Orus-OCR-0.1

Small, CPU-fast OCR for Bangla, English, and mixed Bangla+English pages — handwriting first. For each page it returns the language (bn / en / bn+en), whether it is handwritten (Yes/No), and the text with its structure (line breaks, indentation, blank lines). Text inside photos and charts is ignored.


Recognizer	1.06M-parameter CTC line recognizer (Indra OCR 1.1 architecture, warm-started from Indra), 211 characters (Latin + Bengali block + typographic marks)
Pipeline size	≈ 8.4 MB loaded (`det.onnx` 0.39 + `orus_rec.onnx` 4.27 + `orus_lm.npz` 3.19 + `orus_pic.onnx` 0.16 + `v2_teff.onnx` 0.40)
Speed (4 CPU threads)	1.3–11.7 ms per line; 130–430 ms per page, within ≈20% of Indra OCR 1.1
Line CER	Bangla handwriting 6.9–16.2% (unseen source 25.8%); English handwriting 12.4–19.0%; printed 4.4–4.6%; prescriptions 6.9%
vs Indra OCR 1.1	Bangla handwriting 6.9% vs 84.6% (BN-HTRd); English handwriting 12.4% vs 18.7% (IAM)

Use

pip install onnxruntime opencv-python numpy rapidfuzz indra-ocr torch timm
# torch and timm are only needed for the handwritten flag (MEAN/STD/page_score in orus/train.py);
# inference runs fully on CPU through ONNX Runtime.
python run.py page.jpg           # language, handwritten Yes/No, structured text
python run.py page.jpg --json    # lines with boxes and confidence, timings
python run.py --demo             # self-check (rendered mixed page; expects CER < 5%)

Python:

import sys; sys.path.insert(0, "orus")
from ocr import OCR
r = OCR(release=".")("page.jpg")
print(r["language"], r["handwritten"]); print(r["text"])
for line in r["lines"]:
    if line["conf"] < 0.8: ...
    # unreliable-line cue: conf >= 0.95 has 1-4% CER, < 0.6 has 27-87%

Every line comes back with a confidence value: at 0.95 and above the error rate is 1–4%, and below 0.8 the line is better treated as unreliable.

Where it struggles

Similar-letter confusions (o/a, r/n, ব/র, ল/ন), Bangla conjuncts, digits and punctuation (27–71% wrong inside those classes), receipts and dot-matrix print, and Bangla handwriting from sources it has never seen.

Repository layout

Path	What
`README.md`	this model card
`LICENSE`	MIT License (Non-Commercial)
`run.py`	top-level wrapper that loads the weights from this repo root
`det.onnx`	line detector (Indra OCR 1.1, upstream)
`orus_rec.onnx`	Orus line recognizer, fp32 (INT8 is not shipped: it loses 7–10 CER points on handwriting)
`orus_rec.pt`	same weights, PyTorch checkpoint (for retraining only)
`orus_lm.npz`	order-5 character LM (Indra npz format), used on low-confidence lines
`orus_pic.onnx`	picture-region net (drops text inside photos/charts)
`v2_teff.onnx` + `v2_teff.json`	handwritten / printed page classifier + threshold
`MD5SUMS`	file checksums (verify with `md5sum -c MD5SUMS`)
`orus/`	frozen inference code: `ocr.py` (pipeline), `ocr_model.py` (network), `ocr_pic.py` (picture net), `tiles.py` and `train.py` (handwritten-flag input)
`fonts/`	the Bangla font the `--demo` self-check renders with (SIL OFL 1.1, see `fonts/OFL.txt`)
`inference/`	the same entry point, kept for convenience

License

Released under MIT License (Non-Commercial) — see LICENSE.

The released det.onnx comes from Indra OCR (datalab-to/indra-ocr, modified OpenRAIL-M: research and small-company use; commercial use is not granted by the upstream). Training used research-only datasets (IAM, BanglaWriting, IMGUR5K, IIIT, BN-HTRd, GNHK, CORD, Swapnil and Doctor's prescriptions); consult each dataset's license before redistribution.

Citation / source

This model is the Orus-OCR-0.1 public release accompanying the source code at https://github.com/MahirHamiAbrar/Orus-OCR-V0.

Model on HuggingFace: https://huggingface.co/mhabrar/Orus-OCR-V0
Code on GitHub: https://github.com/MahirHamiAbrar/Orus-OCR-V0

Downloads last month: -; Downloads are not tracked for this model. How to track