Nonlinear noise filtering and texture recognition by an optoelectronic neural network that implements a mean field annealing algorithm

An optoelectronic parallel feedback loop capable of extracting coherent pixel clusters from noisy images is presented. The architecture implements a mean field annealing algorithm and corresponds to a Hopfield-Tank-like neural network with dynamical threshold functions. By combining the noise filtering algorithm with so-called local energy maps specifying the energy contents of specific wavelets at different locations it is possible to accomplish texture recognition in noise. Novel methods have been developed in order to reduce optical misalignment problems and intensity distortions within the optical loop. Experimental results are reported. They are found to be in good agreement with those obtained in computer simulations.