Image processing using an image approximation neural network
E.S. Dunstone · 2002
Discusses a novel neural network architecture for use in image representation and processing. In general methods for using neural networks for image processing have been largely derived from the use of conventional techniques. The neural network demonstrated in this paper, however, provides a new way of abstracting and consequently processing the image data. This is achieved by treating the image as a two-dimensional surface and training a network to learn an approximation to the parametric equation which describes this surface. To achieve this goal an image approximation neural network is proposed. This network has a modular architecture to allow the encoding and integration of several separate image regions. A technique for using IAN networks to perform affine image processing operations quickly and in a scale independent manner is derived and demonstrated.>