SAR图像无监督分割的空间变化混合MAR模型方法

提出SAR（Synthetic Aperture Radar）图像的空间变化混合多尺度自回归（Spatially Variant Mixture Multiscale Auto Regressive，SVMMAR）模型方法，该模型不仅能刻画SAR图像的空间变化性，而且利用了SAR图像多尺度序列的统计特性；采用的分类器是像素标号的极大似然估计，细化的同时简化了传统Bayes分类器；该模型无需预先抑制斑点噪声，就能获得精确分割结果；并且理沧上证明了在图像粗尺度确定分类个数的合理性，在此基础上提出一种在粗尺度确定分类个数的新方法，大大减少了运算量。

SAR Imagery Unsupervised Segmentation Based on Spatially Variant Mixture Multiscale AutoRegressive Model

In this paper, an efficient spatially variant mixture multiscale autoregressive(SVMMAR)model method is presented. The model is capable of not only describing spatially variant characteristics but also exploiting multiscale autoregressive statistical properties of SAR imagery, thus it can describe spatially variant character and filter and reduce the possible effect generated by the presence of speckle noise of SAR images. The classfier is maximum likelihood estimates of the labels themselves, which is refined and the structure is simple compared with Bayes classifier. The model has no use for denoising preprocessing, while precise segmented results can be obtained. Another contribution is a kind of method selecting component number quickly at coarser scale is proposed, and a criterion based on Bayesian Ying-Yang learning theory and system is employed to select component number at coarser scale of SAR imagery, which can reduce computation amount greatly. Experiment results shows that the segmentation results obtained by using the above method are more precise than two popular segmentation methods, the edges are smooth, and the model is less sensitive to speckle noise.