Kohonen's self-organizing maps for contour segmentation of gray level and color images

Abstract

This paper is concerned with contour segmentation of gray level and color images. For segmenting gray level images, the set of gray levels is, in a first step, quantized by a one dimensional self-organizing map. Contour segmentation is the second step: one computes the set of spatially close pixels mapped onto distant cells. Noise reduction in the segmentation can be achived when spatial and gray level pixel components are quantized as a whole: the image is considered a set of points in a three-dimensional space and quantized with a three dimensional map. This way, any two pixels close in gray levels and positions are mapped onto two close map cells. These two methods have straightforward extention to color image segmentation.