SAR图像分割的Bootstrap广义多分辨似然比检验方法

Generalized Multiresolution Likelihood Ratio Test for SAR Imagery Segmentation with Bootstrap Sampling

提出广义多分辨似然比(generalized multiresolution likelihood ratio,简称GMLR)的概念,给出其Bayes准则下的假设检验和判别准则。GMLR能融合待判别信号的多个特征量,增大不同信号的区分度,所以能更精确地对信号进行判别分析。在SAR(synthetic aperture radar)图像分割的应用背景中,首先用弃除图像冗余信息,减小计算量的Bootstrap样本得到GMLR的原假设和备择假设参数的极大似然估计,然后检测GMLR的分割阈值,最后对森林和草地组成的模拟图像和真实SAR图像分割,证明该方法是SAR图像分割的一个有效途径。

A generalized multiresolution likelihood ratio （GMLR） is defined and Bayes test of GMLR is given for making a decision. The GMLR can fuse different characters of signal so that the distinction of different signals is increased for recognition. Because： （i） the choice of independence pixel sample which would allows an estimation of the statistical parameters of the image in the best conditions of independence; （ii） the reduction of redundancy of information connected to the choice of a small representative sample, allows a gain in a factor n/N^2 （n is the number of bootstrap sample, and the N^2 is the pix number of the image） in times of calculation. So in the application of SAR imagery segmentation, the bootstrap sampling is employed to estimate the parameters of null hypothesis and alternative hypothesis in the GMLR. Simulative and experimental results demonstrate that our method performs fairly well,