基于广义多分辨似然比和混合多尺度自回归预报模型的图像无监督分割

Imagery Unsupervised Segmentation Based on GMLR and MMARP Model

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不仅能融合信号的多个特征量,增大不同信号间区分度,而且在融合时无需假定各特征量之间的相互关系,这使得它能进行比较精确而方便的判别分析。在SAR(synthetic aperture radar)图像分割应用背景中,利用混合多尺度自回归预报(mixture multiscale autoregressive prediction简称MMARP)模型估计预报图像的GMLR的原假设和备择假设参数,然后将判别准则应用到预报图像,从而对原SAR图像进行分割。实验与几种流行的SAR图像分割方法进行了比较,结果表明了该理论方法的显著性:不论从分割的精度,对噪声的敏感度,还是从边缘的光滑度考虑,都优于上述通常的分割方法。

A generalized multiresolution likelihood ratio （GMLR） is defined, then the GMI.R test is obtained. The GMLR has the characteristic that can fuse several features which describe different properties, and it can increase distinction between different source outputs, so it is more precise to make a decision. In SAP, imagery segmentation, in order to obtain unsupervised segmentation, an efficient mixture multiscale autoregressive prediction （MMARP） model is applied to estimate the parameters of null hypothesis and alternative hypothesis in the GMLR. Finally we classify each individual pixel based on a test window. The method compared with recent competing methods, demonstrating that our method performs better.