Total-aware suppressed possibilistic c-means clustering

Abstract

Cutset-type possibilistic c-means (CPCM) is an improved version of possibilistic c-means (PCM), and it addresses the issue of consistency clustering in PCM. However, the clustering performance has not significantly improved, and its convergence speed is still slow. Hence, a total-aware suppressed possibilistic c-means (TSPCM) clustering is proposed in this paper. Firstly, from a global perspective, the typicality values of all the non-winners of samples inside the cluster core are decreased by the suppressed factor while increasing the typicality values of all the winners of samples in the same core. Meanwhile, the typicality values of samples outside the cluster core remain unchanged. The corrected typicality values are used to update the clustering centers, and the modified clustering centers are again used to renew the typicality values alternately until the algorithm converges. Secondly, the objective function of the proposed TSPCM is constructed, and Zangwill's theorem firmly establishes its convergence. Finally, two robust total-aware suppressed algorithms are presented when TSPCM is used for the clustering of numerical data with noise and the segmentation of noisy images. Experimental results show that the proposed TSPCM and its robust algorithms are obviously superior to existing state-of-the-art PCM-related algorithms.