1 Introduction
This recent years, Automatic Facial Expression Recognition Research (AFER) has put substantial impact on different areas of human-centric computing, such as emotion analysis, affective computing, and robot control [1]. Since different expressions can be characterized with the appearance changes of the face [2], [3], efficient representation of the expression-related appearance-features is a crucial task in AFER. However, due to the different facial traits and external noise factors, the representation of such features should be, simultaneously, discriminative and robust, which is challenging in practice. Moreover, describing the facial expressions using the most active regions on them is beneficial since not all the regions of the face are active in expression changes [4], [5], [6], [7], [8]. Nevertheless, selecting these active regions is challenging due to the diverse facial appearance and expressions of different individuals.