Pixel-Based versus Correspondence-Based Representations of Human Faces: Implications for Sex Discrimination

Abstract

In human perception, as well as in machine vision, a crucial step in solving any object recognition task is an appropriate description of the object class under consideration. We emphasise this issue when considering the object class ‘human faces’. We discuss different representations that can be characterised by the degree of alignment between the images they provide for. The representations used span the whole range between a purely pixel-based image representation and a sophisticated model-based representation derived from the pixel-to-pixel correspondence between the faces [Vetter and Troje, 1995, in Mustererkennung Eds G Sagerer, S Posch, F Kummert (Berlin: Springer)]. The usefulness of these representations for sex classification was compared. This was done by first applying a Karhunen — Loewe transformation on the representation to orthogonalise the data. A linear classifier was trained by means of a gradient-descent procedure. The classification error in a completely cross-validated simulation ranged from 15% in the simplest version of the pixel-based representation to 2.5% for the correspondence-based representation. However, even with intermediate representations very good performance was achieved.