anpr_sdk 1.0.3

Scan vehicle number plates (license plates) in Flutter from the camera or a photo. On-device ANPR and OCR, works offline after first launch, tested on Indian plates.

anpr_sdk

Plug-and-play number plate detection and recognition (ANPR) for Flutter.

Drop in one widget and get live plate reads from the camera, or pass in a photo and get the plate text back. Everything runs on-device with ONNX models: no per-scan API calls, no server, and it works offline after the first launch.

Screenshots

Demo App

Try the ANPR SDK demo on an Android device.

Download PlateX Demo APK

Note: Android only. Allow installation from this source if prompted. The app requires internet access on first launch to download the recognition models; subsequent use works offline.

Features

• Live scanning: AnprScannerView is a ready-made camera preview with a bounding box and a confirmed-plate callback.
• Headless scanning: AnprScanner gives you the camera and result streams so you can build your own UI.
• Still images: recognizeImage() / recognizeImageFile() for gallery photos, uploads or snapshots.
• Stable results: confidence-weighted voting across frames, so a plate is only reported once it has been read consistently.
• One-time model download: models are fetched on first run, cached on the device, and then work fully offline.
• Typed errors: AnprException with an AnprErrorCode you can switch on.
• Tunable: thresholds, intervals and voting through AnprSdkConfig.

Tested on Indian number plates

Built and tested on Indian plates: standard white, yellow commercial and green EV plates, in daylight, at night and with glare, from both live camera and uploaded photos.

Also tested on images from the public Indian Number Plates Dataset by DataCluster Labs. The dataset images are used for testing only and are not included in this package.

Accuracy depends on lighting, angle and plate condition. See Tips for good results.

Platform support

Android only for now. iOS has not been tested and is not officially supported yet.

Installation

dependencies:
  anpr_sdk: ^1.0.0

flutter pub get

Android setup

Add the permissions to android/app/src/main/AndroidManifest.xml (outside the <application> tag):

<uses-permission android:name="android.permission.INTERNET"/>
<uses-permission android:name="android.permission.CAMERA"/>

INTERNET is only needed for the first-run model download.

If your build fails with a minimum SDK error, raise minSdk in android/app/build.gradle(.kts). The camera plugin needs 21 or higher, and flutter_onnxruntime may need a higher value (24 is a safe choice).

Quick start

Call initialize() once, as early as possible, then show the scanner.

import 'dart:async';

import 'package:anpr_sdk/anpr_sdk.dart';
import 'package:flutter/material.dart';

Future<void> main() async {
  WidgetsFlutterBinding.ensureInitialized();

  // Starts the model download (first run only) and loads the models.
  // No need to await it: AnprScannerView shows progress by itself.
  unawaited(
    AnprSdk.instance
        .initialize()
        .catchError((Object e) => debugPrint('ANPR init failed: $e')),
  );

  runApp(const MaterialApp(home: ScanPage()));
}

class ScanPage extends StatefulWidget {
  const ScanPage({super.key});

  @override
  State<ScanPage> createState() => _ScanPageState();
}

class _ScanPageState extends State<ScanPage> {
  String? _plate;

  @override
  Widget build(BuildContext context) {
    return Scaffold(
      body: Stack(
        children: [
          AnprScannerView(
            onPlate: (r) => setState(() => _plate = r.plateNumber),
          ),
          SafeArea(
            child: Container(
              margin: const EdgeInsets.all(16),
              padding: const EdgeInsets.all(16),
              color: Colors.white,
              child: Text(_plate ?? 'Point the camera at a number plate'),
            ),
          ),
        ],
      ),
    );
  }
}

Usage

1. Initialize

await AnprSdk.instance.initialize();

• Safe to call more than once: it returns immediately when ready, and concurrent calls share a single run.
• The first run downloads the models over the internet and caches them. Later runs load from the cache and work offline.
• Throws an AnprException on failure. Call it again to retry.

Pass a config to tune behaviour:

await AnprSdk.instance.initialize(
  config: const AnprSdkConfig(
    detectionThreshold: 0.40,
    confirmReads: 3,
  ),
  onStatus: (s) => debugPrint('ANPR: $s'),
);

You can also watch the state to build your own loading screen:

ValueListenableBuilder<AnprSdkStatus>(
  valueListenable: AnprSdk.instance.status,
  builder: (context, status, _) {
    // uninitialized, initializing, ready, error
    return Text(status.name);
  },
);

// First-run download progress (null when not downloading)
ValueListenableBuilder<DownloadProgress?>(
  valueListenable: AnprSdk.instance.downloadProgress,
  builder: (context, p, _) => LinearProgressIndicator(value: p?.fraction),
);

2. Live scanning with the widget

AnprScannerView(
  onPlate: (AnprResult r) => print('${r.plateNumber} (${r.ocrConfidence})'),
  showBoundingBox: true,
  boxColor: Colors.green,
  fit: BoxFit.cover,
)

Customise what the user sees:

AnprScannerView(
  onPlate: (r) => save(r.plateNumber!),
  loadingBuilder: (context, progress) =>
      Center(child: CircularProgressIndicator(value: progress)),
  errorBuilder: (context, error, retry) => Center(
    child: Column(
      mainAxisSize: MainAxisSize.min,
      children: [
        Text(error.message),
        TextButton(onPressed: retry, child: const Text('Retry')),
      ],
    ),
  ),
  // Draw anything on top of the preview using the live (per-frame) result.
  overlayBuilder: (context, live) => Align(
    alignment: Alignment.bottomCenter,
    child: Text(live.plateNumber ?? '…'),
  ),
)

onPlate fires only for confirmed plates: the same plate must be read confirmReads times (default 2). A plate that stays in view is reported again only after repeatCooldown (default 5 s); a plate that leaves the view and returns is reported again immediately.

3. Live scanning without the widget (headless)

Use AnprScanner when you want your own camera UI or want to start and stop scanning from code.

final scanner = AnprScanner(
  lensDirection: CameraLensDirection.back,
  resolution: ResolutionPreset.high,
);

// Confirmed plates
final sub = scanner.plates.listen((r) => print(r.plateNumber));

// Per-frame updates (box, current text, timings)
scanner.live.addListener(() => print(scanner.live.value));

await scanner.start();   // never throws: check scanner.status / scanner.error
// ...
await scanner.stop();    // releases the camera, can start() again
await sub.cancel();
await scanner.dispose(); // when finished for good

You can also hand your own scanner to the widget to control it from outside. The view then will not dispose it for you:

final scanner = AnprScanner();
AnprScannerView(scanner: scanner, onPlate: (r) {});

The scanner pauses the camera automatically when the app goes to the background and resumes when it comes back.

4. Still images

final AnprResult? result =
    await AnprSdk.instance.recognizeImage(bytes);      // Uint8List (JPEG/PNG/…)

final AnprResult? fromFile =
    await AnprSdk.instance.recognizeImageFile(File(path));

if (result == null) {
  // No plate found
} else if (result.plateNumber == null) {
  // A plate was found but its text could not be read confidently
} else {
  print(result.plateNumber); // e.g. MH12AB1234
}

The image must be upright. EXIF rotation is not applied, so rotate photos from a camera first if needed.

5. Handle errors

try {
  await AnprSdk.instance.initialize();
} on AnprException catch (e) {
  switch (e.code) {
    case AnprErrorCode.noInternet:
      // First-time setup needs a connection
      break;
    case AnprErrorCode.modelSetupFailed:
      // Download or model loading failed: retry
      break;
    default:
      debugPrint(e.toString());
  }
}

Code	Meaning
notInitialized	initialize() was not called, or has not finished
noInternet	First-run model download needs a connection
modelSetupFailed	Models could not be downloaded or loaded
invalidImage	The bytes passed to recognizeImage() could not be decoded
noCamera	The device has no camera
cameraPermissionDenied	The user denied camera permission
cameraError	Any other camera failure

Reading a result

AnprResult fields:

Field	Type	Description
plateNumber	String?	Upper-case letters and digits only, e.g. MH12AB1234. Null if nothing was read
ocrConfidence	double	Confidence of the latest text read, 0 to 1
detectionConfidence	double	Confidence of the plate box, 0 to 1
box	Rect?	Plate box in pixels of the upright frame
frameSize	Size	Size of the frame the box refers to
isConfirmed	bool	True when the same plate was read enough times
timestamp	DateTime	When the result was produced
detectionMs, ocrMs	double	Timings in milliseconds, handy for debugging

Configuration

All fields of AnprSdkConfig are optional.

Option	Default	Description
detectionThreshold	0.35	Minimum confidence to accept a plate box
detectionInterval	150 ms	Minimum gap between detection runs
ocrInterval	400 ms	Minimum gap between OCR runs
ocrMinConfidence	0.55	Text reads below this are ignored
ocrMinLength	3	Text reads shorter than this are ignored
boxGrace	500 ms	Keep the box on screen this long if a frame misses
textClear	3 s	Clear the plate text if no plate was seen for this long
voteWindow	2500 ms	Window used to vote between noisy reads
voteMaxReads	7	Max recent reads kept for voting
confirmReads	2	Reads needed before a plate is confirmed
repeatCooldown	5 s	A plate that stays in view is reported again after this

Raise confirmReads and ocrMinConfidence for fewer false reads. Lower them for faster results.

How the models are delivered

The package ships no model files. On first run it downloads three files from the project's GitHub release and stores them in the app's support directory (models_<version>):

• best.onnx: plate detector
• anpr_ocr_f16.onnx: plate text recognizer
• en_dict.txt: character dictionary

After that the SDK works offline. If a cached model fails to load, it is deleted and downloaded again once automatically.

Tips for good results

• Keep the plate fully inside the frame, roughly facing the camera.
• Good light matters more than resolution. Avoid strong glare on the plate.
• Use ResolutionPreset.high for distant plates; medium on low-end phones if frames are slow.
• Call AnprSdk.instance.dispose() only if you want to free the native models; call initialize() again to reuse the SDK.

Troubleshooting

Problem	Fix
Spinner never finishes on first run	The device needs internet on first launch. Check AnprSdk.instance.lastError
"Model download failed (404)"	The model files are missing from the release. Check the package version
Black screen or camera error on Android	Add the CAMERA permission and accept the runtime prompt
Wrong text read	Raise confirmReads / ocrMinConfidence, improve lighting

Example

A full runnable app is in the example/ folder.

Issues and contributions

Bug reports and pull requests are welcome at github.com/sohvn/anpr_sdk.

License

The SDK source code is released under the MIT License. See LICENSE.

The package ships no models. The models it downloads at runtime have their own licenses: the plate detector is based on Ultralytics YOLO (AGPL-3.0) and the text recognizer on PaddleOCR (Apache-2.0). See THIRD_PARTY_NOTICES.md and read the model terms before using the SDK in a product.