Publication

Abstract : Handwritten character recognition poses a significant challenge due to the complex nature of the script. The need for accurate Devanagari Handwritten Character recognition is of utmost importance in various applications, like document digitization, text analysis, and language processing tasks. Current recognition techniques still have drawbacks despite recent improvements, especially when it comes to accurately identifying compound characters and preserving high levels of accuracy over a wide range of datasets. This paper introduces a newly created dataset for Devanagari Handwritten Character that includes 10 numerals, 13 vowels, 17 similar-looking consonants and 20 compound characters. The dataset contains a total of 36000 images, with 600 images of each type. Images were collected on an A4 sheet in table format. Collected images are scanned and digitized. Characters which were partially written or incorrectly written were removed before pre-processing. Images are segmented from tables using the contour segmentation method and then stored in 60 different folders. The created dataset is tested with different ML/DL algorithms, with the CNN 2D model demonstrating the highest accuracy of 99.66% among the selected models. In future, the dataset can be used for developing models for Devanagari language recognition and optical character recognition.