基于面部动作单元组合特征的表情识别

Facial expression identification based on combinational feature of facial action units

人脸表情可以被看作是由面部表情编码系统(FACS)定义的不同面部运动单元的组合。不同于人脸图像的灰度、纹理等表象特征,基于面部运动单元的表情混合特征能够更准确地描述表情,然而,面部运动单元很难精确定位,为了避免这个问题,在前人的工作中通过将图像分成许多子块,并从子块中提取面部运动单元信息来组成基于面部运动单元的表情成分特征。在此基础上,本文首先通过对人脸图像的眼睛和口部作粗定位,接着根据眼睛和口部的水平位置,提取眼睛区域、口部区域和鼻子区域的图像子块,然后对每个子块提取Haar特征,并采用错误率最小策略从这些子块中选出面部运动单元组合特征,最后使用组合特征进行学习得出弱分类器,并嵌入到Boost学习结构中构造出强分类器。通过在Cohn-Kanada数据库上的测试,证明本文的方法能够取得很好的表情分类效果。

Facial expressions may be described as combination of facial action units defined by facial action coding system.Unlike appearance features of face images,such as gray and texture,the combinational feature of facial action units can describe the facial expressions better.However,it is difficult to detect facial action units accurately.So,many previous works try to divided face image into local patches,and extract the information of facial action units to compose the compositional features of facial expressions.According to these works,in this paper we firstly locate the position of eye and mouth in face images,and then divide face images into local patches due to the position of eye and mouth,after that extracted Haar features from each patches and use a minimum error based combination strategy to build combinational feature of facial action units from these features of patches,then use combinational feature to build weak learners.Finally boosting learning structure is used to build the final strong learner.In the experiment on Cohn-Kanada database,the method described in this paper has a promising performance.