To promote further research into OCR for scientific documents, we releases databases that may be suitable for research outside of InftyProject development. We have carefully scrutinized the data in these releases, and expect that the databases can be relied upon with high confidence. Nevertheless, if you experience any problems with them, please contact us.
The datasets, called GTDB-1 and GTDB-2, consist of document page images collected from scientific journals and textbooks. The GTDB-1 dataset contains 31 English articles on mathematics. The GTDB-2 dataset contains 16 articles. Diverse font faces and mathematical notation styles are included in these articles. A list of the articles in both datasets is provided in the bottom of this page.
The statistics of each dataset are shown in the table. The article pages were originally scanned at 600 dpi. The ground truth annotations for each math symbol and ordinary character were attached manually.
Statistics of the datasets
GTDB-1 | GTDB-2 | |
---|---|---|
Number of articles | 31 | 16 |
Number of pages | 544 | 343 |
Number of math symbols | 162,406 | 115,433 |
Number of ordinary text characters | 646,714 | 507,412 |
Here, the ground truth annotation is released. Although we could not include the original document images of articles for copyright reasons, we provide hyperlinks to the web pages of the original documents, where the users can obtain the document images.
Annotations for document pages in one article are provided by one CSV file. The format of CSV files are as following:
Infty GT-Data Format Ver.1.1
Sheet,1,AIF_1970_20_493.png,-1
Text,1,116,411,996,559
Line,1,118,411,996,471
Chardata,1,118,416,159,463,0,-1,-1,0141
Chardata,2,164,428,197,462,0,-1,1,016E
Chardata,3,201,428,234,461,0,-1,2,016E
Chardata,4,239,453,249,464,0,-1,3,142E
Sheet
denote the beginning of a new page. Each field (column) in Sheet
record denotes
1:Sheet, 2:Page ID, 3:Filename of the page image, 4:always -1
. Text
or Image
are corresponding to one document layout component. 1:Text or Image, 2:Component ID, 3-6:Coordinates of the bounding box (left, top, right, bottom)
Line
denote the text line. 1:Line, 2:Line ID, 3-6:Coordinates of the bounding box (left, top, right, bottom)
Chardata
denote the characters or symbols. 1:Chardata, 2:Character ID, 3-6:Coordinates of the bounding box (left, top, right, bottom), 7:Text Mode (0:Ordinary text, 1:Math symbol), 8:Link label, 9:ID of parent character, 10:OCR code
The number in Link label
field represents the positional relationship to the preceding character, called parent character.
0: horizontal
1: right-superscript
2: right-subscript
3: left-superscript
4: left-subscript
5: upper
6: lower
-1: the first character in expressions or ordinary texts
GTDB datasets are distributed under CC BY-NC-ND license. Usage for research, development, or testing of OCR systems (not commercial) for scientific documents is permitted, free of charge.
The copyright and license holder of GTDB datasets is
Masakazu Suzuki:
Science Accessibility Net,
Professor emeritus of Kyushu University
Citation:
W. Ohyama, M. Suzuki and S. Uchida, "Detecting Mathematical Expressions in Scientific Document Images Using a U-Net Trained on a Diverse Dataset," in IEEE Access, vol. 7, pp. 144030-144042, 2019, doi: 10.1109/ACCESS.2019.2945825. link
@ARTICLE{8861044,
author={W. {Ohyama} and M. {Suzuki} and S. {Uchida}},
journal={IEEE Access},
title={Detecting Mathematical Expressions in Scientific Document Images Using a U-Net Trained on a Diverse Dataset},
year={2019},
volume={7},
number={},
pages={144030-144042},
doi={10.1109/ACCESS.2019.2945825}}
In GTDB-1 dataset
In GTDB-2 dataset