On robust jump detection in regression surfaces with applications to image analysis

In the one-dimensional case the difference of two one-sided kernel estimators can be used to detect discontinuities in regression functions. In smooth regions, an estimator using only observations on the left side will be similar to the estimator using only observations on the right side. In contrast, near jump points, the difference of these two estimates will be close to the jump height. Based on this method, we use in this thesis the difference of two rotated robust one-sided M-kernel estimators. For a special model, consistency results are shown. For more general situations, statistical tests for detection jump points are derived and it is shown, how these detected points can be further processed. To show the advantages resulting from the use of robust estimators, comparative simulations are performed.