Region merging for image segmentation based on unimodality tests

Abstract

A novel approach is proposed for image segmentation based on region merging. We define a new criterion to decide on whether to merge two regions that does not require the specification of user defined thresholds. The method begins with an image oversegmentation (based on SLIC superpixels) into small homogeneous regions. At each step regions are iteratively merged to form larger regions based on the result of a merge test that measures unimodality as an indication of visual content homogeneity. More specifically, the merge test employs the dip-dist criterion that decides on the unimodality of a set of data objects through the application of the Hardigans' dip-test for unimodality. We propose a fast version of the dip-dist criterion, where the two feature centroids of the regions to be merged are used to decide on the unimodality of the region that results from merging. To demonstrate the performance of the method, we provide segmentation results for several images from well-known image datasets using CIELAB color as the feature vector of each pixel.

Topics

5 Figures and Tables

Download Full PDF Version (Non-Commercial Use)