nagohachi/japanese-str-dataset-v1

---
license: cc-by-4.0
task_categories:
- image-to-text
language:
- ja
tags:
- ocr
- webdataset
size_categories:
- 1M<n<10M
---

# STR Dataset

Japanese STR (Scene Text Recognition) dataset in WebDataset format.

This dataset is composed of:

- Images of Japanese named entities (full names and their affiliations)
- Images of sentences retrieved from Aozora Bunko (青空文庫)

and their corresponding ground truth texts.

All images are synthesized using [TRDG](https://github.com/Belval/TextRecognitionDataGenerator).

## Dataset Structure

| Split | Samples | Shards |
|-------|---------|--------|
| train | 10,000,000 | 1000 |
| valid | 50,000 | 5 |
| test | 50,000 | 5 |
| **Total** | **10,100,000** | **1010** |

## Usage

```python
import webdataset as wds

base_url = "https://huggingface.co/datasets/nagohachi/japanese-str-dataset-v1/resolve/main"

# Load train split
train_dataset = (
wds.WebDataset(base_url + "/train/train-{00000..00999}.tar")
.decode("pil")
.to_tuple("png", "txt")
)

for image, text in train_dataset:
# image: PIL Image
# text: str
pass
```

## Format

Each sample contains:
- `png`: Image file (PNG format)
- `txt`: Ground truth text