German-Brazilian Newspapers (GBN) Dataset

 The GBN Dataset (v1.0)[1] consists of 152 grayscale page images from eight diferent German-Brazilian historical newspapers, along with their binarized images and ground files. It is a dataset created for evaluation of layout analysis methods.

Complex layouts, poor preservation conditions of the paper, old spelling, different font sizes and the presence of multiple languages in the same page are some of the challenges of historical newspaper layout analysis that are present in the GBN Dataset. There are also pages with a slightly transparent paper and some blurs, which makes the dataset more challenging.

The dataset samples belong to the dbp digital collection from the dokumente.br initiative which has as one of the goals to digitally reconstruct the complete collection of these newspapers.

The pages of eight different newspaper titles were scanned at 600 dpi and are available in PNG format. The following table shows the composition and distribution of these images from the dataset.

Newspaper Title Publishing date Script Language Image size (in pixels) # training pages # testing pages
“Der Gemeindebote” 1935-1938 Fraktur German 3850 x 5480  19 05
“Der Jugendfreund” 1911-1917 Fraktur German from 2650 x 3950 to 5320 x 8050
 15  03
“Der Pioner” 1888-1891 Fraktur / Latin Portuguese / German  7100 x 10590  17  05
“Der Sandwirt” 1934-1940 Fraktur German 4250 x 6020  17  05
“Evangelissch-Lutherisches Kirchenblatt” 1916-1919 Fraktur German 2590 x 3690
 17  05
“Kolonie Zeitung” 1863-1889 Fraktur / Latin Portuguese / German

5470 x 7010

6700 x 8400

7050 x 9300

 17  05
“Gemeindeblatt” 1938 Fraktur German 3850 x 5870
 –  08
“Heimatbote” 1936 Fraktur German 3850 x 5480  –  14

The GBN Dataset is composed of:

  • grayscale images of the pages;
  • binarized images of the pages;
  • ground truth files for each of them (XML PAGE[2] and TXT). The XML PAGE files were produced with the Aletheia software and contain the geometric representation and the class (text, image, graphic or separator) of each region of the page. The TXT files contain the coordinates of each foreground pixel along with its corresponding class:
Class label Description
0 text: text with any font size.
1 image: photos.
2 graphic: family crest and other graphic symbols with text inside.
3 separator: column and line separators.

How to obtain the Dataset

The GBN dataset is released for academic research only and is free to researchers from educational or research institutes for non-commercial purposes.

Training GBN v1.0 Dataset: training-GBNv1.zip

References

  1. Araujo, A. B.  “Análise de layout de página em jornais históricos germano-brasileiros,” Federal University of Paraná, Curitiba, 2019. https://acervodigital.ufpr.br/handle/1884/63706
  2. Pletschacher, S. and Antonacopoulos, A. “The PAGE (Page Analysis and Ground-Truth Elements) Format Framework,” in IEEE 20th International Conference on Pattern Recognition (ICPR), 2010, pp. 257–260.

Contact

Please contact Alessandra Belézia Araujo (alessandra.belezia@ufpr.br) with questions or comments.