Abstract
This paper proposes a method for estimating fuzzy membership functions of human emotion from facial expression images. The membership functions are based on the dimensional model of emotion and are defined on the emotion space with coordinate axes of valence and arousal. We discuss two models of membership functions. a) The first model uses results of subjective evaluation experiments for learning facial expression images where mean value and standard deviation for each facial image are respectively adopted to mean and spread parameters of the corresponding membership function. b) The second model is a modification of the first model by introducing information of the distribution of the data in the facial image space. The performance of our method is evaluated by approximation error of the estimated membership functions to the original ones. As statistical techniques, canonical correlation analysis (CCA) and kernel CCA are employed and applied to gray-scale image based data. Using two facial image databases of male and female, our experimental results show valid estimation results which have at most less than twice as much dispersion of that of subjective evaluation by human observers, and in the sense of the dispersion, availability of the proposed method is indicated.
