Abstract
It has been recognised that, within the context of face recognition, angular separation between centred feature vectors is a useful measure of dissimilarity. In this thesis we explore this observation in more detail and compare and contrast angular separation with the Euclidean, Manhattan and Mahalonobis distance metrics. This is applied to 2D, 2.5D and 3D face images and the investigation is done in conjunction with various feature extraction techniques such as local binary patterns (LBP) and linear discriminant analysis (LDA). We also employ error-correcting output code (ECOC) ensembles of support vector machines (SVMs) to project feature vectors non-linearly into a new and more discriminative feature space. It is shown that, for both face verification and face recognition tasks, angular separation is a more discerning dissimilarity measure than the others. It is also shown that the effect of applying the feature extraction algorithms described above is to considerably sharpen and enhance the ability of all metrics, but in particular angular separation, to distinguish inter-personal from extra-personal face image differences. A novel technique, known as angularisation, is introduced by which a data set that is well separated in the angular sense can be mapped into a new feature space in which other metrics are equally discriminative. This operation can be performed separately or it can be incorporated into an SVM kernel. The benefit of angularisation is that it allows strong classification methods to take advantage of angular separation without explicitly incorporating it into their construction. It is shown that the accuracy of ECOC ensembles can be improved in this way. A further aspect of the research is to compare the effectiveness of the ECOC approach to constructing ensembles of SVM base classifiers with that of binary hierarchical classifiers (BHC). Experiments are performed which lead to the conclusion that, for face recognition problems, ECOC yields greater classification accuracy than the BHC method. This is attributed primarily to the fact that the size of the training set decreases along a path from the root node to a leaf node of the BHC tree and this leads to great difficulties in constructing accurate base classifiers at the lower nodes.