Deviation-Sparse Fuzzy C-Means With Neighbor Information Constraint

This paper introduces sparsity in the traditional fuzzy clustering framework and presents two novel clustering methods. The first one is called deviation-sparse fuzzy c-means (DSFCM). When spatial correlation is encountered, the second method is proposed, which is called deviation-sparse fuzzy c-means with neighbor information constraint (DSFCMN). The contributions of this paper are threefold. First, the theoretical values of data, estimated from the measured values, are utilized in the clustering process. This could acquire more accurate cluster centers than the traditional fuzzy c-means. Second, by imposing sparsity on the deviations between measured values and theoretical values, DSFCM and DSFCMN could identify noise and outliers. Finally, with the constraint of neighbor information, the estimation of the deviations between measured values and theoretical values of data would be more reliable than only considering the data itself. Experiments performed on artificial and real-world images show that DSFCMN is effective and efficient, and thus more competitive than other fuzzy clustering methods.