利用支持向量数据描述和递归特征消除的水下慢速小目标轨迹特征选择方法  

Underwater low-speed small target trajectory feature selection using support vector data description and recursive feature elimination

作　　者：赖凯 刘雄厚[1,2,3] 杨益新 LAI Kai;LIU Xionghou;YANG Yixin(School of Marine Science and Technology,Northwestern Polytechnical University,Xi’an 710072;Shaanxi Key Laboratory of Underwater Information Technology,Xi’an 710072;Hanjiang National Laboratory,Wuhan 430061)

机构地区：[1]西北工业大学航海学院,西安710072 [2]陕西水下信息技术重点实验室,西安710072 [3]汉江国家实验室,武汉430061

出　　处：《声学学报》2025年第2期475-485,共11页Acta Acustica

基　　金：国家自然科学基金项目(U2341203,12274346);国家重点研发计划项目(2016YFC1400200)资助。

摘　　要：针对水下慢速小目标已有跟踪轨迹特征性能不优、信息冗余而导致分类识别性能不佳的问题,提出了利用改进支持向量数据描述(ISVDD)和递归特征消除(RFE)的ISVDD-RFE轨迹特征选择方法。首先,所提方法选择单分类SVDD以适配小目标分类识别所面临的小样本、类不平衡问题,通过逐步递归消除实现小目标的轨迹特征优选;其次,为提升SVDD-RFE轨迹特征选择能力,从递归效率、相关性和稳健性三个方面改善递归过程;最后,为克服SVDD缺乏全局信息的固有缺陷,从特征区分性和特征态势两方面评估所选轨迹特征,提升整体分类识别性能。实测数据处理结果表明,采用所提方法进行轨迹特征选择后,蛙人目标的精确率从93.8%提升至94.9%,召回率从84.7%提升至91.1%;无人水下航行器目标的精确率从89.0%提升至94.7%,召回率从83.1%提升至85.2%;小目标平均分类准确率从87.7%提升至91.5%。在小样本、类不平衡条件下,所提方法具有优于传统方法的性能。For underwater low-speed small targets,the poor performance of trajectory feature and information redundancy lead to the degradation of the classification and recognition performance.This paper proposes an improved trajectory feature selection method based on the improved support vector data description(ISVDD)and the recursive feature elimination(RFE),abbreviated as ISVDD-RFE in abbreviation.First,to solve the problem of small sample size and class imbalance in small target classification,the ISVDD is employed,with trajectory features selected through step-by-step recursive elimination.Then,to enhance the trajectory feature selection capability of SVDD-RFE,improvements are made to the recursion process from three aspects:recursion efficiency,feature correlation,and robustness.Finally,to overcome the inherent limitation of SVDD in lacking overall information,the selected trajectory features are evaluated from the perspective of feature distinguishability and feature situation,which improves the overall classification performance.Experimental results show that the proposed method improves the precision of frogman targets from 93.8%to 94.9%,and the recall from 84.7%to 91.1%.For unmanned underwater vehicle(UUV)targets,the precision increases from 89.0%to 94.7%,and the recall rises from 83.1%to 85.2%.The average classification accuracy of small targets is improved from 87.7%to 91.5%.Under the small sample size and class imbalance,the proposed method outperforms traditional methods.

关 键 词：水下慢速小目标 分类识别 轨迹特征 特征选择 支持向量数据描述 递归特征消除 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]