react-native-plate-scanner

On-device Korean license-plate recognition for React Native. Fully offline — no network, no cloud, no license/API key. Models are bundled in the native module.

• Android: implemented (YOLOv11-640 detect + tinyLPR OCR + Korean 8-format post-processing)
• iOS: not implemented yet (calls reject on iOS)
• Accuracy: ~70% exact match on real photos; higher on clean plates

Requires a custom dev client / bare workflow. Does not work in Expo Go (native code). For Expo, use a development build (expo prebuild + expo run:android).

Install

npm install react-native-plate-scanner
# or
yarn add react-native-plate-scanner

Autolinking wires up the Android native module. Then you must tell your app not to compress the bundled model files — add this to android/app/build.gradle:

android {
    androidResources {
        noCompress 'tflite', 'onnx', 'txt'
    }
}

Finally rebuild the app:

npx react-native run-android
# (Expo) npx expo run:android

The module pulls TensorFlow Lite + ONNX Runtime automatically. App size grows by ~15–20 MB (models + inference runtimes).

Usage

import { recognizePlate } from 'react-native-plate-scanner';

// imageUri: a file:// URI from a camera/picker (e.g. expo-camera, vision-camera,
// react-native-image-picker). Any orientation is handled internally.
const result = await recognizePlate(imageUri);

if (result.plate && !result.needsConfirm) {
  console.log('Plate:', result.plate);          // e.g. "대전 89 아3305" / "12가3456"
} else if (result.plate) {
  // low confidence — ask user to confirm result.plate
} else {
  // no plate found
}

Result shape

interface PlateResult {
  plate: string;          // normalized plate ("" if none)
  confidence: number;     // 0..1
  plateType: string;      // ordinary_private | commercial | construction_new | ... | unknown
  vehicleType: string;
  region: string | null;  // e.g. "대전"
  matchedFormat: boolean; // matched one of the 8 Korean formats
  needsConfirm: boolean;  // true when confidence < 0.8
  rawOcr: string;         // OCR text before normalization (debug)
  debug: string;          // orientations tried, chosen degree, confidence
}

Notes & limits

• Run recognition off the JS/UI critical path where possible; a call takes a few hundred ms depending on device and image size.
• confidence >= 0.8 typically means correct. Use needsConfirm to gate auto-accept vs. manual confirmation.
• Free/offline engine; for higher accuracy on hard images combine with an online OCR when connectivity is available.
• iOS is a stub; recognizePlate rejects on iOS until an iOS implementation is added.

Author

HIEP LE VAN

• GitLab: https://gitlab.com/react-native7481714
• Email: [email protected]

License

MIT © HIEP LE VAN