利用TMF和置信传播的无监督SAR图像分割

Unsupervised SAR image segmentation using TMF and belief propagation

针对三重马尔可夫场模型传统统计推断方法无法兼顾分割精度和计算效率的问题,提出了一种高效的利用置信传播的三重马尔可夫场模型统计推断方法,并应用于无监督合成孔径雷达图像分割.该算法结合三重马尔可夫场模型和合成孔径雷达图像统计特性,将图像分割问题转化为三重马尔可夫场的最大后验边缘估计问题.针对三重马尔可夫场中的两个标记场,将置信传播算法推广到二元情形,通过消息传递的方式估计双标记场的联合后验边缘概率,并依据最大后验边缘准则同时实现两个标记场的估计.模拟图像和实测合成孔径雷达图像的仿真实验结果表明,该算法能有效抑制相干斑的影响,能以合理的计算代价获得精确的分割结果.

To solve the problem that the traditional statistical inference approach for the triplet Markov fields （TMF） model cannot balance segmentation accuracy and computational efficiency, an efficient statistical inference approach for the TMF model using belief propagation is proposed, and then applied to unsupervised synthetic aperture radar （SAR） image segmentation. The algorithm combines the TMF model and the statistical property of the SAR image, and translates the segmentation problem into maximization of the posterior marginal （MPM） estimation. For the two label fields in TMF, the belief propagation algorithm is generalized to the bivariate case to estimate the joint posterior marginal probability of the two label fields through message passing. The two label fields can he simultaneously estimated according to the MPM criterion. Experiments on both simulated and real SAR images demonstrate that the proposed algorithm can efficiently suppress the influence of the speckle, and obtain accurate segmentation results with a reasonable computational cost.