融合上下文信息的多尺度贝叶斯图像分割

提出了一种融合上下文信息的多尺度贝叶斯图像分割算法，基于多尺度MRF图像模型，将模型中各结点的邻域结点类别信息抽象为上下文，求得结点的后验边缘概率之后，在各尺度融合表征了同一尺度内及相邻尺度的邻域信息的上下文，结点在相邻结点信息的指导下，得到的分割结果在均匀区域内部及区域边界都大为改善，而且没有增加模型的复杂度，算法仍然是快速的、非迭代的．融合过程中的参数采用EM算法估计．分析和实验结果表明算法是有效的．

Multiscale Bayesian Image Segmentation Fusing Context Information

In this paper, a multiscale Bayesian image segmentation algorithm is proposed. C lass label is modelled using multiscale Markov random field. Causality exists be tween scales. So non-iterative probability inference procedure can be obtained. Based on this image model, the neighboring information of a node in the model i s extracted as context. A node context includes label information of the same sc ale neighboring and adjoining upper scale neighboring nodes. This context inform ation is fused via Bayesian rules to scales after posterior marginal probability derived. Segmentation results are greatly improved both on segmentation accurac y and boundary localization after context information is considered. Context inf ormation compensates the deficiency of the multiscale model. The advantage that no complexity added to the model is important. The derived algorithm is fast and non iterative too. The parameter required in fusion can be estimated by EM algo rithm or specified directly. The analysis and experimental results demonstrate t he effectiveness of this novel method.