Unsupervised texture analysis using a robust stochastic image model

This paper considers a new approach of stochastic image modelling (SIM) for unsupervised textured image segmentation. In robust SIM, an efficient Gibbs Markov random field (GMRF) is derived in order to consider only the relationship between a centre pixel and its neighbourhood without a priori knowledge of the distributions of texture patterns in an observed image. Furthermore, in order to apply the efficient GMRF to real images (>64 graylevels), this paper proposes a robust image model, which analyses the relationships in terms of grayslices instead of actual graylevels in a neighbouring system. Also, the grayslice in the robust image model is defined in terms of relative Euclidean distances of parameter estimates in an observed texture image. The robust SIM in this paper extends the traditional second-order neighbouring system into a quasi third-order neighbouring system without increasing the number of parameters to be estimated. For stable and consistent parameter estimates, the least-square (LS) method is adopted with normalisation of the estimated parameters. Maximum a posteriori (MAP) criterion is used for the segmentation algorithm, with subsequent MRF postprocessing in order to decrease the misclassification errors. Finally the unsupervised segmentation results of a variety of natural texture images are presented.