Abstract
This paper clarifies, by statistic analysis, that the feature distribution extracted from on-line hand-written chinese characters is nonsymmetric and far from the gausian and also presents an adaptive Mahalanobis distance method where, under the assumption that the original distribution has one-side gausians on both sides across its average, an unknown pattern is recognized by the gausian on either side which the unknown pattern belongs to.
Computational results on 160 kinds of on-line hand-written chinese character recognition are reported, which shows that the proposed method significantly improves the rate of recognition over the conventional Mahalanobis distance (7.9% for known patterns, 10.8% for unknown patterns).
