Digitization of Devanagari Handwritten Text Using CNN

Abstract

The Devanagari script used in Nepali, Hindi, Sanskrit, and other languages includes many more characters and modifiers, which makes the task of recognizing handwritten text very difficult. This paper outlines a system for recognizing handwritten Devanagari texts through the use of CNNs. The process begins with the generation of a large data set comprising different types of handwriting. The data is normalized, resized, and filtered to get rid of noise and all the unwanted features in the images to ensure that data quality is as high as possible. To allow for supervised learning, each of the images is associated with the corresponding character. A CNN model is then created and trained on this labeled dataset, which includes convolutional, pooling, and fully connected layers to capture the features of the Devanagari characters. The given model is evaluated using numerous evaluation measures. As per these evaluations, optimization is carried out to improve the precision and dependability of such a system. The refined model is then used to predict the new, unseen handwritten samples to check the generality of the model. Following validation, the model is subjected to more practical applications, and people are shown how the model can be applied in real-life scenarios. This work demonstrates the potential of CNNs to improve HTR technologies and contributes to the process of digitization for the Devanagari script.