A Novel Hybrid Clustering Approach for Unsupervised Grouping of Similar Objects

Abstract

A novel hybrid clustering methodology named CDFISM (Clustering Distinct Features in Similarity Matrix) for grouping of similar objects is implemented in this study to address the unsatisfactory clustering results of current methods. Well-known PCA and a distance measuring method along with a new established algorithm (CISM) are employed to establish CDFISM methodology. CISM embodies both Rk-means method and an agglomerative/contractive/expansive (ACE) method. The CDFISM methodology has been tested on sample face images in three face databases to ensure the viability of the methodology. A high rate of accuracy has been achieved with the methodology, namely 97.5%, 98.75% and 80% respectively regarding the three image databases used in the study, averaging 92%. The hybrid methodology runs effectively for revealing interrelated pattern of similarities among objects.