基于目标特征的动态支持向量机研究

Dynamic Support Vector Machine Study based on Target Feature

研究了将待识别目标特征与SVM相结合的动态SVM。提出一种以目标特征与每个训练样本间的距离度量SVM软间隔优化问题中惩罚参数C的方法,可根据两者间距离大小赋予每个训练样本一个惩罚参数,从而更好地体现了不同训练样本对于待识别目标特征的价值。然后,根据各样本惩罚参数的大小重构动态训练样本集,训练以待识别目标特征的分类为核心任务的动态SVM,寻求以目标特征为中心的局部空间的最优分类面。并对两类水声目标的识别情况进行了比较,实验表明效果好于SVM和k-近邻分类器。

The DSVM（dynamic support vector machine）was researched by integrating the target feature with SVM.To show better importance of each sample to the target feature,a method was put forward firstly that assigned a penalization-parameter Ci to each training sample.Different from SVM whose Ci was a constant,Ci of DSVM was measured by using the distance between the target feature and each training sample.Furthermore,to search the hyperplane of the local space taking the target feature as center,the DSVM based on the target feature was trained after the training sample set was reconstructed according to the penalization-function Ci.At last,the DSVM was applied in the underwater acoustic target recognition.Experiment results show that the DSVM is more robust than the traditional SVM,and the k-nearest neighbors.