Towards exaggerated image stereotypes

Abstract

Given a training set of images and a binary classifier, we introduce the notion of an exaggerated image stereotype for some image class of interest, which emphasizes/exaggerates the characteristic patterns in an image and visualizes which visual information the classification relies on. This is useful for gaining insight into the classification mechanism. The exaggerated image stereotypes results in a proper trade-off between classification accuracy and likelihood of being generated from the class of interest. This is done by optimizing an objective function which consists of a discriminative term based on the classification result, and a generative term based on the assumption of the class distribution. We use this idea with Fisher's Linear Discriminant rule, and assume a multivariate normal distribution for samples within a class. The proposed framework has been applied on handwritten digit data, illustrating specific features differentiating digits. Then it is applied to a face dataset using Active Appearance Model (AAM), where male faces stereotypes are evolved from initial female faces.