多特征提取和多核SVM的舰船目标识别方法

A Ship Target Recognition Method Based on Multi-Feature Extraction and Multi-Kernel SVM

针对单核SVM分类识别SAR图像舰船目标的低精度问题,提出了一种基于多特征提取和多核学习SVM的SAR图像舰船目标识别方法,从特征提取和分类器训练两个方面提升目标识别的准确度。首先选用公开数据集提取舰船目标的多类特征,然后加权融合多个核函数构造多核SVM模型,最后使用多项特征数据训练识别舰船目标。鉴于多组目标特征存在信息冗余问题,采用相关性系数去除某些信息高度冗余的特征,降低特征维度。通过粒子群优化算法解决了SVM核函数的核参数选择难题。实验结果表明,所提方法有效改善了对舰船目标的识别性能,综合识别准确率由传统SVM的87.18%提高至92.31%。

Aiming at the problem of low recognition rate of single-kernel Support Vector Machine(SVM) for ship target classification in Synthetic Aperture Radar(SAR) image, a method for ship target recognition based on multi-feature extraction and Multi-Kernel Learning(MKL) SVM is proposed, which improves the accuracy of target recognition from two aspects of feature extraction and classifier training.First, the public data set is selected to extract the multiple types of features of the ship target, and then weighted fusion is made to the multi-kernel functions to construct a multi-core SVM model.Finally, multi-feature data are used to train and recognize the ship target.In view of the information redundancy problem in multiple sets of target features, the correlation coefficient is used to remove some highly redundant features and reduce the feature dimensions.The Particle Swarm Optimization(PSO) algorithm is used to solve the problem of kernel parameter selecting of the SVM kernel function.The experimental results show that the proposed method effectively improves performance of ship target recognition, and the comprehensive recognition rate is increased from 87.18% of the traditional SVM to 92.31%.