DevChar: An Extensive Dataset for Optical Character Recognition of Devanagari Characters

The advent of cameras has only accelerated the need to digitize content as it helps prevent data corruption by natural processes and enables faster transfer of the data across communities. Handwritten documents and ancient manuscripts form a large part of this data as they call for a need to be translated from the local languages they were written in. The first step into solving this problem is the recognition of handwritten text. Existing handwritten datasets for the Devanagari script can be used for the recognition of individual characters, but they fail to perform well when the text contains matras and conjuncts created by joining character modifiers. This also introduces a dependency between the model and the data source due to required pre-processing for extracting characters recognized by the model from the word itself. These datasets also lack variation in their penmanship which is essential to encompass diversity in the writing style. We present an extensive dataset that addresses these issues. Our dataset has around 4 million characters of varying handwriting styles, complex characters and matras. Training a simple CNN on our data, to detect characters with matras, gave accuracies exceeding 98%. We also show that using this dataset allows a separation of the input data format from the model design, thus allowing researchers to focus on the latter. This dataset is made publicly available at DevChar2020.