The University of Sussex

Automatic face recognition using radial basis function networks

A.J. Howell

It is well known that the task of automatic face recognition in dynamic environments is very hard. The key problem is that everyday influences, such as lighting, head pose and expression, can lead to greater variation between images of the same person than between images of different people. However, there must be some essential invariant set of features that allow us to recognise familiar faces. Automated face recognition systems must be robust with respect to everyday variability and capture essential similarities to identify individuals. This thesis investigates the task of real-time face recognition within a small known group of people, using an example-based probabilistic learning scheme to learn and recognise individuals. The artificial neural network model used, the radial basis function (RBF) network, is an exceptionally fast classifier, both in training and subsequent classification phases. In addition, it provides a level of confidence in its output which allows ambiguous data to be discarded. Comparisons with other techniques using a standard database indicate the suitability of our approach. Methods for view-based face representation are discussed and analysed, with emphasis on normalisation and preprocessing techniques. We then investigate how variations, such as pose and resolution, in face images affect recognition performance with RBF networks and explore the generalisation properties of the RBF network, looking specifically at pose, scale and shift invariance. We present experimental work using a novel variant of the RBF network, the `Face Unit' network, which learns to identify one particular individual. We then apply the RBF network to image sequences taken from a less `constrained' environment to assess the suitability of the proposed approach for real-life applications. Finally, we look at the temporal learning abilities of a Time-Delay variant of the RBF network, focusing on simple behaviours based on head rotation,

Download compressed postscript file