Darryl Hond and Libor Spacek
Department of Computer Science
University of Essex
Colchester
email hondd@essex.ac.uk
Face recognition requires sufficiently accurate measurement. This can be achieved by transforming faces to a normalised form prior to description. This method uses template matching to yield parameters for normalisation. Matching is based on the combined output of four whole face templates, each utilising a different descriptor. Recognition is performed by matching a normalised test face to normalised training faces. The matching employs descriptions based on local statistical analysis of the orientation of the gradient vector. Orientation histograms are produced for small sub-images. Recognition can be enhanced if a subset of sub-images are used. The scheme has been applied to a database of faces in the public domain.
Hond D A