追加型再帰フィッシャー判別による認識性能のオンライン改善

カテゴリ: 部門大会

論文No: GS17-3

グループ名: 【C】平成21年電気学会電子・情報・システム部門大会講演論文集

発行日: 2009/09/03

タイトル(英語): Online Improvement in Recognition Performance by Incremental Recursive Fisher Linear Discreminant

著者名: 太田良平 (神戸大学),小澤 誠一(神戸大学)

著者名(英語): Ryohei Ohta(Kobe University),Seiichi Ozawa(Kobe University)

キーワード: パターン認識|特徴抽出|追加学習|線形判別分析|Pattern Recognition|Feature Extraction|Incremental Learning|Linear Discriminant Analysis

要約(日本語): In conventional Linear Discriminant Analysis (LDA), the number of features is limited to the number of classes minus one. Recursive Fisher Linear Discriminant (RFLD) can eliminate this limitation. RFLD search effective features in the complementary space of a discriminant space recursively. LDA is extended for incremental learning as Incremental LDA (ILDA). In this paper, we proposed Incremental RFLD (IRFLD) which search new features in the complementary space of updated discriminant space. The experimental result shows IRFLD outperforms ILDA in terms of the recognition accuracy in some datasets, but some others worse the accuracy.

PDFファイルサイズ: 3,591 Kバイト