中文摘要：

采用SVM方法估计似然概率，克服了混合高斯模型对观测场建模不准确的问题；通过改进的多尺度MRF模型，在标记场建模时充分考虑了各尺度之间和同一尺度内的相关性，进一步准确描述了标记场的统计特性。最后利用改进的建模方法，在序贯最大后验概率估计算法框架下进行影像分割。通过对人工合成影像和实际遥感影像的分割实验结果分析，证明了本文方法能够有效提高分割效果。

英文摘要：

Current image segmentation methods have two problems in Bayesian frameaork, one is that the estimation of probability distribution for the observed filed is not accurate, the other is that making use of the relevant information of tag field is not enough. This article o- vercomes the problem of inaccurately modeling the observed field by Gauss matured model, through the approach which estimates the likelihood probability by SVM method. Mean- while, this article makes full use of the statistical characteristic of tag field through the im- proved multi-scale MRF model which takes full account of the correlation within various scales and between the same scale. Finally, the proposed image segmentation approach uses the improved modeling approach under the framework of sequential maximum a posteriori probability estimation algorithm （SMAP）. The proposed method is proved to improve the accuracy of segmentation through the experimental results of the artificial synthetic and real remote sensing image.