pdf_ocr_ondevice library

On-device, downloadable OCR for dart_pdf_editor.

Implements PdfOcrEngine so PdfEditor.applyOcr can add a selectable, searchable, invisible text layer over scanned PDF pages - running entirely on the device, with no per-page network call. The (small, ~21 MB) PP-OCR model is downloaded once via PdfOcrModelManager and then runs locally on ONNX Runtime.

Supported on the native platforms (Android, iOS, macOS, Windows, Linux); on the web use the HTTP-backed pdf_ocr_vlm engine instead.

final manager = PdfOcrModelManager();
final model = PdfOcrModels.ppOcrV5Mobile;
if (!await manager.isDownloaded(model)) {
  await manager.download(model, onProgress: (p) => print(p.fraction));
}
final engine = await OnDeviceOcrEngine.fromDownloadedModel(manager, model);
final editor = PdfEditor(PdfDocument.open(bytes));
await editor.applyOcr(0, engine);
await engine.dispose();

Classes

CtcDecoder

Greedy CTC decoder over a fixed character dictionary.

CtcResult

The text decoded from a recognizer's per-timestep scores.

DetectedBox

A text region found by detection, as an axis-aligned box in the original image's pixel space (top-left origin, y down) with the mean detector probability over its pixels.

DetectionResize

The target size and the per-axis scale factors of a detection resize.

IsolateOcrModelRunner

Runs an unloaded, isolate-sendable OcrModelRunner on a long-lived worker isolate.

OcrImage

A decoded RGBA raster in plain Dart memory - the form the OCR pipeline works in, so all of its geometry/cropping/resizing is pure Dart and unit testable off a GPU.

OcrModelRunner

The inference backend behind OnDeviceOcrEngine: turns a rasterized page into recognized text lines.

OnDeviceOcrEngine

A PdfOcrEngine that recognizes pages on device, with no network call at recognition time - the model is downloaded once (see PdfOcrModelManager) and then runs locally.

OnnxOcrModelRunner

An OcrModelRunner that runs a PP-OCR-style detect-then-recognize pipeline on ONNX Runtime.

PdfOcrDownloadCancelToken

Cooperative cancellation handle for PdfOcrModelManager.download.

PdfOcrDownloadProgress

Progress of a PdfOcrModelManager.download, emitted on the download's stream as bytes arrive.

PdfOcrModel

A complete on-device OCR model: a text-detection network, a text-recognition network, the recognizer's character dictionary, and an optional orientation classifier - the four pieces of a classic detect-then-recognize OCR pipeline (PP-OCR family).

PdfOcrModelFile

One downloadable file that makes up an on-device OCR model - typically an ONNX network or a character dictionary.

PdfOcrModelManager

Downloads, caches, verifies, and removes on-device OCR models.

PdfOcrModels

Built-in model descriptors.

RecognizedTextLine

One recognized text line: the characters and where they sit in the page raster's pixel space (top-left origin, y down - the same space the OCR model worked in). OnDeviceOcrEngine maps pixelBounds back to PDF user space.

Functions

detectionResize(int srcWidth, int srcHeight, {int sideLimit = 960, int multiple = 32}) → DetectionResize

Computes the detection input size for an srcWidth x srcHeight page: longest side scaled down to at most sideLimit, each side rounded to a multiple of multiple (and never below it). Pages already within the limit are only rounded, never upscaled.

extractDetectionBoxes(Float32List probMap, int width, int height, {double threshold = 0.3, double boxScoreThreshold = 0.5, double unclipRatio = 1.6, int minSize = 3, double scaleX = 1.0, double scaleY = 1.0}) → List<DetectedBox>

Extracts text boxes from a DB-style detection probability map.

parseDictionary(String contents, {bool addSpace = true}) → List<String>

Parses a PP-OCR character dictionary file: one token per line, blank lines preserved as a single space (some dicts encode the space as an empty line). A trailing space token is appended when the file does not already end with one, matching PP-OCR's use_space_char default.

recognitionInput(OcrImage crop, {int targetHeight = 48, int maxWidth = 320}) → ({Float32List tensor, int width})

Recognition preprocessing for one cropped text line: resize to the network's fixed targetHeight keeping aspect (width clamped to maxWidth), then normalize to NCHW float32 in [-1, 1] (PP-OCR rec convention: (pixel/255 - 0.5) / 0.5).

toNchwFloat32(OcrImage image, {List<double> mean = const [0.485, 0.456, 0.406], List<double> std = const [0.229, 0.224, 0.225]}) → Float32List

Normalizes image into an NCHW float32 tensor ([1, 3, height, width], channel order R, G, B) using per-channel mean and std on the 0..1 range - the standard ImageNet-style detection input.

Exceptions / Errors

PdfOcrModelDownloadCanceled

Thrown when a model download is canceled by PdfOcrDownloadCancelToken.cancel.

PdfOcrModelException

Thrown when a model file cannot be downloaded, fails its integrity check, or is requested on an unsupported platform.