A novel hierarchical statistical model for count data modeling and its application in image classification

Abstract

The problem that we elaborate in this work is developing and comparing statistical models for learning hierarchical image categories from a structural point of view. Previously different statistical models have been proposed based on different statistical schemes for dealing with hierarchical structures. In this work following the lead of the previous models we develop our own hierarchical model and we make a thorough comparison between the existing and the proposed models. Our main contribution in this work is the utilization of Beta-Liouville distribution as a replacement for Dirichlet distribution, which is traditionally used for prior distribution modeling, and deriving the criteria for making it compatible to hierarchical data modeling. For the development of our statistical model, we make extensive use of the Bag of the visual words model and the concept of count data in machine learning. © 2012 Springer-Verlag.