Learning Visual Categories based on Probabilistic Latent Component Models with Semi-supervised Labeling

Abstract

This paper proposes a learning method of object and
scene categories based on probabilistic latent component models
in conjunction with semi-supervised object class labeling. In this
method, a set of object segments extracted from scene images of
each scene category is firstly clustered by the probabilistic latent
component analysis with the variable number of classes, next the
probabilistic latent component tree is generated as a classification
tree of all the object classes of all the scene categories, and
then object classes are incrementally labeled by propagating
prior scene category labels and posterior object category labels
given to representative object instances of some object classes as
teaching signals. Through experiments by using images of plural
categories in an image database, it is shown that the method
works effectively in learning a labeled object category tree and
object category composition of scene categories and achieves high
performance for object and scene recognition.