Simultaneous optimization of class configuration and feature space for object recognition

A new algorithm for object classification based on an extension of the Fisher's discriminant analysis is presented. Object recognition algorithms using the standard Fisher's algorithm, such as the Fisherface, train the classifier using sample-class pairs, where, for the classes, object categories determined in the application systems are used directly. In contrast, the new algorithm automatically produces subclasses, within each predetermined category that are actually used for classification, via unsupervised learning. In order to perform this, we combine the Fisher's discriminant analysis with the Akaike information criterion, optimizing the class configuration, i.e. sample-subclass correspondences, and the feature extraction function simultaneously, thereby improving the potential of linear separability. By applying this new method to face recognition, we show how it outperforms the traditional Fisher-based method.