Authors:
Juha Hanni
;
Esa Rahtu
and
Janne Heikkilä
Affiliation:
University of Oulu, Finland
Keyword(s):
Image categorization, Clustering, Generative model.
Related
Ontology
Subjects/Areas/Topics:
Computer Vision, Visualization and Computer Graphics
;
Image and Video Analysis
;
Statistical Approach
Abstract:
In this paper we show an unsupervised approach how to find the most natural organization of images. Previous methods which have been proposed to discover the underlying categories or topics of visual objects create no structure or at least the structure, usually tree-shaped, is defined in advance.
This causes a problem since the most relevant structure of the data is not always known. It is worthwhile to consider a generic way to find the most suitable structure of images. For this, we apply the model of finding the structural form (among eight natural forms) to automatically discover the best organization of objects in visual domain. The model simultaneously finds the structural form and an instance of that form that best explains the data. In addition, we present a generic structural form, so called meta structure, which can result in even more natural connections between clusters of images. We show that the categorization results are competitive with the state-of-the-art methods
while giving more generic insight to the connections between different categories.
(More)