Datasets:
Tasks:
Image Classification
Formats:
parquet
Sub-tasks:
multi-class-image-classification
Languages:
Japanese
Size:
10K - 100K
License:
language: | |
- ja | |
multilinguality: monolingual | |
task_categories: | |
- image-classification | |
task_ids: | |
- multi-class-image-classification | |
license: apache-2.0 | |
pretty_name: JPCharRecog_v1.1 | |
size_categories: | |
- 10K<n<100K | |
tags: | |
- ocr | |
- japanese | |
- cjk | |
- synthetic | |
- font | |
- rendered | |
- single-character | |
- document-ai | |
# JPCharRecog_v1.1 | |
JPCharRecog v1.1 は、日本語の単文字(CJK を含む)認識を目的とした、合成レンダリング画像のデータセットです。純視覚要因(フォントファミリ/ウェイト/文字サイズ/劣化)を操作可能な設計で、辞書や文脈の寄与を切り分けるためのベンチマーク用途を想定しています。 | |
- Task: Single-character recognition (JPCharRecog-like) | |
- Language: Japanese (CJK) | |
- Families: NotoSansJP, NotoSerifJP | |
- Weights: ExtraLight, Regular, Black | |
- Sizes(px): 12, 64, 96 | |
- Effects: none (PNG) | |
## Purpose | |
- 語彙や文脈による補正を排し、モデルの純粋な「読取り」能力を単文字で測るための基礎ベンチです。 | |
- フォント設計や最小線幅、エイリアシングの影響を、ウェイトやサイズの系統操作で定量化します。 | |
- OCR 系の上限線と、多モーダル生成モデルなどの一般モデルの読み取り限界を同一条件で比較可能にします。 | |
## Expected Experiments | |
- 閾値カーブ: 正答率や ANLS をサイズやウェイト、ファミリに対して描き、視覚的限界を同定。 | |
- モデルスケール比較: 小中大の多モーダルモデルや OCR ベースラインと並走比較。 | |
- フォント交差一般化: ファミリ間での頑健性(例: Sans で調整したプロンプトが Serif に転移するか)。 | |
- 文字種分析: ひらがな/カタカナ/漢字/英数/記号での難易度差、出力正規化の影響(半角全角、類似字形)。 | |
- 劣化耐性(将来拡張): blur や jpeg 劣化条件の段階化と用量–反応関係の確認。 | |
推奨プロトコル例(生成モデルの評価): | |
1. プロンプトは「画像に写っている1文字のみを出力。説明や記号は付けない。」のように拘束を明記。 | |
2. 出力を正規化(前後空白除去、全角半角の統一、大小統一など)後に比較。 | |
3. 指標は Exact Match に加えて ANLS を併記(単文字では二値に近いが、将来の劣化条件で頑健)。 | |
## Structure | |
- Split: test | |
- Features: image(HFImage), character, unicode_hex, char_type, font_family, font_weight, font_px, writing, effect, id | |
## Generation | |
- Canvas: 224×224, white background, black glyphs, centered | |
- Fonts: Noto Sans JP / Noto Serif JP(Google Noto; OFL) | |
- Deterministic seed=42; per-character split; stable ids | |
## Statistics | |
- Total samples: 18,000(Sans/Serif 各 9,000) | |
- Families: NotoSansJP, NotoSerifJP × Weights: ExtraLight, Regular, Black × Sizes(px): 12, 64, 96 | |
- Split: test | |
## Usage | |
```python | |
from datasets import load_dataset | |
ds = load_dataset("Silviase/JPCharRecog_v1.1") | |
test = ds["test"] | |
row = test[0] | |
img = row["image"] # PIL.Image | |
label_char = row["character"] | |
``` | |
## License | |
- Dataset: Apache-2.0(本データセットの配布ライセンス) | |
- Fonts: Google Noto (OFL-1.1)。フォントの著作権表示とライセンスに従ってご利用ください。 | |
## Limitations | |
- 合成画像のため、実世界スキャン/撮影画像とは分布が異なります。 | |
- 劣化効果(blur/jpeg)は v1.1 では未使用(将来拡張)。 | |
## Citation | |
If you use this dataset, please cite this repository and Google Noto fonts: | |
``` | |
@misc{JPCharRecog_v11, | |
title = {JPCharRecog v1.1: Synthetic Japanese Single-Character Recognition Dataset}, | |
author = {Silviase}, | |
year = {2025}, | |
howpublished = {Hugging Face Datasets}, | |
note = {Families: NotoSansJP/NotoSerifJP; Weights: ExtraLight/Regular/Black; Sizes: 12/64/96} | |
} | |
``` | |