基于语义属性的人脸表情识别新方法

Facial Expression Recognition Method Based on Semantic Attribute

针对人脸表情底层视觉特征无法表达高层语义的问题,提出一种基于语义属性的人脸表情识别新方法.该方法利用表情语义属性这一中间人脸表情特征表示方法可在个别类别样本很少的情况下共享情感特征信息的特点,通过统计CK+库中人脸表情AU(Action Unit)编码建立表情语义属性与表情类别矩阵,然后采用SIFT(Scale-Invariant Feature Transform)底层视觉特征训练获得语义属性标注器,最后利用贝叶斯模型识别人脸表情.在CK+和BU-3DFE两个公开人脸表情数据库上的实验结果表明,与其它底层特征提取方法相比,该方法能有效提取表情特征信息并且把8种表情类别的平均识别率提高了4%.

In order to reduce the semantic gap between the low-level visual features of face images and high-level semantic, this paper presents a new facial expression recognition method based on semantic attribute. Semantic attributes as an intermediate representation, which enables parameters sharing between classes, can interpret the expression of a face image when training data is scarce. Firstly, we set up the matrix of semantic attributes and facial expression tag through the statistics of Action Units in the CK ＋ database. Then, the semantic attribute classifier is modeled by extracting the SIFT （ Scale-Invariant Feature Transform ） feature vectors for the static facial expression images. Finally,our model uses the semantic attribute classifier and the matrix for implementing the facial expression classi- fier. Experimental results show that the average recognition rate of the proposed method is increased by nearly 4% when it is compared with low-level visual features using databases of CK ＋ and BU3DFE.