doclens

A document scanner for Flutter. Edge detection runs in native code (Apple Vision on iOS, a Kotlin pipeline on Android), but every pixel of UI is a Flutter widget you build and control.

The detected document outline — a 4-corner quad — is streamed to Dart on every frame. The camera preview is a Flutter Texture. So the overlay, the shutter button, the flash toggle, the labels: all yours. No native UI bleeds through unless you ask for it.

Features

Native edge detection — Apple Vision on iOS, a pure-Kotlin CameraX pipeline on Android. The 4-corner quad is streamed to Dart every frame.
Three ways to scan — a one-line drop-in screen, a fully branded custom UI on the package widget, or a hand-off to the OS-native scanner.
Multi-page / batch scanning — DoclensMultiScreen keeps the camera open, collects a stack of pages with a thumbnail rail, and returns them in order; reorder and delete from a built-in page manager.
Large-document scan — DoclensLargeDocScreen captures a document too big for one frame as overlapping pieces (tap a + on any edge, line the next shot up against an overlap ghost, watch a miniview fill in), then stitches them into a single image.
Auto-capture with confirmation — fires once the document is framed and held still, with a brief "hold still" window you can abort.
Continuous autofocus + tap-to-focus, programmatic focus, flash/torch modes, and camera switch.
Perspective-correct crop — the detected quad is dewarped to a clean, flat document image.
Image enhancement & shadow removal — grayscale, shadow-corrected colour ("magic colour"), and near-bitonal black-and-white for OCR. Runs on-device with no bundled model and no extra dependency.
Auto-orientation & rotate — straighten the crop upright from its detected text direction, plus a manual rotateImage API.
Edit corners after capture — drag-the-corners helper with re-warp on save; every handle and button is overridable.

Three ways to scan

There are three ways to use this package, from "one line and done" to "I'll draw everything myself."

1. Drop-in: one line

DoclensScreen is a finished scanner screen. It handles the live preview, auto-capture, and a review step with retake / edit-corners / accept. You get a cropped JPEG back.

final ScanResult? result = await DoclensScreen.scan(context);
if (result == null) return;            // user cancelled
final croppedJpegPath = result.croppedImagePath;

You can tune it without writing any builders:

final result = await DoclensScreen.scan(
  context,
  accentColor: Theme.of(context).colorScheme.primary,
  autoCaptureStabilityDuration: const Duration(milliseconds: 600),
  jpegQuality: 95,
  useLabel: 'Save',
);

Every parameter has dartdoc explaining its default and when you'd want to change it.

Multi-page / batch. DoclensMultiScreen is the batch sibling of DoclensScreen — keep the camera open and collect a stack of pages in one session. Use it as a one-line route:

final List<ScanResult>? pages = await DoclensMultiScreen.scan(context);
if (pages == null) return;             // user cancelled
for (final page in pages) {
  print(page.croppedImagePath);
}

…or mount it directly as a widget and handle the result yourself via onComplete (the batch analogue of DoclensScreen.onCapture). Unlike the single-page widget, DoclensMultiScreen is self-contained — no DoclensController to wire up:

DoclensMultiScreen(
  maxPages: 20,                    // null = unlimited
  imageEnhancement: ImageEnhancement.enhanced,
  autoOrientation: AutoOrientation.auto,
  onPagesChanged: (pages) {
    // fires on every add / remove / reorder — update a counter, etc.
  },
  onComplete: (pages) {
    // pages: List<ScanResult>, in order, when the user taps "Done"
    for (final page in pages) {
      // page.croppedImagePath  — perspective-warped JPEG
      // page.detectedQuad      — Quad in raw image pixel coords
      // page.warpError         — non-null if the warp failed
    }
    Navigator.of(context).pop();   // with onComplete set, you drive navigation
  },
)

The live preview grows a thumbnail rail and a Done button; the review screen's accept button reads Add. Tap the rail to open a page manager that reorders (drag) and deletes pages, and closing with uncommitted pages prompts a discard confirmation. Pass maxPages to cap the batch; every other DoclensScreen knob (enhancement, auto-orientation, overlay style, review builders, …) carries over.

Large documents. DoclensLargeDocScreen is the drop-in for a page too big to fit in one frame — a long contract, a poster, a whiteboard. The user captures one piece, taps a + on any edge of the composite, and lines the next shot up against a translucent overlap ghost of the previous piece; a miniview shows the whole document forming. On Done the pieces are stitched into one image and the route returns its path:

final String? path = await DoclensLargeDocScreen.scan(context);
if (path == null) return;              // user cancelled
// `path` is the stitched composite image on disk.

Every piece of chrome is overridable, true to the package's "you own the UI" stance — plusButtonBuilder, miniviewBuilder, hintBuilder, captureButtonBuilder, plus accentColor, overlapFraction, and ghostOpacity:

DoclensLargeDocScreen(
  accentColor: Theme.of(context).colorScheme.primary,
  overlapFraction: 0.3,                // how much of the last piece to ghost
  hintBuilder: (ctx, edge) => MyCoachHint(edge: edge),
  miniviewBuilder: (ctx, canvas) => MyPageMap(canvas: canvas),
)

The capture flow is built on injectable pieces you can also use directly for a fully custom UI: LargeDocCanvas (the 2-D piece grid), LargeDocSession (the capture → review → merge state machine), LargeDocAligner (overlap correction; the default ManualPlacementAligner trusts the hand alignment), and LargeDocMerger (CanvasLargeDocMerger pastes pieces at their grid slots). Swap in your own aligner or merger via the constructor.

2. Custom UI — full control with `DoclensView`

Want a scanner that matches your brand? Mount DoclensView yourself and supply builders for the overlay, shutter, and flash button. You own the result flow.

class _MyScannerState extends State<MyScanner> {
  final controller = DoclensController();

  @override void initState() { super.initState(); controller.initialize(); }
  @override void dispose()    { controller.dispose(); super.dispose(); }

  @override
  Widget build(BuildContext context) {
    return DoclensView(
      controller: controller,
      overlayBuilder: DoclensView.defaultOverlayBuilder,
      captureButtonBuilder: DoclensView.defaultCaptureButton,
      flashButtonBuilder: DoclensView.defaultFlashButton,
      onCapture: (ScanResult result) {
        // result.croppedImagePath  — perspective-warped JPEG
        // result.rawImagePath      — full uncropped JPEG
        // result.detectedQuad      — Quad in raw image pixel coords
        // result.rawImageSize      — pixel size of the raw image
        // result.warpError         — non-null if the warp failed
      },
    );
  }
}

Each builder slot takes three things: null to draw nothing, a static default for a quick start, or your own widget.

Just want a different shape on the quad?

That's the most common ask, so the package ships a set of ready-made overlays as named constructors on QuadOverlay:

Variant	Look
`QuadOverlay.outline`	A stroked polygon, nothing else
`QuadOverlay.filled`	Stroked polygon with a tinted fill (this is the default)
`QuadOverlay.corners`	Four corner brackets, no connecting lines
`QuadOverlay.cornersFilled`	Corner brackets plus a tinted fill
`QuadOverlay.dots`	A filled dot at each corner
`QuadOverlay.dotsLine`	Corner dots joined by a hairline
`QuadOverlay.glow`	A blurred halo behind a stroked polygon

Drop one straight into an overlayBuilder:

DoclensView(
  controller: controller,
  overlayBuilder: (ctx, quad, status) => QuadOverlay.corners(
    quad: quad,
    status: status,
    accent: Colors.lime,
  ),
  ...,
);

Or name a style on DoclensScreen.scan(...) and skip the builder entirely:

final result = await DoclensScreen.scan(
  context,
  overlayStyle: QuadOverlayStyle.cornersFilled,
  accentColor: Colors.lime,
);

The overlay color tracks detection status on its own: a brighter accent while confirming, your accent when aligned, the warning color when the doc is tilted / tooClose / tooFar, and muted white while still searching.

3. OS-native: hand it to the system scanner

If you don't need custom branding on the camera, just call the OS scanner.

final List<String>? paths =
    await DoclensPlatform.instance.scanWithNativeUI(
  pageLimit: 20,
  allowGalleryImport: true,
);

iOS opens VNDocumentCameraViewController; Android opens ML Kit's GmsDocumentScanner. Both are full-screen and multi-page. You don't need a DoclensController for this one.

How auto-capture works

Auto-capture is a three-step state machine, modeled on the feel of Apple's native scanner:

The detector finds a document-shaped quad → DetectionStatus.aligned.
The quad holds still for autoCaptureStabilityDuration (800 ms by default) → status flips to confirming and the default overlay turns a brighter green.
It stays still for another autoCaptureConfirmationDelay (350 ms) → the shutter fires.

Move the camera during that window and the capture aborts. Every threshold lives on ScannerConfig and is also a parameter on DoclensScreen.scan(...).

Sharpness gate — no blurry captures

When ScannerConfig.enableSharpnessGate is true (the default), auto-capture also waits for the frame to be in focus before firing, not just geometrically aligned. Once the document aligns, the controller locks focus on the quad's centroid and holds the shutter, and status reports DetectionStatus.focusing so the default overlay shows a "focusing, hold steady" hint.

Focus is judged on the native side from a per-frame variance-of-Laplacian sharpness measured inside the quad, surfaced as DetectionEvent.sharpness. A fixed threshold doesn't work, since sharpness numbers swing with scene and distance, so a frame must clear sharpnessFloor (8.0 default) and then plateau near the top of a short rolling window. If focus hasn't cleared by autoCaptureFocusTimeout (2500 ms default), capture fires anyway so nobody gets stuck; the user can review and re-shoot. Set enableSharpnessGate: false for the old geometry-only behaviour.

Continuous autofocus + tap-to-focus

The native session runs continuous autofocus by default on both platforms (.continuousAutoFocus on iOS, CONTROL_AF_MODE_CONTINUOUS_PICTURE on Android), and on iOS it adds the near-distance hint that suits holding an A4 page at arm's length.

Tap-to-focus is on by default (ScannerConfig.enableTapToFocus). Tap the preview and you get a one-shot focus plus auto-exposure at that point, a focus reticle painted where you tapped, and a return to continuous AF after about 3 seconds.

You can also focus from code:

await controller.focusAt(const Offset(0.5, 0.5));  // centre of frame

Editing corners after capture

If the detected quad isn't quite right, let the user drag the corners and re-warp.

EditCornersScreen(
  imagePath: scan.rawImagePath,
  initialQuad: scan.detectedQuad,
  imageSize: scan.rawImageSize,
  onSave: (finalQuad) => controller.warpImage(scan.rawImagePath, finalQuad),
);

Every handle, line, and button here is overridable through builders too.

While a corner is dragged, a magnifier loupe shows the region under the finger so the point being placed is never hidden. It is on by default (showMagnifier) and tunable via magnifierSize and magnifierScale; pass a magnifierBuilder to supply a custom loupe widget (return null to fall back to the bundled one).

Gallery import — run the pipeline on an existing photo

The whole detect → edit → warp flow also works without the camera, on a photo the user already has. detectInImage runs the same native edge detector the live preview uses on a still image on disk, so you can pick a photo from the gallery and dewarp it like a fresh capture:

// Pick a photo however you like (e.g. the `image_picker` package).
final picked = await ImagePicker().pickImage(source: ImageSource.gallery);
if (picked == null) return;

// 1. Detect — native edge detection on the still image.
final detection = await controller.detectInImage(picked.path);
if (detection == null) return; // couldn't read the image

// 2. Edit — seed EditCornersScreen with the detected corners (or a 10%
//    inset when nothing was found), 3. Warp on save.
final warpedPath = await Navigator.of(context).push<String>(
  MaterialPageRoute(
    builder: (_) => EditCornersScreen(
      imagePath: picked.path,
      initialQuad: detection.quadIn,   // pixel-space, ready to use
      imageSize: detection.imageSize,
      onSave: (quad) => controller.warpImage(picked.path, quad),
    ),
  ),
);

detectInImage returns an ImageDetection with the quad in normalized [0,1] coords (or null when nothing document-like is found) plus the image's EXIF-upright pixel size; quadIn gives you the same quad already scaled into pixel coordinates. Like warpImage / rotateImage / recognizeText, it's a pure file operation — no initialize() or camera session required. The example app's Gallery import entry shows the full flow end to end.

Image enhancement & shadow removal

By default the cropped output is a pure dewarp — the original pixels, straightened, with nothing else touched. Set imageEnhancement to apply a post-warp filter to the cropped image (the raw image is never modified):

final result = await DoclensScreen.scan(
  context,
  imageEnhancement: ImageEnhancement.blackAndWhite, // best for OCR
);

Mode	Effect	Good for
`none` (default)	Pure dewarp, unmodified pixels	Archival, your own preprocessing
`grayscale`	Plain desaturate (no shadow handling)	Neutral look, smaller files
`enhanced`	Shadow removal + background whitening, colour kept ("magic colour")	Photos in uneven light
`blackAndWhite`	Shadow removal + adaptive/Otsu threshold, near-bitonal	Plain text, OCR on faint print

enhanced and blackAndWhite genuinely remove uneven lighting and soft shadows — not just global contrast. The technique is the classic illumination-division ("flatten") used by document scanners: estimate the lighting and divide it out. It runs entirely on-device with no bundled model and no extra dependency:

iOS uses Apple's built-in CIDocumentEnhancer (iOS 16+), falling back to CIHighlightShadowAdjust (local shadow lift) on older OSes; blackAndWhite desaturates then binarises with CIColorThresholdOtsu.
Android estimates the background from a heavily downscaled copy and divides it out per pixel (adaptive-mean thresholding for blackAndWhite).

Enhancement applies to both the capture's cropped output and any re-warp done through EditCornersScreen (it travels on the controller's config). It's a knob on ScannerConfig too, so it works from every entry point.

For the absolute best shadow/glare removal, the OS-native scanners (scanWithNativeUI) apply Apple's / Google's own document cleanup — at the cost of using their full-screen UI instead of this package's custom flow.

Auto-orientation & rotate

A capture only knows a document's in-frame orientation — shoot a page sideways or upside-down and the dewarped crop comes out the same way. Set autoOrientation to detect the page's text direction on-device and rotate the crop in 90° steps so it reads upright:

final result = await DoclensScreen.scan(
  context,
  autoOrientation: AutoOrientation.auto,
);

Mode	Effect
`none` (default)	Keep the crop's in-frame orientation
`auto`	Detect the dominant text direction and rotate the crop upright

Detection reuses the OS text APIs already on each platform — Apple Vision's VNRecognizeTextRequest on iOS and Play-services ML Kit text recognition on Android — so no model is bundled (the Android model is delivered on demand by Google Play services, exactly like scanWithNativeUI). The crop is read at each of the four 90° rotations and turned to whichever reads as the most confident text; a blank or purely graphical page (no confident text) is left untouched. Like enhancement, it runs on the cropped output only — the raw image is never rotated — and travels on ScannerConfig, so it applies to captures and to re-warps via EditCornersScreen.

For a manual rotate control (e.g. a button in your review UI), call the controller directly — quarterTurns is clockwise and normalized modulo 4, so -1 and 3 both turn one step the respective way:

final rotatedPath = await controller.rotateImage(scan.croppedImagePath!, 1);

Platform setup

iOS

Minimum iOS 13.0. Add this to Info.plist:

<key>NSCameraUsageDescription</key>
<string>Used to scan documents</string>

On iOS 15+ detection uses VNDetectDocumentSegmentationRequest, and falls back to a docs-tuned VNDetectRectanglesRequest on iOS 13/14. Capture goes through AVCapturePhotoOutput, and the perspective warp uses CIPerspectiveCorrection. Both pixel and EXIF orientation are handled — the details are in doc/decisions.md. The native flow uses VNDocumentCameraViewController.

Android

Minimum API 21. android.permission.CAMERA merges into your manifest automatically.

Detection uses CameraX with a pure-Kotlin Sobel pipeline on the preview path — no OpenCV, no bundled ML model. Capture uses ImageCapture plus android.graphics.Matrix.setPolyToPoly.

The native flow uses ML Kit's GmsDocumentScanner, delivered on demand by Google Play services. On a device without Play services it throws ScannerUnavailableException rather than crashing.

API reference

DoclensScreen — drop-in single-page scanner. DoclensScreen.scan(ctx) pushes a full-screen route, awaits a ScanResult?, and pops itself when the user accepts or cancels; or mount the widget directly and use onCapture.
DoclensMultiScreen — drop-in multi-page / batch scanner. DoclensMultiScreen.scan(ctx) awaits a List<ScanResult>?; or mount the widget directly and use onComplete. Thumbnail rail, reorder/delete manager, maxPages cap.
DoclensLargeDocScreen — drop-in scanner for a document too big for one frame. DoclensLargeDocScreen.scan(ctx) awaits the stitched image path (String?). Captures overlapping pieces via a +-per-edge UI with an overlap ghost and a miniview. Overridable chrome (plusButtonBuilder, miniviewBuilder, hintBuilder, captureButtonBuilder) and injectable LargeDocAligner / LargeDocMerger; the underlying LargeDocCanvas and LargeDocSession are exported for fully custom UIs.
DoclensController — owns a session. Streams: quadStream, statusStream, autoCaptureStream, lowLightStream, previewSizeStream. Methods: initialize(), capture(), warpImage(), rotateImage(), detectInImage(), focusAt(), setFlashMode(), cycleFlashMode(), switchCamera(), pause(), resume(), dispose().
DoclensView — Flutter widget rendering preview + your overlays. Builder slots: overlayBuilder, captureButtonBuilder, flashButtonBuilder, lowLightHintBuilder, debugOverlayBuilder. Handles tap-to-focus when ScannerConfig.enableTapToFocus is true.
EditCornersScreen — drag-the-corners helper with re-warp on save.
QuadOverlay + QuadOverlayStyle — family of pre-built overlay widgets (outline, filled, corners, cornersFilled, dots, dotsLine, glow) with status-driven colour. Pass the enum via DoclensScreen.overlayStyle or use a constructor directly inside an overlayBuilder.
recognizeText() — on-device OCR over an image path. On DoclensController or directly on DoclensPlatform.instance; needs no camera session. Returns an OcrResult (text, blocks, lines, imageSize) with per-block / per-line bounding boxes and confidence.
detectInImage() — run native edge detection on a still image (e.g. a gallery import). On DoclensController or directly on DoclensPlatform.instance; needs no camera session. Returns an ImageDetection (quad, imageSize, quadIn) to feed EditCornersScreen + warpImage — the full detect → edit → warp pipeline off a picked photo.
scanWithNativeUI() on DoclensPlatform.instance — full native-modal scan, returns List<String>?.
ScannerConfig — every feature flag with a sensible default (auto-capture timing, smoothing window, detection throttle, JPEG quality, image enhancement, auto-orientation, flash, lens, lifecycle, telemetry, tap-to-focus, pinch-to-zoom).
Quad — 4-point TL/TR/BR/BL with area, centroid, contains, interpolate, maxCornerDistance, scaleToSize.
ScanResult — croppedImagePath, rawImagePath, detectedQuad, rawImageSize, warpError.
StabilityTracker + QuadSmoother — pure Dart helpers, exposed for tests or custom pipelines.
DetectionStatus — searching, tooFar, tooClose, tilted, aligned, confirming, noPaper.
Exceptions — ScannerPermissionException, ScannerUnavailableException, ScannerInitializationException, ScannerCaptureException.

On-device OCR (text recognition)

Pull the text out of a scan. recognizeText runs full text recognition on any image on disk — typically a capture's croppedImagePath — and returns the transcript plus per-block / per-line bounding boxes and confidence:

final result = await DoclensScreen.scan(
  context,
  imageEnhancement: ImageEnhancement.blackAndWhite, // cleanest for OCR
);
if (result?.croppedImagePath == null) return;

final ocr = await DoclensPlatform.instance.recognizeText(
  imagePath: result!.croppedImagePath!,
);
print(ocr.text);                                   // the full transcript
for (final line in ocr.lines) {
  print('${line.text}  @ ${line.boundingBox}  (${line.confidence})');
}

It needs no camera session (no initialize()), so the call lives on the platform instance and works on any JPEG/PNG on disk. If you already manage a DoclensController, the same method is on it too:

final ocr = await controller.recognizeText(somePath);

OcrResult exposes text (blocks joined by newlines), blocks (each an OcrBlock with lines, a pixel-space boundingBox, and on Android a recognizedLanguage), and a flattened lines getter (OcrLine — text, boundingBox, confidence). All bounding boxes are in the recognised image's pixel coordinates (origin top-left); use OcrResult.imageSize to map them onto a scaled preview. A blank or purely graphical page yields an empty result (OcrResult.isEmpty) rather than an error.

Like auto-orientation, recognition reuses the OS text APIs already on each platform — Apple Vision's VNRecognizeTextRequest (run at the .accurate level) on iOS and Play-services ML Kit text recognition on Android — so no model is bundled (the Android model is delivered on demand by Google Play services, exactly like the OS-native scanner).

Script coverage follows the recogniser: Android uses ML Kit's default Latin-script model; iOS Vision recognises its full language set. For non-Latin scripts on Android, run a dedicated ML Kit script model on the cropped path yourself.

What this package deliberately does NOT do

Multi-page PDF export — returns image paths; assemble a PDF yourself.
Web or desktop targets.

More docs

Architecture — the Dart ↔ native pipeline and the threading rules.
Decisions — every non-obvious design choice, with links to the Apple and Google docs behind it.

License

MIT — see LICENSE.

doclens

Features

Three ways to scan

1. Drop-in: one line

2. Custom UI — full control with `DoclensView`

Just want a different shape on the quad?

3. OS-native: hand it to the system scanner

How auto-capture works

Sharpness gate — no blurry captures

Continuous autofocus + tap-to-focus

Editing corners after capture

Gallery import — run the pipeline on an existing photo

Image enhancement & shadow removal

Auto-orientation & rotate

Platform setup

iOS

Android

API reference

On-device OCR (text recognition)

What this package deliberately does NOT do

More docs

License

Libraries

doclens package

doclens

Features

Three ways to scan

1. Drop-in: one line

2. Custom UI — full control with DoclensView

Just want a different shape on the quad?

3. OS-native: hand it to the system scanner

How auto-capture works

Sharpness gate — no blurry captures

Continuous autofocus + tap-to-focus

Editing corners after capture

Gallery import — run the pipeline on an existing photo

Image enhancement & shadow removal

Auto-orientation & rotate

Platform setup

iOS

Android

API reference

On-device OCR (text recognition)

What this package deliberately does NOT do

More docs

License

Libraries

doclens package

2. Custom UI — full control with `DoclensView`