A multiscale approach to image segmentation using Kohonen networks

Abstract

An approach is developed to multiscale image segmentation, based on pixel classification by means of a Kohonen network. An image is described by assigning a feature pattern to each pixel, consisting of a scaled family of differential geometrical invariant features. The invariant feature pattern representation of a training image is input to a Kohonen network, in order to obtain a description of the feature space in terms of so-called prototypical feature patterns (the weight vectors of the network). Supervised labeling of these prototypical feature patterns may be accomplished using classes derived from an a priori segmentation of the training image. We can segment any image similar to the training image by comparing the feature pattern representation of each pixel with all weight vectors, and assigning each pixel the class of the best matching weight vector. In our study we evaluated the benefit of applying features at multiple scales, as well as the effects of first and second order information on the results.