Classification of double attributes via mutual suggestion between a pair of classifiers
K. Hiraoka, T. Mishima · 2003
Real-world objects often have two or more significant attributes. For example, face images have attributes of persons, expressions, and so on. Even if you are interested in only one of those attributes, additional informations on auxiliary attributes can help recognition of the main one. The authors have been proposed a method for classification with double attributes. Its main idea is mutual suggestion of hints between a pair of classifiers. In the present paper, we will reexamine the task based on information geometry, and propose a new method of EM-like iterations. We will also show experimentally that the heuristic method in our previous work can be used as a good approximation of the new method which has solid theoretical basis.