On-Line Handwritten KANJI Character Recognition with Mahalanobis' Generalized Distance

筆者らは北陸先端科学技術大学院大学で収集された,オンライン手書き文字データベースのαセットを対象に手書き漢字の認識実験を行っている。漢字の特徴として各画の方向と長さを採用することにし,各文字87文字から各画の長さと方向の平均,分散,方向と長さの共分散を求めた。これを1016文字種全てについて行い特徴辞書とした。認識過程にはマハラノビス距離概念を用いた類似度を使って認識実験を行った。未知文字10,160サンプルについて,実験を行ったところ10,149サンプルを正しく認識し,認識率99.8%を得ることができたので報告する。In this paper, we describe experimental results of handwritten KANJI character recognition using geometrical features with Maharanobis’ Generalized Distance method in which data density is considered. Therefore, many complicated data such as KANJI character images can be easily classified into categories. In this study two experiments were performed. In experiment 1, length and direction features on pen-down strokes were adopted, whereas, in experiment 2, the corresponding features on pen-up strokes were done. Experimental results for IIPL α-set Data Base which JAIST had collected demonstrated that Maharanobis’ Generalized Distance method can show a recognition accuracy rate of 99.8%, and suggested that this method has more advantages than Euclidean Distance in pattern recognition