Issue No. 05 - September (1995 vol. 15)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/38.403830
Human faces are perceived to differ along many dimensions. Some dimensions can be defined by objectively different categories such as old/young or male/female but others reflect more subjective categories such as attractive/unattractive and happy/sad. In this paper we explain a technique for determining the consistent (or "prototypic") visual characteristics (shape and color) of a category of faces. We describe a method for evaluating and visualizing the extent to which prototypes differ. Information derived from prototypes can be used to reconstruct a plausibly colored image of an individual's face originally recorded in black and white. Information from pairs of prototypes can also be used to define shape and color transformations which allow individual face images to be moved in appearance along quantifiable dimensions in "face space." A given face image can thus be made to look older or younger, more or less masculine/feminine, etc. Transformation performed in a given direction can be made either interpolating between two prototypes (e.g. making a male face less masculine) or as an extrapolation beyond the prototype (e.g. enhancing the masculinity of a male face). Although the processing has been developed to analyze and manipulate face images the same principles can be applied to other visually homogeneous classes.
facial recognition systems
D. I. Perrett and D. A. Rowland, "Manipulating Facial Appearance through Shape and Color," in IEEE Computer Graphics and Applications, vol. 15, no. , pp. 70-76, 1995.