用于SVM的RCS统计特征集约减方法

传统的支持向量机分类算法对训练样本数目敏感且不具备增量学习的能力，而空间目标识别的工程应用需要积累样本进行大样本的增量学习。根据样本在特征空间分布，提取几何意义上边缘的样本点，成功约减了应用于支持向量机学习的基于雷达散射截面统计特征的训练样本集。利用中心距离比和特征空间多维高斯分布特性，分别提取两类边界样本集和单类边缘样本集；再采用直推式实验设计方法再采样，根据样本潜在结构分布信息选择最能代表样本集结构分布的高价值样本。实验结果表明：样本初选算法能够在有效约减样本集规模的同时保持支持向量机训练分类的精度。

A Method for Reducing Radar Cross Section Data in Support Vector Machine Classification

Traditional support vector machine (SVM) has difficulty in incremental learning and its algorithm complexity is sensitive to the number of samples, while the space separation target recognition requests the algorithm with large data set and incremental learning ability. A new geometric fast sampling algorithm for support vector machine is proposed and used for the target recognition based on radar cross section (RCS) data. Since the training samples most likely to become the support vectors are located in the geometric edge region, the center distance ratio and the multidimensional Gaussian distribution probability density are used to extract samples from the single label edge section and the classification intersection margin. Transductive experimental design further reduces the number of the samples by selecting the most information data points to represent the distribution information. The narrowband radar experiment data are extracted by the proposed sampling method and classified by the support vector machine. Experimental results demonstrate that this algorithm could select most valuable samples in the original training set and has little influence in classification precision.