Scale-Space Approach for Color Image Segmentation

This paper presents a new method for color image segmentation based on a scale-space clustering of the image pixels. Unlike standard image scale-spaces, which smooth the images, this approach consider an augmented space which combines spatial and color dimensions. The clustering relies on the mean-shift algorithm to find the modes of the position and color distribution of the pixels. In order to produce a non-hierarchical segmentation, we also present a method to prune the computed hierarchy by suppressing nodes that are not stable enough.