基于差分纹理的人脸表情识别

Facial expression recognition based on differential texture features

考虑到自动人脸表情识别背景复杂性问题,提出了一个新的表情识别方法——基于差分纹理的人脸表情识别,该方法在一定程度上能够有效地屏蔽掉个体人脸之间的差异,同时保留住人脸表情信息。首先选定一个标准人脸参考模型,该模型合理分布面部55个基准点,这些基准点主要分布于眼睛、鼻子、嘴和包含表情丰富的外部轮廓上;然后利用Delaunay三角剖分获取这些基准点的相对位置信息。对于人脸表情图像,首先利用主动形状模型(ASM)跟踪定位这55个基准点,然后利用三角剖分获得的相对位置信息,以及应用纹理映射技术将表情图像映射到标准人脸参考模型中,这样中性表情图像(不含表情信息的人脸)和非中性表情(六种基本表情)图像均被映射到同一大小的框架内,最后将它们的差值图像作为表情特征,称为DT(differential texture,差分纹理)特征。最后分别将JAFFE人脸表情库和CK人脸表情库中的部分样本组成混合数据并进行实验,结果表明提出的方法对六种基本表情具有较好的识别率,并且该方法优于传统的Gabor特征和LBP特征方法,并能扩展到动态图像中的表情识别中去。

Considering the problem of automatically recognizing facial expression with complex background, this paper proposed a novel method, which could extract expression features regardless of face information. First, the method selected a standard reference model, in which 55 facial landmark points were reasonably distributed by geometric information of the face. Those landmark points mainly located at facial contour, eyebrows, eyes, nose and lips, which constituted the convex hull of face model. Then it deployed the Delaunay triangulation to get the relative location information of those points in the standard reference model. It got 55 landmark points by using ASM location for neutral expression and non-neutral expression, and applied the relation location information and texture mapping technology to those expression images. After the above processes, all face images were mapped to a standard reference framework. The difference between neutral expression and non-neutral expressions could be formed to one vector as facial expression features called DT features. In order to verify the effectiveness of the proposed method, it conducted 6 kinds of facial expression recognition experiments on JAFFE database and Cohn-Kanade database. The experiments show that this method has higher recognition rate for expression recognition. It also compared this method with other conventional feature extraction method, namely LBP （local binary pattern） features and Gabor features, the recognition rates show that this method outperforms these methods. This method can also be extended to facial expression recognition of dynamic image sequences.