Unsupervised texture segmentation based on the modified Markov random field model

The Gaussian-Markov random field (MRF) model is a very useful technique for image processing, such as feature extraction and data compression. However its strict stability condition makes the model identification complex. The major problem is the choice of a proper support region for the model. In this paper a new model is proposed which is based on the MRF model and called the modified Gaussian-Markov random field model. It is not an optimal MRF model but has a very useful property, namely decorrelation. A stable modified MRF model always exists even if a stable MRF model does not exist on the given support region. Applications to texture segmentation are also presented