Iterative extended mean shift algorithm

An iterative filter for locally estimating the center of the clusters present in multi-spectral images is developed. In each iteration and for each pixel, an energy function computed for windows of increasing size is evaluated. The initial window size is defined as a function of the separability between neighboring pixels. The increments of the window size are a function of the internal to external entropy ratio of discs of consecutive radii. For a given pixel, the minimum value of the energy function is preserved and used as the initial guess for the next iteration. This minimum corresponds to the estimated point in which a state of higher order emerges. The scheme proposed was tested on a set of synthetical images, and compared to the output of the iterative median filter and to the k-Means algorithm, showing better performance than these ones. Results are shown on dermatological and fundus digital images.