Indra OCR 0.3.12026-10-13
Install: pip install indra-ocr. Models: Hugging Face datalab-to/indra-ocr, tag v3.1.
0.3.1
- A model download that fails (no network, no access) or a model folder with
missing files raises indra_ocr.ModelError with one clear message; the CLI
and the viewer print it as one line and exit with 1.
- --version for indra-ocr and indra-ocr-viewer.
- page.skew is rounded to 0.01 degree.
- Project links on PyPI, and absolute image URLs in the PyPI description, so
its images show there.
- README: install from PyPI (pip install indra-ocr); a Limitations section
that leads with complex tables, handwriting and scripts outside the OCR list.
0.3.0 (first release)
Models v3 (Hugging Face datalab-to/indra-ocr, tag v3; v3.1 adds the math models).
- Viewer. A Datalab-styled page with the pipeline options in the page
(source, math, headers and footers, reading order, OCR inside images, dpi,
page limit), show/hide for lines, blocks, block labels and reading order, and
tabs for the rendered markdown, its source, the lines and the JSON.
- convert_many no longer hangs when its worker processes cannot start (with
the spawn start method, code not under if name == "main":, or code
from stdin): it warns and reads the remaining inputs in the calling process.
- Math (opt-in). --math / load(math=True): a line classifier
(mathcls.onnx) and a math recognizer (mathenc / mathdec_init /
mathdecstep.onnx, attention decoder with cached keys and values) write
lines with math as text with inline LaTeX. A reading replaces a line only
with math in it and a mean token probability of 0.9 or more. Omni Parse
Bench: headline 62.1 -> 66.8, text 41.4 -> 57.5, math tests 14.3 -> 56.5,
tables 56.4 -> 54.8; median page ~210 -> ~360 ms (x86, 4 cores, auto mode).
Off by default; the math models download only when it is on.
- Orientation. orient.onnx turns a page that is turned by a quarter or
half turn upright before detection (page.rotation). A turn needs at least
95% from the model. Omni Parse Bench: rotated content 3.5% -> 24.3%.
- All cores. The CLI and the new indraocr.convertmany run one
single-threaded worker process per usable core, one page per task (a single
PDF too); --cores limits them. 40 documents: 30.0 s on 1 core, 8.7 s on 4,
2.8 s on 16 (one process with ORT threads gained at most 1.2x beyond 4
cores). The CLI's --threads is now an alias of --cores.
- Deskew. A page tilted by up to 5 degrees is straightened before
detection (page.skew, about 8 ms a page); boxes are mapped back to the
page as turned. Omni Parse Bench rotation_skew suite: 60.0 -> 76.4.
- Low-confidence lines get no text. A recognized line under 0.5 mean
confidence (empty form fields, checkboxes, rules read as "-", "I", "1")
keeps its box but no text.
- Orientation from a broken text layer. A PDF page OCR'd for a broken font
map is not turned when its glyphs run upright (orient.onnx turns about half
of upright Arabic and Hebrew pages).
- CJK markdown. Lines of Chinese or Japanese join without a space; a
bullet detected as its own line starts a list item.
- Table cells escape only &, < and > (no ').
- Text inside diagrams and charts. The detector finds 51% of the text
inside pictures (12.5% before).
- Right-to-left text layers. Runs in a right-to-left line are ordered right
to left.
- OCR only where it can read. A text layer in a script the recognizer
cannot read is kept even when a trust check fails (not for broken font
maps), and empty lines on such pages are not OCR'd into junk.
- Layout blocks. page.blocks lists the layout blocks (label, box, text)
in reading order; the CLI's JSON output includes them.
- --keep-furniture keeps page headers and footers in the markdown.
- Optional models: script.onnx + rec_cjk.onnx route lines with Chinese
characters to a Chinese recognizer; mask.onnx (CenterMask) masks
neighbouring lines before recognition. Both run only when present.
- bench/opb_indra.py: Omni Parse Bench runner.
- Robust inputs (release test: 4,340 benchmark pages, 1,177 documents with
16,017 pages, 37 edge cases): the file type comes from the file's bytes;
multi-page TIFFs give every page; pages above 100 MP render at a lower dpi;
unreadable, corrupt or password-protected files raise InputError
(password= / --password); a page that fails gives an empty page with
source == "error" instead of ending the document; the CLI continues past
failed files and exits 1. Text outside a PDF's crop box is not read.
- Faster on dense pages: reading-order clustering is O(n^2) instead of
O(n^3) (a 1,984-line page: 53 s -> 0.45 s).
- Bounded memory: ONNX Runtime's CPU arena is off (it grew without bound
on changing batch shapes; a 40-page large scan: 4.9 GB -> 1.4 GB).