多核集成支持向量机合成孔径雷达目标分类

Multi‑kernel Ensemble SVM for SAR Target Classification

雷达数据的复杂性增加了合成孔径雷达(SAR)目标分类和识别的难度。传统的多核分类方法先在单个再生核希尔伯特空间(RKHS)中学习基核的线性组合,再通过学习相关参数对SAR目标进行分类。由于传统方法忽略了多核中的最佳参数选择以及多核之间的结构特征,因此会导致识别结果出现偏差。鉴于此,提出了一种多核集成支持向量机SAR目标分类方法,通过设计一个集成损失函数将多个单独的核分类损失进行集成,从而将多个单独的核分类模型统一成一个整体,并共同优化和学习多个单独模型的最优参数。试验结果表明,该方法与SimpleMKL和SpicyM⁃KL等方法相比,在移动与静止目标搜索与识别(MSTAR)计划的多类别SAR数据集上的分类识别准确率可提升0.5%~10%。

The complexity of radar data increases the difficulty of synthetic aperture radar(SAR)tar⁃get classification and recognition.The traditional multi-kernel classification methods learn the linear combination of basis kernels in a single reproducing kernel Hilbert space(RKHS),and then the SAR target is classified and identified by learning the related parameters.Due to the traditional method ig⁃nores the optimal parameter selection and the structural characteristics of multi-kernel,the result of recognition will be biased.In view of this,a method of multi-kernel ensemble SVM for SAR target classification is proposed,thus it ensembles the multiple independent kernel classification losses by de⁃signing a ensemble loss function,and so as to unify the multiple individual kernel classification models into a whole,and jointly optimize and learn the optimize parameters of multiple individual models.Ex⁃perimental results show that the classification accuracy of this method on the multi-category SAR data set of the moving and stationary target acquisition and recognition(MSTAR)plan is improved by 0.5%to 10%compared with the SimpleMKL method and SpicyMKL method.