dni_peru_ocr

Peruvian DNI OCR helpers for Flutter — denoise the Google ML Kit Latin TextRecognizer output against the Documento Nacional de Identidad and recover clean, structured fields with temporal consensus across frames.

Why

A single ML Kit frame on a Peruvian DNI is noisy: Ñ is encoded as NXX on the MRZ line, address labels (DIRECCIÓN) collide with QR artifacts, civic-box content (CONSTANCIA DE SUFRAGIO) drowns the real address, and document tilt skews block geometry. This package solves these problems with a layered, side-by-side strategy pipeline plus a temporal accumulator — no manual cleanup required at the consumer.

Features

MRZ parsing — ICAO 9303 TD1 with checksum validation via mrz_parser, plus Peruvian-specific Ñ → NXX recovery.
Strategy decomposition — three independent extractors (MrzFieldStrategy, TextOcrFieldStrategy, AddressFieldStrategy) combined by a thin coordinator. Each strategy is stateless and individually testable.
Temporal consensus — OcrConsensusAccumulator votes across frames and emits a deterministic winner even under noisy OCR (address vote consolidation tolerates micro-variants; name votes consolidate by strict prefix containment).
Ubigeo extraction — populates department, province, and district from the back-side administrative line (ANCASH/SANTA/ CHIMBOTE, /CALLAO/VENTANILLA, LIMA/LIMA/VILLA MARIA DEL TRIUNFO).
Pluggable observability — inject your own OcrLogger (Sentry, Crashlytics, Datadog, custom) at the extractor constructor.
Capture widget — DniScanner is the single production capture widget. It ships auto-capture with a centralized capture brain (CaptureCoordinator, internal), a live 3-2-1 countdown, an IMU jolt-reset guard, a live lighting/glare gate, post-shutter blur reject-and-retry, a manual fallback after a configurable timeout, tilt detection, two-sided front→back handoff, and a dispose-safe lifecycle. Capture readiness, countdown, document presence, and the manual fallback all flow through one source of truth. The lifecycle-only DniCameraController is exposed for OCR-consensus and lookup wiring.
Document quad detection (hybrid, non-blocking) — an on-device document-quadrilateral detector (DocumentQuadDetector port, OpenCV adapter via opencv_dart, with a pure-Dart fallback). The quad is an enhancement signal, not a capture gate: auto-capture fires on OCR readiness + frame stability, and the quad raises confidence / enables a cleaner crop when available. It never blocks a capture, so a present DNI is captured even when the quad detector returns no corners on a given frame. See Document quad detection below.

Installation

dependencies:
  dni_peru_ocr: ^1.0.0

Breaking platform floor (v1.0.0). The on-device quad detector depends on opencv_dart, which relies on Dart Native Assets hooks. That raises the host floor to Flutter >=3.38.0 / Dart >=3.10.0. Consumers on older toolchains must upgrade before adopting v1.0.0.

The native binary (libdartcv.so) is trimmed to the modules quad detection needs (core, imgproc, imgcodecs); calib3d and every other module are excluded. Measured per-ABI native cost is ~9.98 MiB on arm64-v8a, and an App Bundle ships a single ABI per device.

Note: dio is a direct runtime dependency of this package (required for pub.dev compliance). You don't need to add it to your own pubspec.yaml.

flutter pub get

Example

A runnable example app is available under example/. It demonstrates the complete two-sided DNI capture flow — DniScanner drives the front→back handoff internally and emits a single DniScanResult — plus result display with per-field confidence indicators, on a real Android or iOS device. See example/README.md for setup instructions and a walkthrough of the recommended integration pattern.

Quick start — headless extraction

import 'package:dni_peru_ocr/dni_peru_ocr.dart';

// Static entry point — no logger, default strategies.
final fields = OcrFieldExtractor.extract(recognizedText);

print(fields.firstName);      // JUAN CARLOS
print(fields.lastName);       // MUÑOZ
print(fields.secondLastName); // PEREZ
print(fields.address);        // AV. SANTA ROSA 1080 MARIATEGUI
print(fields.department);     // LIMA
print(fields.province);       // LIMA
print(fields.district);       // VILLA MARIA DEL TRIUNFO

To route OCR/MRZ mismatch breadcrumbs through your observability stack, use the instance API:

const extractor = OcrFieldExtractor(logger: SentryOcrLogger());
final fields = extractor.extractWith(recognizedText);

DniScanner is a Flutter widget that owns the full capture flow. The host provides a camera plugin CameraController and listens for the final capture.

Two-sided mode (isBackSide == null, the default): the scanner drives the front→back handoff itself and emits a single DniScanResult through onScanComplete.
Single-side mode (isBackSide: false or true): the scanner scans one side and emits a DniSideScanResult through onSideCaptured.

The constructor assert enforces the pairing: two-sided requires onScanComplete; single-side requires onSideCaptured.

import 'package:dni_peru_ocr/dni_peru_ocr.dart';

// Two-sided flow — scanner handles front then back internally.
DniScanner(
  controller: cameraController,
  fields: DniFields.kyc(),
  onScanComplete: (result) {
    print(result.hunt.firstName);
    print(result.hunt.address);
    print(result.frontPhoto.path);
    print(result.backPhoto.path);
  },
)

Tuning is optional — every parameter has a sensible default:

DniScanner(
  controller: cameraController,
  captureMode: DniCaptureMode.auto,   // auto | manual | hybrid
  autoCaptureMs: 1500,                // dwell before auto-capture
  gracePeriodMs: 600,                 // tolerated quality dip mid-countdown
  manualFallbackMs: 30000,            // show manual button after this
  minStableFrames: 3,                 // stable frames required to arm
  scanHints: const DniScanHints(),    // rotating guidance (see below)
  onScanComplete: (result) { /* ... */ },
)

Scan hints

DniScanHints configures the rotating, phase-aware guidance shown along the bottom of the scanner. The copy is intentionally generic — it guides the physical action (focus, hold still, flip) and never names a specific DNI field, because the side a field belongs to is not knowable across the many Peru DNI versions. Defaults are neutral Spanish; pass your own lists to localize or reword:

DniScanner(
  controller: cameraController,
  scanHints: const DniScanHints(
    waitingFront: ['Place the document inside the frame', 'Focus the document'],
    extractingFront: ['Hold the document still', 'Hold steady for a moment'],
    waitingBack: ['Flip the document', 'Place it inside the frame'],
    extractingBack: ['Hold the document still'],
    processing: 'Processing…',
    documentAbsent: 'No document detected',
  ),
  onScanComplete: (result) { /* ... */ },
)

Host-callback safety

DniScanner treats every callback it hands you — onScanComplete, onSideCaptured, and onDniReady — as an untrusted boundary. If your callback throws, the scanner catches the error, keeps the capture flow and camera alive, and never rethrows into the Flutter framework. A thrown host callback can no longer tear down the capture session.

Provide the optional onError callback to observe these failures; when it is omitted, the error is logged through DniLogger and capture continues.

DniScanner(
  controller: cameraController,
  onScanComplete: (result) {
    // Your code may throw here without crashing the scanner.
  },
  onError: (error, stack) {
    // Optional. Surface or report the failure however you like.
  },
)

Integration requirements

DniScanner reads the device IMU through sensors_plus, which raises the host build floor. The capture pipeline also uses isolates and the camera plugin, but the only new platform constraints come from sensors_plus.

Android

sensors_plus requires a modern Android toolchain. Configure your app module with:

Java 17 (compileOptions / kotlinOptions.jvmTarget = "17")
Android Gradle Plugin ≥ 8.12.1 (recommended minimum)
Gradle ≥ 8.13

The example app in this repository currently pins AGP 8.11.1; bump it to 8.12.1 or newer in your own project to satisfy sensors_plus. No runtime permission is needed for the accelerometer or gyroscope.

iOS

No Info.plist entry is required. DniScanner reads only the accelerometer and gyroscope via sensors_plus. It never initializes the barometer (CMAltimeter), which is the only sensor that would require NSMotionUsageDescription. Standard camera plugin keys (NSCameraUsageDescription) still apply for the preview itself.

Tuning parameters and defaults

Parameter	Default	Purpose
`autoCaptureMs`	`1500`	Dwell time the document must stay aligned and well-lit before auto-capture fires.
`gracePeriodMs`	`600`	Tolerated transient quality dip or jolt mid-countdown before the anchor resets.
`manualFallbackMs`	`30000`	Idle time before the manual capture button is surfaced.
`minStableFrames`	`3`	Consecutive stable frames required to arm the countdown.

The phone does not need to be held perfectly still. Auto-capture arms once the document is aligned in frame and well-lit; the dwell countdown is the "hold steady for a moment" step. The IMU is only a jolt guard — it tolerates normal hand tremor and resets the countdown solely on a deliberate shake. Its permissive jolt thresholds (accelerometer ≈ 2.5 m/s², gyroscope ≈ 1.5 rad/s, EMA window ≈ 5 samples) and the lighting thresholds (minLuminance = 40, maxLuminance = 235, maxSaturatedFraction = 0.10) ship as device-tunable defaults. They are calibrated for typical mid-range phones; adjust them per device class if your fleet skews very low- or high-end.

DNI Lookup

Beyond OCR, dni_peru_ocr ships a flexible lookup contract so you can fetch normalized DNI data from any backend. The lookup feature is fully optional — OCR-only consumers never instantiate any lookup service and pay no behavioral cost.

Recommended composition — caching + service

Implement DniCache for your storage layer (Hive, Isar, SharedPreferences, or just in-memory), then compose with a concrete service:

import 'package:dio/dio.dart';
import 'package:dni_peru_ocr/dni_peru_ocr.dart';

// Minimal in-memory cache — replace with Hive / Isar in production.
class InMemoryDniCache implements DniCache {
  final Map<String, DniData> _store = {};

  @override
  Future<DniData?> get(String dni) async => _store[dni];

  @override
  Future<void> set(String dni, DniData data) async => _store[dni] = data;

  @override
  Future<void> evict(String dni) async => _store.remove(dni);
}

final lookup = CachingDniLookupService(
  delegate: ReniecSunatLookupService(
    httpClient: DioDniHttpClient(Dio()),
    baseUrl: 'https://your-reniec-sunat-backend.example.com',
  ),
  cache: InMemoryDniCache(),
  ttl: const Duration(minutes: 5),
);

final result = await lookup.lookup('43005787');
switch (result) {
  case DniLookupSuccess(:final data):
    print(data.nombreCompleto);
  case DniLookupNotFound():
    print('DNI not found');
  case _:
    print('Lookup failed');
}

Multi-backend fallback

For setups where you want to try a primary service and fall back to a secondary, use FallbackDniLookupService. It stops the chain on DniLookupInvalidToken by default so you do not hammer a service with bad credentials:

final lookup = FallbackDniLookupService(
  services: [primaryService, secondaryService],
);

final result = await lookup.lookup('43005787');

The retry predicate is configurable — pass retryOn to override which result types allow the chain to continue.

Field Selection

If your app only needs a subset of fields, configure the scanner with a DniFields to reduce CPU usage.

final scanner = DniScanner(
  controller: cameraController,
  fields: DniFields.kyc(),
  onScanComplete: (result) { ... },
);

Built-in presets:

DniFields.minimal() — 4 fields (dni, firstName, lastName, secondLastName)
DniFields.kyc() — 7 fields for KYC flows
DniFields.full() — all 19 fields (default behavior when omitted)

Or define a custom set: DniFields.required({DniField.documentNumber, DniField.firstName, DniField.address}).

Document quad detection

v1.0.0 ships an on-device document-quadrilateral detector. It is a hybrid, non-blocking enhancement — be precise about what that means:

Aspect	Behavior
What fires a capture	OCR readiness + frame stability (via `HuntStateMachine`).
Role of the quad	An annotation: it raises confidence and enables a cleaner perspective crop when a clean 4-corner card boundary is found.
Can the quad block a capture?	No. A present, OCR-confirmed DNI is captured even when the detector returns zero corners on a frame. The quad never vetoes.
Native vs fallback	`DocumentQuadDetector` is the domain port. The OpenCV adapter (`opencv_dart`) is selected when the native binary loads; otherwise a pure-Dart `FallbackQuadDetector` is used. A one-time runtime probe picks the adapter and never throws.

Honest status. On real, text-dense DNI frames the native detector frequently returns no clean quad (text edges are not a card boundary). That is why capture is driven by OCR + stability, not by the quad. Treat quad corners as a best-effort enhancement signal, not a reliable edge detector. Improving quad quality (clean corners on real frames) and post-capture perspective crop are tracked as future work — see Roadmap.

The port surface, for consumers who want to supply their own detector:

abstract interface class DocumentQuadDetector {
  bool get isNativeAvailable;
  QuadDetectionResult detectQuad(QuadFrame frame);
  Uint8List? rectify({required Uint8List imageBytes, required List<QuadCorner> corners});
}

QuadFrame carries a luminance plane + dimensions; QuadDetectionResult carries framingValid and an ordered (TL, TR, BR, BL) corner list that is empty when no quad is found.

Public API

Extraction & consensus

Type	Purpose
`OcrFieldExtractor`	Static + instance extraction coordinator.
`OcrExtractedFields`	Mutable field bag (document number, names, address, ubigeo).
`OcrConsensusAccumulator`	Per-field vote accumulator across frames.
`OcrConsensusResult`	Immutable snapshot of the accumulator.
`MrzFieldStrategy`	MRZ-only extractor (checksum-valid).
`TextOcrFieldStrategy`	Label-anchored text extractor.
`AddressFieldStrategy`	Address + ubigeo extractor with multi-line stitching.
`OcrFieldStrategy`	Interface for custom strategies.
`OcrFieldNormalizer`	Pure normalization helpers (`Ñ` recovery, document, date).
`AddressNoiseFilter`	Peruvian address vocabulary + noise-token filter.
`StringSimilarity`	Levenshtein utilities.
`OcrLogger` / `NoOpOcrLogger`	Observability hook (default no-op).

Field selection & hunt

Type	Purpose
`DniField` (enum)	The 19 extractable DNI fields.
`DniFields`	Immutable field selection (`minimal()`, `kyc()`, `full()`, `required({...})`).
`FieldHunter`	Per-frame field extraction pipeline honoring a `DniFields` selection.
`HuntStateMachine`	OCR-readiness state machine that drives auto-capture eligibility.
`HuntResult` / `ExtractedFields`	Per-side hunt output.

Type	Purpose
`DniScanner`	The single production capture widget.
`DniScanHints`	Configurable, phase-aware rotating guidance (neutral Spanish default).
`DniScanResult` / `DniSideScanResult`	Two-sided / single-side capture payloads.
`DniCaptureMode` (enum)	`auto` \| `manual` \| `hybrid`.
`DniCameraController`	Lifecycle-only controller: OCR-consensus accumulator + lookup pipeline wiring (no capture-state subsystem).
`DniCaptureOrchestrator`	Auto-capture countdown logic.
`DniCaptureState` (sealed)	Capture state hierarchy (`Scanning`, `CountingDown`, `InFlight`, `Expired`, `Done`).
`MotionStillnessGate`	IMU jolt-guard contract (default `SensorsMotionGate`).
`LightingGate`	Mean-luminance + glare scorer for live frames.
`ImageQualityGate`	Post-shutter blur (Laplacian) sharpness gate.
`DocumentValidationResult`	Geometric + OCR validation gate.
`ValidationGate` (enum)	Exhaustive failing-gate cases (incl. `lighting`, `glare`).
`ValidationGateColors`	Presentation-side gate → color mapping.
`KycTheme` / `KycThemeProvider`	Inject visual identity into the capture widget.
`UserVerificationData`	Pre-scan user context for OCR-vs-user matching.

Document quad detection

Type	Purpose
`DocumentQuadDetector`	Domain port for the hybrid, non-blocking quad detector.
`OpenCvQuadDetector`	`opencv_dart`-backed adapter (used when the native binary loads).
`FallbackQuadDetector`	Pure-Dart fallback adapter (used when native is unavailable).
`QuadFrame` / `QuadCorner` / `QuadDetectionResult`	Quad detection value types.

DNI lookup

Type	Purpose
`DniLookupService`	Lookup contract for external DNI data sources.
`ApisPeruLookupService` / `ReniecSunatLookupService`	Built-in backend adapters.
`CachingDniLookupService`	Cache-aside decorator (TTL, consumer-provided `DniCache`).
`FallbackDniLookupService`	Ordered service chain with configurable retry predicate.
`DniCache` / `InMemoryDniCache`	Cache contract + in-memory implementation.
`DniData` / `DniLookupResult`	Lookup model + sealed result type.
`DniHttpClient` / `DioDniHttpClient`	HTTP contract + Dio adapter.

The capture brain — CaptureCoordinator, FrameInput, FramingSignal, and CaptureDecision — is internal and intentionally not exported. It is the single source of truth that owns capture readiness, the 3-2-1 countdown, document presence, and the manual fallback behind DniScanner.

Logging adapter example

import 'package:dni_peru_ocr/dni_peru_ocr.dart';
import 'package:sentry_flutter/sentry_flutter.dart';

class SentryOcrLogger implements OcrLogger {
  const SentryOcrLogger();

  @override
  void breadcrumb(String category, String message, {Map<String, Object?>? data}) {
    Sentry.addBreadcrumb(
      Breadcrumb(
        category: category,
        message: message,
        data: data,
        level: SentryLevel.info,
      ),
    );
  }
}

Architecture

lib/src/
├── domain/           — entities + ports, pure Dart (no Flutter import)
│   ├── entities/     (UserVerificationData, ValidationGate, DocumentSide…)
│   ├── capture/      (DocumentQuadDetector port, MotionStillnessGate, StabilityState)
│   ├── extraction/   (DniField, DniFields, FieldHunter, HuntStateMachine, HuntResult)
│   └── interfaces/   (OcrLogger)
├── data/             — extraction strategies + accumulator
│   ├── strategies/   (Mrz / TextOcr / Address / OcrFieldStrategy)
│   ├── ocr_consensus.dart
│   ├── ocr_field_extractor.dart
│   ├── ocr_field_normalizer.dart
│   ├── address_noise_filter.dart
│   └── string_similarity.dart
├── extraction/       — per-field extractors (dni number, MRZ, dates, ubigeo…)
├── infrastructure/   — ML Kit / camera lifecycle + quad adapters
│   ├── opencv_quad_detector.dart    (opencv_dart adapter; only dartcv4 importer)
│   └── fallback_quad_detector.dart  (pure-Dart fallback)
├── lookup/           — DNI lookup services, cache, decorators, http
└── presentation/     — Flutter widgets + capture brain
    ├── coordinators/ (CaptureCoordinator, CaptureDecision, FrameInput — internal)
    ├── framing/      (FramingSignal — internal)
    ├── controllers/  (DniCameraController — lifecycle-only)
    ├── orchestrators/(DniCaptureOrchestrator + sealed DniCaptureState)
    ├── widgets/      (DniScanner + DniScanHints + sub-widgets)
    └── theme/        (KycTheme + provider)

Follows Clean Architecture (domain has no Flutter import; the quad detector is a domain port with infrastructure adapters). Each layer depends only on its inner neighbours. Strategies follow the Strategy pattern; consensus follows the Accumulator pattern.

The capture subsystem is centralized: a single internal CaptureCoordinator (pure Dart, in presentation/coordinators/) owns capture readiness, the 3-2-1 countdown, document presence, and the manual fallback. It consumes a normalized FrameInput, runs the real readiness path (DocumentSideDetector → HuntStateMachine), unifies framing through one FramingSignal, and emits a CaptureDecision. DniScanner is a thin renderer that acts on those decisions; DniCameraController is lifecycle-only. None of these capture-brain types are exported — only DniScanner and its public configuration are.

Roadmap

v1.0.0 (current)

Platform floor raised to Flutter >=3.38.0 / Dart >=3.10.0 for the opencv_dart Native Assets dependency (the MAJOR bump).
Document quad detection (DocumentQuadDetector port + OpenCV adapter + pure-Dart fallback) shipped as a hybrid, non-blocking enhancement.
Capture subsystem redesign: a single internal CaptureCoordinator owns capture readiness, the 3-2-1 countdown, document presence, and the manual fallback. Cured the front auto-capture reliability, the false "no document" banner on a held front DNI, and the stay-still-after-countdown symptoms.
DniScanner is the single, canonical capture widget; the legacy DniCameraMask was removed.
Configurable DniScanHints (neutral Spanish default).

Future work

Quad quality — real DNI frames frequently return corners = 0 (text edges are not a card boundary). Improving clean-corner detection on real frames is open work; until then the quad stays a non-blocking enhancement and capture is driven by OCR + stability.
Post-capture perspective crop — rectify the captured still using the quad corners when a clean quad is available.

Earlier history

See CHANGELOG.md for the full pre-1.0 history (Clean Architecture refactor, Strategy + Accumulator decomposition, the ubigeo feature, the DNI lookup contract, and the v0.6.x real-world bug-fix cycle).

Planned — sibling library

face_validator_peru: extract face validation + selfie capture into a separate package mirroring this one's structure. Face logic currently lives in the consumer app.

Testing

flutter test                # ~1260 test/testWidgets calls, 1301 cases pass
flutter analyze             # 0 issues in lib/ and example/

The suite includes a device-faithful capture harness that drives real CameraImage-equivalent frame sequences through the real DocumentSideDetector → HuntStateMachine → CaptureCoordinator path, plus a golden capture oracle that pins the sacred "both sides auto-capture" behavior end-to-end. CI runs flutter test and flutter analyze on every push and PR (see .github/workflows/ci.yaml).

License

MIT — see LICENSE.

dni_peru_ocr

Why

Features

Installation

Example

Quick start — headless extraction

Quick start — capture widget

Scan hints

Host-callback safety

Integration requirements

Android

iOS

Tuning parameters and defaults

DNI Lookup

Recommended composition — caching + service

Multi-backend fallback

Field Selection

Document quad detection

Public API

Extraction & consensus

Field selection & hunt

Capture widget & gates

Document quad detection

DNI lookup

Logging adapter example

Architecture

Roadmap

v1.0.0 (current)

Future work

Earlier history

Planned — sibling library

Testing

License

Libraries

dni_peru_ocr package