trocr-base-handwritten-hist-swe-2

Swedish Lion Libre

An HTR model for historical Swedish developed by the Swedish National Archives in collaboration with the Stockholm City Archives, the Finnish National Archives and Jämtlands Fornskriftsällskap. The model is trained on Swedish handwriting from the period 1600-1900.

Model Details

Model Description

• Developed by: The Swedish National Archives
• Model type: TrOCR base handwritten
• Language(s) (NLP): Historical Swedish handwriting
• License: apache-2.0
• Finetuned from model: trocr-base-handwritten

Uses

The model is trained on Swedish running-text handwriting dating from the start of the 17th century to the end of the 19th century. Like most current HTR models it operates on a text-line level, so its intended use is within an HTR pipeline that segments the text into text lines, which are transcribed by the model.

Direct Use

The model can be used without fine-tuning on all handwriting but performs best on the type of handwriting it was trained on, which is Swedish handwriting from 1600-1900. See below for detailed test and evaluation results.

Downstream Use

The model can be fine-tuned on other types of handwriting, or if you plan to use it to transcribe some specific material that is within it's domain but not included in the training data, for instance if you got a large letter collection dating from the 17th century, it can be fine-tuned on a small amount of manually transcribed in-domain data, say 20-50 letters, and then used to transcribe the entire collection.

Out-of-Scope Use

The model won't work well out-of-the-box for other languages than Swedish, and it won't work well for printed text.

How to Get Started with the Model

Use the code below to get started with the model, but bear in mind that the image has to be a single text line.

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image
import requests

img_path = 'path/to/image'
image = Image.open(img_path)

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
model = VisionEncoderDecoderModel.from_pretrained('Riksarkivet/trocr-base-handwritten-hist-swe-2')
pixel_values = processor(images=image, return_tensors="pt").pixel_values

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

If you want to transcribe entire pages, consider using HTRflow, a package developed by the Swedish National Archives and intended for streamlining large and small scale HTR/OCR-projects. Install the package, write a pipeline config yaml, where you specify the models to use by their huggingface id, add preprocessing or post-processing steps, and then run the pipeline with htrflow pipeline <path/to/yaml> <path/to/image>. A .yaml file for an entire pipeline, transcribing full pages, could look like this:

```yaml

Demo pipeline for running text

steps:

Region segmentation

• step: Segmentation settings: model: yolo model_settings: model: Riksarkivet/yolov9-regions-1 generation_settings: conf: 0.3 batch_size: 32

Line segmentation

• step: Segmentation settings: model: yolo model_settings: model: Riksarkivet/yolov9-lines-within-regions-1 generation_settings: conf: 0.3 batch_size: 16

• step: TextRecognition settings: model: WordLevelTrocr model_settings: model: Riksarkivet/trocr-base-handwritten-hist-swe-2 generation_settings: batch_size: 16 num_beams: 1

• step: ReadingOrderMarginalia settings: two_page: always

Export to Alto and Page XML

• step: Export settings: dest: outputs/alto format: alto

• step: Export settings: dest: outputs/page format: page ```

See the documentation for the HTRflow package for further instructions on specific steps and customizations.

Training Details

Training Data

We cannot publically release all data the model was trained on, since we ourselves haven't created all the data, but below are links to the datasets we can release publically:

Göteborgs poliskammare 1850-1900
Krigshovrättens domböcker
Svea hovrätt
Bergskollegium
Frihetstidens utskottshandlingar
Carl-Fredrik Påhlmans resejournaler
Trolldomskommissionen
Göta hovrätt
Bergmästaren i Nora
Älvsborgs lösen
Jönköpings rådhusrätt magistrat

Training Procedure

Preprocessing

The text line polygons were masked out and placed against a white background, with dimensions decided by the polygon's bounding box.

Training Hyperparameters

See config.json. - training regime: bf16 - learning rate: 5e-5
- weight decay: 0.01

Evaluation

In-Domain Evaluation Data (Sorted by CER)

These are the character and word error rates on evaluation data taken from the same archives that was included in the training set. The evaluation samples are not part of the training data. The number of samples included in the training set give an indication of how the model improves by fine-tuning it on some specific material within the model's range.

Testing Data

Out-of-Domain Test Data (Sorted by CER)

These are the model's CER and WER on the eval_htr_out_of_domain_lines evaluation set, which contains lines from archives that were not at all included in the training data. So these are the results one would expect if one uses this model out-of-the-box on just any running text document within the model's time span.

Metrics

Character Error Rate (CER)

Character Error Rate (CER) is a metric used to evaluate the performance of a Handwritten Text Recognition (HTR) system by comparing the recognized text to the reference (ground truth) text at the character level.

The CER is calculated using the following formula:

$$ \text{CER} = \frac{S + D + I}{N} $$

Where: - S is the number of substitutions (incorrect characters) - D is the number of deletions (missing characters) - I is the number of insertions (extra characters) - N is the total number of characters in the reference text

A lower CER indicates better recognition accuracy.

Word Error Rate (WER)

Word Error Rate (WER) is a metric used to assess the accuracy of an HTR system at the word level by comparing the recognized text to the reference text.

The WER is calculated using the following formula:

$$ \text{WER} = \frac{S + D + I}{N} $$

Where: - S is the number of substitutions (incorrect words) - D is the number of deletions (missing words) - I is the number of insertions (extra words) - N is the total number of words in the reference text

Similar to CER, a lower WER indicates better word-level accuracy.

Technical Specifications

Model Architecture

See config.json.

More details can be found in the original TrOCR paper.