Offline Handwritten Devanagari Character Recognition Using Modified Capsule Neural Network

Recognizing offline handwritten characters is challenging due to the uniqueness of different handwriting styles. Over the past few decades, this task has extensively used neural networks. Convolutional Neural Networks (CNNs) struggle with relative position and angle changes in Devanagari characters. A deep capsule network (CapsNet) approach is studied to address these issues, which encapsulates multiple entity attributes using groups of neurons called capsules, preserving information about the entity’s position and posture. The dynamic routing algorithm enhances information exchange between capsules and accurately determines the entity’s pose. This method overcomes the limitations of CNNs, combining their strengths while considering relative positions, thus improving recognition performance. The modifications in CapsNet are performed to recognize Devanagari’s characters. This modified network is designed and tested using the four datasets. The first and second datasets are self-created and comprise 59 Hindi character classes (10 numerals, 36 consonants, and 13 vowels). The third dataset is the Devanagari Handwritten Character Dataset (DHCD), comprised of 36 and 10 character classes of consonants and numerals, respectively, and the fourth dataset is an extension of DHCD (DHCDExtended) comprising 59 classes of Hindi characters. 96.17%, 97.26%, 99.30%, and 99.34% accuracies are obtained from the first, second, third, and fourth datasets, respectively.