Correlation-based criterion for the most discriminative principal component selection in normalized cut segmentation

Image segmentation is a fundamental problem in computer vision. Normalized Cut (Ncut) scheme uses second smallest eigenvector for solving this problem, while such eigenvectors may be sensitive to undesired changes in image. In this paper, firstly, we point out that optimization of Ncut is equivalent to optimization of Fisher-Rao criterion in classification. Then we look at the classification experience to gain a new perspective on the selection of eigenvectors in Ncut approach. Experimental results on image segmentation, demonstrate the truth about this alternative view of eigenvector selection for image segmentation.