Improvised eigenvector selection for spectral Clustering in image segmentation

Abstract

General spectral Clustering(SC) algorithms employ top eigenvectors of normalized Laplacian for spectral rounding. However, recent research has pointed out that in case of noisy and sparse data, all top eigenvectors may not be informative or relevant for the purpose of clustering. Use of these eigenvectors for spectral rounding may lead to bad clustering results. Self-tuning SC method proposed by Zelnik and Perona [1] places a very stringent condition of best alignment possible with canonical coordinate system for selection of relevant eigenvectors. We analyse their algorithm and relax the best alignment criterion to an average alignment criterion. We demonstrate the effectiveness of our improvisation on synthetic as well as natural images by comparing the results using Berkeley segmentation and benchmarking dataset.