CArDIS: A Swedish Historical Handwritten Character Dataset with 116,000 characters, 30,000 Swedish names, and 1000 region names
CArDIS dataset, the collection of Swedish historical documents were acquired from Arkiv Digital Company’s document image dataset which has more
than 88 million images created from Swedish Church, census, court records and
other historical documents and these documents were penned between 1800 to
1900 by various anonymous priests in different counties. From which, 64,084
sample Swedish historical handwritten character images were selected to
create the dataset. Each sample of Swedish birth record handwritten document
image has information about a newly born child’s name, born date, baptized date,
born place, father’s name, and mother’s name.
Using the above 64,084 samples of Swedish historical handwritten character document images, 116,000 characters, 30,000 names and 1000 region names have been cropped to generate CArDIS dataset were each character has 4000 samples. This dataset has multiple unique characteristics explained below:
• Degradation: The age of the documents, quality of the paper and ink used play an important role in degradation of the characters in CArDIS dataset. Since, the documents were penned in 19th century by various priests using ink pen on old brown papers, the letters in the dataset are deteriorated and fainted. Additionally, there are multiple disturbances in the data like variation of the background, show-through, weak text stroke, merged characters and bleed-through.
• Handwriting Styles: Every individual is different from another and unique in their own way same like every individual will possess his/her own handwriting style. Considering this fact, the handwriting style, the way of holding pen with different pressure, their choice of choosing materials like using ink dips and ball pens result in distinct writing styles, shapes, sizes, orientations and directions. Ultimately, leading to a endless inter-class variations and new font.
• Presence of Special characters: Since, the documents are written in Swedish language, there are special characters (å, ä, ö) other than standard Latin alphabets causing intra and inter class variations. However, the above characteristics become dominant over the existing handwritten datasets overcoming the restrictions over existing datasets. Thus, CArDIS dataset form a new type of dataset with new and unique characteristics.
This collection of dataset consists of isolated handwritten Latin characters and special Swedish characters. Latin characters consists of alphabets a to z where as Swedish special characters consists of alphabets, å,ä,ö.Note that all the characters collected are lowercase letters only and also each and every character have been cropped manually from 64, 084 historical Swedish birth record handwritten document images. This particular dataset contains 29 classes in which 26 classes represent a to z and the other 3 classes represent å, ä, ö. Each class has 4000 images which in total will be 116,000 lowercase letter images (4000*29). Since, there is no historical handwritten Swedish lowercase letters dataset that provides more than 1,00,000 image samples in RGB color space available till date, this point stands as a motive behind generating this huge dataset. This eventually paves path to further researchers who would continue their research in handwritten lowercase letter recognition systems in Swedish as well as English while promoting segmentation for OCR systems in historical document images. (a). This dataset is in contrast with the existing publicly available handwritten digit datasets (e.g. EMNIST (b))
(a)
(b)
Character dataset 116,000 [Swedish name dataset] (will be available shortly) [Region names of Sweden] (will be available shortly)
If you use any of these datasets, please cite:
Reference:
• Amir Yavariabdi, Huseyin Kusetogullari, Turgay Celik, Sakib Rijwan, Shivani Thummanapally, Johan Hall, “CArDIS: A Swedish Historical Handwritten Character Dataset”, IEEE Access (under review).
• Amir Yavariabdi, Huseyin Kusetogullari, Turgay Celik, Sakib Rijwan, Shivani Thummanapally, Johan Hall, CArDIS: A Swedish Historical Handwritten Character Dataset with 116,000 characters, 30,000 Swedish names, and 1000 region names, March 2022. Accessed on: March 20, 2022. Available: Link.
BibTeX:
• @article{CARDIS2022, title={CArDIS: A Swedish Historical Handwritten Character and Word Dataset, author={Amir Yavariabdi, Huseyin Kusetogullari, Turgay Celik, Sakib Rijwan, Shivani Thummanapally, Johan Hall}, journal={IEEE Access}, year={2022} }
• @misc{CARDIS2022, author = {Amir Yavariabdi, Huseyin Kusetogullari, Turgay Celik, Sakib Rijwan, Shivani Thummanapally, Johan Hall}, title = {CArDIS: A Swedish Historical Handwritten Character Dataset with 116,000 characters, 30,000 Swedish names, and 1000 region names}, howpublished = {\url{https://cardisdataset.github.io/CARDIS/}}, note = {Accessed: 2022-03-20} }