DeepOCR v3

A line-level text recognition model for historical print and handwriting: Fraktur, Antiqua, Kurrent, and medieval and modern Latin script, in one checkpoint. Code, training and architecture: DeepOCR.

Files

File	What it is
`deepocr-v3.pt`	The model weights (EMA), with every architecture setting stored inside
`union_bytes_v3.bin`	Byte-level 6-gram KenLM, fused into the beam search
`union_bytes_v3.dict`	The LM's token list, index-parallel to the model's 261-token byte vocabulary

The LM belongs to this checkpoint. Do not pair it with another one.

Use

from pathlib import Path

import torch
from PIL import Image
import infer
from infer import load_model, transcribe_beam

infer.KENLM_PATH = Path('union_bytes_v3.bin')
infer.LM_DICT_PATH = Path('union_bytes_v3.dict')
infer.LM_WEIGHT = 1.0
infer.BEAM_WIDTH = 10

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model, tokenizer = load_model('deepocr-v3.pt', device)
print(transcribe_beam(model, tokenizer, [Image.open('line.png')], device)[0])

Input is one text line as an RGB image; it is scaled to 256 px high and read up to 6144 px wide. Three flags choose the transcription convention: long_s (ſ), oblique_hyphen (⸗) and typographic_punctuation. Set them in infer.CONVENTION; all off by default.

The model is also the OCR builtin of DeepLabel.

Model

170.5M parameters. A grouped CNN backbone feeds four mixture-of-experts transformer layers (64 experts, 8 active per column, plus one shared expert) and a CTC head over bytes, so it can emit any Unicode character. Trained from scratch.

Evaluation

CER on the held-out validation split, 25,000 lines:

Decoding	CER
Greedy	3.68%
Beam 10	3.57%
Beam 10 + KenLM, weight 1.0	3.03%

Training data

Public datasets of historical print and handwriting. The language model was built from the same training split.

Limitations

Reads single lines only; it needs a line detector or segmentation in front of it.
Reads best on material like its training data: European print and handwriting from the medieval to the modern period. Accuracy drops on other scripts and unusual hands.
The language model favours the training text; for mostly numeric or tabular text, lower LM_WEIGHT or decode without it.

License

Apache-2.0.

Citation

@software{fleischhacker_deepocr,
  author = {Fleischhacker, David},
  title  = {DeepOCR},
  year   = {2026},
  url    = {https://github.com/m2key1/DeepOCR}
}

Downloads last month: -; Downloads are not tracked for this model. How to track