CSM-178 - Learning to Recognise Human Faces

Recognition of human faces is an ambitious problem, being currently attacked by psychologists, cognitive scientists, and - to a limited extent - also by AI-community. Nevertheless, computer programs solving this task are still rare. Most of them rely on the artificial neural nets, which are not used in our approach to the problem. The presented paper reports a successful attempt to extract a reliable set of stable intrinsic features from the images by using edge-detection, boundary grouping, and boundary characterisation. Particular attention is paid to the local properties of the boundaries at junction points. No attempt to attach high-level meaning to the individual features is made. The resulting symbolic descriptions are processed by a simple Machine-Learning program constructing a recognition scheme in the form of a decision tree. In spite of some constraints - frontal head-on view, limited training set - the results, as measured by predictive accuracy, are promising for dealing with larger numbers of individuals.