MULTISPECTRAL IMAGE CLASSIFICATION USING GABOR FILTERS AND STOCHASTICRELAXATION NEURAL-NETWORK

In this article, we propose a supervised classification scheme for mul tispectral image data based on the spectral as well as textural featur es. A filter. bank consisting of Gabor wavelets is used to extract the features from the multispectral imagery. The classification model con sists of three three random processes, namely, feature formation, part ition and label competition. The feature formation process models the multispectral texture features from the Gabor filter bank as a multiva riate Gaussian distribution. The partition process and the label compe tition process represent a set of label constraints. These constraints are represented on a Hopfield neural network model, and a stochastic relaxation strategy is used to evolve a global minimum energy state of the network, corresponding to the maximum a posteriori (MAP) probabil ity. The performance of the scheme is demonstrated on a variety of mul tispectral multipolar images obtained from SIR-C/X-SAR. (C) 1997 Elsev ier Science Ltd.