Applying scaled differential invariant features to image segmentation with Kohonen feature maps

An approach to image segmentation has been developed on the basis of scaled differential geometrical invariant features for describing pixels properties and utilizing Kohonen networks for probabilistic pixel classification. The invariant feature pattern representation of a training image is input to a Kohonen network, of which the weight vectors tend to form a so-called Kohonen feature map. Supervised labeling of the weight vectors in the map is accomplished using classes derived from an a priori segmentation of the training image. Any image similar to the training image can be segmented by presenting the feature pattern representation of each pixel to the map and interpreting the caused excitation pattern. The applied features yield a mathematically thorough and complete description of arbitrary image structures up to any desired order. The Kohonen map has successfully been applied although the classification problem is nonlinear. Furthermore, the map provides a means to obtain probabilistic segmentations.