Clustering through SOM Consistency

Clustering is a classical tool in image analysis, with wide applications. Yet, most of its algorithmic solutions include a considerable amount of stochasticity, e.g. due to different initialisations. Here, we introduce a clustering method rooted on self organizing maps, that exploits the maps' intrinsic variability, to produce reliable clustering. Although only a subset of the data is consistently clustered, we show that this set is trustworthy, and can be used for posterior classification.