Effective measurement selection in truncated kernel density estimator: Voronoi mean shift algorithm for truncated kernels

The Gating/Truncation technique is adapted to choose relatively significant measurements rather than all measurements to speed up mean shift algorithm which is one of the well-known clustering algorithms in the field of computer vision. The conventional mean shift algorithm can be sensitive to selecting measurements since the measurements are truncated with a Gaussian window of a fixed size. In particular when a small gating window is selected, it cannot properly cluster data points located far from major clusters and thus it generates unwanted, small clusters. We present a robust gating technique for truncated mean shift algorithm based on a geometric structure called Voronoi diagram of a given data set. Unlike conventional gating/truncation techniques our proposed truncation technique can provide nonlinear truncation windows with variable sizes constructed by using the Voronoi diagram to effectively identify outlier points in clusters. We also demonstrate the feasibility of this technique by applying it on synthetic and real-world image data sets. The experimental results show that the proposed truncation technique provides a more robust clustering result compared to the conventional truncation techniques. The proposed algorithm can be effectively applied to denoising of images by removing background noise.