Texture mixture modelling using the neural network approach

We propose a hybrid image model based on a mixed neuro-Markovian approach. An image is supposed to be composed of homogeneous regions separated by intermediary zones. The homogeneous regions are filled with textures and the intermediary zones are transition areas where an interpolation between two bordering regions is done. The models adopted for homogeneous regions are Markov random field (MRF) type, and the intermediary zones are generated by applying two different neural networks (NN). It is possible to generate pixels which belong to several classes at the same time. Each MRF is described by a marginal probability distribution function, and correlation coefficients defined in two orthogonal directions. The NN is characterized by the intermediary zone order and a transition function defining the spatial dynamics of parameters changes