Unsupervised Segmentation Using Cluster Ensembles

Abstract

We propose a novel framework for automatic image segmentation. In this approach, a mixture of several over-segmentation methods are used to produce superpixels and then aggregation is achieved using a cluster ensemble method. Generated by different existing segmentation algorithms, superpixels can describe the manifold patterns of a natural image such as color space, smoothness and texture. We use them as the initial superpixels. Grouping cues which affect the performance of segmentation can also be captured. After the oversegmentation, the simultaneous collection of superpixels is expected to achieve synergistic effects and ensure the accuracy of the segmentation. For this purpose, cluster ensemble methods are used to process the initial segmentation results and produce the final result. Our method achieves significantly better performance on the Berkeley Segmentation Database compared to state-of-theart techniques.