概算距離の精度向上による近似最近傍探索の高速化

摘要

This paper presents a new framework based on tempos and rhythms of facial expressions. In order to clarify the relationship between psychological stress and facial expressions. We specifically examine time-series changes of arousal levels obtained from the iterations of spontaneous facial expressions. In this study, we defined spontaneous facial expressions as the actions of human's own choice. We define time-series changes of arousal levels from the maximum expressions to neutral expressions as one tempo. Moreover, we define some tempos as one rhythm. We obtain facial expressions tempos and rhythms using left-to-right Hidden Markov Models (HMMs) . We consider that facial expressions tempos and rhythms are strongly affected by transitory stress with external stimulus in psychological stress. For the evaluation experiment, we created an original facial expressions dataset in three states (normal, pleasant, unpleasant) of twenty subjects. Subsequently, we demonstrate expressions processes affected by transitory stress for extracting and analyzing facial expressions tempos and rhythms in each subject.%登録されたデータからクエリに最も近いものを探し出す最近傍探索問題では，探索誤りを許容することで計算時間を大幅に削減することができ，これを近似最近傍探索問題と呼ぶ．近似最近傍探索は一般に，最近傍点となる確率の高い点を選択し，それらとクエリとの距離を計算するという2段階の処理で実現され，前者が手法の良し藩しを決定する．本稿では，この処理で用いる「概算距離」を計算量を増やすことなく．より高精度に推定することにより，高精度かつ高速な近似最近傍探索，を実現する手法を提案する．実験の結果，50％の精度で比較すると従来手法［1］と比べて，64次元のデータで約4倍，256次元のデータで約2．5倍の処理速度を得ることが確認できた．