NNEngine/Captcha-OCR

---
license: mit
task_categories:
- text-classification
language:
- en
size_categories:
- 1M<n<10M
---

# Synthetic CAPTCHA OCR Dataset (1M)

## Overview
This dataset contains **synthetically generated CAPTCHA images** designed for training and benchmarking Optical Character Recognition (OCR) models. Each image contains a randomly generated alphanumeric string rendered in CAPTCHA style with noise, distortions, and visual artifacts to simulate real-world conditions.

The dataset is created entirely using automated rendering pipelines and therefore contains perfectly accurate ground-truth labels.

---

## Dataset Characteristics

- **Dataset size:** 1,000,000 images
- **Image format:** PNG
- **Image resolution:** 160 × 60 pixels
- **Text length:** 5–10 characters
- **Character set:**
- Uppercase letters (A–Z)
- Lowercase letters (a–z)
- Digits (0–9)

Each file is named using the ground-truth label:

```

<text>.png

```

Example:

```

A7kD3.png
pQ82Lm.png

```

Thus, labels can be directly extracted from filenames without requiring an additional annotation file.

## Generation Methodology

Images were generated using a synthetic rendering pipeline that includes:

- Random font selection
- Character position perturbations
- Random background noise
- Random line interference
- Gaussian pixel noise

This process improves robustness and helps OCR models generalize to real-world CAPTCHA images.

## Intended Use

This dataset is suitable for:

- Training deep learning OCR systems
- CAPTCHA recognition research
- Sequence recognition benchmarking
- Synthetic data pretraining for document OCR systems
- Curriculum learning before fine-tuning on real-world datasets

## Limitations

- Images are synthetically generated and may not capture every real-world CAPTCHA style.
- Domain adaptation may still be required for specific CAPTCHA systems.
- Distribution of character sequences is random rather than language-based.

## Citation

If you use this dataset in academic work, please cite:

```

Synthetic CAPTCHA OCR Dataset (1M), 2026

```