Cluster Analysis and Optimization in Color-Based Clustering for Image Abstract

Jing He1,3, Guangyan Huang2, Yanchun Zhang1, and Yong Shi3 1School of Computer Science and Mathematics, Victoria University, Australia 2Institute of Software, Chinese Academy of Sciences, Beijing 100080, P.R.China 3Research Center on Fictitious Economy and Data Science, Chinese Academy of Sciences, Beijing 100080, P.R.China hejing@gucas.ac.cn, huanggy@ercist.iscas.ac.cn, yzhang@csm.vu.edu.au, yshi@gucas.ac.cn Abstract Cluster analysis has been identified as a core task in data mining. What constitutes a cluster, or a good clustering, may depend on the background of researchers and applications. This paper proposes two optimization criteria of abstract degree and fidelity in the field of image abstract. To satisfy the fidelity criteria, a novel clustering algorithm named Global Optimized Color-based DBSCAN Clustering (GOC- DBSCAN) is provided. Also, non-optimized local color information based version of GOC-DBSCAN, called HSV-DBSCAN, is given. Both of them are based on HSV color space. Clusters of GOC-DBSCAN are analyzed to find the factors that impact on the performance of both abstract degree and fidelity. Examples show generally the greater the abstract degree is, the less is the fidelity. It also shows GOC- DBSCAN outperforms HSV-DBSCAN when they are evaluated by the two optimization criteria.