biglam/early_printed_books_font_detection

Early Printed Books Font Detection

Photographs of 35,623 pages from books printed between the mid-15th and the end of the 18th century, each labelled by experts with the font group or groups used on the page. This is a mirror of Dataset of Pages from Early Printed Books with Multiple Font Groups by Mathias Seuret, Saskia Limbach, Nikolaus Weichselbaumer, Andreas Maier and Vincent Christlein, deposited on Zenodo in August 2019 and described in their HIP'19 paper.

The page images come from books held by the British Library, Bayerische Staatsbibliothek München, Staatsbibliothek zu Berlin, Universitätsbibliothek Erlangen, Universitätsbibliothek Heidelberg, Staats- und Universitätsbibliothek Göttingen, Stadt- und Universitätsbibliothek Köln, Württembergische Landesbibliothek Stuttgart and Herzog August Bibliothek Wolfenbüttel.

Font group is a practical routing problem before it is a scholarly one: OCR engines need different models for blackletter and for roman type, so knowing which is on the page decides which model to run. The dataset was built for that, and is also of use to book historians tracing which types were in use where and when.

What the labels mean

A font group is a family of types sharing a letterform style, across different sizes and cuts. Ten of them are labelled here, plus two extra classes.

label	what it is
antiqua	Roman type — the upright humanist letterform that modern Latin-script typefaces descend from
italic	Sloped humanist type
textura	Angular blackletter; the style of Gutenberg's 42-line Bible
rotunda	Rounder blackletter, common in Italy and Spain
bastarda	Cursive-influenced blackletter
schwabacher	Rounder, more cursive German blackletter, widespread in the early 16th century
fraktur	German blackletter, dominant from the later 16th century and in use into the 20th
gotico_antiqua	Transitional incunabula type, between blackletter and roman
greek	Greek type
hebrew	Hebrew type
other_font	A font outside the ten groups above
not_a_font	Non-textual content

Structure

Two fields: image (the page photograph, JPEG, resolutions vary) and labels (a sequence of one to five class labels).

The task is multi-label, but most pages are not. 85.2% of training pages and 89.9% of test pages carry exactly one label; five labels occur on 8 training pages and 1 test page.

Label counts, as a percentage of pages in each split:

label	train (24,866)	test (10,757)
antiqua	5,736 (23.1%)	2,282 (21.2%)
not_a_font	5,280 (21.2%)	2,454 (22.8%)
fraktur	5,083 (20.4%)	2,250 (20.9%)
rotunda	3,429 (13.8%)	1,659 (15.4%)
gotico_antiqua	2,319 (9.3%)	270 (2.5%)
italic	2,198 (8.8%)	689 (6.4%)
schwabacher	1,844 (7.4%)	796 (7.4%)
other_font	1,032 (4.2%)	438 (4.1%)
textura	874 (3.5%)	419 (3.9%)
hebrew	709 (2.9%)	337 (3.1%)
bastarda	661 (2.7%)	313 (2.9%)
greek	394 (1.6%)	113 (1.1%)

Counted from labels-training.csv and labels-test.csv in the Zenodo deposit.

Things to check before you train on it

gotico_antiqua is skewed across the splits — 9.3% of training pages against 2.5% of test pages. Per-class test results for it rest on 270 pages and are noisier than the other classes.

The pages were chosen, not sampled. From the paper: "data was cherry-picked, thus it is not statistically representative of what can be found in libraries. For example, as we had a small amount of Textura at the start, we specifically looked for more pages containing this font group, so we can expect that less than 3.6 % of randomly selected pages from libraries would contain Textura." Class frequencies here are not frequencies in the print record.

not_a_font is not simply "no text". It co-occurs with a font label on 567 training and 97 test pages, so it marks non-textual content on a page rather than the absence of type. Two training rows (bsb00031003_15.jpg, bsb10199589_13.jpg) have no label at all and load with an empty labels sequence.

from datasets import load_dataset

ds = load_dataset("biglam/early_printed_books_font_detection", split="train")
row = ds[0]
row["image"]      # the page scan
row["labels"]     # one to five font groups, e.g. [1, 5]
ds.features["labels"].feature.int2str(row["labels"])

Pages are high-resolution scans, so stream the dataset if you do not want the whole thing on disk: load_dataset(..., streaming=True).

Licence

The Zenodo deposit is CC-BY-NC-SA 4.0 — non-commercial, share-alike. The underlying books are held by the libraries listed above and their own reuse terms for the digitised images are not restated in the deposit.

Citation

@dataset{seuret_mathias_2019_3366686,
  author       = {Seuret, Mathias and
                  Limbach, Saskia and
                  Weichselbaumer, Nikolaus and
                  Maier, Andreas and
                  Christlein, Vincent},
  title        = {{Dataset of Pages from Early Printed Books with
                   Multiple Font Groups}},
  month        = aug,
  year         = 2019,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.3366686},
  url          = {https://doi.org/10.5281/zenodo.3366686}
}

The accompanying paper: Mathias Seuret, Saskia Limbach, Nikolaus Weichselbaumer, Andreas Maier and Vincent Christlein, "Dataset of Pages from Early Printed Books with Multiple Font Groups", Proceedings of the 5th International Workshop on Historical Document Imaging and Processing (HIP'19), 2019, 10.1145/3352631.3352640.