A New Adaptive Zoning Technique for Handwritten Digit Recognition

In this paper we present a new adaptive zoning technique based on Voronoi tessellation for the task of handwritten digit recognition. This technique extracts features according to an optimal zoning distribution, obtained by an evolutionary-strategy based search. Several experiments have been conducted on the MNIST and the USPS datasets to investigate the proposed approach. Comparisons with regular square zoning reveal that the presented zoning strategy achieves better results for any type of SVM classifier. Furthermore, the proposed zoning method shows that the combination of the adaptive zoning strategy with the Voronoi topology leads to find a distribution of zones able to improve accuracy significantly. As a matter of fact reached accuracies are close to the best algorithms