Real-time Devanagari Numeral Recognition using Capsule Neural Network

Abstract

Handwritten Devanagari Numeral Recognition is very useful in the field of automation of postal
services, automated cheque processing, automatic number plate detection and digitization of
handwritten government records. Few years back recognition of handwritten documents were
challenging task, but today due to various machine learning algorithms makes this task easy. Still
there is problem related to time, these machine learning algorithm require more time to train as well
as need to send explicit features. Hence automation with less time and less memory is important issue
especially for real-time applications. To solve this purpose we proposed technique based on Capsule
Neural Networks (CapsNet) for the recognition of Handwritten Devanagari numerals, which is
advancement over the Convolutional Neural Networks (CNN) in real-time environment. A CapsNet is
capable of handling a variety of vital features like the pose (position, orientation, and size), velocity,
deformation, texture, hue, albedo, and other features which may get lost or ignored in the deep CNN.
It introduces the concept of dynamic routing between capsules and can improve the recognition
accuracy even when the images are not in their most conventional format. The accuracy achieved by
proposed model is 99.13% which is most sophisticated on our Handwritten Devanagari numeral
dataset.