False-peaks-avoiding mean shift method for unsupervised peak-valley sliding image segmentation

Abstract

The mean shift (MS) algorithm is sensitive to local peaks. In this paper, we show both empirically and analytically that when using sample data, the reconstructed PDF may have false peaks. We show how the occurrence of the false peaks is related to the bandwidth h of the kernel density estimator, using a one-dimensional example motivated by gray-level image segmentation. It is well known that in MS-based approaches, the choice of h is important. However, we provide a quantitative relationship between the appearance of false peaks and the value of h. For the gray-level image segmentation problem, we not only show how to avoid the false peak problem, but also we provide a complete unsupervised peak-valley sliding algorithm for gray-level image segmentation. However, the main contribution of the paper remains the characterization of the false peak problem and the questions it raises regarding this issue in more general settings (e.g. higher dimensional problems).