Synthetic face expressions generated by self organizing maps

We describe a system for reproducing and generating face expressions by representing the significant segments on low-dimensional topology conserving maps. Our aim is to develop a representation of face movements which is derived exclusively from image data and contains no physiological information. The target of this research is real-time animation of face images which can be applied in advanced man-machine interfaces or face recognition tasks. There are several hard problems associated with this project: First, the significant segments of face expressions have to be be identified and a representation has to be found which allows to play back a sequence of facial expressions. Second, the system has to be trainable for the dependencies among the significant segments in order to develop an internal model for effects like variable lighting or small changes in orientation. We show that topology conserving maps and their extension to local linear mapping (LLM) networks allow a very efficient representation of critical face segments and that they are able to produce a sequence of realistic expressions in real time.