基于区域确定的分层马尔可夫模型及其MPM算法

基于四叉树的分层马尔可夫随机场 (Markov random field, MRF) 模型在层间存在因果性, 不需要像非因果马尔可夫随机场模型那样的迭代算法, 但是传统的分层 MRF 模型常常导致分割结果具有块状现象和非连续边缘. 本文提出一种新的基于区域确定的半树分层 MRF 算法, 并推导出它的最大后验边缘概率 (Maximizer of the posteriori marginal, MPM) 算法. 在流域算法过分割结果的基础上, 该模型将层间的点概率转换为区域概率, 采用区域概率实现各层图像分割. 从 SAR 图像的监督分割实验结果来看, 本文提出的模型较好地克服了基于像素分层模型和单分辨率 MRF 模型带来的块现象和非连续边界, 因而具有更好的分割结果.

A Region-determined Hierarchical Markov Model and Its MPM Algorithm

The noniterative algorithm of discrete hierarchical Markov random field(MRF)model has much lower com- puting complexity and better result than its iterative counterpart of noncausal MRF model,since it has causality property between layers.However,traditional hierarchical MRF model always results in the block artifacts and discontinuous edges. In this paper,a new region-determined half tree hierarchical MRF model is proposed and its region-determined maximizer of the posteriori marginals(MPM)algorithm is inferred. Based on over-segmentation of the watershed algorithm,the proposed model converts pixel probabilities between layers into region probabilities and obtains the final segmentation. The experiments on supervised SAR image segmentation demonstrate that the proposed method performs better than the pixel-based hierarchical model as well as the Gibbs sampler with the single resolution model.