Expand description
Read documents. OCR from naina’s C++ core.
Detect text, recognise it, work out the reading order, and get markdown back — all from one compiled core, with no Python and no OpenCV.
use naina::{Reader, Tier};
let reader = Reader::builder().tier(Tier::Small).build()?;
let page = reader.read_rgb(&rgb, width, height)?;
println!("{}", page.markdown());
for line in page.lines() {
println!("{:.3} {}", line.confidence, line.text);
}§Scripts
Latin and CJK by default; Devanagari (Hindi, Marathi, Nepali, Sanskrit) on request:
let reader = Reader::builder().language("devanagari").build()?;Reading Devanagari with the default alphabet does not fail — it returns plausible-looking Latin at high confidence. Confidence measures certainty within the model’s own alphabet and cannot express “wrong alphabet”, so choose deliberately. An unrecognised language is an error, never a fallback.
§Building
naina links ONNX Runtime. Point at it with ONNXRUNTIME_ROOT, or link an
existing build of the core with NAINA_LIB_DIR.
Structs§
- Line
- One recognised line of text.
- Page
- The result of reading one image.
- Reader
- An open naina context.
- Reader
Builder - Builds a
Reader. - Region
- A layout region.
Enums§
- Error
- Anything the native core can refuse to do.
- Region
Kind - What kind of thing a layout region is.
- Tier
- Model size. Not a licence or capability axis — every model naina ships is Apache-2.0 and every tier does detection, recognition and layout.