Business Card Scanning

Point the camera at a business card to extract the name, company, job title, phone numbers, email, website, and address. Korean and English cards are supported.

How it works

camera frame → native OCR (Apple Vision / ML Kit) → text lines → parseBusinessCard() (heuristics)

Like MRZ and cards, the native side does OCR only and returns text lines; parseBusinessCard runs on the JS thread. It has no extra native dependency — it reuses the same scan(frame).

Parsing API

import { parseBusinessCard, type BusinessCardResult } from '@jieonist/vision-camera-ocr-scanner';

const result = parseBusinessCard(ocrTextLines); // BusinessCardResult | null

BusinessCardResult

interface BusinessCardResult {
  name: string | null;
  company: string | null;
  jobTitle: string | null;
  department: string | null;   // org unit ("Technical R&D", "기술개발팀")
  phones: BusinessCardPhone[]; // { type: 'mobile' | 'tel' | 'fax' | 'unknown', number: string }
  email: string | null;
  website: string | null;
  address: string | null;
  lines: string[];             // the OCR lines that were scanned
}

Contact info is the anchor — parseBusinessCard returns null when no email, phone, or website is found. A phone's type is only set from strong signals (an explicit label like M. / Tel: / Fax, or a Korean 010 mobile prefix); otherwise it stays 'unknown' rather than guessing.

Titles aren't dictionary-bound. Known keywords (CEO, 팀장, Engineer…) match directly, and anything else is caught by elimination: a short line that is none of the strongly-patterned fields (contact, address, company, name) and sits next to the name is read as the role. Lines carrying an org-unit marker (R&D, Team, 본부, 팀…) land in department; slogans and sentence-shaped lines are rejected.

Pass lineItems for layout signals. parseBusinessCard(lines, ocr.lineItems) turns text size into evidence — the tallest line is the name (big top text reads as the brand), an unexplained line taller than the name is the company logo, and a line printed larger than the name can't be a title.

No valid field

Unlike MRZ (check digits) and payment cards (Luhn), a business card carries nothing to checksum, so there is no valid flag — every field is best-effort. Judge a result by the fields your flow needs (e.g. require email or a phone), and use the scan session below to reject one-frame misreads.

Multi-frame session

import { createBusinessCardScanSession } from '@jieonist/vision-camera-ocr-scanner';

const session = createBusinessCardScanSession(); // minReads defaults to 2

// in your onLines handler:
const final = session.push(parseBusinessCard(lines));
if (final != null) {
  // confirmed — the same contact identity repeated across frames
}

The session anchors on the contact identity (email, or the set of phone numbers) repeating across frames. Frames without an email or phone are dropped. Unchecked fields (name, company, title, address) are taken by majority vote, and a name is only trusted once it has repeated.

Notes & limitations

• Heuristics favor Korean and English card layouts; other languages will mostly fall back to contact fields only.
• Stylized logos often OCR poorly — company may come from the legal-suffix line (주식회사 / Co., Ltd) or a line matching the email domain instead.
• Decorative fonts and vertical layouts reduce accuracy. Treat results as autofill hints and let users edit.
• In detectDocument(), a business card is the guarded last resort: it only detects when an email is present, after MRZ and card fail — a phone number alone would make almost any text block (receipts, posters) "detect" as a business card.