Handwritten numeral recognition based on multiple feature extractors

A new method for handwritten numeral recognition based on four feature extractors (ranging from pure statistical to pure structural) is proposed. This set of features is transformed into a 209-variable feature vector. This transformation has led us to resolve the problem of taking into account structural features as the vector must contain as continuous as possible numerical values. Two feature-evaluation criteria, based on the inter-class/intra-class inertia ratio and the linear correlation matrix, have been investigated for the feature selection phase which makes it possible to reduce the feature space dimensionality to only 157 components instead of the 209 originals. Large-scale statistically-significant samples of handwritten well-segmented numerals, extracted from the NIST Data Base, have shown that this method provides good results. >