Facial Image Synthesis

Facial expression has been central to the study of emotion < cross-ref to emotion > for over a hundred years (Darwin, 1872/1998). Much of what we have learned was made possible by technological breakthroughs: photography in the nineteenth century and film and later video in the twentieth. Today, two new technologies are just beginning to make their potential felt. These are automated facial image analysis and facial image synthesis. A recent review of the former can be found in (Cohn & Kanade, in press). Here we summarize major approaches to facial image synthesis of identity and static and dynamic changes in facial expression. Synthesis of dynamic sequences is referred to as animation. The anatomical structure of the face consists of layers of bone, muscle, subcutaneous fat, and skin. Facial expression results from complex, non-linear interactions among these layers. For the animator attempting to simulate facial appearance or expression, the question is how best to represent these layers and model interactions between them. These questions can be approached from three perspectives. One utilizes computer graphics, a second image processing, and a third integrates both. Graphics-based approaches offer a high degree of flexibility, are computationally inexpensive, and can be integrated easily into animation systems for full-bodied avatars. Most commercial systems use this approach. Image-based approaches are computationally more expensive, more limited in the range of facial expression, and cannot animate full-body avatars. They can, however, produce stunning realism. Hybrid approaches that combine computer graphics and image processing are just emerging.