International Journal of Computer Theory and Engineering

Abstract

Handwriten Character Recognition (HCR) for Indian Languages is an important problem where there is relatively little work has been done. Particularly difficult is the problem of recognition of Kagunita – the compound characters resulting from the consonant and vowel combination. To recognize a Kagunita, we need to identify the vowel and the consonant present in the Kagunita character image. In this paper, we investigate the use of moments features on Kannada Kagunita. Kannada characters are curved in nature with some symmetry observed in the shape. This information can be best extracted as a feature if we extract moment features from the directional images. So we are finding 4 directional images using Gabor wavelets from the dynamically preprocessed original image. We analyze the Kagunita set and identify the regions with vowel information and consonant information and cut these portions from the preprocessed original image and form a set of cut images. Moments and statistical features are extracted from original images, directional images and cut images. These features are used for both vowel and consonant recognition on Multi Layer Perceptron with Back Propagation Neural Network. The recognition result for vowels are average 86% and consonants are 65% when tested on separate test data. The confusion matrices for both vowels and consonants are analyzed.