๐Ÿ‡ฏ๐Ÿ‡ต

Japanese-Specialized OCR โ€” Manga, Vertical Text, Furigana, and Historical Documents

manga-ocr, Sarashina2.2-OCR, PaddleOCR, KuroNet โ€” every option for Japanese OCR

Why Japanese OCR Is Hard

Unlike English, Japanese has: vertical text (tategaki), furigana (small reading aids), 3 scripts mixed (hiragana + katakana + kanji), manga fonts, historical cursive (kuzushiji).

Tool Comparison

Tool Use Case Vertical Size
manga-ocr Manga Yes 400MB
Sarashina2.2-OCR Documents (best) Yes 3B
PaddleOCR General CJK Yes 10-150MB
KuroNet Historical (kuzushiji) Yes Web API

manga-ocr โ€” Manga Best

from manga_ocr import MangaOcr
mocr = MangaOcr()
text = mocr("speech_bubble.png")  # โ†’ "ใŠๅ‰ใฏใ‚‚ใ†ๆญปใ‚“ใงใ„ใ‚‹"

Handles vertical/horizontal, furigana, various manga fonts, low-quality images.

KuroNet โ€” Historical Documents

The only production tool for pre-Meiji kuzushiji (cursive) โ†’ modern Japanese conversion.

Key Concepts

1

manga-ocr (โ˜…โ˜†โ˜†โ˜†โ˜†) โ€” pip install + 2 lines. Best for manga text

2

Sarashina2.2-OCR (โ˜…โ˜…โ˜…โ˜†โ˜†) โ€” 3B model. Best Japanese document accuracy. GPU needed

3

PaddleOCR (โ˜…โ˜…โ˜†โ˜†โ˜†) โ€” use_textline_orientation=True for vertical text auto-detection

4

KuroNet (โ˜…โ˜…โ˜†โ˜†โ˜†) โ€” kuzushiji (historical) โ†’ modern Japanese. Only tool

Use Cases

Manga translation โ€” extract text with manga-ocr โ†’ translate โ†’ replace in speech bubbles Japanese business doc digitization โ€” Sarashina2.2 to convert scanned docs to Markdown Historical research โ€” decode Edo/Meiji era documents with KuroNet