Image segmentation and labeling using SOMs

Autoren Mario Drobics
TitelImage segmentation and labeling using SOMs
TypTechnischer Bericht
NummerSCCH-TR-0106
OrtHagenberg, Austria
InstitutionSCCH
AbteilungKVS
Jahr2001
SCCH ID#106
Abstract

Image segmentation deals with the problem, of finding segments of related pixels in an image. We use a Self-Organizing Map to create a representation of the feature space of the image. This mapping conceives the topology of the pixels coordinates, as well as the topology of the space, spanned over the color dimensions. By applying a fuzzy clustering method, we are able to identify regions of similarity in this mapping. To find descriptions of this regions, we use an inductive learning method afterwards. Finally, we compute the fuzzy class label for each pixel in the original picture, to create a segmentation of the whole image.