Non Negative Matrix Factorization Clustering Capabilities - Application on Multivariate Image Segmentation

Cosmin Lazar, Danielle Nuzillard, Patrice Billaudel et Sorin Curila

The clustering capabilities of the Non Negative Matrix Factorization algorithm is studied. The basis images are considered like the data membership degree to a particular class. A hard clustering algorithm is easily derived based on these images. This algorithm is applied on a multivariate image to perform image segmentation. The results are compared with those obtained by Fuzzy K-means algorithm and better clustering performances are found for NMF based clustering. We also show that NMF performs well when we deal with uncorrelated clusters but it cannot distinguish correlated clusters. This is an important drawback when we try to use NMF to perform data clustering.