TROCR engineering drawinf training dataset

This dataset consists of synthetically generated images containing text and symbols commonly found in engineering drawings, blueprints, and CAD schematics. It was created to train Optical Character Recognition (OCR) models (specifically Transformer-based models like TrOCR) to recognize technical dimensioning standards under various orientations and noise conditions.

2. Dataset Content

The text labels are procedurally generated to simulate real-world engineering annotations. The vocabulary includes:

Linear Dimensions: Decimal numbers (e.g., 12.5, 450).

Geometric Symbols: Diameter (Ø, Φ) and Radius (R).

Tolerances: Explicit tolerance values (e.g., ±0.05, 10.0±0.2).

Angles: Degree measurements (e.g., 45°, 90°).

Thread Specifications: Metric thread callouts (e.g., M8×1.25).

Multipliers: Count indicators (e.g., 4×12.5).

3. Image Specifications

Resolution: 512 × 128 pixels.

Format: PNG images.

Color Space: RGB (though visually grayscale/monochrome).

Background: Randomized light gray (Pixel values: 235–255).

Text Color: Black.

Fonts Used:

DejaVu Sans

Liberation Sans (Regular)

Liberation Mono (Regular)

4. Augmentations & Variations

To ensure robustness, the dataset includes heavy augmentations applied via GPU processing:

Rotation: Text is not just horizontal. Images are rotated at specific angular increments: 0°, 15°, 30°, 45°, 60°, 75°, 90°, 105°, 135°. This simulates the multi-directional text found in technical drawings.

Gaussian Noise: Applied to 70% of the images (σ=10) to simulate scan grain.

Gaussian Blur: Applied to 30% of the images to simulate low-resolution scans or printing defects.

5. Dataset Structure & Statistics

Total Samples: 2,000 image-text pairs.

Split:

Training Set: 1,800 samples (90%)

Validation Set: 200 samples (10%)

File Organization:

dataset/train/images/: Images named img_00000.png.

dataset/train/labels/: Corresponding text files img_00000.txt.

tokenizer/: Custom RoBERTa-compatible tokenizer files (vocab.json, merges.txt) generated from the specific dataset vocabulary.