hwr_text_ocr_rus

Handwritten word-level OCR (HWR) model for Russian.

This model is intended for recognizing cropped text snippets / single words from handwritten notebook images (not full-page OCR, use e.g. kotmayyaka/hwr_text_detection_rus).

For best results, feed tight word crops (or short token crops) with minimal surrounding background.

What’s inside

• Checkpoint: ocr_model.ckpt
• Inference helper code:
  • hwr_ocr.py — HWRTextOCR class (load + preprocess + decode)
  • inference.py — CLI example

Intended use

• ✅ Word-level handwritten recognition (Russian)
• ✅ Small cropped regions of text (one token / short piece)
• ❌ Not a full-page OCR pipeline (you need word/line detection & cropping)
• ❌ Not guaranteed to generalize to very different handwriting styles, paper types, or scanning conditions

Quickstart (inference)

1) Install dependencies

pip install torch torchvision pillow

2) Run CLI inference

python inference_ocr.py --image /path/to/word_crop.png --checkpoint ocr_model.ckpt

3) Use from Python

from PIL import Image
from hwr_ocr import HWRTextOCR

ocr = HWRTextOCR(checkpoint_path="ocr_model.ckpt", device="cpu")

img = Image.open("word_crop.png").convert("RGB")
text = ocr.predict(img)

print(text)

Input recommendations

• Prefer tight crops around a single word.
• Avoid large margins; background clutter reduces accuracy.
• If you have a full line/page image, run a detector/segmenter first and then recognize each crop.

Output

• The model outputs a single string (recognized word/text snippet).

Evaluation

Metrics reported in the model card header were obtained on an internal mixed validation split based on:

• ai-forever/school_notebooks_RU
• ai-forever/school_notebooks_EN

License

• MIT