GENERATION OF COMPLEX STOCHASTIC TEXTURES USING CELLULAR NEURAL NETWORKS

Abstract

The following paper explores possibilities of an application of Cellular Neural Networks for spontaneous generation of complex stochastic textures. The considered class of textures includes images, which are combinations of large-scale and fine-scale stochastic patterns. Accurate modeling of properties of such complex stochastic fields by means of a single 1-neighborhood CNN poses great difficulties. Therefore, we propose a method that exploits texture decomposition into two elements and involves an application of two networks, designated for performing distinct partial tasks. The first step of the proposed CNN-based texture generation procedure focuses on modeling of a coarse, large-scale textural pattern. The second step refines this pattern into a final image that is expected to closely imitate a target texture.