Color-Based Image Segmentation by Means of a Robust Intuitionistic Fuzzy C-means Algorithm

To yield well-suited image segmentation results, conventional clustering algorithms depend on customized hand-crafted features as well as an appropriate initialization process. This latter aspect is a mandatory pre-requisite for convergence of the algorithm, in other words, its efficiency impacts the quality of the result. In this work, we introduce an Intuitionistic Fuzzy C-Means clustering algorithm enhanced by means of Robust Statistics, which develops an outstanding image segmentation based on a basic feature such as the color information, and it requires a reduced iteration number to converge. The non-parametric Lorentzian Redescending M-estimator is used both at initialization and iterative stages of the clustering algorithm; since, it behaves such as a robust location estimator when the centroid vector is computed, and as a weighting when the membership matrix is updated. With the fusion of both techniques, we can guarantee that the introduced clustering algorithm can efficiently develop the task of segmentation of color images and pattern recognition processes. The robustness and effectiveness of this proposal is verified by experiments on the natural color images BSDS500 dataset, as well as a simulated dataset corrupted with atypical data.