yezay/myanmar-ocr-dataset

---
dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
splits:
- name: train
num_bytes: 27613382844.995
num_examples: 7576785
- name: test
num_bytes: 74400308.776
num_examples: 22964
download_size: 24805635073
dataset_size: 27687783153.771
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
language:
- my
pretty_name: Myanmar OCR Dataset (Synthetic)
license: mit
task_categories:
- image-to-text
---

# Myanmar OCR Dataset

A synthetic dataset for training and fine-tuning Optical Character Recognition (OCR) models specifically for the Myanmar language.

## Dataset Description

This dataset contains synthetically generated OCR images created specifically for Myanmar text recognition tasks. The images were generated using [myanmar-ocr-data-generator](https://github.com/chuuhtetnaing/myanmar-ocr-data-generator), a fork of [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) with fixes for proper Myanmar character splitting.

## Direct Download

Available from [UT Austin SharePoint](https://utexas-my.sharepoint.com/:u:/g/personal/cn23349_my_utexas_edu/EZcWC2NEbKNMkeJ_WByugx0BMTSE6qKYQjVp62fn6m0toA?e=L9AKko)
- The downloadable data is specifically formatted for PaddleOCR
- Includes organized folder structure optimized for PaddleOCR training
- Labels are provided in tab-separated format
- Ready to fine-tuned for PaddleOCR

### Key Features

- Synthetic Myanmar text images with corresponding text labels
- Maximum text length of 100 characters per image
- Various fonts, styles, skews, various distortion types, blur levels, and backgrounds to improve model robustness

## Generation Method

The dataset was created using a forked and customized version of the TextRecognitionDataGenerator:
- Original repository: [Belval/TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator)
- Customized fork: [chuuhtetnaing/myanmar-ocr-data-generator](https://github.com/chuuhtetnaing/myanmar-ocr-data-generator)

### Text Sources

The text content used to generate the OCR images was extracted from:
- [chuuhtetnaing/myanmar-wikipedia-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-wikipedia-dataset): A collection of Myanmar language text from Wikipedia
- [chuuhtetnaing/myanmar-facebook-flores-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-facebook-flores-dataset): A subset of [Meta's FLORES dataset](https://huggingface.co/datasets/facebook/flores) containing only Myanmar language text
- [chuuhtetnaing/myanmar-instruction-tuning-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-instruction-tuning-dataset): Compilation of question-answer pairs in the Myanmar language
- [chuuhtetnaing/mm-lib-book-dataset](https://huggingface.co/datasets/chuuhtetnaing/mm-lib-book-dataset): A dataset of 437 books extracted from MM-Lib (Only 421 unicode font books are used for OCR dataset)

## Usage

This dataset is particularly useful for:

- Training OCR models for Myanmar language text recognition
- Fine-tuning existing OCR models to improve performance on Myanmar script
- Developing applications for Myanmar text digitization

## License

This dataset is free to use for any purpose - personal, commercial, or educational.

dataset_info:
特征：
- 名称：image（图像），数据类型：image（图像）
- 名称：text（文本），数据类型：string（字符串）
划分：
- 名称：train（训练集），字节数：27613382844.995，样本数：7576785
- 名称：test（测试集），字节数：74400308.776，样本数：22964
下载大小：24805635073，数据集总大小：27687783153.771
配置项：
- 配置名称：default（默认配置），数据文件：
- 划分：train，路径：data/train-*
- 划分：test，路径：data/test-*
语言：
- my（缅甸语）
展示名称：缅甸语OCR合成数据集（Myanmar OCR Dataset (Synthetic)）
许可证：MIT许可证
任务类别：
- 图像到文本（image-to-text）

# 缅甸语OCR数据集
本数据集为专门针对缅甸语的光学字符识别（Optical Character Recognition, OCR）模型的训练与微调任务构建的合成数据集。

## 数据集说明
本数据集包含专为缅甸语文本识别任务生成的合成OCR图像。这些图像通过[myanmar-ocr-data-generator](https://github.com/chuuhtetnaing/myanmar-ocr-data-generator)生成，该项目是[TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator)的分支版本，针对缅甸语字符拆分问题进行了修复。

## 直接下载
可从[德克萨斯大学奥斯汀分校SharePoint](https://utexas-my.sharepoint.com/:u:/g/personal/cn23349_my_utexas_edu/EZcWC2NEbKNMkeJ_WByugx0BMTSE6qKYQjVp62fn6m0toA?e=L9AKko)获取：
- 该可下载数据专门针对PaddleOCR优化格式
- 包含适配PaddleOCR训练的标准化文件夹结构
- 标签采用制表符分隔格式提供
- 可直接用于PaddleOCR的微调任务

### 核心特性
- 带有对应文本标签的合成缅甸语文本图像
- 单张图像最大文本长度为100个字符
- 涵盖多种字体、样式、倾斜角度、各类畸变类型、模糊等级与背景，以提升模型鲁棒性

## 生成方式
本数据集基于经过定制修改的TextRecognitionDataGenerator分支版本构建：
- 原始仓库：[Belval/TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator)
- 定制分支：[chuuhtetnaing/myanmar-ocr-data-generator](https://github.com/chuuhtetnaing/myanmar-ocr-data-generator)

### 文本来源
用于生成OCR图像的文本内容提取自以下数据集：
- [chuuhtetnaing/myanmar-wikipedia-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-wikipedia-dataset)：来自维基百科的缅甸语文本合集
- [chuuhtetnaing/myanmar-facebook-flores-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-facebook-flores-dataset)：[Meta's FLORES数据集](https://huggingface.co/datasets/facebook/flores)的子集，仅包含缅甸语文本
- [chuuhtetnaing/myanmar-instruction-tuning-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-instruction-tuning-dataset)：缅甸语问答对合集
- [chuuhtetnaing/mm-lib-book-dataset](https://huggingface.co/datasets/chuuhtetnaing/mm-lib-book-dataset)：从MM-Lib提取的437本图书数据集（本OCR数据集仅使用其中421本支持Unicode字体的图书）

## 应用场景
本数据集尤其适用于：
- 训练用于缅甸语文本识别的OCR模型
- 微调现有OCR模型以提升其对缅甸文字符的识别性能
- 开发缅甸语文本数字化相关应用

## 许可证
本数据集可免费用于个人、商业或教育等任意场景。