面向视频序列表情分类的LSVM算法

Facial Expression Recognition from Image Sequences with LSVM

为了提高基于视频序列的表情识别精度,在KNN-SVM算法的基础上提出局部SVM分类机制,并将其用于视频序列中的表情分类.对于一个待分类的几何特征样本,首先在训练集中寻找该样本的k个近邻样本,然后根据这k个近邻样本和待分类样本的相似度信息,重新构建局部最优的SVM分类决策超平面,用来对该几何特征样本进行分类.在Cohn-Kanade数据库中的对比实验表明,该分类器有效地提高了表情分类的精度.

In order to improve the accuracy of video based facial expression recognition, we propose a local SVM based on KNN-SVM algorithm, applied in facial expression recognition. For a geometric feature test sample, we first select k nearest neighboring training samples. A local optimal SVM decision hyper-plane is rebuilt based on the similarity between the test sample and the k neighboring samples for classifying the geometric feature test sample. Four different classifiers, KNN, SVM, KNN-SVM and LSVM, were compared, and the comparison results on the Cohn-Kanade database show the effectiveness of the method.