基于SFS的多模态生理信号情感识别

Multi-modal physiological signal emotion recognition based on SFS

多模态生理信号通常具有较高的维度,高维度的特征集不仅包括噪声数据和多余数据,影响分类结果,并且在情感识别过程中将花费大量的计算机开销。因此,从高维度的特征集中选取质量较优,影响因子较大的特征具有重要意义。文中主要提取血容量搏动信号,肌电信号,呼吸信号和皮肤电反应信号四种生理信号的平均值、标准偏差、一阶差值的绝对值的平均值、归一化信号的一阶差值的绝对值的平均值、二阶差值的绝对值的平均值、归一化信号的二阶差值的绝对值的平均值这6种统计特征,采用序列前向选择算法进行生理信号特征选择,最后用支持向量机和K近邻作为分类器对选取的特征子集进行样本集的分类,分类的精度作为衡量特征子集好坏的标准。实验结果表明,采用序列前向选择算法,可以选出比原始特征集维度更低且更优的特征子集。

Multi-modal physiological signals usually have higher dimensions.The high-dimensional features not only in clude noise and redundant data,but also affect the classification results.And it takes a lot of computer overhead in the emotion recognition process.It is a question how to select features with better quality and larger impact factors from high-dimensional feature sets.In this paper,we extract six statistical features from the blood flow pulsation,electromyo gram,respiratory signal and skin electrical response.There are extracted mean,standard deviation,the mean of first dif ference absolute value,the mean of second difference absolute value,the mean of normalized first difference absolute value and the mean of normalized second difference absolute value in this experiment.Forward selection algorithm is used for feature selection of physiological signals.Finally,the wrapper is used to classify the selected feature subsets,which the classification accuracy is used as a criterion to measure the quality of the feature subset.The experimental re sults show that the sequence forward selection algorithm can select lower dimensions and better feature subsets than the original features.