テスト特徴法の多値パタン認識への拡張とその実験的評価

Abstract

An extended Test Feature Classifier (TFC) for classification of patterns with many-valued features is proposed, which is a non-parametric method that is based on combination of features called prime-test features and on voting-based discrimination. In the binary-valued TFC, the features should be coded into some binary representations. Using the classifier was difficult in cases having many features because the long binary representation leads an increase in the computational cost of learning and discrimination. By introducing the Euclidean distance as a metric to define prime-test features, the classifier can directly handle manyvalued features. To confirm the effectiveness of the new prime-test features and of the proposed TFC several quantitative evaluations are performed using an artificial dataset having heavily overlapped distributions and non-linear separability, as well as a real dataset having a large number of patterns and difficult classification tasks.