OCR Chinese, Japanese, and Korean PDFs

CJK (Chinese, Japanese, Korean) OCR is a different beast from Latin OCR. The recognizer has to discriminate between thousands of distinct ideographs per page rather than 52 letters, and layout conventions like vertical text or right-to-left columns break naive line-detection logic. This guide explains where the accuracy comes from, what document types need extra care, and what to verify on output.

Why CJK Is Different

Character set size

Latin OCR has roughly 100 distinct shapes to learn (upper, lower, digits, punctuation). Simplified Chinese has around 7,000 commonly used characters; Traditional Chinese has more. Japanese mixes ~2,000 kanji with two phonetic alphabets (hiragana and katakana, ~50 characters each). Korean writes syllable blocks (~11,000 unique blocks formed from 24 base letters). The recognizer needs to discriminate among all of them — and many look very similar at small sizes (土 vs 士).

Vertical and mixed layouts

Traditional Japanese books, Chinese classical texts, and Korean historical documents are often typeset vertically, top to bottom, right to left. Modern documents are usually horizontal but a single page can mix orientations — vertical body text with horizontal Latin captions, for example. OCR engines need to detect orientation per text region rather than assuming the whole page.

Simplified vs Traditional Chinese

Mainland China uses Simplified characters; Taiwan, Hong Kong, and most overseas communities use Traditional. They share a lot of vocabulary but the same word can have different shapes (国 vs 國, 学 vs 學). Tell the recognizer which variant your document uses — auto-detection sometimes guesses wrong on documents with few characters.

Japanese script mixing

Japanese mixes kanji (Chinese-origin characters), hiragana (cursive phonetic), katakana (angular phonetic for foreign words and emphasis), and frequently Latin letters and Arabic numerals all in one paragraph. The recognizer has to identify each script per character, not per line.

Korean syllable blocks

Korean Hangul is alphabetic but visually grouped into square syllable blocks. Each block contains 2-4 letters arranged in a specific layout. Recognizers either treat each block as a unit (works for clean printed text) or decompose it into its letters (works for stylized fonts but more error-prone).

What Determines Accuracy

1. Resolution. CJK characters need more pixels than Latin letters to be reliably distinguished. 300 DPI is the minimum for body text; 400 DPI is better for dense classical documents.
2. Font style. Recognizers are trained heavily on standard typefaces (Heiti, Mincho, Batang). Calligraphic or decorative fonts drop accuracy substantially.
3. Image quality. JPEG compression artifacts on dense scripts merge adjacent strokes. If you have a high-quality TIFF or PNG, use that — don't pre-convert to JPEG.
4. Engine choice. Cloud OCR (Google Vision, AWS Textract) currently outperforms open-source Tesseract on dense CJK by a wide margin. The gap is largest on documents with stylized fonts or mixed scripts.

What to Check on Output

1. Search a 3-character word. If Ctrl-F finds a phrase like 中国人 or 学生, characters were correctly identified and the text layer was written with proper Unicode.
2. Copy a sentence. Paste into a CJK-capable editor. You should see real characters, not boxes or placeholder squares. Boxes mean the embedded font lacks CJK coverage — common with older OCR tools that default to a Latin-only text-layer font.
3. Compare a tricky pair. Pick a character that has a Simplified/Traditional variant in your document (e.g. 学/學) and verify the recognizer kept the form your original used.
4. Check punctuation. CJK uses fullwidth punctuation (，。！？) distinct from halfwidth (,.!?). The output should match the source — mixing them up is a sign the recognizer is misclassifying.

Common Document Types

• Business documents — invoices, contracts, internal reports. Usually horizontal, mixed CJK/Latin/numbers.
• Academic papers — dense formulas mixed with CJK body. Math layout is often the failure mode.
• Government records and historical archives — older fonts, vertical layouts, Traditional Chinese in mainland archives or Simplified in Taiwan archives.
• Manga and printed novels — vertical right-to-left, often handwritten-style fonts. Hard mode for OCR.

convert2searchablepdf for CJK

We use Google Cloud Vision's document text detection, which ranks among the best for CJK accuracy. The text layer in the searchable PDF is drawn with a Unicode-aware font (Noto family) that includes the full Han, Hiragana, Katakana, and Hangul blocks — copy-paste returns proper characters, not boxes. The layout-preserving text export handles vertical layouts and multi-column tables.

Try it on your own document — drop the PDF onto the homepage and you'll get both the searchable PDF and a plain-text export back within seconds.