation angle
5. Shape type (if any)

Features

1. Realistic Background Patterns
   Grid, dots, crosses, or diagonal lines are used to simulate cluttered or technical drawings.

2. Scalable Generation
   With default settings on a single CPU core, approximately 2,000 samples can be generated per minute—enough to exceed 3 million samples in 24 hours, ensuring coverage of more than 500k samples well within a single day.

3. Alphanumeric & Symbolic Text
   Generations include customizable separators ( -, _, ., etc. ), varied lengths, and random fonts to mimic mechanical plan labels.

4. Rotation and Underlines
   Text may appear at 0°, 90°, or 270° orientation and often comes with optional underlines, as is common in floor plan notations.

5. Icons & Shapes
   Simple geometric shapes (rectangles, triangles, circles, etc.) are drawn behind or around text to replicate equipment tags and icons often found on mechanical drawings.

Installation and Requirements

1. Clone or download this repository.
2. Install Python 3.8 or higher.
3. Install dependencies:

pip install -r requirements.txt

4. Fonts
   Place any .ttf or .otf fonts in a fonts/ directory. The code references fonts/arial.ttf, fonts/times.ttf, etc. Adjust these paths or add alternative fonts as desired.

Usage

Generating the Synthetic Dataset

1. Command-line Arguments

python generate_synthetic_data.py --num-samples 500000 --output-dir synthetic_dataset --workers 4

• --num-samples: Total number of images to create.
• --output-dir: Output directory for images and annotations.
• --workers: Number of parallel processes for faster generation.

2. Configurable Parameters
   Update the global constants in generate_synthetic_data.py as needed:

   • IMAGE_SIZE: Resolution of the synthetic images.
   • TEXT_SIZE_RANGE: Range of text font sizes.
   • ROTATIONS: Possible text rotations (e.g., [0, 90, 270]).
   • UNDERLINE_PROB: Probability of underlining text.
   • PATTERNS: Types of background patterns to draw.
   • SHAPES: Shapes to simulate icons.

3. Performance
   With default settings, approximately 2,000 samples per minute can be generated on a single CPU core. Parallelizing with multiple cores increases throughput substantially.

Evaluating with an OCR Model

1. Run the Evaluation

python evaluate_ocr.py

• Loads each generated PNG and its corresponding JSON annotation.
• Applies a configured OCR model to extract text and bounding boxes.
• Computes metrics:
  • Character Error Rate (CER)
  • Word Error Rate (WER)
  • Intersection over Union (IoU) for bounding boxes
• Saves results to an ocr_results/ directory.

2. OCR Model Integration
   The example script references placeholders (surya.ocr, surya.model.detection.segformer, surya.model.recognition.model). Adjust these to integrate with any custom OCR solution (e.g., Tesseract, EasyOCR, or a proprietary model).

3. Metrics

   • CER: Measures character-level accuracy.
   • WER: Evaluates word-level accuracy.
   • IoU: Compares ground-truth bounding boxes and detected bounding boxes.

Notes

• The dataset intentionally focuses on textual labels and icons found within floor plans.
• Mechanical detail diagrams (complex engineering drawings) are out of scope; only text-centric elements are generated.
• The generation approach is highly configurable, enabling diverse text scenarios, rotation angles, background patterns, and shapes to rigorously train in-domain OCR models.

Contributing

Contributions are welcome. For bug reports or feature requests, open an issue or submit a pull request with a clear explanation of proposed changes.