基于去相关邻域保持判别投影的声目标特征提取

Feature extraction of acoustic targets based on uncorrelated neighborhood preserving discriminant projections

提出了一种基于去相关邻域保持判别投影的声目标特征提取算法(UNPDP),在邻域保持投影(NPP)算法保持局部线性结构的基础上,通过引入类别信息,在增强局部类内几何关系的同时最大化类间距离,提高了其低维嵌入的区分性;通过加入去相关限制,使得其得到的特征向量具有统计不相关特性,去除了冗余信息。在SensIT实验数据和外场实际采集数据上的实验结果表明基于去相关邻域保持判别投影的特征提取方法可以更好的表征声目标信号,识别的准确性和鲁棒性得到较大的提高。

A new method called uncorrelated neighborhood preserving discriminant projections（UNPDP） is proposed for feature extraction of acoustic targets.Based on neighborhood preserving projections（NPP）,UNPDP takes the class label information into account with maximizing the between-class distance while preserving the local linear structure of samples.Moreover,the low-dimension embeddings of UNPDP are statistically uncorrelated via some uncorrelated constraints,which makes the features contain minimum redundancy.The experimental results prove that the new algorithm can achieve significant advancement over the former method in accuracy and robustness,which can improve the acoustic target recognition system performance effectively.