Curvelet Entropy for Facial Expression Recognition

Abstract

This paper proposes the use of curvelet entropy for classifying facial expressions from still images. The idea behind this work is that the expressions impose non-rigid motions on the face thereby changing the orientations of facial curves occurring due to different types of expressions. Hence a multiresolution transform like curvelet which refines its domain by using orientation information may be applied for the task of expression classification. Since similarity of facial expressions has earlier been studied using Gabor wavelet which uses filters oriented in different directions on specific feature points in images, the orientation selectivity and information content of curvelet subbands at specific facial points are used here. The information at selected facial points are gathered using the entropy of the corresponding pixel at various subbands. The proposed method is evaluated in the JAFFE and Cohn-Kanade databases without and with cross-validations. Experimental results show that the curvelet subband entropy at selected points may be used to form effective features for classifying facial expressions.