Face recognition : two-dimensional and three-dimensional techniques

2 ABSTRACT We explore the field of automated face recognition. Beginning with a survey of existing methods applied to two-dimensional (2D) and three-dimensional (3D) face data, we focus on subspace techniques, investigating the use of image pre-processing applied as a preliminary step in order to reduce error rates. We implement the eigenface and Fisherface methods of face recognition, computing False Acceptance Rates (FAR) and False Rejection Rates (FRR) on a standard test set of images that pose typical difficulties for recognition. Applying a range of image processing techniques we demonstrate that performance is highly dependant on the type of pre-processing used and that Equal Error Rates (EER) of the eigenface and Fisherface methods can be reduced from 25.5% to 20.4% and 20.1% to 17.8% respectively, using our own specialised methods of image processing. However, with error rates still too high for use in many proposed applications we identify the use of 3D face models as a potential solution to the problems associated with lighting conditions and head orientation. Adapting the appearance-based subspace methods previously examined, for application to 3D face surfaces, introducing the necessary orientation normalisation and format conversion procedures, we show that low error rates can be achieved using surface shape alone, despite variations in head orientation and expression. In addition, these techniques are invariant to lighting conditions as no colour or texture information is used in the recognition process. We introduce a 3D face database providing 3D texture mapped face models, as well as 2D images captured at the same instant. This database facilitates a direct comparison of 3D and 2D techniques, which has not previously been possible. Contrasting the range of face recognition systems we explore methods of combining multiple systems in order to exploit the advantage of several methods in a single unified system. Various methods of system combination are tested, including combination by dimensional accumulation, elimination and genetic selection. This research leads to an innovative multi-subspace face recognition method capable of combining 2D and 3D data, producing state-of-theart error rates, with a clear advantage over single subspace systems: The lowest EER achieved using 2D, 3D and 2D Projection methods are 9.55%, 10.41% and 7.86% respectively, yet multi-subspace combination reduces this error down to 4.50% on the same test data.