Abstract
The authors have developed an automatic caricature system using the principal component analysis for facial feature points. Because of the various influences caused by differences in individual persons and shooting conditions, automatically extracted facial feature points sometimes contain the wrong results. This paper proposes a new method to probabilistically evaluate the wrong extraction of feature points of mouth based on the shapes represented by each of principal components and statistical distribution of principal component scores. Furthermore the misdetections of feature points are corrected by reflecting the characteristics of principal components. The experimental results show the effectiveness of the proposed method for the mouth.
