Image generation and inversion based on a probabilistic recurrent neural model

The iterated function system composed of contraction mappings can generate various intricate images with very few parameters in simple and iterative computations. A recurrent neural model with probabilistically weighted connections is proposed as a nonlinear iterated function system. To find connection weights of the recurrent neural model that generates an approximate version of a given gray scale image, an adaptive function estimation method, using the square error criteria, is proposed. Its coding efficiency is evaluated.<>