基于核方法和主成分分析(PCA)的探地雷达目标特征提取新方法

常用探地雷达目标特征提取方法LDA(又称为FDA)直接在低维的探地雷达数据空间提取探地雷达目标特征, 提取的探地雷达目标特征的区分度小;常用的正则化技术存在正则化参数选取困难的问题。本文提出先通过基于核方法的 非线性变换把低维的探地雷达样本数据投影到高维空间,然后在高维空间中用PCA对奇异的核矩阵降维重建,最后对重 建后的非奇异核矩阵用LDA提取探地雷达目标特征。对实测数据的对比处理分析表明,本文所提探地雷达目标特征提取 方法优于其它方法。

A New GPR Target Feature Extraction Method Based on Kernel method and PCA

The method of the conventional ground penetrating radar (GPR) target feature (GPRTF) extraction is LDA. Due to it extracts the GPRTF in the low-dimension space directly, the discriminability between the GPRTF extracted is limited. The kernel method is proposed to map the low-dimension GPR data into the high-dimension space by nonlinear transformation, before performing the LDA. Due to the kernel matrix is always singular and the LDA can't be used directly, and it is difficult for us to fix the regularization parameter in the conventional regularization method. The PCA is proposed to reduce the dimension and reconstruct the kernel matrix. Then the LDA is performed in the kernel matrix reconstructed. The treatment comparisons of the measurement data show that the method proposed is batter than the conventional methods.