Contour grouping by clustering with multi-feature similarity measure

Abstract

Contour grouping is a key task in computer vision domain. It extracts the meaningful objects information from low-level image features and provides the input for the further processing. There have been many techniques proposed over the decades. As a useful data analysis method in machine learning, clustering is a natural way for doing the grouping task. However, due to many complicated factors in natural images, such as noises and clutter in background, many clustering algorithms, which just use pairwise similarity measure, are not robust enough and always fail to generate grouping results that are consistent with the visual objects perceived by human. In this article, we present how the grouping performance is improved by utilizing multi-feature similarity under the information based clustering framework compared with other clustering methods using pairwise similarity. © 2010 Springer-Verlag Berlin Heidelberg.