WAVELET DESCRIPTORS FOR MULTIRESOLUTION RECOGNITION OF HANDPRINTED CHARACTERS

We present a novel set of shape descriptors that represents a digitized pattern in a concise way and that is particularly well-suited for the recognition of handprinted characters. The descriptor set is derived from the wavelet transform of a pattern's contour. The approach is closely related to Feature extraction methods by Fourier series expansion. The motivation to use an orthonormal wavelet basis rather than the Fourier basis is that wavelet coefficients provide localized frequency information, and that wavelets allow us to decompose a function into a multiresolution hierarchy of localized frequency bands. We describe a character recognition system that relies upon wavelet descriptors to simultaneously analyze character shape at multiple levels of resolution. The system was trained and tested on a large database of more than 6000 samples of handprinted alphanumeric characters. The results show that wavelet descriptors are an efficient representation that can provide for reliable recognition in problems with large input variability.