Subspace classifiers in recognition of handwritten digits

Abstract

This thesis consists of two parts. The first part reviews the general structure of a pattern recognition system and, in particular, various statistical and neural classification algorithms. The presentation then focuses on subspace classification methods that form a family of semiparametric methods. Several improvements on the traditional subspace classification rule are presented. Most importantly, two new classification techniques, here named the Local Subspace Classifier (LSC) and the Convex Local Subspace Classifier (LSC+), are introduced. These new methods connect the subspace principle to the family of nonparametric prototype-based classifiers and, thus, seek to combine the benefits of both approaches.

The second part addresses the recognition of handwritten digits, which is the case study of this thesis. Special attention is given to feature extraction methods in optical character recognition systems. As a novel contribution, a new method, here named the error-corrective feature extraction, is presented. The prototype recognition system developed for the experiments is described and various options in the implementation are discussed.

For the background of the experiments, thirteen well-known statistical and neural classification algorithms were tested. The results obtained with two traditional subspace methods and ten novel techniques presented in this thesis are compared with them. The results show that the Convex Local Subspace Classifier performs better than any other classification algorithm in the comparison.

The conclusions of this thesis state that the suggested enhancements make the subspace methods very useful for tasks like the recognition of handwritten digits. This result is expected to be applicable in other similar cases of recognizing two-dimensional isolated visual objects.