在变长模式识别中利用测地距离的非线性插值进行特征选取

讨论了变长模式识别中的特征选择问题。采用基于测地距离（Geodesic Distance）的非线性插值来进行特征选择．使得变长的模式映射为等长的模式，从而可以使用传统的等长模式的方法来解决变长模式识别问题。用非特定说话人的汉语孤立词识别来验证提出方法的性能，并采用支持向量机（Support Vector Machine，SVM）作为基本的分类方法。实验结果表明，提出的方法可以获得比传统方法诸如线性插值更好的性能，而计算量仅有很少增加。

Feature Selection in Length-variant Pattern Recognition Using Non-linear Interpolation Based on Geodesic Distance

Discuss the feature selection in length-variant pattern recognition.Non-linear interpolation based on geodesic distances is used in feature selection.Length-variant patterns are mapped into Length-invariant patterns by a non-linear interpolation;then traditional pattern recognition methods for length-invariant patterns can be used in solving the rec- ognition problem.Experiments of speaker-independent Mandarin isolated word recognition were performed to evaluate the performance of the proposed method.And support vector machine(SVM)is used as basic classification method. Experimental result shows that the proposed method has achieved better performance than traditional method such as linear interpolation,and computational complexity increased slightly.