Multispectral image classification using Gabor filters and stochastic relaxation neural network

In this article, we propose a supervised classification scheme for multispectral image data based on the spectral as well as textural features. A filter bank consisting of Gabor wavelets is used to extract the features from the multispectral imagery. The classification model consists of three random processes, namely, feature formation, partition and label competition. The feature formation process models the multispectral texture features from the Gabor filter bank as a multivariate Gaussian distribution. The partition process and the label competition 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) probability. The performance of the scheme is demonstrated on a variety of multispectral multipolar images obtained from SIR-C/X-SAR.