Image clustering based on a shared nearest neighbors approach for tagged collections

Browsing and finding pictures in large-scale and heterogeneous collections is an important issue, most particularly for online photo sharing applications. Since such services are experiencing rapid growth of their databases, the tag-based indexing strategy and the results displayed in a traditional matrix representation may not be optimal for browsing and querying image collections. Naturally, unsupervised data clustering appeared as a good solution by presenting a summarized view of an image set instead of an exhaustive but useless list of its element. We present a new method for extracting meaningful and representative clusters based on a shared nearest neighbors (SNN) approach that treats both content-based features and textual descriptions (tags). We describe, discuss and evaluate the SNN method for image clustering and present some experimental results using the Flickr collections showing that our approach extracts representative information of an image set.