A classifier based on distance between test samples and average patterns of categorical nearest neighbors

The recognition rate of the typical nonparametric method "k-nearest neighbor rule (kNN)" is degraded when the dimensionality of feature vectors is large. Another nonparametric method "linear subspace methods" cannot represent the local distribution of patterns, so recognition rates decrease when pattern distribution is not normal distribution. This paper presents a classifier that outputs the class of a test sample by measuring the distance between the test sample and the average patterns, which are calculated using nearest neighbors belonging to individual categories. A kernel method can be applied to this classifier for improving its recognition rates. The performance of those methods is verified by experiments with handwritten digit patterns and two class artificial ones.