Classiication of Facial Expressions with Domain Gaussian Rbf Networks

This paper examines the problem of categorisation of facial expressions through the use of a receptive eld neural network model, based upon novel domain Gaussian network units trained through error backpropagation. Such networks are trained upon images derived from the Ekman and Friesen \Pictures of Facial AAect" database, and they are subsequently able to successfully generalise to images of unseen subjects, and provide qualitative replication of the perceptual confusions common to previous studies. By using digital morphing techniques to produce intermediate frames between the existing stills, we are able to study the space of transitions between endpoint expressions, our results suggest that expressions unrelated to the endpoint images may be perceived during certain transitions, a path far more complex than direct translation through a neutral expression.