Cellular Neural Networks (CNN) are excellently suited for image processing. A big challenge thereby is the determination of CNN templates for special image processing tasks. In many cases appropriate templates can only be found by a parameter optimization. Unfortunately, especially in the context of image processing, such an optimization is frequently a difficult task due to a lot of local minima in the error measure. In this contribution we present a new method of optimization that detects a global minimum of an error measure even if the function contains many local minima. To prove this allegation we constructed a number of multidimensional test functions, which have not only a global minimum but also many local minima. We present a comparison between the introduced Iterative Annealing method and other analytical and statistical optimization methods. Furthermore, by using the new optimization method we realized a feature point extractor with CNN.