Simple-to-complex discriminative clustering for hierarchical image segmentation

Abstract

We propose a novel discriminative clustering algorithm with a hierarchical framework for solving unsupervised image segmentation problems. Our discriminative clustering process can be viewed as an EM algorithm, which alternates between the learning of image visual appearance models and the updates of cluster labels (i.e., segmentation outputs) for each image segment. In particular, we advance a simple-to-complex strategy during the above process, which allows the learning of a series of classifiers with different generalization capabilities from the input image, so that consecutive image segments can be well separated. With the proposed hierarchical framework, improved image segmentation can be achieved even if the shapes of the segments are complex, or the boundaries between them are ambiguous. Our work is different from existing region or contour-based approaches, which typically focus on either separating local image regions or determining the associated contours. Our experiments verify that we outperform state-of-the-art approaches on unsupervised image segmentation.