Gabor-based Coding and Facial Similarity Perception

Abstract

Little is known about how facial representation in the face-selective areas of inferotemporal cortex is built up from the output of earlier visual areas such as primary visual cortex (area V1). We report work on a model of perceived facial similarity based on topographically ordered aggregates of localised, oriented, and spatial-frequency-selective receptive fields characteristic of V1 cells. The receptive fields are approximated with a set of Gabor filters. This Gabor-based code allows representation of the fine differences in texture and configuration needed for facial discrimination processes. Lyons and Morikawa (1996 Investigative Ophthalmology and Visual Science 37 910), showed that Gabor-based similarity is a good predictor of facial similarity for comparisons of fairly similar faces but not sufficient to model experience-dependent effects such as the ‘other race effect’.

Here we report results of a study on the effects of image negation on facial similarity perception. Negation of image gray levels interferes with face recognition (Bruce and Lanton, 1994 Perception 23 803 – 822) while preserving 2-D facial-shape information. The Gabor similarity measure models non-endstopped complex cells of V1 and is not affected by image negation. One group of subjects judged similarity among a set of normal gray-scale facial images while another group judged similarity between negative images of the same stimuli. Agreement between the model and human subjects did not decrease with image negation. Moreover, human similarity ratings between negative faces were strongly correlated with those between positives. These results support Gabor-based similarity as a model for facial similarity perception.