indra-ocr-js

The indra-ocr ONNX models in JavaScript — line detection, recognition (Latin + CJK), reading order, orientation, math — in the browser on onnxruntime-web (wasm), in Node on onnxruntime-node. Models stream at runtime from the Hugging Face Hub (~18 MB, cached) or any URL / local folder.

Install

npm install @ducng99/indra-ocr-js

ESM only, Node ≥ 24.

Usage

Node — import "@ducng99/indra-ocr-js"; browser — import "@ducng99/indra-ocr-js/browser":

import { loadIndra, readGray } from "@ducng99/indra-ocr-js";

const ocr = await loadIndra("datalab-to/indra-ocr"); // HF repo, URL or local dir
const page = await ocr.run(await readGray("page.png"));

page.text;      // every line, in reading order
page.markdown;  // headers, lists, tables, math
page.lines;     // [{ box: [x0, y0, x1, y1], text, conf }]
page.blocks;    // [{ label, bbox, text }] — Text, Section-Header, Table, ...

readGray(source, { maxSide }) takes a path or Buffer in Node, a URL, Blob or ImageBitmap (anything createImageBitmap accepts) in the browser, and returns the grayscale { data, width, height } the pipeline takes. Transparent pixels are composited over white. maxSide optionally downscales (aspect kept, never upscaled) so the longer side is at most that many pixels — e.g. { maxSide: 2048 } for large phone photos. In the browser it uses OffscreenCanvas, so it also works inside a Web Worker.

loadIndra(base, opts) options:

• {order, orient, cjk}: false — skip those models (a missing file also degrades gracefully: XY-cut reading order, no rotation, Latin-only)
• math: true — read math lines as inline LaTeX (extra models)
• tableHeads, furniture, adaptive — table building from the layout heads, keep headers/footers, second detection pass
• progress — download callback, {file, loaded, total} per chunk
• providers (Node, default ["cpu"]), wasmPaths (browser) — backend tuning

ocr.run(gray, turn = true) — pass turn = false to skip orientation/deskew. Lower-level: ocr.detect(gray), ocr.recognize(gray, boxes), ocr.orientation(gray); custom runtimes can call IndraOCR.create(backend) with an injected {Tensor, meta, session}.

Tests

npm test                # browser entry, headless Chrome
npm run test:node       # Node entry
npm test page.png ...   # real images; + INDRA_OCR_BIN=/path/to/indra-ocr for a parity check vs the Python CLI

First run downloads the pinned models into test/models/ and puppeteer's Chrome.

Porting

A 1:1 port of the Python indra-ocr package (det.mjs ← det_post.py, rec.mjs ← rec_preproc.py + line_mask.py, order.mjs ← order_post.py, math.mjs ← math.py, layout.mjs ← layout_post.py, vocab.mjs ← vocab.py, rtl.mjs ← rtl.py with unicode.mjs ← Python's unicodedata, lm.mjs ← lm.py, pipeline.mjs ← pipeline.py, imaging.mjs ← cv2), down to half-even rounding, cv2's resize taps and grayscale conversion, so model inputs stay numerically identical.

demo/ is a client-side playground (bun run build), deployed to GitHub Pages and a HF Space by the Deploy demo workflow.

License

Apache-2.0