Texturerkennung und Segmentation mittels eines neuronalen Netzwerks zur optimalen Berechnung von Waveletentwicklungskoeffizienten

In recent years a broad interest in coupled space-frequency representation envolved as a possible coding for special image analyis problems. The features of these representation enable a significant description of the intrinsic properties of textures. For homogeneous textures a convenient representation is the local fractal dimension. We proved, that a Kohonen feature map network is able to segment image intensity surface values at different resolutions according to their fractal dimension. But for oriented inhomogeneous textures Gabor functions and the functional class of wavelets are a better approach. These functions depend on parameters, which in general are fitted to a given picture analysis problem heuristically. Therefore we suggest to calculate them by a Kohonen feature map in analogy to the fractal dimension.