Application of suboptimal Bayesian classification to handwritten numerals recognition

Non-parametric estimation of probability densities provides a useful way to realise Bayesian classifiers that may be used for example in OCR problems. The complexity of conventional kernel estimators is however far beyond the acceptable limits for performant systems. We present in this paper a novel learning vector quantization technique (IRVQ) which allows to strongly decrease the complexity of kernel estimators. We apply this original technique to the recognition of handwritten numerals and we prove its interest through high recognition rates coupled with low memory and computational requirements.<
>