Abstract
Biometric person identification has been an active research area in the recent decades. Among the many biometric features, such as finger prints, face, and voice; the human face offers particular advantages such as the ability to recognise an individual from a distance, and without their cooperation. Although face recognition using 2D images taken under controlled conditions has reached high performance rates, its performance declines under non-cooperative scenarios such as surveillance using CCTV cameras where the scene can have arbitrary illumination conditions and the subject can have arbitrary pose with respect to the camera and be at a far distance. The focus of this thesis is on the problem of recognising individuals from a 2D facial image with low resolution and arbitrary pose and illumination. We investigate the use of 3D information in order to boost the performance of 2D face recognition in such scenarios. A 3D Morphable Face Model is used to extract 3D shape and facial texture information from a 2D low-resolution facial image with arbitrary pose and illumination. To this end, the 3D model is fitted to the input image using a novel low-resolution fitting algorithm proposed in this thesis. It is shown that the fitting algorithm is able to extract reliable 3D shape and texture information across a large range of variations in pose and illumination. It is shown, through extensive experimental evaluation, that the model parameters obtained using our fitting algorithm are reliable enough to be directly used for face recognition in low-resolution under varying poses and illuminations. Furthermore, we propose a novel approach to using 3D information in order to enhance the low-resolution facial texture. More specifically, we propose a 3D-assisted facial texture super-resolution framework which uses the 3D information extracted from an LR image to map the facial texture to a shape- and pose-normalised domain. The facial texture is then enhanced by applying texture super-resolution in this domain. Through this procedure, a high-resolution estimate of the facial texture is obtained which can then be used to render the face in a normalised pose and illumination and with high-resolution texture. It is shown that this procedure can inject relevant and discriminative high-resolution information to the facial texture thereby boosting the performance of a conventional 2D face recognition engine.