Title: Energy-based features for Kannada handwritten digit recognition

Authors: Gururaj Mukarambi; B.V. Dhandra

Addresses: Department of Computer Science, School of Computer Science, Central University of Karnataka, Kadaganchi, Aland Road, Kalaburagi, India ' Department of Statistics, Christ (Deemed University), Bangalore, Karnataka, India

Abstract: In this paper, Kannada handwritten digit recognition system is proposed based on energy features. Ground truth datasets are not available to test the performance of proposed features. Hence, own dataset of Kannada handwritten digits are collected from schools, colleges, business persons and professionals. The digital images are pre-processed using morphological opening operation for removing the noise and bilinear operation is used for normalisation. The normalised image is divided into 16 blocks, and then wavelet filters were applied for each of the 16 blocks and computed the standard deviation for each of them. In this process, a total of 64 standard deviation of the wavelet coefficients are generated of which 48 coefficients are selected as potential features. The average recognition accuracy of 94.80% is achieved using nearest neighbour classifier. The proposed algorithm is free from skew and thinning and it is novelty of the paper.

Keywords: OCR; DWT; nearest neighbour; SVM.

DOI: 10.1504/IJCVR.2020.105684

International Journal of Computational Vision and Robotics, 2020 Vol.10 No.2, pp.156 - 166

Received: 07 Jan 2019
Accepted: 17 Apr 2019

Published online: 04 Mar 2020 *

Full-text access for editors Access for subscribers Purchase this article Comment on this article