The Teachable Letter Recognizer

Abstract

The ' Teachable Letter Recognizer" (TLR) is a program for letter learning which uses a method not yet described in the character recognition literature. TLR has two main characteristics: (I) It uses a discrimination tree as a knowledge representation. The discrimination tree is a quadtree with some additions. (2) The two ty pes of global features that are used to characterize a letter are the density of pixels and the overall "color", white, black or grey. TLR is invariant to shifts and shows several interesting effects which are related to human behavior, for example occasionally it becomes confused when learning new letters. TLR's importance as a letter recognition program lies in its ability to recognize some distortions of letters which it has never seen before, and for which it also does not have a transformation algorithm .