A method for improving the segmentation of images is presented. It involves taking an initial segmentation provided by some other means, and modifying the region boundaries depending on the estimated region models until an equilibrium is reached. The advantages of this technique are: (1) no parameters are required, (2) it is invariant under constant scalings of the image intensities, and (3) it is relatively insensitive to the position and topology of the initial segmentation. Examples are given of its application to single and multi-scale intensity images, textured images, range images and multi-band satellite images.
Multi-Scale multi-band densenets for audio source separation
