Arjuna-OCR [Kn&En]

Layout-aware Kannada + English document OCR: a layout model, a text detector and a CTC recognizer in ONNX, with tables, reading order, per-line confidence and review flags. No language model, no GPU required.

pip install arjuna-ocr          # CPU, works anywhere


arjuna doctor                   # what this machine can do and what the pipeline will use
arjuna ocr page.pdf -o out/     # files, folders, globs or PDFs -> JSON / Markdown / hOCR / ALTO / TSV
arjuna serve --port 8080        # HTTP server

from kanen_infer import KanEnOCR
ocr = KanEnOCR()            # detects hardware, configures itself
doc = ocr.page("page.png")  # kanen-ocr/1.0 JSON
print(ocr.render(doc, "md"))

The models (~170 MB) download once from anandkaman/arjuna-ocr-kn-en-inference at the revision this package was tested against (models v1.2.1) and are cached. The Python package version can move ahead of the model version for packaging fixes; the version stamped into every document is always the model's. Point ARJUNA_MODELS at a local copy for air-gapped installs.

Speed and accuracy. 12.8 pages/s on an RTX 5060 Ti with TensorRT, ~4.6 on the CUDA provider, seconds per page on CPU. On a frozen 40-page private-document set: word F1 0.967, page CER 3.1 %; 96.3 % word accuracy on Mozhi, 93.7 % on MILE. TensorRT engines are built in the background on first use — the pipeline serves immediately on CUDA and swaps each engine in only after it reproduces the output of the session it replaces.

Full documentation, hosted-inference instructions and the evaluation record: the model repositories linked above. Licences: code, recognizer and detector Apache-2.0; the layout model is AGPL-3.0 (see its NOTICE).