特徴抽出過程におけるブースティングの適用による相補的な識別器の生成およびその統合  [in Japanese] Combining Complementary Classifiers generated by Boosting in Feature Transformation  [in Japanese]

Access this Article

Search this Article

Author(s)

Abstract

本稿では、特徴変換にブースティングの枠組を適用した識別器統合手法を提案する。一般的に、複数の識別器を統合するとき、識別性能は向上することが期待できる。しかし、識別器の統合にあたって、二つの重要な課題がある。一点目は、統合する識別器各々の誤り傾向が異なっていなければ(相補性がなければ)、わずかな性能の改善しか得られない点、二点目は、相補的な識別器が生成されたとしても、各々の識別器が与える情報の統合手法が適切でない場合、やはりわずかな性能の向上しか得られないという点である。そこで本稿では、上述した二点を考慮した上で、相補的な識別器の生成手法と、その統合手法について検討を行う。相補的な識別器を生成するにあたっては、Heteroscedastic linear discriminant analysis (HLDA) に基づく特徴変換の過程でブースティングの枠組を適用した。また、統合においては、各々の識別器から出力される尤度の情報を特徴ベクトルとし、このベクトルが張る空間上で Support vector machine (SVM) に基づくパターン認識を行った。提案手法により識別器を統合することで、孤立単語音声認識実験において、統合前と比較し74%の誤りが削減されることがわかった。A framework of system combination using boosting in a feature transformation is proposed. In general, the combination of multiple classifiers improves the classification performance of each classifier. However, there are two important issues in such a system combination. First, the classification performance is not necessarily improved if the classifiers are not complementary. Second, an inappropriate combination makes the performance worse even if the complementary classifiers can be obtained. In this paper, we attempt to solve how to generate and how to combine the complementary classifiers. Aiming at generating the complementary classifiers, the boosting was applied in HLDA based feature transformation. At the combination stage, a pattern recognition using support vector machine was performed, in which a pair of the likelihoods emitted by the classifiers of the first stage was used as a feature parameter. Experimental results showed the effectiveness of proposed method: it reduced the errors by 74% compared to the case without any system combination.

A framework of system combination using boosting in a feature transformation is proposed. In general, the combination of multiple classifiers improves the classification performance of each classifier. However, there are two important issues in such a system combination. First, the classification performance is not necessarily improved if the classifiers are not complementary. Second, an inappropriate combination makes the performance worse even if the complementary classifiers can be obtained. In this paper, we attempt to solve how to generate and how to combine the complementary classifiers. Aiming at generating the complementary classifiers, the boosting was applied in HLDA based feature transformation. At the combination stage, a pattern recognition using support vector machine was performed, in which a pair of the likelihoods emitted by the classifiers of the first stage was used as a feature parameter. Experimental results showed the effectiveness of proposed method: it reduced the errors by 74% compared to the case without any system combination.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 2006(136(2006-SLP-064)), 203-208, 2006-12-22

    Information Processing Society of Japan (IPSJ)

References:  9

Codes

  • NII Article ID (NAID)
    110006248343
  • NII NACSIS-CAT ID (NCID)
    AN10442647
  • Text Lang
    JPN
  • Article Type
    Technical Report
  • ISSN
    09196072
  • NDL Article ID
    8600954
  • NDL Source Classification
    ZM13(科学技術--科学技術一般--データ処理・計算機)
  • NDL Call No.
    Z14-1121
  • Data Source
    CJP  NDL  NII-ELS  IPSJ 
Page Top