Faces can vary in terms of size, location in an image and orientation about the z-axis. Such variation can be removed by normalising the face. Descriptions for recognition can then be obtained.
The eyes are commonly detected for normalisation , and some effective eye detectors have been produced . Active Shape Models have been used to locate a large number of characteristic points on the face which are used as reference points . The variables are sometimes constrained by capturing faces in a consistent manner.
Eye detection cannot always be successfully applied to faces; glasses or other obstacles can hide the eyes. This paper presents a whole face approach to face normalisation which is more robust.
A system which normalises prior to recognition must tackle two recognition problems: discriminating a face from its background, and identifying the face. The key is to find descriptions that allow these distinctions. This method uses four descriptors for face normalisation: ranked grey-levels, ranked gradient vector magnitudes, the x-component of the unit gradient vector and the y-component of the unit gradient vector. The normalisation parameters can be found effectively by applying an average whole face template based on any one of these descriptors. The combined templates outperform any single template.
Several recognition schemes have made at least partial use of intensity information: ,,  and . The gradient magnitude has also been shown to be effective .
Descriptions for recognition in this scheme are based on the orientation of the gradient vector.