OCR17+ - Layout analysis and text recognition for 17th c. French prints

This repo contains training data and models for Layout analysis and text recognition for 17th c. French prints

This repo is an updated version of the OCR17 repo.

How to use

Training data is organised per print: * Balzac1624_Lettres_btv1b86262420_corrected * Boyer1697_Meduse_cb30152139c_corrected *

To train a model, all the data needs to added to a single file, prior to the repartition between train, validation and test. To do so: 1. git clone https://github.com/Heresta/datasetsOCRSegmenter17 2. cd datasetsOCRSegmenter17 3. bash build_train_alto_Seg17.sh creates a trainingDataSeg17 directory 4. python train_val_prep.py ./trainingDataSeg17/*.xml creates two new files train.txt (with training data) and val.txt (validation data).

The test.txt file is already prepared for the reproducibility of the test, and evaluate the improvement over time. It was created with 3 title pages, 14 pages containing damage, 2 pages with margin, 14 with decoration, 19 with rubric or signatures (or both), 1 with a running title on bottom of page, 3 pages with decorated drop capitals, 7 with basic drop capitals and 28 basic pages. This test file can also be used for an HTR training test.

Structure

The structure of the repo is the following:

``` Data Print1 alto4eScriptorium pageXmlTranskribus pagexmlTranskribusCorrected png Print2 alto4eScriptorium pageXmlTranskribus pagexmlTranskribusCorrected png

Models | HTR | | bleu.mlmodel | | cheddar.mldmodel | | dentduchat.mldmodel | | README.md | Segment | appenzeller.mlmodel | README.md buildtrainaltoSeg17.sh filesinformations.csv partsdataset.csv trainvalprep.py test.txt segmontoAltoValidator.xsd validatoralto.py README.md ```

The Data directory contains excerpts of 17<sup>th</sup> century books, i.e. scans of selected pages and their encoding in PageXML and ALTO-4 files. Regarding the difference between all these directories, cf. infra, Data production.

The Models directory contains several trained models, three for HTR (more information here) and the second for segmentation (more information here).

• The files propose not only the transcription of the text but also an annotation of the layout using the SegmOnto vocabulary.
• Some old prints have been selected in the OCR17 repo, and are all described individually in their respective folder.

build_train_alto_Seg17.sh is a script to create a .png + ALTO4 dataset from all the print.

files_informations.csv contains all names of dataset's files sorted depending on their main specificity.

parts_dataset.csv contains the percentage of each specificity in this dataset.

segmontoAltoValidator and validator_alto.py are two files to help to validate any xml data added in this repository thanks to a workflow on GitHub. They ahave been taken from HTR-United/cremma-medieval repository.

Data production

Some of used data come from the OCR17 repo, the composition of which started with Transkribus, which needs to be adapted for eScriptorium. Therefore, for each print, we propose 1. export format (pageXmlTranskribus) 2. its prepared form for eScriptorium (pagexmlTranskribusCorrected) 3. the final version exported from eScriptorium (alto4eScriptorium)

About files' segmentation

Types of zones

``` Title: 44 (1.12%)

Main: 1041 (26.45%)

Damage: 236 (6.0%)

Decoration: 357 (9.07%)

DropCapital: 261 (6.63%)

Margin: 45 (1.14%)

Numbering: 802 (20.38%)

RunningTitle: 851 (21.62%)

Signatures: 270 (6.86%)

Stamp: 29 (0.74%) ```

Type of lines

``` Default: 25024 (97.79%)

DropCapitalLine: 461 (1.8%)

Rubric: 104 (0.41%)

```

Type of lines/type of zone

``` Title:

• Default: 218

Main:

• Default: 22 801

• DropCapitalLine: 381

• Rubric: 102

Damage:

• Default: 2

• Rubric: 1

Decoration:

• Default: 5

DropCapital:

• Default: 11

• DropCapitalLine: 79

Margin:

• Default: 134

• DropCapitalLine: 1

Numbering:

• Default: 734

RunningTitle:

• Default: 853

Signatures:

• Default: 261

• Rubric: 1

Stamp:

• Default: 5 ```

  

Contacts

Claire Jahan : claire.jahan[at]chartes.psl.eu

Simon Gabay : Simon.Gabay[at]unige.ch

Cite this dataset

Claire Jahan and Simon Gabay, 17<sup>th</sup> century printed books (ALTO, PAGE-XML and png), 2021, Paris: ENS Paris,
https://github.com/Heresta/datasetsOCRSegmenter17.

Licence

Data is CC-BY, except images which come from Gallica (cf. conditions d'utilisation).