If we consider an n x n image as an n2-dimensional vector, then images
of faces can be considered as points in this n2-dimensional image spa
ce. Our previous studies of physical transformations of the face, incl
uding translation, small rotations, and illumination changes, showed t
hat the set of face images consists of relatively simple connected sub
regions in image space. Consequently linear matching techniques can be
used to obtain reliable face recognition. However, for more general t
ransformations, such as large rotations or scale changes, the face sub
regions become highly non-convex. We have therefore developed a scale-
space matching technique that allows us to take advantage of knowledge
about important geometrical transformations and about the topology of
the face subregion in image space. While recognition of faces is the
focus of this paper, the algorithm is sufficiently general to be appli
cable to a large variety of object recognition tasks. (C) 1994 Academi
c Press, Inc.