Recognition of Handprinted Digits


Thomas M. Breuel
IDIAP, C.P. 609, 1920 Martigny, Switzerland
tmb@idiap.ch

  
This paper describes a system that recognizes hand-printed digits.
The system is based on optimal bounded error matching, a technique
already in common use in general-purpose 2D and 3D visual object
recognition systems in cluttered, noisy scenes. In this paper, we
demonstrate that the same techniques achieve high recognition rates
(up to 99.2\%) on real-world data (the NIST database of hand-printed
census forms and the CEDAR database of digits extracted from U.S.
mail ZIP codes).

As part of the system, we describe a post-processing step for
$k$-nearest neighbor classifiers based on decision trees that can be
used (in place of the usual heuristic methods) for setting
thresholds and improves recognition rates significantly.