Adaptive Classifier Construction: An Approach to Handwritten Digit Recognition

Abstract

Optical Character Recognition (OCR) is a classic example of decision making problem where class identities of image objects are to be determined. This concerns essentially of finding a decision function that returns the correct classification of input objects. This paper proposes a method of constructing such functions using an adaptive learning framework, which comprises of a multilevel classifier synthesis schema. The schema’s structure and the way classifiers on a higher level are synthesized from those on lower levels are subject to an adaptive iterative process that allows to learn from the input training data. Detailed algorithms and classifiers based on similarity and dissimilarity measures are presented. Also, results of computer experiments using described techniques on a large handwritten digit database are included as an illustration of the application of proposed methods.