表情识别的ICA两种架构分析

Application of ICA's Two Architectures in Emotion Recognition

比较了ICA方法进行面部表情识别的两种架构ICA1和ICA2。采用欧式、城区、余弦KNN和6种核函数的SVM算法进行识别,比较了不同的距离函数和核函数对整体识别率和单个表情识别率的影响。实验表明:ICA1整体上优于ICA2;对于KNN算法,在ICA1下KNN+城区距最优,t检验不显著,在ICA2下,KNN+余弦距最优,t检验显著;SVM算法对ICA1有效,对ICA2失效;在ICA1下,对SVM算法,线性、径向基和Sigmoid核取得相同的识别率;惊奇是最好识别的表情,高兴是最难识别的表情。最后利用神经科学对视觉脑区的最新研究,得出稀疏的特征比稀疏的编码能够取得更好的表情识别率。

An independent comparison of emotion recognition by ICA1 and ICA2 is proposed in this paper. ICA1 and ICA2 are two basic architectures of independent component analysis. ICA1 searches statistically independent basis image, while ICA2 finds independent factorial code representation. ICA1 and ICA2＇ feature sparseness and representation sparseness is compared by kurtosis. Then average emotion recognition accuracy and different kinds of emotion recognition accuracy are attained under KNN and support vector machine in each architecture. Three distance metric （Euclidean,Citybloek,Cosine） for KNN and six different kernel for SVM are considered. Based on experiments, some conclusions are introduced： ICA1 outperforms ICA2. for KNN algorithm, KNN＋Cityblock is the best for ICA1, but the t- test is not significant, while the KNN＋Cosine is the best for ICA2, t-test is significant; SVM algorithm is effective for ICAl,but bad for I- CA2; the linear,radical basis, sigmoid kernel archive the same accuracy for ICA1 ;Surprise is the easiest emotion to recognize while happy is the hardest one. at last, by latest research on visual cortex, we provide a hypnosis that the sparse feature outperforms sparse representation.