Chinese, Japanese and Korean OCR: Common Pitfalls

Chinese and Japanese are written without spaces between words. Older tools often inserted spurious spaces between characters, producing text that looked broken in every downstream tool. Modern engines keep contiguous runs intact, which is what you want — the lack of spaces is correct, not an error to fix.

CJK characters can be genuinely similar at small sizes: thousands of glyphs share radicals and differ in a few strokes. At low resolution, characters like 未/末 or 己/已 become coin flips. Context and language models rescue most of them; the residue concentrates in rare characters and proper nouns — exactly the ones you cannot predict.

A model trained on both may normalize output toward one convention — converting traditional text to simplified or vice versa — which is wrong for anyone quoting a source. Check a few known traditional-only or simplified-only characters if the distinction matters for your document.

Japanese interleaves kanji, hiragana, and katakana. Recognition must track all three simultaneously, and it mostly does. The common artifact is furigana (small reading characters above kanji) being inlined into the text, where it reads as random kana. If your document has furigana, plan to strip those lines.

Korean uses spaces like Latin languages, but particles attach to words, and modern tech documents freely mix Korean with English identifiers. Both are handled well. Watch for hangul being split at line-wrap points: the engine should rejoin syllables that wrap across lines, and most do.

Shoot at higher resolution than you would for English — CJK glyphs carry more strokes per character, so blur costs proportionally more. Avoid aggressive image compression before uploading. And always spot-check names and numbers: they concentrate the remaining errors.