Unsupervised fuzzy model-based image segmentation

Abstract

This paper presents a novel unsupervised fuzzy model-based image segmentation algorithm. The proposed algorithm integrates color and generalized Gaussian density (GGD) into the fuzzy clustering algorithm and incorporates their neighboring information into the learning process to improve the segmentation accuracy. In addition, a membership entropy term is used to make the algorithm not sensitive to initial clusters. To optimize the objective function of the proposed segmentation model, we define the dissimilarity measure between GGD models using the Kullback–Leibler divergence, which evaluates their discrepancy in the space of generalized probability distributions via only the model parameters. We also present mathematical analysis that proves the existence of the cluster center for the GGD parameters, thus establishing a theoretical basis for its use. Experimental results show that our proposed method has a promising performance compared with the current state-of-the-art fuzzy clustering-based approaches.