@tacman1123/tiny-ocr

In-browser OCR using PP-OCRv6 Tiny on onnxruntime-web. The same det.onnx / rec.onnx / dict.txt that imagor's optional ocr() filter (rusto) uses on the server, so a browser result and a server result have the same shape and can be compared directly.

Two calls:

• probe(frame) is detection only. It is cheap enough for a live camera loop and answers "does this look like it has readable text?"
• read(frame) is detection + recognition. It returns an object shaped like imagor's ocr metadata.

import { createTinyOcr } from '@tacman1123/tiny-ocr';

const ocr = await createTinyOcr({ modelBase: '/models/' }); // where det.onnx, rec.onnx, dict.txt are served

const p = await ocr.probe(videoElement);   // { hasText, count, coverage, score, sharpness, boxes, ms }
if (p.hasText) {
  const r = await ocr.read(photoBlob);     // { text, text_score, char_count, regions: [{ text, score, box }], timing }
}

Inputs can be a <video>, <canvas>, OffscreenCanvas, <img>, ImageBitmap, or a Blob/File (EXIF orientation is applied). Box coordinates are in the source image's pixels.

Models

The weights are not in this repo or the package (6.3 MB). Serve them yourself with long cache headers. Get them from byrizki/rusto-rs:

bash scripts/download_models.sh --model ppocrv6 --tier tiny --output-dir models
cd models && cp tiny/* . && shasum -a 256 -c SHA256SUMS

models/SHA256SUMS lists the checksums of the Tiny files this was built and tested against.

The text gate

probe() returns hasText from two thresholds, minBoxes (default 2) and minCoverage (default 0.01 of the frame). Those defaults are starting points, not a validated classifier. Tune them on your own photos (product boxes, envelopes, no-text shots) before relying on them. sharpness (Laplacian variance) is returned for a blur check but is not part of the decision.

Options

createTinyOcr({ modelBase, ort, canvasFactory, numThreads, executionProviders }). ort defaults to a global ort (script tag) or import('onnxruntime-web'). Threaded wasm needs COOP/COEP headers; without them it runs single-threaded. probe()/read() take { maxSide, thresh, boxThresh, unclip, dilate, minTextScore, minBoxes, minCoverage }.

Known limits

• Boxes are axis-aligned. Tilted text gets looser crops than rusto's rotated boxes.
• Tall boxes (height/width ≥ 1.5) are rotated 90° counter-clockwise before recognition. Upside-down text is not corrected.
• No orientation detection, no multi-page handling, no table or layout.
• Tiny's text-presence recall has not been validated on dense or unusual scripts.

Develop

npm i && npm test        # Node: onnxruntime-node + canvas shim, needs models/ populated
python3 -m http.server   # then open /demo/ (camera needs HTTPS or localhost)

MIT. Model weights are PaddleOCR's; see their license.

test/sample.png is from byrizki/rusto-rs (MIT); the pre/post-processing here follows its implementation.