I. Introduction
Facial expression recognition plays important role in a variety of applications such as 2D or 3D face animation, intelligence computer system, psychology research and cognitive research. More and more researchers are joined this area. There are many aspects which pose a number of technical challenges: 1) head pose 2) face difference 3) the ambiguous definition of expression. The first point we can deal with through regularizing, but the other two are very difficult to eliminate by preprocessing. On the other hand, in light of recent advances in image and video processing, various features have been proposed for recognizing facial expression since early 1990s [1], [2], [3], [4], [5] and a thorough survey of the existing works can be found in [6], [7], [8]. In general, there are two kinds of features are extracted: local and holistic. Holistic features are appearance based and the face difference will definitely affect the recognition performance. Local features for facial expression recognition are extracted only from local face information, such eyes, mouth etc which only reflect expression changes.