Determining the Clustering Centers by Slope Difference Distribution

Clustering is fundamental in image processing, machine learning, pattern recognition, and data analysis. For robust clustering, the prerequisite condition is determining the clustering centers robustly. In this paper, we propose to determine the clustering centers from the slope difference distribution of the data. The proposed method comprises two parts: 1) computation of the slope difference distribution from the original data distribution and 2) selection of the peaks of the slope distribution as the clustering centers. We tested the proposed method with two different types of synthesized data sets: 1) data with Gaussian noise and 2) data with salt and pepper noise. Experimental results show that the proposed method is significantly more accurate than the state-of-the-art methods: K-means method, expectation maximization method, and fuzzy C-means method. The significance of determining the clustering centers robustly is also verified by a demonstration.