针对基于单一极化特性增强的极化SAR图像目标检测方法的缺陷,该文将DP(Dirichlet Process)混合隐变量SVM模型(DPLVSVM)应用于极化SAR图像舰船目标检测,提出一种基于多极化散射机理的检测方法。该方法通过联合Dirichlet过程混合与BayesSVM模型,将信号空间划分成若干局部区域,然后在每一局部区域学习一个独立的极化检测器,并将各局部检测器进行组合实现全局多极化散射机理的目标检测。模型采用非参数化Bayes方法自动确定局部区域数量,在完全Bayes框架下,将局部区域划分及检测器学习进行联合优化,保证了各局部区域样本的可分性。另外,为了降低极化特征冗余,该文进一步提出带特征选择功能的稀疏提升DP混合隐变量SVM模型(SPDPLVSVM),提高模型的推广能力。该模型由于采用共轭先验分布,因而可以利用Gibbs采样方法进行高效求解。在RADARSAT-2数据上进行的实验验证了所提方法的有效性。
Considering the shortcoming of detection method based on polarimetric contrast enhanced with single polarimetric scattering mechanism, a PolSAR detection method based on multiple polarimetric mechanisms called Dirichlet Process mixture of Latent Variable SVM (DPLVSVM) is proposed. By assembling a set of local polarimetric detectors that based on single polarimetric scattering mechanism, a global multiple polarimetric scattering mechanisms detector is obtained. With a fully Bayes treatment, DPLVSVM learns the clustering and the local detectors jointly. Taking the advantage of Bayes nonparametric, DPLVSVM handles the model selection problem flexibly. Further, in order to reduce the redundancy of polarimetric feature and improve the model generalization, a model with feature selection, Sparsity-Promoting Dirichlet Process mixture of Latent Variable SVM (SPDPLVSVM), is proposed. Thanks to the conjugate property, the parameters in both of models can be inferred efficiently via the Gibbs sampler. Finally, the proposed models on RADARSAR-2 dataset is implemented to validate their effectiveness.