Principal component and neural network analyses of face images: What can be generalized in gender classification? (Q1384550)

One of the major problems in modeling face processing is to find a way of representing faces that allows for the wide range of tasks typical of human performance. Traditionally, computational models of face recognition represent faces in terms of geometric descriptors that include distances, angles, and areas between elementary features such as eyes, nose, or chin or in terms of template parameters, or isodensity lines. Although these approaches economically represent faces in a way that is relatively insensitive to variations in scale, tilt, or rotation of the faces, they are not without problems. The general purpose of the present work is to analyze the robustness of the kind of face representation proposed by the principal components (PCA) approach. The human face is a complex visual pattern that contains general categorical information as well as idiosyncratic, identity specific information. By categorical information, we mean that some aspects of a face are not specific to that particular face but are shared by subsets of faces (e.g., female faces share some visual characteristics such as smoothness of the skin, prominence of the cheeks, or roundness of the face). These aspects can be used to assign both unfamiliar and familiar faces to general semantic categories such as gender or race. The present paper is organized as follows. The PCA approach is presented briefly first followed by a discussion of the interpretation of eigenvectors as ``macro-features''. Next, the usefulness of this approach for analyzing the perceptual information in faces is discussed along with its relationship to some earlier work on the role of different spatial frequencies in face processing. Finally, three series of simulations concerning the statistical properties of eigenvectors derived from a set of face images are described.