Model Card for Sprakbanken/trocr_smi_nor_pred_synth

This is a TrOCR-model for OCR (optical character recognition) of Sámi languages.
It can be used to recognize text in images of printed text (scanned books, magazines, etc.) in North Sámi, South Sámi, Lule Sámi, and Inari Sámi.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image

processor = TrOCRProcessor.from_pretrained("Sprakbanken/trocr_smi_nor_pred_synth")
model = VisionEncoderDecoderModel.from_pretrained("Sprakbanken/trocr_smi_nor_pred_synth")

image = Image.open("path_to_image.jpg").convert("RGB")

pixel_values = processor(image, return_tensors="pt").pixel_values
generated_ids = model.generate(pixel_values)

generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

Model Details

This model is microsoft/trocr-base-printed trained on Sprakbanken/synthetic_sami_ocr_data for 5 epochs, and then fine-tuned on manually annotated and automatically transcribed Sámi data, and manually annotated Norwegian.
See our paper for more details.

Model Description

• Developed by: The National Library of Norway
• Model type: TrOCR
• Languages: North Sámi (sme), South Sámi (sma), Lule Sámi (smj), and Inari Sámi (smn)
• License: CC BY 4.0
• Finetuned from model : microsoft/trocr-base-printed

Model Sources

• Repository: https://github.com/Sprakbanken/nodalida25_sami_ocr
• Paper: "Enstad T, Trosterud T, Røsok MI, Beyer Y, Roald M. Comparative analysis of optical character recognition methods for Sámi texts from the National Library of Norway. Accepted for publication in Proceedings of the 25th Nordic Conference on Computational Linguistics (NoDaLiDa) 2025." preprint

Collection details

This model is a part of our collection of OCR models for Sámi languages.

The following TrOCR models are available:

• Sprakbanken/trocr_smi: microsoft/trocr-base-printed fine-tuned on manually annotated Sámi data
• Sprakbanken/trocr_smi_nor: microsoft/trocr-base-printed fine-tuned on manually annotated Sámi and Norwegian data
• Sprakbanken/trocr_smi_pred: microsoft/trocr-base-printed fine-tuned on manually annotated and automatically transcribed Sámi data
• Sprakbanken/trocr_smi_nor_pred: microsoft/trocr-base-printed fine-tuned on manually annotated and automatically transcribed Sámi data, and manually annotated Norwegian data
• Sprakbanken/trocr_smi_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then on manually annotated Sámi data
• Sprakbanken/trocr_smi_pred_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then fine-tuned on manually annotated and automatically transcribed Sámi data
• Sprakbanken/trocr_smi_nor_pred_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then fine-tuned on manually annotated and automatically transcribed Sámi data, and manually annotated Norwegian

Sprakbanken/trocr_smi_pred_synth is the model that achieved the best results (of the TrOCR models) on our test dataset.

Uses

You can use the raw model for optical character recognition (OCR) on single text-line images in North Sámi, South Sámi, Lule Sámi, and Inari Sámi.

Out-of-Scope Use

The model only works with images of lines of text. If you have images of entire pages of text, you must segment the text into lines first to benefit from this model.

Citation

APA:

Enstad, T., Trosterud, T., Røsok, M. I., Beyer, Y., & Roald, M. (2025). Comparative analysis of optical character recognition methods for Sámi texts from the National Library of Norway. Proceedings of the 25th Nordic Conference on Computational Linguistics (NoDaLiDa).