Segmentation of Natural Images by Non-linear Scale-Space linking

A new method of image segmentation based on building up a hierarchical tree through non-linear scale-space is proposed. Non-linear scale space creates from the original image the whole family of blurred images but keeps sharp region boundaries. This characteristic of preserving edges turns out to be extremely useful in choosing a linking strategy through scale. Segmentation is realized in two stages. First, the set of blurred images of the scale-space are split into a number of spatial regions composed of pixels with similar intensity levels. Second, a parent-child linking between the regions of successive layers is performed based on tracking level sets through scale. Such process will generate a scale-space hierarchical tree that induces a segmentation without a priori knowledge. Experimental results based on natural images with respect to the hierarchy and segmentation are given. Keywords— Scale-Space, PDE , unsupervised segmentation, edge detection, hierarchical tree.