A hybrid approach to unconstrained handwritten numerals recognition

Unconstrained handwritten numeral recognition using self-organizing maps (SOM), and self-organizing principal component analysis (PCA) is presented. In the feature-extraction phase, we develop the methods to acquire nonlinear normalization of the numeral image. In the classifying phase, we construct the classifier by two layers: PCA and SOM. To acquire the ability of real-time self-learning, the algorithm of the PCA and SOM are combined together. Experiments on 48000 handwritten numerals show that our technique achieves satisfactory results in terms of the classification accuracy and time