Understanding NFC-Net: a deep learning approach to word-level handwritten Indic script recognition

Abstract

This paper presents a deep learning architecture modified for resource-constrained environments, called Non-Fully-Connected Network or NFC-Net, based on convolutional neural network architecture in order to solve the problem of Indic script recognition from handwritten word images. NFC-Net mainly targets resource constraint environment where there is a limited computation power or inadequate training samples or restricted training time. Previous approaches to handwritten script recognition included handcrafted features such as structure-based features and texture-based features. In contrast, here our model learns relatively different features from raw input pixels using NFC-Net. Various parameters of the NFC-Net are adjusted to present a vast and comprehensive study of the neural net in the domain of handwritten script recognition. In order to evaluate the performance of the NFC-Net with suitable parameter estimation, a dataset of 18,000 handwritten multiscript word images consisting of 1500 text words from each of the 12 officially recognized Indic scripts has been considered and a maximum script recognition accuracy of 96.30% is noted. Our proposed model also performs better than some of the recently published script recognition methods in bi-script, tri-script, tetra-script and 12-script scenarios. It has been additionally tested on the RaFD and BHCCD datasets with improved results to prove dataset independency of our model.