Unsupervised segmentation of color images based on k-means clustering in the chromaticity plane

Presents an original technique for unsupervised segmentation of color images which is based on an extension (for use in the u'v' chromaticity diagram) of the well-known k-means algorithm, which is widely adopted in cluster analysis. We suggest exploiting the separability of color information which, represented in a suitable 3D space, may be "projected" on to a 2D chromatic subspace and on to a 1D luminance subspace. One can first compute the chromaticity coordinates (u', v') of colors and find representative clusters in such a 2D space by using a 2D k-means algorithm, and then associate these clusters with appropriate luminance values by using a 1D k-means algorithm, which is a simple dimensionally-reduced version of the 2D one. Experimental evidence of the effectiveness of our technique is reported.