Capsule-Net for Urdu Digits Recognition

Abstract

A capsule is formed when a group of additional neurons is added to existing convolutional layer in a typical convolutional neural network (CNN). Capsules have activity vector that represents instantiation parameters of an object or part of an object. Capsule network has recently been introduced by Hinton to overcome the shortcomings of typical CNN model trained with back-propagation. In this work, we investigate the use of capsule networks for recognition of handwritten digits of Urdu. Our results show that a multi-layer capsule network achieves better results (98.5% accuracy) than deep auto-encoder (97.3% accuracy) and a convolutional neural network (96%), especially when we have digits that are highly overlapped.