A Study of Representation Learning for Handwritten Numeral Recognition of Multilingual Data Set

Abstract

Handwritten numeral recognition, a subset of handwritten character recognition is the ability to identify the numbers correctly by the machine from a given input image. Compared to the printed numeral recognition, handwritten numeral recognition is more complex due to variation in writing style and shape from person to person. The success in handwritten digit recognition can be attributed to advances in machine-learning techniques. In the field of machine learning, representation-based learning in deep learning context is gaining popularity in the recent years. Representative deep learning methods have successfully implemented in image classification, action recognition, object tracking, etc. The focus of this work is to study the use of representation learning for dimensionality reduction, in offline handwritten numeral recognition. An experimental study is carried out to compare the performance of the handwritten numerals recognition using SVM-based classifier on raw features as well as on learned features. Multilingual handwritten numeral data set of English and Devanagari numbers is used for the study. The representation learning method used in the experiment is restricted Boltzmann machine (RBM).