在雷达目标识别中, 利用核主分量分析(KPCA)方法来进行目标特征提取, 忽略了高分辨率距离像(HRRP)的本身特性。提取一种平移不变特征-中心矩作为特征向量, 采用KPCA进行特征降维; 由于BP神经网络易陷入局部极小, 采用遗传算法(GA)对BP网络节点权值和阀值进行优化选择。基于雷达实测数据的实验结果表明: 平移不变的KPCA特征提取方法实现了平移不变和降维的结合, 同时, 利用GA优化BP神经网络提高了分类器稳定性改善易陷入局部最小的缺陷, 提高了雷达目标识别的性能。

When using the method of kernel principal component analysis(KPCA) to extract feature of target in radar target recognition, the HRRP characteristic is ignored. A translation invariant features-central moments was extracted as feature vector, KPCA was used to reduce the dimensionality; The BP neural network was easy to fall into local minimum, the genetic algorithm(GA) was used to optimize the BP network node weights and threshold. The experimental results based on the measured radar data show that the translation invariant KPCA feature extraction method achieve the combination of translation invariant and descending dimension, and the BP neural network optimized by GA improves the stability of classifier and improves the defect of falling into local minimum easily.