Purchase this article with an account.
Adrian Nestor, Michael J. Tarr; Region-based representations of faces. Journal of Vision 2006;6(6):1065. doi: https://doi.org/10.1167/6.6.1065.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
A great deal of face recognition research is concerned with finding a robust method for identifying facial features and uncovering a plausible encoding scheme. We develop and test a representational scheme based on the idea of facial feature segmentation. Region-based segmentation (Fowlkes, Martin and Malik, 2003) is used to parse human faces into features, i.e., parts, at any given level of detail. The segmentation algorithm tested closely mimics human generated segmentations for front-view adult Caucasian faces. The validity of this scheme is further evaluated by using the surface properties, luminance and chrominance, of the facial features thus identified to predict facial gender. Both automatic gender discrimination and our ability to account for human gender discrimination are superior when we consider the properties of multiple facial features compared to average face properties, e.g. mean face luminance (Tarr et al., 2001). The representational scheme is thus relevant to the study of higher-level aspects of face processing, e.g., the involvement of surface properties in face categorization. In particular, automatic segmentation of faces offers a model of how face representations are constructed as well as a means to investigate some lower-level aspects of face processing, e.g., the role of color in face segmentation. In sum, region-based models show significant advantages in automatic face processing, as well as psychological plausibility in accounting for human face processing.
This PDF is available to Subscribers Only