表情识别中支持向量机核函数选取研究

SVM Kernel Function Selection In Emotion Recognition

支持向量机表情识别的准确率和时间消耗取决于核函数选取和特征数目。该文讨论了支持向量机的表情分类和核函数的实验方法,并进一步探讨了核和特征数目与识别准确率和时间消耗的关系。基于JAFFE数据库和LibSVM2.86的实验表明,随着特征数目的增加,训练时间呈指数增长,交叉验证准确率先增加后降低,表现为某种单峰分布。同时表明,线性核时间消耗最小,径向基核在特征数目较小时,具有最好的识别率,而在特征数目较大时,线性核最优。综合时间和识别率考虑,在低维时,优先选用径向基核,高维优先选用线性核。

SVM kernel function and feature number are critical to time consumption and accuracy in emotion recognition.This article discusses the classification of emotion recognition and the experiment methods of kernel function.With the relationship of recognition accuracy and the time consumption,kernel and feature num are explored using JAFFE dataset and LibSVM2.86.The experiment shows that:with the feature number increasing,time consumption grows exponentially,and the cross validation accuracy increases to top fast then descends slowly,representing as some single peak distribution.The experiment also shows that,linear kernel has the smallest time consumption;radical basis kernel has the best accuracy with small feature num,which is lower than linear kernel with large feature num.In respect to time consumption and accuracy,in low dimension,the best option is radical basis kernel;in contrast,the linear kernel is the best option in high dimension.