Design of morphological filters by learning

A method for designing morphological filters for image processing is proposed. This method regards morphological filters as a sort of the neural network formulation. However, our formulation has no connecting weights and treats only extents of structural elements. The basic nonlinear operations of our network are morphological erosion/dilation. The learning procedure optimizes the shape and size of the elements. The simulated annealing method is utilized to realize this style of learning. The optimization techniques of the grayscale morphological filters are developed.