Adaptive two-dimensional neuron grids

In the last decade many early-vision tasks have been cast into the form of global optimization principles: their solution is obtained by the minimization of appropriate cost functions. The minimization procedure, which consists in most cases of a simple gradient descent, often yields a two-dimensional particle model with local exchange interaction. Our starting point is a quite general representative of such a model, a two-dimensional neuron grid, which is based on a standard neuron model. The optimization principles enter our model via a backpropagation like adaption scheme for the weights. In the case of edge detection the results we arrive at so far are similar to those obtained by the gradient descent methods. So the formalism proposed here may form an alternative basis for more sophisticated image preprocessing algorithms.