Cascaded neural networks based image classifier

The authors present a texture image classification system based upon the use of two cascaded multilayer feedforward neural networks (MFNNs). The first network transforms a set of high-dimensional and correlated feature images into another set of uncorrelated principal feature images with its dimensionality being significantly compressed while minimizing the information lost. The second accomplishes the task of feature pattern classification by using only those principal features obtained by the former. A synthesized training system for synchronously learning the weights of these two networks is also presented. Important advantages of both the classification system and the associated training system are described. They are further demonstrated by detailed examples.<>