Entropy-based fuzzy C-means with weighted hue and intensity for color image segmentation

Image segmentation is a task of grouping pixels based on similarity. In this paper the problem of segmentation of color images, especially noisy images, is studied. In order to improve the speed of segmentation and avoid redundant calculations, our method only uses two color components, hue and intensity, which are chosen rationally. These two color components are combined in a specially defined cost function. The impact of each color component (hue and intensity) is controlled by weights (called hue weight and intensity weight). These weights lead to focusing on the color component that is more informative and consequently the speed and accuracy of segmentation is improved. We have also used entropy maximization in the core of the cost function to improve the performance of segmentation. Furthermore we have suggested a fast initialization scheme based on peak finding of two dimensional histogram that prevents Fuzzy C-means from converging to a local minimum. Our experiments indicate that the proposed method performs superior to some related state-of-the-art methods.