---
license: other
license_name: polyform-noncommercial-1.0.0
license_link: https://polyformproject.org/licenses/noncommercial/1.0.0/
language:
- en
- fr
- de
- es
- ru
- el
library_name: rten
pipeline_tag: image-to-text
tags:
- ocr
- handwriting
- htr
- latin
- cyrillic
- greek
- crnn
- ctc
- onnx
- rten
- text-recognition
---

# gigapdf-ocr-handwriting — Latin/Cyrillic/Greek handwriting OCR (CRNN + CTC)

A **handwriting** text-line recognizer (HTR) for the [`gigapdf-lib`](https://www.npmjs.com/package/@qrcommunication/gigapdf-lib)
OCR engine. PaddleOCR is printed-text only, so this model adds **handwriting** for the Latin /
Cyrillic / Greek alphabets. It runs on **[RTen](https://github.com/robertknight/rten)** (pure-Rust
ONNX, no C++, no Tesseract) and is invoked **opt-in** (a handwriting model is overconfident on printed
input, so it is kept out of the engine's automatic printed-script selection).

## Architecture

- **CRNN + CTC**, standard ops only: conv backbone (W/4 downsample) → height collapse → 2-layer
  **bidirectional `nn.LSTM`** (256 hidden) → CTC. The standard LSTM exports to a **dynamic-width** ONNX
  `LSTM` op, so the engine feeds each line at its **natural width** — no fixed-width padding.
- **Input**: **grayscale**, height **32**, ink `= 1 − gray` (dark text → 1) on a 0 background,
  tight-cropped to the ink, tensor `[1, 1, 32, W]` (dynamic width).
- **Output**: `[1, T, K+1]` CTC logits. Charlist: classes `0..K-1` = the `dict.txt` alphabet
  (Latin-extended + Cyrillic + Greek, one char per line), class `K` = **CTC blank (last)**.

## Training data

Trained on **real handwriting line corpora** (~108k lines) via the Hugging Face datasets-server
mirrors — **IAM, RIMES, NorHand, NewsEye, Belfort, POPP, Esposalles** (Latin) + a synthetic Cyrillic
handwriting set — **plus** synthetic lines rendered from text corpora × handwriting/print fonts
for breadth on glyphs the real corpora under-cover. Trainer:
`crates/ocr-rten/tools/train_handwriting.py`. The lineage traces to gigapdf's first handwriting model,
which beat Tesseract on IAM (CER 0.309); this is the clean, dynamic-width RTen re-train.

## Files

| File | Use |
|------|-----|
| `model.onnx` | ONNX graph (dynamic batch/width), opset 17 |
| `model.rten` | Converted for the RTen runtime (`rten-convert`) |
| `dict.txt` | alphabet, one char per line (blank is the implicit last class) |

## Usage (RTen, via gigapdf-ocr-rten)

```rust
use gigapdf_ocr_rten::OcrEngine;
// Drop model.rten + dict.txt into <models_dir>/latin_hw/, then call it EXPLICITLY:
let eng = OcrEngine::load_models_dir("models")?;
let lines = eng.recognize_page_handwriting(&rgb_image)?;          // opt-in
// or: eng.recognize_page_with(&rgb_image, gigapdf_ocr_rten::HANDWRITING_MODEL)?;
```

## License

**PolyForm Noncommercial 1.0.0.** Copyright 2025 Rony Licha / QR Communication. Commercial use requires
a separate license.