In designing more accurate character recognition, revealing the differ
ences with other categories in distance function is important. In this
paper, I propose Learning by Discriminant Analysis (LDA) as a method
to learn distance functions. With a weighted Euclidean distance and a
quadratic discriminant function as the original distance functions, LD
A learns parameters by superposing the decision function for searching
on the pattern set of the noticed category. The results for handwritt
en numeral recognition rate improved dramatically and its effectivenes
s was verified. In addition, when the values of the parameters after l
earning are changed and applied in a weighted Euclidean distance so th
at the misread patterns before learning are efficiently segmented and
strong correlations exist between features, appropriate category bound
aries are obtained. When applied to the quadratic discriminant functio
n, the effect of the offset from the normal distribution of features i
s reduced.