ANPR Scanner Flutter

A Flutter package for Automatic Number Plate Recognition (ANPR) using TFLite models. Runs a two-stage ML pipeline (detection + OCR) entirely on-device for real-time license plate scanning.

BYO Models — the package does not bundle any TFLite models. You provide your own detection and OCR .tflite files.

Features

  • Real-time license plate detection and OCR on camera frames
  • GPU-accelerated inference via TFLite GPU delegate (falls back to CPU)
  • Drop-in AnprScannerWidget with built-in camera, permissions, loading UI, and scan flow
  • Low-level LicensePlateDetector API for custom integrations
  • Stability tracking with configurable quality gates for auto-capture
  • Optimized YUV-to-RGB conversion with lookup tables and isolate offloading

Quick Start

1. Add the dependency

dependencies:
  anpr_scanner_flutter:
    path: ../  # or git/pub reference

2. Provide model files

Place your TFLite model files in your app's assets:

your_app/
  assets/
    models/
      detection.tflite   # YOLOv8-style plate detection model
      ocr.tflite          # Character detection model

Register them in pubspec.yaml:

flutter:
  assets:
    - assets/models/

3. Platform setup

Android — Add to android/app/src/main/AndroidManifest.xml:

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

iOS — Add to ios/Runner/Info.plist:

<key>NSCameraUsageDescription</key>
<string>Camera access is needed to scan license plates.</string>

4. Use the scanner

import 'package:anpr_scanner_flutter/anpr_scanner_flutter.dart';

// Drop-in widget — handles everything automatically
AnprScannerWidget(
  detModelPath: 'assets/models/detection.tflite',
  ocrModelPath: 'assets/models/ocr.tflite',
  onPlateRecognized: (result) {
    print('Plate: ${result.fullPlate}');
    print('Code: ${result.code}, Number: ${result.number}');
  },
  onError: (error) => print('Error: $error'),
)

Two Abstraction Levels

High-Level: AnprScannerWidget

A self-contained widget that manages the full scan lifecycle:

loading -> scanning -> capturing -> recognizing -> result

AnprScannerWidget(
  detModelPath: 'assets/models/detection.tflite',
  ocrModelPath: 'assets/models/ocr.tflite',
  onPlateRecognized: (LicensePlateResult result) {
    Navigator.pop(context, result);
  },
  // Optional customization:
  delegateType: DelegateType.gpu,          // gpu (default), cpu, nnapi, auto
  captureQualityConfig: CaptureQualityConfig(
    minConfidence: 0.70,
    requiredStableFrames: 3,
    maxCentreMoveFraction: 0.08,
  ),
  preloadedDetector: myDetector,           // skip loading if already loaded
)

Includes:

  • Camera permission handling
  • Model loading overlay with progress
  • Live bounding box overlay
  • Pinch-to-zoom and tap-to-focus
  • Torch toggle
  • Capture flash animation
  • Result card with plate image and metrics

Low-Level: LicensePlateDetector

Direct access to the two-stage inference pipeline. Use this when you want full control over the image source and UI.

final detector = LicensePlateDetector();
await detector.initialize(
  detModelPath: 'assets/models/detection.tflite',
  ocrModelPath: 'assets/models/ocr.tflite',
  delegateType: DelegateType.gpu,
);

// Run on any img.Image (from camera, gallery, file, etc.)
final result = await detector.recognizePlate(image);

if (result != null) {
  print('Full plate: ${result.fullPlate}');
  print('Code: ${result.code}');
  print('Number: ${result.number}');
  print('Confidence: ${result.plateBox?.confidence}');
  print('Detection time: ${result.metrics.detectionMs}ms');
  print('OCR time: ${result.metrics.ocrMs}ms');
  print('Total time: ${result.metrics.totalMs}ms');
}

detector.dispose();

ML Pipeline

The package runs a two-stage pipeline:

Stage 1: Detection

  • Input: 640x640 float32 letterboxed image
  • Model: YOLOv8-style object detection
  • Output: Bounding boxes with confidence scores
  • Post-processing: Non-Maximum Suppression (NMS)

Stage 2: OCR

  • Input: 160x160 float32 letterboxed crop of the detected plate
  • Model: Character-level detection (0-9, A-Z)
  • Output: Character positions with class IDs and confidences
  • Post-processing: Left-to-right sorting, then code/number splitting via:
    • Letter/number regex separation (when alphabetic characters are present)
    • Largest-gap heuristic (for all-numeric plates)

GPU Delegate Strategy

  • The detection model uses the GPU delegate for maximum performance (it's the heavier model).
  • The OCR model uses CPU threads. Running two simultaneous GPU delegates crashes the native GPU driver on many Qualcomm Adreno and ARM Mali devices.

Model Requirements

Detection Model

Property Value
Input shape [1, 640, 640, 3]
Input type float32 (0.0-1.0)
Output YOLOv8 transposed format
Classes 1 (license plate)

OCR Model

Property Value
Input shape [1, 160, 160, 3]
Input type float32 (0.0-1.0)
Output YOLOv8 transposed format
Classes 36 (0-9, A-Z)

Key Classes

Class Purpose
AnprScannerWidget Drop-in scanner widget with full UI
LicensePlateDetector Core TFLite inference engine
ModelServiceManagerOptimized Singleton model cache with async loading
DetectionOnlyProcessor Real-time camera frame pipeline with stability tracking
LicensePlateResult Recognition result (code, number, metrics, cropped plate)
CaptureQualityConfig Stability gate thresholds for auto-capture
RealtimeConfig Global tuning for frame pacing, conversion, thresholds
DetectionOverlay Animated bounding box overlay widget
ModelLoadingOverlay Model loading progress widget

Configuration

CaptureQualityConfig

Controls when auto-capture triggers during real-time scanning:

CaptureQualityConfig(
  minConfidence: 0.55,          // Min detection confidence per frame
  requiredStableFrames: 3,      // Consecutive passing frames before capture
  maxCentreMoveFraction: 0.08,  // Max box centre drift between frames
  maxAreaChangeFraction: 0.06,  // Max box area change between frames
  minPlateAreaFraction: 0.005,  // Min plate area relative to frame
)

RealtimeConfig

Static constants for global real-time tuning:

  • downsampleFactor -- Frame downsampling during conversion (default: 1)
  • useIsolateConversion -- Offload YUV-to-RGB to background isolate (default: true)
  • minFrameIntervalMs -- Min time between processed frames (default: 50ms)
  • detectionConfidence -- Live detection threshold (default: 0.50)
  • staleFrameThreshold -- Frames before clearing overlay (default: 3)
  • useLookupTables -- Pre-computed YUV tables (default: true)

Package Structure

lib/
  anpr_scanner_flutter.dart          # Public API barrel file
  src/
    config/
      realtime_config.dart           # Frame pacing, conversion, stability thresholds
    models/
      detection_box.dart             # DetectionBox, CharDetection
      license_plate_result.dart      # LicensePlateResult, InferenceMetrics
      captured_frame.dart            # CapturedFrame, CaptureQualityConfig
    services/
      license_plate_detector.dart    # Core TFLite inference (detection + OCR)
      model_service_manager_optimized.dart  # Singleton async model loader/cache
      detection_only_processor.dart  # Real-time camera frame processor
    utils/
      camera_image_converter_optimized.dart  # YUV420/BGRA -> img.Image
    widgets/
      anpr_scanner_widget.dart       # High-level drop-in scanner widget
      detection_overlay.dart         # Animated bounding box overlay
      scan_result_overlay.dart       # Frozen frame, result card, flash, recognizing overlay
      loading_widgets.dart           # Model loading overlay

Dependencies

Package Purpose
tflite_flutter TFLite model inference
camera Camera stream access
image Image manipulation (resize, crop, encode)
permission_handler Camera permission requests
image_picker Gallery image selection (low-level usage)
path_provider File system paths

License

See LICENSE for details.

Libraries

anpr_scanner_flutter