Abstract
Biometrics axe unique to each individual and are therefore valuable for identification purposes. One of such biometrics is the individual’s face; consequently the existence of the research field of face recognition. Although face recognition technology using 2D images taken in controlled environments has reached high performance rates, its reliability declines when variations in head pose, lighting, and facial expression, are introduced. More recently, the field of face recognition has expanded into the 3D domain of shape geometry. Since 3D shape does not change under different illumination conditions or viewpoint, 3D face recognition could become more reliable than its 2D counterpart. However 3D acquisition systems are expensive and have limited acquisition volumes, whereas 2D imaging devices are cheap and are established as a popular method for surveillance of large areas. Therefore this work addresses two main questions: i) how can 3D shape geometry be used for reliable 3D face recognition? and ii) can 3D shape geometry be used to boost the performance of 2D face recognition systems? This thesis introduces a new dense registration algorithm, based on a deformable surface framework, for 3D shape data. The algorithm registers 3D raw data via fitting a deformable model, with a mean fitting error below 0.2mm. Registered 3D data is then manipulated as a collection of vectors of the same dimensionality and in correspondence, thus enabling the application of any pattern classification technique to 3D shape geometry data. The registration algorithm is also used as the framework for the implementation of a novel model based lossy compression algorithm tailored for 3D face data. This compression algorithm achieves compression ratios of more than 35 without a significant loss the of discriminatory information of the face. Furthermore, the reliability of the registration algorithm for establishing accurate dense correspondences is exploited in this work to build a morphable model. Morphable models are powerful tools for the synthesis and reconstruction of 3D faces. Additionally they can be fitted to 2D images, thus extracting 3D information and enabling to render novel 2D views. This thesis proposes extensions to the morphable model framework that allow to handle texture maps, and facial expressions via a model built from displacement vectors. The morphable model is then used as a tool for correcting pose and eliminating facial expressions from 2D images of faces. The effect of using pose and expression correction to aid a to 2D face recognition algorithm is then evaluated in two mainstream databases. The evaluation shows there is a real potential in using 3D information, in this case via a morphable model, to boost the performance of 2D face recognition systems. However before a definite conclusion can be reached, more research in morphable model building and fitting reliability will be required in the future.