Abstract
This paper proposes a method for measuring and estimating emotion level of a facial expression image by using canonical correlation analysis (CCA) and Kernel CCA (KCCA). According to well known circumplex model of emotion by J.A.Russel, we adopt a scheme of representing emotion by two coordinate axes of valence and arousal. Our CCA and KCCA analyze two groups of variables: 1) gray scale images of facial expressions, and 2) their subjectively evaluated values of emotion in the two coordinate axes. In order to reduce error or uncertainty of estimation for emotion level, several weighting functions for reading facial expression are assembled from human gazing property of eye tracking data. Due to insufficient number of training image samples, CCA method often gets into difficulty of singular covariance matrices. A solution for the difficulty is proposed in this paper by introducing intermediate variables via principal component analysis (PCA). Property of nonlinear KCCA method with Gaussian kernel is compared to that of CCA method. The method is applied to databases of still images and movie images of facial expressions. The experimental results show that the proposed method is effective in the sense that it can estimate the level of emotion from facial expression images with comparable estimation uncertainty to human observers.
