IrakliJani/ka-ocr-v1-line

ka-ocr v1 line OCR

This repository contains a single-line OCR model for Georgian-first typed text.

It is designed for machine-rendered screenshot-style text, not general document or camera OCR.

What this model is for

Use this model when your input is already close to a single horizontal text line, for example:

• cropped app or web screenshots
• quiz-card text lines
• social/news screenshot text lines
• other clean typed UI text

It works best on typed, machine-rendered text.

What this model is not for

This model is not a full multiline OCR system.

It does not do:

• page layout analysis
• paragraph or block segmentation
• multiline reading by itself
• handwriting OCR
• photo OCR for natural camera images
• scene text detection in arbitrary photos

If you want to use it on multiline screenshots, documents, or photos, you should first:

1. detect the text regions yourself
2. break the input into individual lines
3. crop / deskew / normalize those lines
4. run the model on each line separately
5. join the line outputs yourself

In other words: treat this as a line recognizer, not a complete OCR pipeline.

Expected input

Best results come from inputs that are:

• screenshots or screenshot-like crops
• single-line
• reasonably horizontal
• reasonably sharp
• typed text, not handwriting
• high enough contrast between text and background

Output character support

The model is Georgian-first, but it can also emit Latin letters, digits, and punctuation from the included vocabulary.

See:

• config.json
• vocab.json
• specs.json

Files

• best.safetensors — model weights
• config.json — minimal runtime config
• vocab.json — public vocabulary definition
• specs.json — model specs and validation metrics

Training data

Training for this project used:

• Darsala/english_georgian_corpora — English-Georgian Parallel Corpus by Luka Darsalia
• a private author-created dataset

Most of the training data behind this released model came from the private dataset. The private dataset is not included in this repository.

Model summary

• architecture: CRNN + CTC single-line OCR
• image height: 48
• max width: 2048
• hidden size: 128
• width downsample: 4
• parameters: 1,200,414

Main validation snapshot from specs.json:

• CER: 0.00350
• WER: 0.01751

Limitations

This model can fail when:

• the input contains multiple lines at once
• the input is a photo instead of a screenshot-like crop
• the text is handwritten
• text is heavily warped, skewed, blurred, or perspective-distorted
• non-text graphics/icons are mixed into the same crop
• line detection/cropping is poor

License

This model is released under CC BY-NC 4.0.

That means attribution is required, and commercial use is not allowed.

See LICENSE for the short form and the official link.

Citation

If you use this model, please cite this model release. If relevant, also cite the public dataset used in the project.

Model citation

@misc{janiashvili2026ka_ocr_v1_line,
  title={ka-ocr v1 line OCR},
  author={Irakli Janiashvili},
  year={2026},
  url={https://huggingface.co/IrakliJani/ka-ocr-v1-line},
  note={Single-line Georgian-first screenshot OCR model}
}

Dataset citation

@dataset{darsalia2025georgian_corpus,
  title={English-Georgian Parallel Corpus},
  author={Luka Darsalia},
  year={2025},
  url={https://huggingface.co/datasets/Darsala/english_georgian_corpora},
  note={Bachelor thesis project - Tbilisi University}
}