Cyrillic-MNIST: a Cyrillic Version of the MNIST Dataset

Cyrillic-MNIST: a Cyrillic Version of the MNIST Dataset Bolat Tleubayev, Zhanel Zhexenova, Kenessary Koishybay, Anara Sandygulova Department of Robotics and Mechatronics, School of Engineering and Digital Sciences Nazarbayev University, Nur-Sultan, Kazakhstan {bolat.tleubayev, zhanel.zhexenova, kenessary.koishybay, anara.sandygulova}@nu.edu.kz Abstract This paper presents a new handwritten dataset, Cyrillic-MNIST, a Cyrillic version of the MNIST dataset, comprising of 121,234 samples of 42 Cyrillic letters. The performance of Cyrillic-MNIST is evaluated using standard deep learning approaches and is compared to the Extended MNIST (EMNIST) dataset. The dataset is available at https://github.com/bolattleubayev/cmnist Keywords: dataset, handwriting recognition, optical character recognition, Cyrillic, letters 1. Introduction Optical Character Recognition (OCR) is the use of computer vision solutions to distinguish printed or handwritten characters in digital images of either scanned documents or photos. Once a text is recog- nized and converted into American Standard Code for Information Interchange (ASCII), it can be edited or searched as if it was created in a word processor. OCR is a language-specific problem, thus the state of the art in one language does not imply that there exists any working usable solution in another. Many languages have their datasets for the use by the researchers in computer vision to advance the field. For example, the Modified National Institute of Standards and Technol- ogy (MNIST) dataset (LeCun et al., 1998) has become a standard benchmark for learning, classification and computer vision systems. An extension of the MNIST, the Extended MNIST (EMNIST) dataset contains dig- its, uppercase and lowercase handwritten letters of En- glish (Cohen et al., 2017). Other examples of handwrit- ten datasets are CASIA (Liu et al., 2011) for Chinese, CMATERdb (Das et al., 2014) for Bangla and Arabic, and HODA (Khosravi and Kabir, 2007) for Farsi. Unfortunately, there is no such data for the handwritten Cyrillic script. According to Iliev (2013), there exist 112 languages that use or have used the Cyrillic alpha- bet. Currently 0.3 billion people, which is 4% of the world’s population use the Cyrillic alphabet on a daily basis (WorldStandards, 2020). It is used exclusively for several languages including Russian, Belorussian, Ukrainian, Bulgarian, Kazakh, Kyrgyz, Macedonian, Serbian, Tajik, Turkmen, Uzbek and others. Originally Cyrillic alphabet comprised of 43 letters in addition to modified and combined Greek and Hebrew letters (in the case of the Cyrillic letters for ts, sh, and ch) (Britan- nica, 2019). The modern Cyrillic alphabets have fewer letters than the original Cyrillic: Russian language has 33 letters, Bulgarian - 30, Serbian - 30, Ukranian - 33. Modern Russian Cyrillic has also been adapted to many non-Slavic languages with an addition of special let- ters. OCR has numerous use cases ranging from postcard address recognition (Palumbo et al., 1992) to automatic license plate recognition in modern traffic control sys- tems (Du et al., 2012; Draghici, 1997). OCR has al- lowed to efficiently exploit the content of low qual- ity historical documents by digitizing them across the globe (Gatos et al., 2004; Kusetogullari et al., 2019). Gordienko et al. (2018) used OCR for the recognition of the historical writings …

Content truncated to 500 words