Robustness of computer vision algorithms requires stability of the computed results against variations in the input data caused by noise or modelling uncertainty. In unsupervised image processing tasks like texture segmentation the extracted image partition should not depend on the specific texture data but should extract reliable model estimates of the different texture types. Overfitting of texture models has to be avoided by robustness against within--class texture variability, i.e., segmentation solutions have to generalize from the given texture samples to new instances of the same texture type.

In this talk I will present an extension of the Empirical Risk Minimization induction principle to distributional clustering for texture segmentation. This analysis yields determinististic annealing algorithms with a finite stopping temperature. Overfitting phenomena for segmentation of mondrians of Brodatz textures are documented empirically.