CORREIA, S.E.N.; http://lattes.cnpq.br/8845965627299767; CORREIA, Suzete Élida Nóbrega.
Resumo:
The automatic recognition of handwritten numerical characters has been a research
topic extensively studied in the last two decades because both of its theoretical value in
pattern recognition and its potential for applications, such as in automatically processing
postal ZIP codes from mail pieces and money amount from bankchecks. However,
it still remains a difficult problem due to the large degree of variability the handwritten
data may exihibit. The wavelet transform is a new tool developed in recent years, with
the localization properties in both time and frequency domains, which simultaneously
provides global and directional features about the numerals images. In this work, a
novel approach for recognition of handwritten numerals is proposed, consisting of three
stages: pre-processing, feature extraction and classification. Preprocessing deals with
image normalization. Feature extraction aims to represent the normalized images by
wavelets coefficients. Classification performs the final decision using a multilayer cluster
neural network trained with the backpropagation algorithm. Experiments were realized
with the wavelet families Haar, Daubechies, Coifiets and Cohen-Daubechies-Feauveau
using the characters of the numerical database of CENPARMI. Results obtained show
that the proposed method yields good performance.