摘要
首先在多分辨四叉树上定义了一个广义多分辨似然比,刻画并且累积了SAR(syntheticaperture radar)图像中目标与背景在不同分辨率上的差异,从而增大了目标与背景之间的区分度。为了达到图像无监督分割目的,提出一个有效的空间变化混合多尺度自回归(spatially variantm ixture m u ltiscale autoregressive简称SVMMAR)模型方法,利用该模型分别估计出每个分辨率上广义多分辨似然比中一组密度函数的参数。为了考虑被分类象素与周围象素之间的M arkov性,减弱对噪声的敏感性,利用开窗技术来确定中心象素点的类别。实验中与通常的分割技术作了比较,也表明该方法不论从分割的精度,对噪声的敏感度,还是从边缘的光滑度都表明该方法具有较强优势。
In the full paper we begin with a somewhat detailed analysis of the shortcomings of existing algorithms for segmentation of synthetic aperture radar （SAR） imagery. We aim at ameliorating some of these shortcomings with a mathematically more efficient new algorithm. We define a generalized multiresolution likelihood ratio （GMLR） which generalizes the classic likelihood ratio, and which, after identifying the difference between background and target for each of the multiresolutions of SAR imagery, sums up all the differences. So the GMLR increases the distinction between background and target. In order to get unsupervised segmentation, an efficient spatially variant mixture multiscale autoregressive （SVMMAR） model is proposed and applied to estimating the parameters of the GMLR easily. Then, we classify each individual pixel based on a test window. Finally, we give experimental results, which show preliminarily that our method improves unsupervised segmentation of SAR imagery in three aspects： （1） it segments the background and targets precisely; （2） it gets clear and smooth edges between background and targets; （3） it is not sensitive to speckle noise.
出处
《西北工业大学学报》
EI
CAS
CSCD
北大核心
2005年第5期666-670,共5页
Journal of Northwestern Polytechnical University
基金
国家自然科学基金(60375003)
航空基础科学基金(03I53059)资助
关键词
广义多分辨似然比
空间变化混合多尺度自回归模型
无监督分割
generalized multiresolution likelihood ratio （GMLR）, spatially variant mixture multiscale autoregressive （SVMMAR） model, unsupervised segmentation
