基于自编码神经网络的高分辨率距离像降维法

为了提高支持向量机(SVM)分类效率,大幅减少以高分辨率距离像(HRRP)功率谱为特征的支持向量机目标识别分类器的计算量,采用自编码神经网络深度学习方法,实现高维、非线性HRRP功率谱的数据降维。在此基础上,提出了Autoencoder-SVM模型,综合利用自编码神经网络的特征提取能力和SVM的分类能力。仿真结果显示,在HRRP功率谱降维方面,自编码神经网络的降维效果远好于核主成分分析和等距映射算法,其降维结果对SVM分类结果影响甚微,但大幅缩短了SVM的计算时间;同时,在隐层节点数相同的情况下,随着隐含层数的增加或者深度的增加,自编码神经网络数据降维或特征提取效果更好。

Dimension reduction method of high resolution range profile based on Autoencoder

To reduce the computational load of support vector machine(SVM) target recognition classifier based on power spectrum feature of high resolution range profile (HRRP) and improve its classification efficiency, the data dimension of HRRP power spectrum with high dimension and nonlinearity was reduced by using a depth learning method Autoencoder. On this basis, the Autoencoder-SVM model was put forward, which combines Autoencoder feature extraction and SVM classification ability. Simulation results show that the dimension reduction effect of Autoencoder is much better than that of kernel principal component analysis(KPCA) and isometric mapping(ISOMAP) in the dimension reduction of HRRP power spectrum, with little effect on the SVM classification results, but much reduction of the calculation time of SVM. In the case of the same number of hidden nodes, the experimental results also show that with the increase of the number of hidden layers or the increase of the depth, the Autoencoder data dimension reduction or the feature extraction is better.