SBB/page_extraction_dataset

Title

Page Extraction Dataset

Description

In digitised cultural heritage items such as books, newspapers and archival records, a problem that can negatively affect OCR are black margins around a page caused by document scanning. In order to enable document layout analysis (DLA), these black margins need to be cropped and the pages need to be extracted correctly. To enable the training of a machine learning model capable of extracting pages, a dataset was created. The machine learning task for which this dataset was collected falls into the domain of image segmentation and, more generally, of computer vision. The dataset was compiled by Vahid Rezanezhad within the research project "Mensch.Maschine.Kultur – Künstliche Intelligenz für das Digitale Kulturelle Erbe" at the Staatsbibliothek zu Berlin – Berlin State Library (SBB). The research project was funded by the Federal Government Commissioner for Culture and the Media (BKM), project grant no. 2522DIG002. The Minister of State for Culture and the Media is part of the German Federal Government.

Homepage

Eynollah – Document Layout Analysis with Deep Learning and Heuristics

Publisher

Staatsbibliothek zu Berlin – Berlin State Library

Dataset Curators

Dr. Vahid Rezanezhad, postdoctoral researcher, Staatsbibliothek zu Berlin – Berlin State Library, Vahid.Rezanezhad@sbb.spk-berlin.de, ORCID: 0009−0007−8041−6506. Vahid Rezhanezhad has studied computational physics at the University of Potsdam, where he received a Ph.D. in 2016. Between 2019 and 2022, he has worked in the QURATOR project and is now employed in the research project "Mensch.Maschine.Kultur" at Berlin State Library; he was responsible for preparing the dataset, labeling and model training.

Dr. Jörg Lehmann, postdoctoral researcher, Staatsbibliothek zu Berlin – Berlin State Library, Joerg.Lehmann@sbb.spk-berlin.de, ORCID: 0000-0003-1334-9693. Jörg Lehmann has studied history and comparative literature and works in the research project "Mensch.Maschine.Kultur"; he was responsible for preparing the data for publication and drafting the datasheet.

Other Contributors

Contributions: The Europeana Newspapers Project Dataset; Poznán Supercomputing and Networking Center (PSNC) Digital Libraries in the frame of the IMPACT project; Dataset for dhSegment (Ares Oliveira & Seguin); Vahid Rezanezhad, project collaborator in the "Human.Machine.Culture" project at Staatsbibliothek zu Berlin – Berlin State Library.

Point of Contact

Clemens Neudecker, Staatsbibliothek zu Berlin – Berlin State Library, Clemens.Neudecker@sbb.spk-berlin.de, ORCID: 0000-0001-5293-8322

Papers and/or Other References

Clausner, C., Papadopoulos, C., Pletschacher, S., & Antonacopoulos, A. (2015). The ENP image and ground truth dataset of historical newspapers. 2015 13th International Conference on Document Analysis and Recognition (ICDAR), 931–935. https://doi.org/10.1109/ICDAR.2015.7333898

Oliveira, S. A., Seguin, B., & Kaplan, F. (2018). dhSegment: A Generic Deep-Learning Approach for Document Segmentation. 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), Niagara Falls, NY, USA, 2018, 7–12. https://doi.org/10.1109/ICFHR-2018.2018.00011

Papadopoulos, C., Pletschacher, S., Clausner, C., & Antonacopoulos, A. (2013). The IMPACT dataset of historical document images. Proceedings of the 2nd International Workshop on Historical Document Imaging and Processing, 123–130. https://doi.org/10.1145/2501115.2501130

Rezanezhad, V., Baierer, K., Gerber, M., Labusch, K., & Neudecker, C. (2023). Document Layout Analysis with Deep Learning and Heuristics. Proceedings of the 7th International Workshop on Historical Document Imaging and Processing (HIP ‘23) Association for Computing Machinery, New York, 73–78. https://doi.org/10.1145/3604951.3605513

Supported Tasks and Shared Tasks

This dataset was not part of a shared task.

AI Category

image segmentation

Type of Cultural Heritage Application

computer vision

(Cultural Heritage) Application Example

page extraction

Distribution

This dataset is distributed by the two named data curators. It is planned to cite the dataset in upcoming publications. The dataset fully complies with the European General Data Protection Regulation GDPR.

Data Access URL

https://doi.org/10.5281/zenodo.15094542

Licensing Information

Creative Commons Attribution 4.0 International CC BY 4.0. Except for the data coming from Berlin State Library (SBB), all the original image files used for creating this dataset have previously been published separately under individual terms and conditions and have therefore not been included in this data publication.

File Format

image/jpg, image/png

Citation Information

@dataset{rezanezhad_2026_15094542,
  author       = {Rezanezhad, Vahid and
                  Lehmann, Jörg},
  title        = {Page Extraction Dataset},
  month        = jan,
  year         = 2026,
  publisher    = {Staatsbibliothek zu Berlin - Berlin State Library},
  version      = 1,
  doi          = {10.5281/zenodo.15094542},
  url          = {https://doi.org/10.5281/zenodo.15094542},
}

Composition

Data Category

content

Media Category

image

Object Type

books, newspapers, historical documents

Dataset Structure

The complete training dataset consists of 3.820 original image files in both .tif and .jpg format, in 1.000 binarized images in .png format and in 3.820 labeled images in .png format. 2.029 original image files were taken from the dataset for dhSegment, 505 original image files were taken from the Europeana Newspapers Project (ENP), 286 from the Poznán Supercomputing and Networking Center (PSNC) dataset compiled within the frame of the IMPACT Project, and 1.000 images from the Berlin State Library (SBB). However, as all the original image files used for creating this dataset have previously been published separately except for the SBB images, this data publication consists only of the 1.000 SBB images as well as the labeled image files and does not comprise the original image files provided by the above-named sources, where the original image files can be requested or downloaded via the links provided above. Moreover, this data publication includes a .csv file containing all the names of the original images as well as their provenance that were used as training data in order to enable a replication of the model training.

Data Instances

Not applicable.

Data Fields

Not applicable.

Compliance with Standard(s)

Not applicable.

Data Splits

Not applicable.

Languages

deu, eng, pol, fra, lat, ita (ISO 639-2)

Descriptive Statistics

Data can be found in three directories: The first directory is named "images" and contains the 1.000 previously unpublished image files from SBB in .jpg format; the second directory is named "images_bin" and contains the same 1.000 images like the first directory, only in binarized form and in .png format; the third directory is named "labels" and contains 3.820 annotated files in .png format. The compressed file amounts up to 1,89 GB. Metadata are not part of the data publication.

Data Collection Process

Curation Rationale

The motivation for creating this dataset was to compile examples of digitised historical documents with black margins around a page with the aim to train a machine learning model capable of extracting the pages from these document scans.

Source Data

Initial Data Collection

The source data were collected by three independent projects focusing on optimising the performance of optical character recognition (OCR) and image segmentation on historical documents: The Europeana Newspapers Project Dataset; the Poznán Supercomputing and Networking Center (PSNC) Digital Libraries dataset, and the dataset for dhSegment (Ares Oliveira & Seguin). To these 2.820 images, 1.000 images from a German-language newspaper were added from the collections held by SBB. The collected documents were printed in Antiqua and Gothic or were handwritten and are in several languages; in part, they were provided alongside with ground truth for OCR. However, this ground truth was not used for the dataset provided here. Rather, from the datasets published by these projects, individual files were selected and annotated in order to provide a broad range of examples suitable to train a model performing the task of page extraction. No modification or normalisation of the selected files was performed. The complete training dataset can be reassembled by collecting the data from the named sources and selecting the individual files using the file listing.

Source Data Producers

The source data were produced by humans within the named research projects. The initial intent of source data collection was to provide ground truth for OCR and image segmentation. However, in this case, the ground truth data were not used.

Digitisation Pipeline

Not applicable.

Preprocessing and Cleaning

No preprocessing of the original image files was performed.

Annotations

As described above in the section "Initial Data Collection" and "Who are the source data producers?", the published dataset consists mainly of annotations. In the case of the ENP and IMPACT-PSNC files, those annotations were not part of the original datasets.

Annotation Process

All 3.820 files have been annotated by Vahid Rezanezhad. A part of the labels were reused from the dhSegment dataset where the labels are available in .png-files. For the rest of the images, Vahid has annotated bounding boxes of page elements using the tool labelme, which were then written in a .json-file; then they were again encoded in a .png-file. The coordinates of the pages are extracted from the page-xml files and encoded in the .png-files. Each class element in the label like background or page is encoded with an integer.

Annotators

The only annotator was the project collaborator Vahid Rezhanezhad; see the information provided about him above.

Crowd Labour

Not applicable.

Data Provenance

The data of the ENP and the IMPACT-PSNC projects as well as the dhSegment dataset can be obtained via the websites linked above. ENP data were published under their own terms and conditions, IMPACT-PSNC data under a CC-BY-3.0 licence, dhSegment data without any dedicated licence. Republication of these datasets was abdicated; rather, they may be requested by users who are interested in establishing a dataset for comparable tasks or who want to replicate the model training. The 1.000 images from SBB have not yet been published, but are now made available as part of this data publication under a CC-BY 4.0 licence.

Use of Linked Open Data, Controlled Vocabulary, Multilingual Ontologies/Taxonomies

Not applicable.

Version Information

There is no previous version of this dataset.

Release Date

2026-01-22

Date of Modification

Not applicable.

Checksums

MD5 checksum of the page_extraction.tar.gz: c15cd826cd43d6401f50e7cf2a892b7a

SHA-256 checksum of the page_extraction.tar.gz: fe0e0a68ccbda11d80d6b8c3ba552b8155379d8291cd5996ccdf608f6a2514e1

Maintenance Plan

Maintenance Level

Limited Maintenance – The data will not be updated, but any technical issues will be addressed

Update Periodicity

It is not foreseen to update the dataset.

Examples and Considerations for Using the Data

This dataset consists of image files and .png files with annotations marking the bounding boxes of page elements. The original images contain text which may be relevant for ethical considerations. However, this text was not used for the training of the model, nor was OCR performed to extract the text. The following sections of the datasheet are therefore not applicable.

Ethical Considerations

Personal and Other Sensitive Information

Not applicable.

Discussion of Biases

Not applicable.]

Potential Societal Impact of Using the Dataset

Not applicable.

Examples of Datasets, Publications and Models that (re-)use the Dataset

So far, this dataset has not yet been reused.

Known Non-Ethical Limitations

Not applicable.

Unanticipated Uses made of this Dataset

Not applicable.