Adaptive Anisotropic Parameter Estimation in the Weak Membrane Model

The weak membrane model uses Markov Random Fields within the Bayesian inference framework for image reconstruction and seg-mentation problems. Recently, the model has been extended for the 4D Gabor feature vector space and was applied to texture segmen-tation. A limitation of this technique is that the parameters in the model have to be adjusted for each diierent input image and they are xed throughout the image. This paper proposes a technique to alleviate this limitation by estimating the parameters using local feature statistics. The technique has the following desirable properties: 1) the whole segmentation process is done in an unsupervised fashion, 2) robustness to noise and contrast variation, and 3) increased connectivity of boundaries.