Vijayaraghavan, Prashanth, and Misha Sra. "Handwritten Tamil Recognition using a Convolutional Neural Network.", NEML Poster 2015

Abstract

We classify characters in Tamil, a south Indian language, using convolutional neural networks (ConvNets) into 35 different classes. ConvNets are biologically inspired neural networks. Unlike other vision learning approaches where features are hand designed, ConvNets can automatically learn a unique set of features in a hierarchical manner. We augment the ConvNetJS library for learning features by using stochastic pooling, probabilistic weighted pooling, and local contrast normalization to establish a new state-of-theart of 94.4% accuracy on the IWFHR-10 dataset. Furthermore, we describe the different pooling and normalization methods we implemented and show how well they work in our experiments.