Statistical Discrimination of Facial Images with Verbal Feature Variables

Abstract

The purpose of this study is to examine the statistical methods for discriminating human faces with the ratings on feature scales and the descriptions of feature key-words given by observers. For that purpose, multinomial and linear discriminant analyses of the rating variables, and multinomial and Bernoulli ones of the key-word variables are presented. Those two classes of the analyses are integrated into the method for discriminating faces with both the rating and key-word variables. In the multinomial and Bernoulli discriminant analyses, the distributions of the feature variables are smoothed by either the method using the random guessing probability or the kernel method.
A training sample of feature data for a hundred faces was used for the parameter estimation in the discrimination methods and another sample was used to validate the methods. The results indicated the superiority of the discrimination with both the rating and key-word variables to the discrimination with either of those. Especially, the integration between the linear discriminant analysis of the rating variables and the Bernoulli analysis of the key-word ones yielded the most accurate discrimination. It was also shown that the use of the kernel method was not so effective in some cases.