Abstract:
Our aim is to detect homogeneously colored regions invariant to
surface orientation change, illumination, shadows and highlights in
multi-spectral images where the spectral range corresponds to the
visible wavelength interval. To this end, the influence of
multi-spectral sensor space, normalized multi-spectral sensor space,
and hue color space are examined, in theory, for the dichromatic
reflection model and, in practice, for segmentation techniques based
on k-means clustering. We show that homogeneously colored regions can
be detected invariant to surface orientation change, shadow and
highlights under the condition of equal-energy illumination where
= e = constant.
In this paper, we first present a method that achieves, in theory and in practice, an approximation of equal-energy illumination. The method requires that the spectral distribution of the illuminant is known. Secondly, we derive in theory three cluster models: points, lines and planes and show the invariance for each model to surface orientation change, illumination, shadows and highlights. We then present segmentation algorithms which incorporate these models and which are based on the k-means clustering technique. Experiments are conducted on multi-spectral images taken from colored objects in real-world scenes.
On the basis of the theoretical and experimental results on multi-spectral images it is concluded that the line and plane model detect regions invariant to a change in surface orientation, viewpoint of the camera, and illumination intensity. Furthermore, the plane model also detect regions independent of highlights. The point model provides segmentation results which is sensitive to surface orientation and illumination intensity.