Novel hybrid object-based non-parametric clustering approach for grouping similar objects in specific visual domains

Abstract Current widely employed clustering approaches may not yield satisfactory results with regard to the characteristics and distribution of datasets and number of clusters to be sought, especially for visual domains in multidimensional space. This study establishes a novel clustering methodology using a pairwise similarity matrix, Clustering Visual Objects in Pairwise Similarity Matrix (CVOIPSM), for grouping similar objects in specific visual domains. A dimensionality reduction and feature extraction technique, along with a distance measuring method and a newly established algorithm, Clustering in Pairwise Similarity Matrix (CIPSM), are combined to develop the CVOIPSM methodology. CIPSM utilizes both Rk-means and an agglomerative, contractible, expandable (ACE) technique to calculate a membership degree based on maximizing inter-class similarity and minimizing intra-class similarity. CVOIPSM has been tested on several datasets, with average success rates on downsized subsamples between 87.5% and 97.75% and between 81% and 87% on the larger datasets. The difference in the success rates for small and large datasets is not statistically significant (p > 0.01). Moreover, this method automatically determines the likely number of clusters without any user dictation. The empirical results and the statistical significance test on these results ensure that CVOIPSM performs effectively and efficiently on specific visual domains, disclosing the interrelated patterns of similarities among objects.