A Convolution Neural Networks Based Character and Word Recognition System for Similar Script Languages Kannada and Telugu

Abstract

This paper presents a cross language platform to recognize characters and words of low resource scripts i.e. scripts which do not have standard dataset and the datasets are not available for public access. Indic scripts come from common origin and some of the scripts have a common 3 zonal structure. Recognition of such scripts can be done with the help of other scripts having similar structure. To recognize these characters the model is trained with source language Kannada with zone-wise training and testing is done with both Kannada and the target language Telugu. An accuracy of 88% for Kannada and 62% for Telugu characters is achieved by using Inception Model which is built using Convolution Neural Networks (CNN) image classifier. The dataset consists of 10700 Kannada characters. The model is also tested for 100 words of Telugu and Kannada with an accuracy of 72% and 82% respectively.