Spectral clustering for data categorization based on self-organizing maps

The exploration and categorization of large and unannotated image collections is a challenging task in the field of image retrieval as well as in the generation of appearance based object representations. In this context the Self-Organizing Map (SOM) has shown to be an efficient and scalable tool for the analysis of image collections based on low level features. Next to commonly employed visualization methods, clustering techniques have been recently considered for the aggregation of SOM nodes into groups in order to facilitate category specific data exploration. In this paper, spectral clustering based on graph theoretic concepts is employed for SOM based clustering and data categorization. The results are compared with those from the Neural Gas algorithm and hierarchical agglomerative clustering. Using SOMs trained on an eigenspace representation of the Columbia Object Image Library 20 (COIL20), the correspondence of the results to a semantic reference grouping is calculated. Based on the Adjusted Rand Index it is shown that independent from the number of selected clusters, spectral clustering achieves a significantly higher correspondence to the reference grouping than any of the other methods.