Voronoi Tessellation for Effective and Efficient Handwritten Digit Classification

Abstract

The aim of this paper is to explore the\ud properties of a new zoning technique based on Voronoi\ud tessellation for the task of handwritten digit recognition.\ud This technique extracts features according to an optimal\ud zoning distribution, obtained by an evolutionary-strategy\ud based search. Extensive experiments have been conducted\ud on the MNIST dataset to investigate strengths and\ud weakness of the proposed approach. Comparisons with\ud regular square zoning reveal that the presented zoning\ud strategy achieves better results with any type of features.\ud Furthermore, the proposed zoning method, jointly with a\ud suitable choice of features, allows a low complexity\ud classifier to reach excellent performances both in terms of\ud accuracy and speed

Handwritten Digit Recognition Based on Pooling SVM-Classifiers Using Orientation and Concavity Based Features

“…Finally, our results also show that our feature extractor techniques, that are based on rectangular spatial division, are competitive with the result of using Voronoi based zoning [10], whose best result in the MNIST dataset is 0.77%. Our best result in that dataset is 0.81% but without requiring an additional cost for zoning.…”

“…These are based on orientations and concavities, using rectangular spatial division for representing locality. In this vein, we show that a simple spatial division allow us to get competitive results with lower computational cost, in contrast with those based on Voronoi tessellation [10]. We conducted diverse experiments using two standard public databases.…”

“…In the case of handwriting recognition, this technique consists in dividing a pattern image into zones or regions and extracting relevant characteristics from each zone. Many zoning methods have been proposed based on different partitioning strategies [10]. In this regard, Impedovo et al [10] recently proposed a zoning strategy using Voronoi tessellation.…”

“…Many zoning methods have been proposed based on different partitioning strategies [10]. In this regard, Impedovo et al [10] recently proposed a zoning strategy using Voronoi tessellation. They showed that a Voronoi based division is better that a classical rectangular division.…”