Deriving rules from evolutionary adapted texture filters by neural networks

This work is motivated by the recently proposed 2D-lookup framework for the evolutionary and data-driven adaptation of texture filters. A class of images, the 2D-lookup matrices, appears to play an important role for the performance of the adapted texture filters. Two approaches for approximating these 2D-lookup matrices by neural networks are presented, one based on the multilayer backpropagation neural network (MBPN), and the other based on the unit RBF network. While the MBPN approach gives only a rough approximation of the 2D-lookup matrices, the unit RBF approach approximates these images better, especially for specific details at a lower scale. Also, the unit RBF approach is faster and more simple to handle, and its outcome serves a texture model based on fuzzy rules.