Stochastic Learning Networks For Texture Segmentation*
Bangalore S Manjunath, Rama Chellappa · 2005
In this paper we describe Neural Network based algorithms for the segmentation of textured gray level images. We for- mulate the problem as one of minimizing an energy function, derived through the representation of textures as Markov Random Fields (MRF). The texture intensity array is modelled as a Gauss Markov Random Field (GMRF) and an Ising model is used to characterize the label distribution. The resulting nonconvex energy function is minimized using a Hopfield neural network. The solution obtained is a local optimum in general and may not be satisfactory in many cases. Although stochastic algorithms like simulated annealing have a potential of finding the global optimum, they are computationally expensive. We suggest an alternate approach based on the theory of learning automata which introduces stochastic learning into the iterations of the Hopfield network. A probability distribution over the possible label configurations is defined and the probabilities are updated depending on the final stable states reached by the neural network. This iterated hill climbing algorithm combines the fast convergence of the deterministic relaxation with the sustained exploration of the stochastic algorithms. The perfor- mance of this rule in classifying some real textured images is given.