Wavelet-based feature extraction from character images

Abstract

A wavelet-based feature extraction algorithm is proposed for character images. The contours of character contain most of information that discriminates the different classes of characters. The proposed algorithm is primarily based on the notion that the wavelet transformation decomposes a 2-dimenional image into three high-frequency sub-bands representing the vertical, horizontal, and diagonal edges. Several schemes of partitioning the sub-bands into blocks are presented that are meshlike or slice-like. The moments are extracted from each of the blocks and they are taken as the features. The low-frequency sub-band is also used to compute the moments. Experiments performed with two character recognition problems showed promising results and the comparison with other features revealed a superior recognition rate of the proposed algorithm. © Springer-Verlag 2003.