magwrap/cnn-ctc-ocr-sme

CNN-CTC OCR for North Sámi (sme)

Lightweight CNN + CTC line-level OCR model for North Sámi (sme). Trained from scratch on the Språkbanken synthetic North Sámi corpus.

• Architecture: 7-layer SimpleCNN backbone → adaptive column pooling → linear projection → CTC head. No RNN encoder.
• Parameters: 5,785,100 (~23 MB checkpoint)
• Vocabulary: 395 characters + CTC blank
• Input: grayscale line image, 32 × 800 px
• Training data: 276,649 lines (Språkbanken synthetic, validation 30,738)
• Optimiser: AdamW, lr=1e-4, weight_decay=1e-4, batch=32, 100 epochs, ReduceLROnPlateau, grad-clip 5.0

Results

Split	CER	WER	Char acc
Språkbanken synthetic (val, 30,738 lines)	2.04%	7.60%	97.96%
Benchmark test set (1,048 lines)	12.24%	37.35%	87.76%

Character accuracy is reported as 100 − CER. The 12.24% CER figure is the headline number reported in the accompanying papers. The lower 2.04% CER is the in-distribution validation result during training (best epoch, 97/100). For reference, sequence-level (full-line exact-match) accuracy at the best epoch was 80.76%.

Usage

The checkpoint is a plain torch.save pickle containing the state dict, character vocabulary, and architecture config. Loading requires the model code from the source repo: https://github.com/magwrap/north-sami-ocr.

import torch
from huggingface_hub import hf_hub_download
from src.ocr.models import load_checkpoint

ckpt_path = hf_hub_download(
    repo_id="magwrap/cnn-ctc-ocr-sme",
    filename="checkpoint_best.pt",
)
model, char_to_idx, idx_to_char, config = load_checkpoint(ckpt_path, device="cpu")

For preprocessing and greedy CTC decoding see src/ocr/pipeline.py in the source repo.

Training

Full per-epoch loss / CER / WER curves are in train.log. The training command is in config.json. Best checkpoint was saved at epoch 97 (val CER 0.0204).

Intended use and limitations

Intended use. Line-level OCR of pre-segmented printed North Sámi text — heritage digitization, transcription pipelines, and OCR → MT workflows. The 23 MB footprint and ~30 ms/image CPU inference make it suited to resource-constrained deployment (mobile, edge, batch processing without GPU).

Limitations. Trained only on synthetic Språkbanken renderings, so robustness to physical scan degradation (faded ink, paper damage, skew) is untested. The fixed 32 × 800 px input degrades on long lines — sentences past ~80 characters benefit from conjunction-based pre-splitting. Expects pre-segmented line images (no layout analysis). No morphological post-correction, so rare inflected forms and out-of-distribution fonts (historical blackletter, handwriting) are out of scope.

Responsible use. Follows the CARE Principles for Indigenous Data Governance — please use this model in support of, not as a substitute for, Sámi community-driven preservation work.

Citation

If you use this model, please cite the accompanying paper(s) — see https://github.com/magwrap/north-sami-ocr for the current reference.

License

Apache-2.0. Training data: Språkbanken synthetic North Sámi corpus — see source dataset for its own terms.