Contents I. Introduction
Image segmentation is usually an important issue prior to the manipulation and analysis of remotely sensed images, such as multispectral SPOT images, aerial photos, and synthetic aperture radar (SAR) images. Bayesian approaches to image segmentation have been proven efficient in integrating both image features and prior contextual properties, where maximum a posteriori (MAP) estimation is usually involved. In [1], [2], a Markov random field (MRF) was developed to model contextual behavior of image data, and Bayesian segmentation becomes the MAP estimate of the unknown MRF from the observed data. Since the MRF model usually favors the formation of large uniformly classified regions, it may over-smooth texture boundaries and wipe off small isolated areas. Moreover, the noncausal dependence structure of MRFs typically results in high computational complexity. To reduce computational complexity and improve classification accuracy, researchers proposed multiscale techniques that apply contextual behavior in the coarser scale to guide the decision in the finer scale and to retain the underlying MRF model in each fixed scale, e.g., [3], [4]. In particular, in [4], Markovian dependencies are assumed across scales to capture interscale dependencies of multiscale class labels with a causal MRF structure, so that a noniterative segmentation algorithm was developed where a sequential MAP (SMAP) estimator replaces the MAP estimator.
