Image segmentation by selecting eigenvectors based on extended information entropy

Abstract

AbstractFor spectral clustering algorithm, the quality of eigenvectors of graph affinity matrix is very important for the clustering result. So, how to obtain high‐quality eigenvectors is crucial. In this paper, the authors aim to propose some new measurement methods to evaluate each eigenvector of affinity matrix for spectral selection. Based on extended information entropy, three criteria, i.e. Spectral Distinguishability (SD), Spectral Distinguishability Validity (SDV) and pectral Distinguishability ‐Degree (SDD), are defined respectively. The compactness of clusters for each eigenvector is measured by SD; SDV is used to remove the inefficient eigenvectors for clustering; SDD is used to evaluate the contribution of eigenvectors to clustering and is exploited to build a selective spectral ensemble scheme. To indicate the merits of the authors’ algorithm, the authors consider varied artificial data and natural images, including Berkeley image segmentation data set as benchmark data set. The authors’ simulation results confirm the superior performance of the proposed method in developing spectral clustering compared to conventional clustering methods and recent eigenvectors‐selection‐based algorithms.