Mkz-Prog/gost-ru-technical-ocr-dataset

Dataset Card for Dataset Name

Synthetic OCR dataset of Russian technical drawings text generated with TextRecognitionDataGenerator.

Dataset Details

Dataset Description

Texts include engineering terms, GOST standards, dimensions, specifications, and full Cyrillic alphabet coverage.

Uses

This synthetic OCR dataset is specifically designed for training and evaluating text recognition models on technical documentation, with a strong focus on Russian language and the unique characteristics of GOST-standardized documents.

Direct Use

Training OCR Models: Ideal for developing robust OCR systems capable of accurately reading complex technical drawings, blueprints, and engineering specifications. Language & Font Specificity: Perfectly suited for models that need to handle Russian text (Cyrillic) and the distinctive GOST font, which is commonly used in industrial and construction documentation. Real-World Scenario Simulation: The dataset includes realistic variations such as:

• Mixed alphanumeric strings (10500.31230.123СБ, 1:10)
• Complex symbols and punctuation (70x5=350, п.7 TT)

Benchmarking: Useful for comparing the performance of different OCR algorithms when processing technical content.

Advanced Use Cases

Building Specialized Datasets: Can serve as a foundation for creating larger, more diverse datasets by combining it with real-world data. Model Robustness Testing: The controlled generation process allows for systematic testing of how well a model handles specific distortions (skew, blur, noise). Research in Document Analysis: Provides a standardized resource for research into automated document processing, particularly in engineering and manufacturing domains.

Dataset Structure

The dataset is organized into two main components:

1. Images (images/)

Total Count: 1,000 images. File Formats: A mix of JPEG and PNG formats to simulate real-world document storage and exchange scenarios. Image Size: Automatically resized to a standard resolution of approximately 640x480 pixels (with minor variations). Content: Each image contains one or more lines of text generated using the GOST.TTF font. The text represents realistic engineering and technical documentation, including:

• Part numbers and codes (e.g., 10500.31230.123СБ)
• Technical terms (e.g., "Каркас площадки", "Сборочный чертеж")
• Measurements and scales (e.g., 16,0, 1:10, 70x5=350)
• Special symbols and punctuation (e.g., п.7 TT, * Размеры для справок.)

Visual Properties: Images are synthetically generated with controlled variations to enhance realism:

• Background: Gaussian noise resembling a light white background.
• Distortions: Subtle random skewing (up to 5°), minimal sinusoidal distortion, and slight random blurring.
• Alignment: Text is centered horizontally.

2. Labels (labels.csv)

Format: A plain CSV file with comma as the delimiter. Structure:

• filename: A string representing the name of the corresponding image file (e.g., image_0001.jpg).
• words: A string containing the complete text found on that image, preserving all original characters, spaces, and punctuation (e.g., "10500.31230.123СБ").