Neural network recognition of textured images using third order cumulants as functional links

The proposed texture classification method involves parameter estimation using a recursive instrumental variable approach. A linking of the output third-order cumulants of the textured image data to the estimation of the autoregressive (AR) model parameters is proposed. Through one approach, the instrumental variable is generated as the output of a moving average (MA) linear filter whose weights are chosen such that the MA filter output matches the third-order cumulants of the texture data being modeled. Generating the instrument variable in such a manner is attractive since it will be uncorrelated with the AR model measurement noise; therefore, the extracted feature vector will tend toward the optimum parameter vector. The proposed method also yields a pattern vector representing a model that captures the third-order statistics of the data. The length of the feature vector chosen for classification corresponds to the AR and MA model orders. The classification of the feature vectors was performed by a multilayer perceptron artificial neural network (ANN) classifier. The network paradigm chosen was a feedforward neural net, using backpropagation learning