Dominique Valentin

Connectionist models of face processing : a survey

Abstract Connectionist models of face recognition, identification, and categorization have appeared recently in several disciplines, including psychology, computer science, and engineering. We present a review of these models with the goal of complementing a recent survey by Samal and Iyengar [Pattern Recognition25, 65–77 (1992)] of nonconnectionist approaches to the problem of the automatic face recognition. We concentrate on models that use linear autoassociative networks, nonlinear autoassociative (or compression) and/or heteroassociative backpropagation networks. One advantage of these models over some nonconnectionist approaches is that analyzable features emerge naturally from image-based codes, and hence the problem of feature selection and segmentation from faces can be avoided.

Structural aspects of face recognition and the other-race effect

The other-race effect was examined in a series of experiments and simulations that looked at the relationships among observer ratings of typicality, familiarity, attractiveness, memorability, and the performance variables ofd’ and criterion. Experiment 1 replicated the other-race effect with our Caucasian and Japanese stimuli for both Caucasian and Asian observers. In Experiment 2, we collected ratings from Caucasian observers on the faces used in the recognition task. A Varimax-rotated principal components analysis on the rating and performance data for the Caucasian faces replicated Vokey and Read’s (1992) finding that typicality is composed of two orthogonal components, dissociable via their independent relationships to: (1) attractiveness and familiarity ratings and (2) memorahility ratings. For Japanese faces, however, we fond that typicality was related only to memorahility. Where performance measures were concerned, two additional principal components dominated by criterion and byd’ emerged for Caucasian faces. For the Japanese faces, however, the performance measures ofd’ and criterion merged into a single component that represented a second component of typicality, one orthogonal to thememorability-dominated component. A measure offace representation quality extracted from an autoassociative neural network trained with a majority of Caucasian faces and a minority of Japanese faces was incorporated into the principal components analysis. For both Caucasian and Japanese faces, the neural network measure related both to memorability ratings and to human accuracy measures. Combined, the human data and simulation results indicate that the memorahility component of typicality may be related to small, local, distinctive features, whereas the attractiveness/familiarity component may be more related to the global, shape-based properties of the face.

The perception of face gender: the role of stimulus structure in recognition and classification.

The perception of face gender was examined in the context of extending “face space” models of human face representations to include the perceptual categories defined by male and female faces. We collected data on the recognizability, gender classifiability (reaction time to classify a face as male/female), attractiveness, and masculinity/femininity of individual male and female faces. Factor analyses applied separately to the data for male and female faces yielded the following results. First, for both male and female faces, the recognizability and gender classifiability of faces were independent—a result inconsistent with the hypothesis that both recognizability and gender classifiability depend on a face’s “distance” from the subcategory gender prototype. Instead, caricatured aspects of gender (femininity/masculinity ratings) related to the gender classifiability of the faces. Second, facial attractiveness related inversely to face recognizability for male, but not for female, faces—a result that resolves inconsistencies in previous studies. Third, attractiveness and femininity for female faces were nearly equivalent, but attractiveness and masculinity for male faces were not equivalent. Finally, we applied principal component analysis to the pixel-coded face images with the aim of extracting measures related to the gender classifiability and recognizability of individual faces. We incorporated these model-derived measures into the factor analysis with the human rating and performance measures.

Low-dimensional representation of faces in higher dimensions of the face space

Faces can be represented efficiently as a weighted linear combination of the eigenvectors of a covariance matrix of face images. It has also been shown [ J. Opt. Soc. Am.4, 519– 524 ( 1987)] that identifiable faces can be made by using only a subset of the eigenvectors, i.e., those with the largest eigenvalues. This low-dimensional representation is optimal in that it minimizes the squared error between the representation of the face image and the original face image. The present study demonstrates that, whereas this low-dimensional representation is optimal for identifying the physical categories of face, like sex, it is not optimal for recognizing the faces (i.e., discriminating known from unknown faces). Various low-dimensional representations of the faces in the higher dimensions of the face space (i.e., the eigenvectors with smaller eigenvalues) provide better information for face recognition.

More about the Difference between Men and Women: Evidence from Linear Neural Networks and the Principal-Component Approach

The ability of a statistical/neural network to classify faces by sex by means of a pixel-based representation has not been fully investigated. Simulations with pixel-based codes have provided sex-classification results that are less impressive than those reported for measurement-based codes. In no case, however, have the reported pixel-based simulations been optimized for the task of classifying faces by sex. A series of simulations is described in which four network models were applied to the same pixel-based face code. These simulations involved either a radial basis function network or a perceptron as a classifier, preceded or not by a preprocessing step of eigendecomposition. It is shown that performance comparable to that of the measurement-based models can be achieved with pixel-based input (90%) when the data are preprocessed. The effect of the eigendecomposition preprocessing of the faces is then compared with spatial-frequency analysis of face images and analyzed in terms of the perceptual information it captures. It is shown that such an examination may offer insight into the facial aspects important to the sex-classification process. Finally, the contribution of hair information to the performance of the model is evaluated. It is shown that, although the hair contributes to the sex-classification process, it is not the only important contributor.

Can a linear autoassociator recognize faces from new orientations

An often noted limitation of computational models of faces operating on two-dimensional pixel intensity representations is that they cannot handle changes in orientation. We show that this limitation can be overcome by the use of multiple views of a given face instead of a single view to represent the face. Specifically, we show that a linear autoassociator trained to reconstruct multiple views of a set of faces is able to recognize the faces from new view angles. An analysis of the internal representation of the memory (i.e., eigenvectors of the between-unit-connection weight matrix) shows a dissociation between two kinds of perceptual information: orientation and identity information.